feat(dnn/cuda): generate cutlass kimpls using cmake and bazel

GitOrigin-RevId: da3bcfb85a

.gitattributes

@@ -1,5 +1,6 @@
 # Mark generated files as binary, ignore them in git diff.
 # dnn
+dnn/scripts/cutlass_generator/list.bzl binary
 dnn/src/cuda/conv_bias/int4/kimpl/* binary
 dnn/src/cuda/conv_bias/int8/kimpl/* binary
 dnn/src/cuda/conv_bias/int8_imma/kimpl/* binary

dnn/scripts/cutlass_generator/BUILD

@@ -0,0 +1,18 @@
+load("list.bzl", "cutlass_gen_list")
+
+genrule(
+    name = "cutlass_kimpls",
+    outs = cutlass_gen_list,
+    cmd = """GEN=$(location //brain/megbrain/dnn/scripts/cutlass_generator:generator.py)
+          pwd > /tmp/a
+          echo $(@D) > /tmp/b
+          python3 $$GEN --operations gemm --type simt $(@D)
+          python3 $$GEN --operations gemv --type simt $(@D)
+          python3 $$GEN --operations deconv --type simt $(@D)
+          python3 $$GEN --operations conv2d --type simt $(@D)
+          python3 $$GEN --operations conv2d --type tensorop8816 $(@D)
+          python3 $$GEN --operations conv2d --type tensorop8832 $(@D)
+          """,
+    tools = ["//brain/megbrain/dnn/scripts/cutlass_generator:generator.py"],
+    visibility = ["//visibility:public"],
+)

dnn/scripts/cutlass_generator/README.md

@@ -0,0 +1,19 @@
+# Generate device kernel registration code for CUTLASS kernels
+## Usage
+```bash
+python3 generator.py [--operations {gemm, gemv, conv2d, deconv}] [--type {simt, tensorop8816, tensorop8832}]
+                     output
+```
+- operations: operation kind, including gemm|gemv|conv2d|deconv
+- type: opcode class, simt|tensorop8816|tensorop8832
+- output: the output directory for CUTLASS kernels
+
+## Generate file list for bazel
+
+We generate `list.bzl` because the `genrule` method of bazel requires that the output file list be specified in the analysis phase.
+
+Please call `gen_list.py` when new operations are added.
+
+```bash
+python3 gen_list.py
+```

dnn/scripts/cutlass_generator/conv2d_operation.py

@@ -0,0 +1,614 @@
+#
+# \file generator.py
+#
+# \brief Generates the CUTLASS Library's instances
+#
+#
+
+import enum
+import os.path
+import shutil
+from typing import Tuple, List
+
+from lazy_file import LazyFile
+from library import *
+
+###################################################################################################
+
+#
+class Conv2dOperation:
+  #
+  def __init__(self, conv_kind, conv_type, arch, tile_description, src, flt, bias, dst, element_epilogue, \
+    epilogue_functor = EpilogueFunctor.LinearCombination, swizzling_functor = SwizzlingFunctor.Identity4, \
+    need_load_from_const = True, implicit_gemm_mode = ImplicitGemmMode.GemmNt):
+
+    self.operation_kind = OperationKind.Conv2d
+    self.conv_kind = conv_kind
+    self.arch = arch
+    self.tile_description = tile_description
+    self.conv_type = conv_type
+    self.src = src
+    self.flt = flt
+    self.bias = bias
+    self.dst = dst
+    self.element_epilogue = element_epilogue
+    self.epilogue_functor = epilogue_functor
+    self.swizzling_functor = swizzling_functor
+    self.need_load_from_const = need_load_from_const  
+    self.implicit_gemm_mode = implicit_gemm_mode
+  #
+  def accumulator_type(self):
+    accum = self.tile_description.math_instruction.element_accumulator
+
+    return accum
+
+  #
+  def core_name(self):
+    ''' The basic operation kind is prefixed with a letter indicating the accumulation type. '''
+
+    intermediate_type = ''
+
+    if self.tile_description.math_instruction.opcode_class == OpcodeClass.TensorOp:
+      inst_shape = "%d%d%d" % tuple(self.tile_description.math_instruction.instruction_shape)
+      if self.tile_description.math_instruction.element_a != self.flt.element and \
+        self.tile_description.math_instruction.element_a != self.accumulator_type():
+        intermediate_type = DataTypeNames[self.tile_description.math_instruction.element_a]
+    else:
+      inst_shape = ''
+
+    unity_kernel = ''
+    if not self.need_load_from_const:
+        unity_kernel = '_1x1'
+
+    return "%s%s%s%s%s_%s" % (ShortDataTypeNames[self.accumulator_type()], \
+      inst_shape, intermediate_type, ConvKindNames[self.conv_kind], unity_kernel, \
+      ShortEpilogueNames[self.epilogue_functor])
+
+  #
+  def extended_name(self):
+    if self.dst.element != self.tile_description.math_instruction.element_accumulator:
+      if self.src.element != self.flt.element:
+        extended_name = "${element_dst}_${core_name}_${element_src}_${element_flt}"
+      elif self.src.element == self.flt.element:
+        extended_name = "${element_dst}_${core_name}_${element_src}"
+    else:
+      if self.src.element != self.flt.element:
+        extended_name = "${core_name}_${element_src}_${element_flt}"
+      elif self.src.element == self.flt.element:
+        extended_name = "${core_name}_${element_src}"
+ 
+    extended_name = SubstituteTemplate(extended_name, {
+      'element_src': DataTypeNames[self.src.element], 
+      'element_flt': DataTypeNames[self.flt.element], 
+      'element_dst': DataTypeNames[self.dst.element],
+      'core_name': self.core_name()
+      })
+
+    return extended_name
+
+  #
+  def layout_name(self):
+    if self.src.layout == self.dst.layout:
+      layout_name = "${src_layout}_${flt_layout}"
+    else:
+      layout_name = "${src_layout}_${flt_layout}_${dst_layout}"
+
+    layout_name = SubstituteTemplate(layout_name, {
+      'src_layout': ShortLayoutTypeNames[self.src.layout], 
+      'flt_layout': ShortLayoutTypeNames[self.flt.layout], 
+      'dst_layout': ShortLayoutTypeNames[self.dst.layout], 
+    })
+    
+    return layout_name
+
+#
+  def configuration_name(self):
+    ''' The full procedural name indicates architecture, extended name, tile size, and layout. '''
+
+    opcode_class_name = OpcodeClassNames[self.tile_description.math_instruction.opcode_class]
+    
+    warp_shape = [int(self.tile_description.threadblock_shape[idx] / self.tile_description.warp_count[idx]) for idx in range(3)]
+
+
+    threadblock = "%dx%dx%d_%dx%dx%d_%d" % (
+      self.tile_description.threadblock_shape[0],
+      self.tile_description.threadblock_shape[1],
+      self.tile_description.threadblock_shape[2],
+      warp_shape[0], 
+      warp_shape[1], 
+      warp_shape[2], 
+      self.tile_description.stages, 
+    )
+
+    configuration_name = "cutlass_${opcode_class}_${extended_name}_${threadblock}_${layout}"
+
+    return SubstituteTemplate(
+      configuration_name,
+      {
+        'opcode_class': opcode_class_name,
+        'extended_name': self.extended_name(),
+        'threadblock': threadblock,
+        'layout': self.layout_name(),
+      }
+    )
+
+  #
+  def procedural_name(self):
+    ''' The full procedural name indicates architecture, extended name, tile size, and layout. '''
+    return self.configuration_name()
+
+###################################################################################################
+#
+# Emits single instances of a CUTLASS device-wide operator
+#
+###################################################################################################
+
+class EmitConv2dInstance:
+  def __init__(self):
+    self.template = """
+// kernel instance "${operation_name}" generated by cutlass generator
+using Convolution = 
+  typename cutlass::conv::device::Convolution<
+    ${element_src}, 
+    ${layout_src},
+    ${element_flt}, 
+    ${layout_flt},
+    ${element_dst}, 
+    ${layout_dst},
+    ${element_bias}, 
+    ${layout_bias}, 
+    ${element_accumulator}, 
+    ${conv_type}, 
+    ${opcode_class},
+    ${arch},
+    cutlass::gemm::GemmShape<${threadblock_shape_m}, ${threadblock_shape_n}, ${threadblock_shape_k}>,
+    cutlass::gemm::GemmShape<${warp_shape_m}, ${warp_shape_n}, ${warp_shape_k}>,
+    cutlass::gemm::GemmShape<${instruction_shape_m}, ${instruction_shape_n}, ${instruction_shape_k}>,
+    ${epilogue_functor}<
+      ${element_dst},
+      ${epilogue_vector_length},
+      ${element_accumulator}, 
+      ${element_bias}, 
+      ${element_epilogue}
+    >,
+    ${swizzling_functor},     
+    ${stages},
+    ${alignment_src}, 
+    ${alignment_filter}, 
+    ${nonuninity_kernel}, 
+    ${math_operator},
+    ${implicit_gemm_mode}>;
+"""
+
+
+  def emit(self, operation):
+
+    warp_shape = [int(operation.tile_description.threadblock_shape[idx] / operation.tile_description.warp_count[idx]) for idx in range(3)]
+
+    epilogue_vector_length = int(min(operation.dst.alignment * DataTypeSize[operation.dst.element], 128) / DataTypeSize[operation.dst.element])
+
+    values = {
+      'operation_name': operation.procedural_name(),
+      'conv_type': ConvTypeTag[operation.conv_type], 
+      'element_src': DataTypeTag[operation.src.element],
+      'layout_src': LayoutTag[operation.src.layout],
+      'element_flt': DataTypeTag[operation.flt.element],
+      'layout_flt': LayoutTag[operation.flt.layout],
+      'element_dst': DataTypeTag[operation.dst.element],
+      'layout_dst': LayoutTag[operation.dst.layout],
+      'element_bias': DataTypeTag[operation.bias.element],
+      'layout_bias': LayoutTag[operation.bias.layout],
+      'element_accumulator': DataTypeTag[operation.accumulator_type()], 
+      'opcode_class': OpcodeClassTag[operation.tile_description.math_instruction.opcode_class],
+      'arch': "cutlass::arch::Sm%d" % operation.arch,
+      'threadblock_shape_m': str(operation.tile_description.threadblock_shape[0]),
+      'threadblock_shape_n': str(operation.tile_description.threadblock_shape[1]),
+      'threadblock_shape_k': str(operation.tile_description.threadblock_shape[2]),
+      'warp_shape_m': str(warp_shape[0]),
+      'warp_shape_n': str(warp_shape[1]),
+      'warp_shape_k': str(warp_shape[2]),
+      'instruction_shape_m': str(operation.tile_description.math_instruction.instruction_shape[0]),
+      'instruction_shape_n': str(operation.tile_description.math_instruction.instruction_shape[1]),
+      'instruction_shape_k': str(operation.tile_description.math_instruction.instruction_shape[2]),
+      'epilogue_vector_length': str(epilogue_vector_length),
+      'epilogue_functor': EpilogueFunctorTag[operation.epilogue_functor],
+      'element_epilogue': str(DataTypeTag[operation.element_epilogue]),
+      'swizzling_functor': SwizzlingFunctorTag[operation.swizzling_functor],
+      'stages': str(operation.tile_description.stages),
+      'alignment_src': str(operation.src.alignment), 
+      'alignment_filter': str(operation.flt.alignment), 
+      'nonuninity_kernel': str(operation.need_load_from_const).lower(),  
+      'math_operator': MathOperationTag[operation.tile_description.math_instruction.math_operation],
+      'implicit_gemm_mode': ImplicitGemmModeTag[operation.implicit_gemm_mode]
+    }
+
+    return SubstituteTemplate(self.template, values)
+
+class EmitDeconvInstance:
+  def __init__(self):
+    self.template = """
+// kernel instance "${operation_name}" generated by cutlass generator
+using Deconvolution = 
+  typename cutlass::conv::device::Deconvolution<
+    ${element_src}, 
+    ${layout_src},
+    ${element_flt}, 
+    ${layout_flt},
+    ${element_dst}, 
+    ${layout_dst},
+    ${element_bias}, 
+    ${layout_bias}, 
+    ${element_accumulator}, 
+    ${opcode_class},
+    ${arch},
+    cutlass::gemm::GemmShape<${threadblock_shape_m}, ${threadblock_shape_n}, ${threadblock_shape_k}>,
+    cutlass::gemm::GemmShape<${warp_shape_m}, ${warp_shape_n}, ${warp_shape_k}>,
+    cutlass::gemm::GemmShape<${instruction_shape_m}, ${instruction_shape_n}, ${instruction_shape_k}>,
+    ${epilogue_functor}<
+      ${element_dst},
+      ${epilogue_vector_length},
+      ${element_accumulator}, 
+      ${element_bias}, 
+      ${element_epilogue}
+    >,
+    ${swizzling_functor},     
+    ${stages},
+    ${alignment_src}, 
+    ${alignment_filter}, 
+    ${nonuninity_kernel}, 
+    ${math_operator},
+    ${implicit_gemm_mode}>;
+"""
+
+
+  def emit(self, operation):
+
+    warp_shape = [int(operation.tile_description.threadblock_shape[idx] / operation.tile_description.warp_count[idx]) for idx in range(3)]
+
+    epilogue_vector_length = int(min(operation.dst.alignment * DataTypeSize[operation.dst.element], 128) / DataTypeSize[operation.dst.element])
+
+    values = {
+      'operation_name': operation.procedural_name(),
+      'element_src': DataTypeTag[operation.src.element],
+      'layout_src': LayoutTag[operation.src.layout],
+      'element_flt': DataTypeTag[operation.flt.element],
+      'layout_flt': LayoutTag[operation.flt.layout],
+      'element_dst': DataTypeTag[operation.dst.element],
+      'layout_dst': LayoutTag[operation.dst.layout],
+      'element_bias': DataTypeTag[operation.bias.element],
+      'layout_bias': LayoutTag[operation.bias.layout],
+      'element_accumulator': DataTypeTag[operation.accumulator_type()], 
+      'opcode_class': OpcodeClassTag[operation.tile_description.math_instruction.opcode_class],
+      'arch': "cutlass::arch::Sm%d" % operation.arch,
+      'threadblock_shape_m': str(operation.tile_description.threadblock_shape[0]),
+      'threadblock_shape_n': str(operation.tile_description.threadblock_shape[1]),
+      'threadblock_shape_k': str(operation.tile_description.threadblock_shape[2]),
+      'warp_shape_m': str(warp_shape[0]),
+      'warp_shape_n': str(warp_shape[1]),
+      'warp_shape_k': str(warp_shape[2]),
+      'instruction_shape_m': str(operation.tile_description.math_instruction.instruction_shape[0]),
+      'instruction_shape_n': str(operation.tile_description.math_instruction.instruction_shape[1]),
+      'instruction_shape_k': str(operation.tile_description.math_instruction.instruction_shape[2]),
+      'epilogue_vector_length': str(epilogue_vector_length),
+      'epilogue_functor': EpilogueFunctorTag[operation.epilogue_functor],
+      'element_epilogue': str(DataTypeTag[operation.element_epilogue]),
+      'swizzling_functor': SwizzlingFunctorTag[operation.swizzling_functor],
+      'stages': str(operation.tile_description.stages),
+      'alignment_src': str(operation.src.alignment), 
+      'alignment_filter': str(operation.flt.alignment), 
+      'nonuninity_kernel': str(operation.need_load_from_const).lower(),
+      'math_operator': MathOperationTag[operation.tile_description.math_instruction.math_operation],
+      'implicit_gemm_mode': ImplicitGemmModeTag[operation.implicit_gemm_mode]
+    }
+
+    return SubstituteTemplate(self.template, values)
+
+
+###################################################################################################
+#
+# Generator functions for all layouts
+#
+###################################################################################################
+
+#
+def GenerateConv2d(conv_kind, tile_descriptions, src_layout, flt_layout, dst_layout, dst_type, min_cc, src_align = 32, flt_align = 32, dst_align = 128, \
+  skip_unity_kernel = False, implicit_gemm_mode = ImplicitGemmMode.GemmNt):
+  operations = []
+
+  element_epilogue = DataType.f32 
+  if conv_kind == ConvKind.Fprop:
+    if src_layout == LayoutType.TensorNHWC:
+      swizzling_functor = SwizzlingFunctor.ConvFpropNHWC
+    else:
+      swizzling_functor = SwizzlingFunctor.ConvFpropNCxHWx
+  else:
+    swizzling_functor = SwizzlingFunctor.ConvDgradNCxHWx
+
+  # skip rule
+  def filter_tile_with_layout(tile: TileDescription, layout: LayoutType) -> bool:
+    return layout == LayoutType.TensorNC32HW32 and \
+           tile.threadblock_shape[0] % 32 != 0
+  
+  # rule for bias_type and epilogues
+  def get_bias_type_and_epilogues(tile: TileDescription, \
+                                  out_dtype: DataType) -> Tuple[DataType, List[EpilogueFunctor]]:
+    if tile.math_instruction.element_accumulator == DataType.s32 and \
+        out_dtype != DataType.f32:
+      bias_type = DataType.s32
+      if tile.math_instruction.element_b == DataType.u4:
+        epilogues = [EpilogueFunctor.BiasAddLinearCombinationClamp, EpilogueFunctor.BiasAddLinearCombinationReluClamp]
+      else:
+        epilogues = [EpilogueFunctor.BiasAddLinearCombinationClamp, EpilogueFunctor.BiasAddLinearCombinationReluClamp, \
+                     EpilogueFunctor.BiasAddLinearCombinationHSwishClamp]
+    elif tile.math_instruction.element_accumulator == DataType.f32 or \
+        out_dtype == DataType.f32:
+      bias_type = DataType.f32
+      epilogues = [EpilogueFunctor.BiasAddLinearCombination, EpilogueFunctor.BiasAddLinearCombinationRelu, \
+                   EpilogueFunctor.BiasAddLinearCombinationHSwish]
+    return bias_type, epilogues
+
+  # rule for filter alignment
+  def get_flt_align(tile: TileDescription) -> int:
+    nonlocal flt_align
+    if tile.math_instruction.opcode_class == OpcodeClass.Simt \
+            and tile.math_instruction.element_accumulator == DataType.s32:
+      thread_num = tile.warp_count[0] * tile.warp_count[1] * tile.warp_count[2] * 32
+      flt_block = tile.threadblock_shape[0] * tile.threadblock_shape[2] \
+                                            * DataTypeSize[tile.math_instruction.element_a]
+      load_per_thread = flt_block//thread_num 
+      if load_per_thread >= 128:
+        flt_align = 128
+      elif load_per_thread >= 64:
+        flt_align = 64
+      else:
+        assert load_per_thread >= 32
+        flt_align = 32
+    return flt_align
+
+  def get_dst_align(tile: TileDescription, out_layout: LayoutType) -> int:
+    nonlocal dst_align
+    if tile.math_instruction.opcode_class == OpcodeClass.TensorOp \
+      and dst_layout == LayoutType.TensorNC4HW4:
+      dst_align = 32
+    return dst_align
+
+  def filter_epilogue_with_conv_kind(epilogue: EpilogueFunctor, conv_kind: ConvKind) -> bool:
+    return conv_kind == ConvKind.Dgrad \
+        and epilogue != EpilogueFunctor.BiasAddLinearCombinationClamp
+
+  # loop over all tile descriptions
+  for tile in tile_descriptions:
+    if filter_tile_with_layout(tile, dst_layout):
+      continue
+
+    bias_type, epilogues = get_bias_type_and_epilogues(tile, dst_type)
+
+    flt_align = get_flt_align(tile)
+
+    dst_align = get_dst_align(tile, dst_layout)
+
+    for epilogue in epilogues:
+      if filter_epilogue_with_conv_kind(epilogue, conv_kind):
+        continue
+
+      if dst_type == DataType.f32:
+        bias_type = DataType.f32
+      #
+      src = TensorDescription(tile.math_instruction.element_b, src_layout, int(src_align / DataTypeSize[tile.math_instruction.element_b]))
+      flt = TensorDescription(tile.math_instruction.element_a, flt_layout, int(flt_align / DataTypeSize[tile.math_instruction.element_a]))
+      bias = TensorDescription(bias_type, dst_layout, max(1, int(32 / DataTypeSize[bias_type])))
+      dst = TensorDescription(dst_type, dst_layout, int(dst_align / DataTypeSize[dst_type])) 
+
+      new_operation = Conv2dOperation(conv_kind, ConvType.Convolution, min_cc, tile, src, flt, bias, dst, element_epilogue, epilogue, swizzling_functor, True, implicit_gemm_mode)
+      operations.append(new_operation)
+      if not skip_unity_kernel:
+        new_operation = Conv2dOperation(conv_kind, ConvType.Convolution, min_cc, tile, src, flt, bias, dst, element_epilogue, epilogue, swizzling_functor, False, implicit_gemm_mode)
+        operations.append(new_operation)
+  return operations
+
+###################################################################################################
+#
+# Emitters functions for all targets
+#
+###################################################################################################
+
+class EmitConv2dConfigurationLibrary:
+  def __init__(self, operation_path, configuration_name):
+    self.configuration_name = configuration_name
+    self.configuration_path = os.path.join(operation_path, "%s.cu" % configuration_name)
+
+    self.instance_emitter = EmitConv2dInstance()
+
+    self.instance_template = """
+${operation_instance}
+
+// Derived class
+struct ${operation_name} : 
+  public ${operation_name}_base { };
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+"""
+    self.header_template = """
+/*
+  Generated by conv2d_operation.py - Do not edit.
+*/
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+#include "cutlass/cutlass.h"
+#include "cutlass/library/library.h"
+#include "cutlass/library/manifest.h"
+
+#include "library_internal.h"
+#include "conv2d_operation.h"
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+"""
+
+    self.configuration_header = """
+
+namespace cutlass {
+namespace library {
+
+// Initialize all instances
+void initialize_${configuration_name}(Manifest &manifest) {
+
+"""
+
+    self.configuration_instance = """
+  using Operation_${operation_name} = cutlass::conv::device::ImplicitGemmConvolution<
+    ${operation_name}>;
+
+  manifest.append(new cutlass::library::Conv2dOperation<
+    Operation_${operation_name}>(
+      "${operation_name}"));
+
+"""
+
+    self.configuration_epilogue = """
+}
+"""
+    self.epilogue_template = """
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+} // namespace library
+} // namespace cutlass
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+"""
+
+  #
+  def __enter__(self):
+    self.configuration_file = open(self.configuration_path, "w")
+    self.configuration_file.write(SubstituteTemplate(self.header_template, {
+      'configuration_name': self.configuration_name
+      }))
+    self.operations = []
+    return self
+
+  #
+  def emit(self, operation):
+    self.operations.append(operation)
+    self.configuration_file.write(SubstituteTemplate(self.instance_template, {
+      'configuration_name': self.configuration_name,
+      'operation_name': operation.procedural_name(),
+      'operation_instance': self.instance_emitter.emit(operation)
+      }))
+
+  #
+  def __exit__(self, exception_type, exception_value, traceback):
+
+    self.configuration_file.write(SubstituteTemplate(self.configuration_header, {
+      'configuration_name': self.configuration_name
+      }))
+
+    for operation in self.operations:
+      self.configuration_file.write(SubstituteTemplate(self.configuration_instance, {
+        'configuration_name': self.configuration_name,
+        'operation_name': operation.procedural_name()  
+      }))
+
+    self.configuration_file.write(self.configuration_epilogue)
+    self.configuration_file.write(self.epilogue_template)
+    self.configuration_file.close()
+
+###################################################################################################
+###################################################################################################
+
+# Emitters for Conv Kernel Wrapper
+#
+###################################################################################################
+
+class EmitConvSingleKernelWrapper():
+  def __init__(self, kernel_path, operation, wrapper_path):
+    self.kernel_path = kernel_path
+    self.wrapper_path = wrapper_path
+    self.operation = operation
+
+    self.conv_wrappers = { \
+      ConvKind.Fprop: """
+template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
+  const typename Convolution::ElementSrc* d_src, 
+  const typename Convolution::ElementFilter* d_filter, 
+  const typename Convolution::ElementBias* d_bias, 
+  const typename Convolution::ElementDst* d_z, 
+  typename Convolution::ElementDst* d_dst, 
+  int* workspace, 
+  typename Convolution::ConvolutionParameter const& conv_param, 
+  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
+  cudaStream_t stream, 
+  typename Convolution::ExtraParam extra_param);
+""", \
+      ConvKind.Dgrad: """
+template void megdnn::cuda::cutlass_wrapper::cutlass_deconvolution_wrapper<Deconvolution>(
+  const typename Deconvolution::ElementSrc* d_src, 
+  const typename Deconvolution::ElementFilter* d_filter, 
+  const typename Deconvolution::ElementBias* d_bias, 
+  const typename Deconvolution::ElementDst* d_z, 
+  typename Deconvolution::ElementDst* d_dst, 
+  int* workspace, 
+  typename Deconvolution::ConvolutionParameter const& conv_param, 
+  typename Deconvolution::EpilogueOutputOp::Params const& epilogue, 
+  cudaStream_t stream);
+""", \
+    }
+    
+    if self.operation.conv_kind == ConvKind.Fprop:
+      self.instance_emitter = EmitConv2dInstance()
+    else:
+      assert self.operation.conv_kind == ConvKind.Dgrad
+      self.instance_emitter = EmitDeconvInstance()
+
+    self.header_template = """
+#if !MEGDNN_TEGRA_X1
+// ignore warning of cutlass
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wunused-parameter"
+#pragma GCC diagnostic ignored "-Wstrict-aliasing"
+#include "${wrapper_path}"
+"""
+    self.instance_template = """
+${operation_instance}
+"""
+    self.wrapper_template = """
+${wrapper_instance}
+"""
+
+    self.epilogue_template = """
+#pragma GCC diagnostic pop
+#endif
+"""
+
+  #
+  def __enter__(self):
+    self.kernel_path = os.path.join(self.kernel_path, "%s.cu" % self.operation.procedural_name()) 
+    self.kernel_file = LazyFile(self.kernel_path)
+    self.kernel_file.write(SubstituteTemplate(self.header_template, {
+      'wrapper_path': self.wrapper_path, 
+      }))
+    return self
+
+  #
+  def emit(self):
+    self.kernel_file.write(SubstituteTemplate(self.instance_template, {
+      'operation_instance': self.instance_emitter.emit(self.operation),
+      }))
+
+    # emit wrapper
+    wrapper = SubstituteTemplate(self.wrapper_template, {
+      'wrapper_instance': self.conv_wrappers[self.operation.conv_kind], 
+    })
+    self.kernel_file.write(wrapper)
+
+  #
+  def __exit__(self, exception_type, exception_value, traceback):
+    self.kernel_file.write(self.epilogue_template)
+    self.kernel_file.close()
+
+
+###################################################################################################
+###################################################################################################
+

dnn/scripts/cutlass_generator/gen_list.py

@@ -0,0 +1,38 @@
+from generator import (
+    GenerateGemmOperations,
+    GenerateGemvOperations,
+    GenerateConv2dOperations,
+    GenerateDeconvOperations,
+)
+
+
+class GenArg:
+    def __init__(self, gen_op, gen_type):
+        self.operations = gen_op
+        self.type = gen_type
+
+
+def write_op_list(f, gen_op, gen_type):
+    if gen_op == "gemm":
+        operations = GenerateGemmOperations(GenArg(gen_op, gen_type))
+    elif gen_op == "gemv":
+        operations = GenerateGemvOperations(GenArg(gen_op, gen_type))
+    elif gen_op == "conv2d":
+        operations = GenerateConv2dOperations(GenArg(gen_op, gen_type))
+    elif gen_op == "deconv":
+        operations = GenerateDeconvOperations(GenArg(gen_op, gen_type))
+    for op in operations:
+        f.write('    "%s.cu",\n' % op.procedural_name())
+
+
+if __name__ == "__main__":
+    with open("list.bzl", "w") as f:
+        f.write("# Generated by dnn/scripts/cutlass_generator/gen_list.py\n\n")
+        f.write("cutlass_gen_list = [\n")
+        write_op_list(f, "gemm", "simt")
+        write_op_list(f, "gemv", "simt")
+        write_op_list(f, "deconv", "simt")
+        write_op_list(f, "conv2d", "simt")
+        write_op_list(f, "conv2d", "tensorop8816")
+        write_op_list(f, "conv2d", "tensorop8832")
+        f.write("]")

dnn/scripts/cutlass_generator/generator.py

@@ -0,0 +1,651 @@
+#
+# \file generator.py
+#
+# \brief Generates the CUTLASS Library's instances
+#
+
+import enum
+import os.path
+import shutil
+import argparse
+
+from library import *
+from manifest import *
+###################################################################################################
+
+#
+def CudaToolkitVersionSatisfies(semantic_ver_string, major, minor, patch = 0):
+
+  # by default, use the latest CUDA Toolkit version
+  cuda_version = [11, 0, 132]
+
+  # Update cuda_version based on parsed string
+  if semantic_ver_string != '':
+    for i, x in enumerate([int(x) for x in semantic_ver_string.split('.')]):
+      if i < len(cuda_version):
+        cuda_version[i] = x
+      else:
+        cuda_version.append(x)
+  return cuda_version >= [major, minor, patch]
+
+
+###################################################################################################
+###################################################################################################
+
+#
+def CreateGemmOperator(manifest, layouts, tile_descriptions, data_type, \
+  alignment_constraints, complex_transforms = None, epilogue_functor = EpilogueFunctor.LinearCombination, \
+  swizzling_functor = SwizzlingFunctor.Identity8):
+
+  if complex_transforms is None:
+    complex_transforms = [(ComplexTransform.none, ComplexTransform.none),]
+
+  element_a, element_b, element_c, element_epilogue = data_type
+  
+  operations = []
+
+  # by default, only generate the largest tile and largest alignment
+  if manifest.args.kernels == '':
+    tile_descriptions = [tile_descriptions[0],]
+    alignment_constraints = [alignment_constraints[0],]
+
+  for layout in layouts:
+    for tile_description in tile_descriptions:
+      for alignment in alignment_constraints:
+        for complex_transform in complex_transforms:
+  
+            alignment_c = min(8, alignment)
+ 
+            A = TensorDescription(element_a, layout[0], alignment, complex_transform[0])
+            B = TensorDescription(element_b, layout[1], alignment, complex_transform[1])
+            C = TensorDescription(element_c, layout[2], alignment_c)
+
+            new_operation = GemmOperation(GemmKind.Universal, tile_description.minimum_compute_capability, \
+              tile_description, A, B, C, element_epilogue, epilogue_functor, swizzling_functor)
+
+            manifest.append(new_operation)
+            operations.append(new_operation)
+
+  return operations
+
+###########################################################################################################
+#   ConvolutionOperator support variations
+#        ____________________________________________________________________
+#         ConvolutionalOperator |        Analytic      |      Optimized
+#        ____________________________________________________________________
+#        |       Fprop          |     (strided)        |    (strided)
+#        |       Dgrad          |   (strided, unity*)  |     (unity)
+#        |       Wgrad          |     (strided)        |    (strided)
+#        ____________________________________________________________________
+#
+# Note :  Operator marked (*) are supported but not generated to keep the instantiated kernel count low
+###########################################################################################################
+# Convolution for 2D operations
+def CreateConv2dOperator(manifest, layout, tile_descriptions, data_type, alignment, \
+  conv_kinds = [ConvKind.Fprop, ConvKind.Dgrad, ConvKind.Wgrad], epilogue_functor = EpilogueFunctor.LinearCombination):
+  
+  element_a, element_b, element_c, element_epilogue = data_type
+  
+  # one exceptional case
+  alignment_c = min(8, alignment)
+  
+  # iterator algorithm (analytic and optimized)
+  iterator_algorithms = [IteratorAlgorithm.Analytic, IteratorAlgorithm.Optimized]
+
+  # by default, only generate the largest tile size
+  if manifest.args.kernels == '':
+    tile_descriptions = [tile_descriptions[0],]
+
+  operations = []
+
+  for tile in tile_descriptions:
+    for conv_kind in conv_kinds:
+      for iterator_algorithm in iterator_algorithms:
+        A = TensorDescription(element_a, layout[0], alignment)
+        B = TensorDescription(element_b, layout[1], alignment)
+        C = TensorDescription(element_c, layout[2], alignment_c)
+
+        # unity stride only for Optimized Dgrad
+        if (iterator_algorithm == IteratorAlgorithm.Optimized) and (conv_kind == ConvKind.Dgrad):
+          new_operation = Conv2dOperation(conv_kind, iterator_algorithm, tile.minimum_compute_capability, tile,\
+            A, B, C, element_epilogue, StrideSupport.Unity, epilogue_functor)
+
+          manifest.append(new_operation)
+          operations.append(new_operation)
+
+        # strided dgrad is not supported by Optimized Dgrad
+        if (iterator_algorithm == IteratorAlgorithm.Optimized) and (conv_kind == ConvKind.Dgrad):
+          continue 
+
+        # strided support for Fprop (Analytic/Optimized), Dgrad (Analytic), and Wgrad (Analytic)
+        new_operation = Conv2dOperation(conv_kind, iterator_algorithm, tile.minimum_compute_capability, tile,\
+         A, B, C, element_epilogue, StrideSupport.Strided, epilogue_functor)
+
+        manifest.append(new_operation)
+        operations.append(new_operation)
+
+  return operations
+
+###################################################################################################
+###################################################################################################
+
+def GenerateConv2d_Simt(args):
+  operations = []
+
+  layouts = [
+    (LayoutType.TensorNC4HW4, LayoutType.TensorC4RSK4), 
+  ]
+    
+  math_instructions = [
+    MathInstruction(                                  \
+      [1, 1, 4],                                      \
+      DataType.s8, DataType.s8, DataType.s32,         \
+      OpcodeClass.Simt,                               \
+      MathOperation.multiply_add),
+  ]
+
+  dst_layouts = [
+      LayoutType.TensorNC4HW4, 
+      LayoutType.TensorNC32HW32, 
+      LayoutType.TensorNHWC, 
+      LayoutType.TensorNHWC, 
+      LayoutType.TensorNCHW
+  ]
+
+  dst_types = [
+      DataType.s8, 
+      DataType.s8, 
+      DataType.u4, 
+      DataType.s4, 
+      DataType.f32, 
+  ]
+
+  max_cc = 1024
+
+  for math_inst in math_instructions:
+    for layout in layouts:
+      for dst_type, dst_layout in zip(dst_types, dst_layouts):
+        if dst_type == DataType.s4 or dst_type == DataType.u4:
+          min_cc = 75
+          skip_unity_kernel = True
+        else:
+          min_cc = 61
+          skip_unity_kernel = False
+        tile_descriptions = [
+          TileDescription([128, 128, 32], 2, [2, 4, 1], math_inst, min_cc, max_cc),
+          TileDescription([128,  64, 32], 2, [2, 2, 1], math_inst, min_cc, max_cc),
+          TileDescription([ 64, 128, 32], 2, [1, 4, 1], math_inst, min_cc, max_cc),
+          TileDescription([ 64,  64, 32], 2, [1, 2, 1], math_inst, min_cc, max_cc),
+          TileDescription([128,  32, 32], 2, [2, 1, 1], math_inst, min_cc, max_cc),
+          TileDescription([ 32, 128, 32], 2, [1, 2, 1], math_inst, min_cc, max_cc),
+          TileDescription([ 32,  64, 32], 2, [1, 1, 1], math_inst, min_cc, max_cc), 
+          TileDescription([ 64,  32, 32], 2, [1, 1, 1], math_inst, min_cc, max_cc), 
+          TileDescription([ 32,  32, 32], 2, [1, 1, 1], math_inst, min_cc, max_cc), 
+          TileDescription([ 16, 128, 16], 1, [1, 1, 1], math_inst, min_cc, max_cc), 
+          TileDescription([ 16,  64,  8], 2, [1, 1, 1], math_inst, min_cc, max_cc), 
+        ] 
+        operations += GenerateConv2d(ConvKind.Fprop, tile_descriptions, layout[0], layout[1], 
+                                     dst_layout, dst_type, min_cc, 32, 32, 32, 
+                                     skip_unity_kernel) 
+  return operations
+
+
+def GenerateConv2d_TensorOp_8816(args):
+  operations = []
+
+  layouts = [
+    (LayoutType.TensorNC32HW32, LayoutType.TensorC32RSK32), 
+  ]
+    
+  math_instructions = [
+    MathInstruction(                                  \
+      [8, 8, 16],                                     \
+      DataType.s8, DataType.s8, DataType.s32,         \
+      OpcodeClass.TensorOp,                           \
+      MathOperation.multiply_add_saturate),
+  ]
+
+  dst_layouts = [
+      LayoutType.TensorNC32HW32, 
+      LayoutType.TensorNC4HW4, 
+  ]
+
+  dst_types = [
+      DataType.s8, 
+      DataType.s8, 
+  ]
+
+  min_cc = 75
+  max_cc = 1024
+
+  for math_inst in math_instructions:
+    for layout in layouts:
+      for dst_type, dst_layout in zip(dst_types, dst_layouts):
+        if dst_layout == LayoutType.TensorNC32HW32:
+          tile_descriptions = [
+            TileDescription([256, 128, 64], 2, [4, 2, 1], math_inst, min_cc, max_cc),
+            TileDescription([128, 256, 64], 2, [2, 4, 1], math_inst, min_cc, max_cc),
+            TileDescription([128, 128, 64], 2, [2, 2, 1], math_inst, min_cc, max_cc),
+            TileDescription([ 64, 128, 64], 2, [2, 2, 1], math_inst, min_cc, max_cc),
+            TileDescription([128,  64, 64], 2, [2, 2, 1], math_inst, min_cc, max_cc),
+            TileDescription([ 64,  64, 64], 2, [2, 2, 1], math_inst, min_cc, max_cc),
+            TileDescription([ 32,  64, 64], 2, [1, 4, 1], math_inst, min_cc, max_cc),
+          ] 
+        else:
+          assert dst_layout == LayoutType.TensorNC4HW4
+          tile_descriptions = [
+            TileDescription([256, 128, 64], 2, [4, 2, 1], math_inst, min_cc, max_cc),
+            TileDescription([128, 256, 64], 2, [2, 4, 1], math_inst, min_cc, max_cc),
+            TileDescription([128, 128, 64], 2, [2, 2, 1], math_inst, min_cc, max_cc),
+            TileDescription([ 64, 128, 64], 2, [2, 2, 1], math_inst, min_cc, max_cc),
+            TileDescription([128,  64, 64], 2, [2, 2, 1], math_inst, min_cc, max_cc),
+            TileDescription([ 64,  64, 64], 2, [2, 2, 1], math_inst, min_cc, max_cc),
+            TileDescription([ 32,  64, 64], 2, [2, 2, 1], math_inst, min_cc, max_cc),
+          ]
+        operations += GenerateConv2d(ConvKind.Fprop, tile_descriptions, layout[0], layout[1], 
+                                     dst_layout, dst_type, min_cc, 128, 128, 64,  
+                                     False) 
+  return operations
+
+def GenerateConv2d_TensorOp_8832(args):
+  operations = []
+
+  layouts = [
+    (LayoutType.TensorNC64HW64, LayoutType.TensorC64RSK64), 
+  ]
+    
+  math_instructions = [
+    MathInstruction(                                  \
+      [8, 8, 32],                                     \
+      DataType.s4, DataType.s4, DataType.s32,         \
+      OpcodeClass.TensorOp,                           \
+      MathOperation.multiply_add_saturate),           \
+    MathInstruction(                                  \
+      [8, 8, 32],                                     \
+      DataType.s4, DataType.u4, DataType.s32,         \
+      OpcodeClass.TensorOp,                           \
+      MathOperation.multiply_add_saturate)
+  ]
+
+  dst_layouts = [
+      LayoutType.TensorNC64HW64, 
+  ]
+
+  min_cc = 75
+  max_cc = 1024
+
+  for math_inst in math_instructions:
+    for layout in layouts:
+      for dst_layout in dst_layouts:
+        dst_type = math_inst.element_b
+        tile_descriptions = [
+          TileDescription([256, 128, 128], 2, [4, 2, 1], math_inst, min_cc, max_cc),
+          TileDescription([128, 128, 128], 2, [2, 2, 1], math_inst, min_cc, max_cc),
+        ] 
+        operations += GenerateConv2d(ConvKind.Fprop, tile_descriptions, layout[0], layout[1], 
+                                     dst_layout, dst_type, min_cc, 128, 128, 64,  
+                                     True) 
+
+  layouts_nhwc = [
+    (LayoutType.TensorNHWC, LayoutType.TensorNC8HW8, 32), 
+    (LayoutType.TensorNHWC, LayoutType.TensorNC16HW16, 64), 
+    (LayoutType.TensorNHWC, LayoutType.TensorNC32HW32, 128), 
+  ]
+
+  dst_layouts_nhwc = [
+      LayoutType.TensorNHWC, 
+  ]
+
+  for math_inst in math_instructions:
+    for layout in layouts_nhwc:
+      for dst_layout in dst_layouts_nhwc:
+        dst_type = math_inst.element_b
+        tile_descriptions = [
+          TileDescription([128, 32, 64], 2, [2, 1, 1], math_inst, min_cc, max_cc),
+          TileDescription([128, 64, 64], 2, [2, 1, 1], math_inst, min_cc, max_cc),
+        ] 
+        operations += GenerateConv2d(ConvKind.Fprop, tile_descriptions, layout[0], layout[1], 
+                                     dst_layout, dst_type, min_cc, layout[2], layout[2], 32,  
+                                     False, ImplicitGemmMode.GemmTn) 
+  return operations
+
+def GenerateDeconv_Simt(args):
+  operations = []
+
+  layouts = [
+    (LayoutType.TensorNC4HW4, LayoutType.TensorK4RSC4), 
+  ]
+    
+  math_instructions = [
+    MathInstruction(                                  \
+      [1, 1, 4],                                      \
+      DataType.s8, DataType.s8, DataType.s32,         \
+      OpcodeClass.Simt,                               \
+      MathOperation.multiply_add),
+  ]
+
+  dst_layouts = [
+      LayoutType.TensorNC4HW4, 
+  ]
+
+  dst_types = [
+      DataType.s8, 
+  ]
+
+  min_cc = 61
+  max_cc = 1024
+
+  for math_inst in math_instructions:
+    for layout in layouts:
+      for dst_type, dst_layout in zip(dst_types, dst_layouts):
+        tile_descriptions = [
+          TileDescription([64, 128, 32], 2, [1, 4, 1], math_inst, min_cc, max_cc),
+          TileDescription([32, 128, 32], 2, [1, 2, 1], math_inst, min_cc, max_cc),
+          TileDescription([16, 128, 16], 2, [1, 2, 1], math_inst, min_cc, max_cc),
+          TileDescription([16, 128, 16], 1, [1, 1, 1], math_inst, min_cc, max_cc),
+          TileDescription([16,  64,  8], 2, [1, 1, 1], math_inst, min_cc, max_cc),
+        ] 
+        operations += GenerateConv2d(ConvKind.Dgrad, tile_descriptions, layout[0], layout[1], 
+                                     dst_layout, dst_type, min_cc, 32, 32, 32, 
+                                     True) 
+  return operations
+
+################################################################################
+# parameters
+# Edge - for tiles, the edges represent the length of one side
+# Ratio - the maximum ratio between 2 edges, limits the skinnyness of tiles
+# MaxEdge - maximum length of each edge
+# Min/Max - minimum/maximum of the product of edge lengths
+################################################################################
+
+warpsPerThreadblockEdge = [1, 2, 4, 8, 16]
+warpsPerThreadblockRatio = 2
+warpsPerThreadblockMax = 16
+# NOTE 1x32 and 2x16 warp tile shapes fail validation for ~10% of cases
+
+warpShapeEdges = [8, 16, 32, 64, 128, 256]
+warpShapeRatio = 4
+warpShapeMax = 64*64
+warpShapeMin = 8*8
+
+threadblockEdgeMax = 256
+
+#   char,         type             bits/elem, max tile,    L0 threadblock tiles
+precisions = {
+    "c" : [ "cutlass::complex<float>",   64,  64*128, [ [ 64, 128], [ 64,  32]             ] ],
+    "d" : [ "double",                    64,   64*64, [ [ 64,  64], [ 32,  32]             ] ],
+    "h" : [ "cutlass::half_t",           16, 128*256, [ [256, 128], [ 64, 128], [ 64,  32] ] ],
+    "i" : [ "int",                       32, 128*128, [ [128,  64], [ 16, 32]              ] ],
+    "s" :  [ "float",                     32, 128*128, [ [128, 256], [128, 128], [ 64,  64] ] ],
+    "z" : [ "cutlass::complex<double>", 128,   64*64, [ [ 32,  64], [ 16,  32]             ] ],
+}
+# L1 will have a single kernel for every unique shape
+# L2 will have everything else
+def GenerateGemm_Simt(args):
+  ################################################################################
+  # warps per threadblock
+  ################################################################################
+  warpsPerThreadblocks = []
+  for warpsPerThreadblock0 in warpsPerThreadblockEdge:
+      for warpsPerThreadblock1 in warpsPerThreadblockEdge:
+          if warpsPerThreadblock0 / warpsPerThreadblock1 <= warpsPerThreadblockRatio \
+         and warpsPerThreadblock1 / warpsPerThreadblock0 <= warpsPerThreadblockRatio \
+         and warpsPerThreadblock0 * warpsPerThreadblock1 <= warpsPerThreadblockMax:
+              warpsPerThreadblocks.append([warpsPerThreadblock0,
+                  warpsPerThreadblock1])
+  
+  ################################################################################
+  # warp shapes
+  ################################################################################
+  warpNumThreads = 32
+  warpShapes = []
+  for warp0 in warpShapeEdges:
+      for warp1 in warpShapeEdges:
+          if warp0 / warp1 <= warpShapeRatio \
+         and warp1 / warp0 <= warpShapeRatio \
+         and warp0 * warp1 <= warpShapeMax \
+         and warp0*warp1 > warpShapeMin:
+              warpShapes.append([warp0, warp1])
+  
+  # sgemm
+  precisionType, precisionBits, threadblockMaxElements, threadblockTilesL0 = precisions["s"]
+  
+  layouts = [
+    (LayoutType.ColumnMajor, LayoutType.ColumnMajor, LayoutType.RowMajor), # nn 
+    (LayoutType.ColumnMajor, LayoutType.RowMajor, LayoutType.RowMajor),    # nt
+    (LayoutType.RowMajor, LayoutType.ColumnMajor, LayoutType.RowMajor),    # tn
+    (LayoutType.RowMajor, LayoutType.RowMajor, LayoutType.RowMajor),       # tt
+  ]
+  
+  math_instructions = [
+    MathInstruction(                                  \
+      [1, 1, 1],                                      \
+      DataType.f32, DataType.f32, DataType.f32,       \
+      OpcodeClass.Simt,                               \
+      MathOperation.multiply_add),
+  ]
+  
+  min_cc = 50
+  max_cc = 1024
+ 
+  operations = []
+  for math_inst in math_instructions:
+    for layout in layouts:
+      data_type = [
+        math_inst.element_a,
+        math_inst.element_b,
+        math_inst.element_accumulator,
+        math_inst.element_accumulator,
+      ]
+      tile_descriptions = [
+        TileDescription([64,  256, 8], 2, [2, 4, 1], math_inst, min_cc, max_cc),
+        TileDescription([256,  64, 8], 2, [4, 2, 1], math_inst, min_cc, max_cc),
+        TileDescription([ 32, 256, 8], 2, [2, 4, 1], math_inst, min_cc, max_cc),
+        TileDescription([256,  32, 8], 2, [4, 2, 1], math_inst, min_cc, max_cc),
+        TileDescription([128, 128, 8], 2, [4, 2, 1], math_inst, min_cc, max_cc),
+        TileDescription([128,  64, 8], 2, [2, 2, 1], math_inst, min_cc, max_cc),
+        TileDescription([ 64, 128, 8], 2, [2, 2, 1], math_inst, min_cc, max_cc),
+        TileDescription([128,  32, 8], 2, [2, 1, 1], math_inst, min_cc, max_cc),
+        TileDescription([ 32, 128, 8], 2, [1, 2, 1], math_inst, min_cc, max_cc),
+        TileDescription([ 64,  64, 8], 2, [2, 1, 1], math_inst, min_cc, max_cc),
+        TileDescription([ 32,  64, 8], 2, [1, 1, 1], math_inst, min_cc, max_cc),
+        TileDescription([ 64,  32, 8], 2, [1, 1, 1], math_inst, min_cc, max_cc),
+        TileDescription([ 32,  32, 8], 2, [1, 1, 1], math_inst, min_cc, max_cc),
+        TileDescription([  8,  32, 8], 2, [1, 1, 1], math_inst, min_cc, max_cc),
+        TileDescription([ 16,  32, 8], 2, [1, 1, 1], math_inst, min_cc, max_cc),
+        TileDescription([ 16,  64, 8], 2, [1, 1, 1], math_inst, min_cc, max_cc),
+        TileDescription([ 16, 128, 8], 2, [1, 2, 1], math_inst, min_cc, max_cc),
+      ]
+      for warpsPerThreadblock in warpsPerThreadblocks:
+        for warpShape in warpShapes:
+          warpThreadsM = 0
+          if warpShape[0] > warpShape[1]:
+              warpThreadsM = 8
+          else:
+              warpThreadsM = 4
+          warpThreadsN = warpNumThreads / warpThreadsM
+  
+          # skip shapes with conflicting rectangularity
+          # they are unlikely to be fastest
+          blockG = warpsPerThreadblock[0] > warpsPerThreadblock[1]
+          blockL = warpsPerThreadblock[0] < warpsPerThreadblock[1]
+          warpG = warpShape[0] > warpShape[1]
+          warpL = warpShape[0] < warpShape[1]
+  
+          blockG2 = warpsPerThreadblock[0] > warpsPerThreadblock[1]*2
+          blockL2 = warpsPerThreadblock[0]*2 < warpsPerThreadblock[1]
+          warpG2 = warpShape[0] > warpShape[1]*2
+          warpL2 = warpShape[0]*2 < warpShape[1]
+  
+          if blockG2 and warpL: continue
+          if blockL2 and warpG: continue
+          if warpG2 and blockL: continue
+          if warpL2 and blockG: continue
+  
+          # check threadblock ratios and max
+          threadblockTile = [warpShape[0]*warpsPerThreadblock[0],
+                  warpShape[1]*warpsPerThreadblock[1]]
+          if threadblockTile[0] * threadblockTile[1] > threadblockMaxElements: continue
+          if threadblockTile[0] > threadblockEdgeMax: continue
+          if threadblockTile[1] > threadblockEdgeMax: continue
+          totalThreads = warpNumThreads*warpsPerThreadblock[0]*warpsPerThreadblock[1]
+  
+          # calculate unroll
+          # ensure that every iteration at least a full load of A,B are done
+          unrollMin = 8
+          unrollMin0 = totalThreads // threadblockTile[0]
+          unrollMin1 = totalThreads // threadblockTile[1]
+          unroll = max(unrollMin, unrollMin0, unrollMin1)
+  
+          threadTileM = warpShape[0] // warpThreadsM
+          threadTileN = warpShape[1] // warpThreadsN
+          if threadTileM < 2 or threadTileN < 2: continue
+          if threadTileM*threadTileN*precisionBits > 8*8*32: continue
+  
+          # epilogue currently only supports N < WarpNumThreads
+          if threadblockTile[1] < warpNumThreads: continue
+  
+          # limit smem
+          smemBitsA = threadblockTile[0]*unroll*2*precisionBits
+          smemBitsB = threadblockTile[1]*unroll*2*precisionBits
+          smemKBytes = (smemBitsA+smemBitsB)/8/1024
+          if (smemKBytes > 48): continue
+  
+          tile = TileDescription([threadblockTile[0], threadblockTile[1], unroll], \
+                                 2, \
+                                 [threadblockTile[0]//warpShape[0], threadblockTile[1]//warpShape[1], 1], \
+                                 math_inst, min_cc, max_cc)
+          
+          def filter(t: TileDescription) -> bool:
+            nonlocal tile
+            return t.threadblock_shape[0] == tile.threadblock_shape[0] and \
+                   t.threadblock_shape[1] == tile.threadblock_shape[1] and \
+                   t.threadblock_shape[2] == tile.threadblock_shape[2] and \
+                   t.warp_count[0] == tile.warp_count[0] and \
+                   t.warp_count[1] == tile.warp_count[1] and \
+                   t.warp_count[2] == tile.warp_count[2] and \
+                   t.stages == tile.stages
+          if not any(t for t in tile_descriptions if filter(t)): continue
+
+          operations += GeneratesGemm(tile, data_type, layout[0], layout[1], layout[2], min_cc)
+  return operations
+
+#
+def GenerateGemv_Simt(args):
+  threadBlockShape_N = [128, 64, 32]
+  ldgBits_A = [128, 64, 32]
+  ldgBits_B = [128, 64, 32]
+
+  layouts = [
+    (LayoutType.RowMajor, LayoutType.RowMajor, LayoutType.RowMajor), 
+  ]
+
+  math_instructions = [
+    MathInstruction(                                  \
+      [1, 1, 1],                                      \
+      DataType.f32, DataType.f32, DataType.f32,       \
+      OpcodeClass.Simt,                               \
+      MathOperation.multiply_add),
+  ]
+ 
+  min_cc = 50
+
+  operations = []
+  for math_inst in math_instructions:
+    for layout in layouts:
+      data_type = [
+        math_inst.element_a,
+        math_inst.element_b,
+        math_inst.element_accumulator,
+        math_inst.element_accumulator,
+      ]
+      for threadblock_shape_n in threadBlockShape_N:
+        for align_a in ldgBits_A:
+          for align_b in ldgBits_B:
+            ldg_elements_a = align_a // DataTypeSize[math_inst.element_a]
+            ldg_elements_b = align_b // DataTypeSize[math_inst.element_b]
+            threadblock_shape_k = (256 * ldg_elements_a) // (threadblock_shape_n // ldg_elements_b)
+            threadblock_shape = [1, threadblock_shape_n, threadblock_shape_k]
+            thread_shape = [1, ldg_elements_b, ldg_elements_a]
+
+            operations.append(GeneratesGemv(math_inst, \
+                                            threadblock_shape, \
+                                            thread_shape, \
+                                            data_type, \
+                                            layout[0], \
+                                            layout[1], \
+                                            layout[2], \
+                                            min_cc, \
+                                            align_a, \
+                                            align_b))
+  return operations
+
+#
+def GenerateConv2dOperations(args):
+  if args.type == "simt":
+    return GenerateConv2d_Simt(args)
+  elif args.type == "tensorop8816":
+    return GenerateConv2d_TensorOp_8816(args)
+  else:
+    assert args.type == "tensorop8832", "operation conv2d only support" \
+        "simt, tensorop8816 and tensorop8832. (got:{})".format(args.type)
+    return GenerateConv2d_TensorOp_8832(args)
+
+def GenerateDeconvOperations(args):
+  assert args.type == "simt", "operation deconv only support" \
+        "simt. (got:{})".format(args.type)
+  return GenerateDeconv_Simt(args)
+
+def GenerateGemmOperations(args):
+  assert args.type == "simt", "operation gemm only support" \
+        "simt. (got:{})".format(args.type)
+  return GenerateGemm_Simt(args)
+
+def GenerateGemvOperations(args):
+  assert args.type == "simt", "operation gemv only support" \
+        "simt. (got:{})".format(args.type)
+  return GenerateGemv_Simt(args)
+
+###################################################################################################
+###################################################################################################
+
+if __name__ == "__main__":
+
+  parser = argparse.ArgumentParser(description="Generates device kernel registration code for CUTLASS Kernels")
+  parser.add_argument("--operations", type=str, choices=['gemm', 'gemv', 'conv2d', 'deconv'], 
+                      required=True, help="Specifies the operation to generate (gemm, gemv, conv2d, deconv)")
+  parser.add_argument("output", type=str, help="output directory for CUTLASS kernel files")
+  parser.add_argument("--type", type=str, choices=['simt', 'tensorop8816', 'tensorop8832'], 
+                      default='simt', help="kernel type of CUTLASS kernel generator")
+
+  operation2wrapper_path = {
+    "gemm": "src/cuda/matrix_mul/cutlass_matrix_mul_wrapper.cuinl", \
+    "gemv": "src/cuda/matrix_mul/cutlass_matrix_mul_wrapper_batched_gemv_strided.cuinl", \
+    "conv2d": "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl", \
+    "deconv": "src/cuda/convolution/backward_data/implicit_gemm_deconv_cutlass_wrapper.cuinl", \
+  }
+
+  args = parser.parse_args()
+
+  wrapper_path = operation2wrapper_path[args.operations] 
+  if args.operations == "gemm":
+    operations = GenerateGemmOperations(args)
+  elif args.operations == "gemv":
+    operations = GenerateGemvOperations(args)
+  elif args.operations == "conv2d":
+    operations = GenerateConv2dOperations(args)
+  elif args.operations == "deconv":
+    operations = GenerateDeconvOperations(args)
+  
+
+  if args.operations == "conv2d" or args.operations == "deconv":
+    for operation in operations:
+      with EmitConvSingleKernelWrapper(args.output, operation, wrapper_path) as emitter:
+        emitter.emit()
+  elif args.operations == "gemm" or args.operations == "gemv":
+    for operation in operations:
+      with EmitGemmSingleKernelWrapper(args.output, operation, wrapper_path) as emitter:
+        emitter.emit()
+  
+#
+###################################################################################################

dnn/scripts/cutlass_generator/lazy_file.py

@@ -0,0 +1,27 @@
+#
+# \file lazy_file.py
+#
+# \brief LazyFile updates the target file only when the content is changed
+#        in order to avoid generating new cutlass kimpls each time cmake is called
+#
+
+import io
+import os
+
+class LazyFile:
+    def __init__(self, filename):
+        self.filename = filename
+        self.buffer = io.StringIO()
+
+    def write(self, data):
+        self.buffer.write(str(data))
+
+    def close(self):
+        if os.path.isfile(self.filename):
+            old_data = open(self.filename).read()
+        else:
+            old_data = ""
+        new_data = self.buffer.getvalue()
+        if old_data != new_data:
+            with open(self.filename, "w") as f:
+                f.write(new_data)

dnn/scripts/cutlass_generator/library.py

@@ -0,0 +1,614 @@
+#
+# \file generator.py
+#
+# \brief Generates the CUTLASS Library's instances
+#
+
+import re
+
+###################################################################################################
+
+import enum
+
+# The following block implements enum.auto() for Python 3.5 variants that don't include it such
+# as the default 3.5.2 on Ubuntu 16.04.
+# 
+# https://codereview.stackexchange.com/questions/177309/reimplementing-pythons-enum-auto-for-compatibility
+
+try:
+  from enum import auto as enum_auto
+except ImportError: 
+  __cutlass_library_auto_enum = 0
+  def enum_auto() -> int:
+    global __cutlass_library_auto_enum
+    i = __cutlass_library_auto_enum
+    __cutlass_library_auto_enum += 1
+    return i
+
+###################################################################################################
+
+#
+class GeneratorTarget(enum.Enum):
+  Library = enum_auto()
+#
+GeneratorTargetNames = {
+  GeneratorTarget.Library: 'library'
+}
+#
+
+###################################################################################################
+
+#
+class DataType(enum.Enum):
+  b1 = enum_auto()
+  u4 = enum_auto()
+  u8 = enum_auto()
+  u16 = enum_auto()
+  u32 = enum_auto()
+  u64 = enum_auto()
+  s4 = enum_auto()
+  s8 = enum_auto()
+  s16 = enum_auto()
+  s32 = enum_auto()
+  s64 = enum_auto()
+  f16 = enum_auto()
+  bf16 = enum_auto()
+  f32 = enum_auto()
+  tf32 = enum_auto()
+  f64 = enum_auto()
+  cf16 = enum_auto()
+  cbf16 = enum_auto()
+  cf32 = enum_auto()
+  ctf32 = enum_auto()
+  cf64 = enum_auto()
+  cs4 = enum_auto()
+  cs8 = enum_auto()
+  cs16 = enum_auto()
+  cs32 = enum_auto()
+  cs64 = enum_auto()
+  cu4 = enum_auto()
+  cu8 = enum_auto()
+  cu16 = enum_auto()
+  cu32 = enum_auto()
+  cu64 = enum_auto()
+  invalid = enum_auto()
+
+#
+ShortDataTypeNames = {
+  DataType.s32: 'i',
+  DataType.f16: 'h',
+  DataType.f32: 's',
+  DataType.f64: 'd',
+  DataType.cf32: 'c',
+  DataType.cf64: 'z',
+}
+
+#
+DataTypeNames = {
+  DataType.b1: "b1",
+  DataType.u4: "u4",
+  DataType.u8: "u8",
+  DataType.u16: "u16",
+  DataType.u32: "u32",
+  DataType.u64: "u64",
+  DataType.s4: "s4",
+  DataType.s8: "s8",
+  DataType.s16: "s16",
+  DataType.s32: "s32",
+  DataType.s64: "s64",
+  DataType.f16: "f16",
+  DataType.bf16: "bf16",
+  DataType.f32: "f32",
+  DataType.tf32: "tf32",
+  DataType.f64: "f64",
+  DataType.cf16: "cf16",
+  DataType.cbf16: "cbf16",
+  DataType.cf32: "cf32",
+  DataType.ctf32: "ctf32",
+  DataType.cf64: "cf64",
+  DataType.cu4: "cu4",
+  DataType.cu8: "cu8",
+  DataType.cu16: "cu16",
+  DataType.cu32: "cu32",
+  DataType.cu64: "cu64",
+  DataType.cs4: "cs4",
+  DataType.cs8: "cs8",
+  DataType.cs16: "cs16",
+  DataType.cs32: "cs32",
+  DataType.cs64: "cs64", 
+}
+
+DataTypeTag = {
+  DataType.b1: "cutlass::uint1b_t",
+  DataType.u4: "cutlass::uint4b_t",
+  DataType.u8: "uint8_t",
+  DataType.u16: "uint16_t",
+  DataType.u32: "uint32_t",
+  DataType.u64: "uint64_t",
+  DataType.s4: "cutlass::int4b_t",
+  DataType.s8: "int8_t",
+  DataType.s16: "int16_t",
+  DataType.s32: "int32_t",
+  DataType.s64: "int64_t",
+  DataType.f16: "cutlass::half_t",
+  DataType.bf16: "cutlass::bfloat16_t",
+  DataType.f32: "float",
+  DataType.tf32: "cutlass::tfloat32_t",
+  DataType.f64: "double",
+  DataType.cf16: "cutlass::complex<cutlass::half_t>",
+  DataType.cbf16: "cutlass::complex<cutlass::bfloat16_t>",
+  DataType.cf32: "cutlass::complex<float>",
+  DataType.ctf32: "cutlass::complex<cutlass::tfloat32_t>",
+  DataType.cf64: "cutlass::complex<double>",
+  DataType.cu4: "cutlass::complex<cutlass::uint4b_t>",
+  DataType.cu8: "cutlass::complex<cutlass::uint8_t>",
+  DataType.cu16: "cutlass::complex<cutlass::uint16_t>",
+  DataType.cu32: "cutlass::complex<cutlass::uint32_t>",
+  DataType.cu64: "cutlass::complex<cutlass::uint64_t>",
+  DataType.cs4: "cutlass::complex<cutlass::int4b_t>",
+  DataType.cs8: "cutlass::complex<cutlass::int8_t>",
+  DataType.cs16: "cutlass::complex<cutlass::int16_t>",
+  DataType.cs32: "cutlass::complex<cutlass::int32_t>",
+  DataType.cs64: "cutlass::complex<cutlass::int64_t>",
+}
+
+DataTypeSize = {
+  DataType.b1: 1,
+  DataType.u4: 4,
+  DataType.u8: 4,
+  DataType.u16: 16,
+  DataType.u32: 32,
+  DataType.u64: 64,
+  DataType.s4: 4,
+  DataType.s8: 8,
+  DataType.s16: 16,
+  DataType.s32: 32,
+  DataType.s64: 64,
+  DataType.f16: 16,
+  DataType.bf16: 16,
+  DataType.f32: 32,
+  DataType.tf32: 32,
+  DataType.f64: 64,
+  DataType.cf16: 32,
+  DataType.cbf16: 32,
+  DataType.cf32: 64,
+  DataType.ctf32: 32,
+  DataType.cf64: 128,
+  DataType.cu4: 8,
+  DataType.cu8: 16,
+  DataType.cu16: 32,
+  DataType.cu32: 64,
+  DataType.cu64: 128,
+  DataType.cs4: 8,
+  DataType.cs8: 16,
+  DataType.cs16: 32,
+  DataType.cs32: 64,
+  DataType.cs64: 128,
+}
+
+###################################################################################################
+
+#
+class ComplexTransform(enum.Enum):
+  none = enum_auto()
+  conj = enum_auto()
+
+#
+ComplexTransformTag = {
+  ComplexTransform.none: 'cutlass::ComplexTransform::kNone',
+  ComplexTransform.conj: 'cutlass::ComplexTransform::kConjugate',
+}
+
+#
+RealComplexBijection = [
+  (DataType.f16, DataType.cf16),
+  (DataType.f32, DataType.cf32),
+  (DataType.f64, DataType.cf64),
+]
+
+#
+def is_complex(data_type):
+  for r, c in RealComplexBijection:
+    if data_type == c:
+      return True
+  return False
+
+#
+def get_complex_from_real(real_type):
+  for r, c in RealComplexBijection:
+    if real_type == r:
+      return c
+  return DataType.invalid
+
+#
+def get_real_from_complex(complex_type):
+  for r, c in RealComplexBijection:
+    if complex_type == c:
+      return r
+  return DataType.invalid
+
+#
+class ComplexMultiplyOp(enum.Enum):
+  multiply_add = enum_auto()
+  gaussian = enum_auto()
+
+###################################################################################################
+
+#
+class MathOperation(enum.Enum):
+  multiply_add = enum_auto()
+  multiply_add_saturate = enum_auto()
+  xor_popc = enum_auto()
+  multiply_add_fast_bf16 = enum_auto()
+  multiply_add_fast_f16 = enum_auto()
+  multiply_add_complex = enum_auto()
+  multiply_add_complex_gaussian = enum_auto()
+
+#
+MathOperationTag = {
+  MathOperation.multiply_add: 'cutlass::arch::OpMultiplyAdd', 
+  MathOperation.multiply_add_saturate: 'cutlass::arch::OpMultiplyAddSaturate',
+  MathOperation.xor_popc: 'cutlass::arch::OpXorPopc',
+  MathOperation.multiply_add_fast_bf16: 'cutlass::arch::OpMultiplyAddFastBF16',
+  MathOperation.multiply_add_fast_f16: 'cutlass::arch::OpMultiplyAddFastF16',
+  MathOperation.multiply_add_complex: 'cutlass::arch::OpMultiplyAddComplex',
+  MathOperation.multiply_add_complex_gaussian: 'cutlass::arch::OpMultiplyAddGaussianComplex',
+}
+
+###################################################################################################
+
+#
+class LayoutType(enum.Enum):
+  ColumnMajor = enum_auto()
+  RowMajor = enum_auto()
+  ColumnMajorInterleaved2 = enum_auto()
+  RowMajorInterleaved2 = enum_auto()
+  ColumnMajorInterleaved32 = enum_auto()
+  RowMajorInterleaved32 = enum_auto()
+  ColumnMajorInterleaved64 = enum_auto()
+  RowMajorInterleaved64 = enum_auto()
+  TensorNHWC = enum_auto()
+  TensorNDHWC = enum_auto()
+  TensorNCHW = enum_auto()
+  TensorNGHWC = enum_auto()
+  TensorNC4HW4 = enum_auto()
+  TensorC4RSK4 = enum_auto()
+  TensorNC8HW8 = enum_auto()
+  TensorNC16HW16 = enum_auto()
+  TensorNC32HW32 = enum_auto()
+  TensorNC64HW64 = enum_auto()
+  TensorC32RSK32 = enum_auto()
+  TensorC64RSK64 = enum_auto()
+  TensorK4RSC4 = enum_auto()
+
+#
+LayoutTag = {
+  LayoutType.ColumnMajor: 'cutlass::layout::ColumnMajor',
+  LayoutType.RowMajor: 'cutlass::layout::RowMajor',
+  LayoutType.ColumnMajorInterleaved2: 'cutlass::layout::ColumnMajorInterleaved<2>',
+  LayoutType.RowMajorInterleaved2: 'cutlass::layout::RowMajorInterleaved<2>',
+  LayoutType.ColumnMajorInterleaved32: 'cutlass::layout::ColumnMajorInterleaved<32>',
+  LayoutType.RowMajorInterleaved32: 'cutlass::layout::RowMajorInterleaved<32>',
+  LayoutType.ColumnMajorInterleaved64: 'cutlass::layout::ColumnMajorInterleaved<64>',
+  LayoutType.RowMajorInterleaved64: 'cutlass::layout::RowMajorInterleaved<64>',
+  LayoutType.TensorNHWC: 'cutlass::layout::TensorNHWC',
+  LayoutType.TensorNDHWC: 'cutlass::layout::TensorNDHWC',
+  LayoutType.TensorNCHW: 'cutlass::layout::TensorNCHW',
+  LayoutType.TensorNGHWC: 'cutlass::layout::TensorNGHWC',
+  LayoutType.TensorNC4HW4: 'cutlass::layout::TensorNCxHWx<4>',
+  LayoutType.TensorC4RSK4: 'cutlass::layout::TensorCxRSKx<4>',
+  LayoutType.TensorNC8HW8: 'cutlass::layout::TensorNCxHWx<8>',
+  LayoutType.TensorNC16HW16: 'cutlass::layout::TensorNCxHWx<16>',
+  LayoutType.TensorNC32HW32: 'cutlass::layout::TensorNCxHWx<32>',
+  LayoutType.TensorC32RSK32: 'cutlass::layout::TensorCxRSKx<32>',
+  LayoutType.TensorNC64HW64: 'cutlass::layout::TensorNCxHWx<64>',
+  LayoutType.TensorC64RSK64: 'cutlass::layout::TensorCxRSKx<64>',
+  LayoutType.TensorK4RSC4: 'cutlass::layout::TensorKxRSCx<4>', 
+}
+
+#
+TransposedLayout = {
+  LayoutType.ColumnMajor: LayoutType.RowMajor,
+  LayoutType.RowMajor: LayoutType.ColumnMajor,
+  LayoutType.ColumnMajorInterleaved2: LayoutType.RowMajorInterleaved2,
+  LayoutType.RowMajorInterleaved2: LayoutType.ColumnMajorInterleaved2,
+  LayoutType.ColumnMajorInterleaved32: LayoutType.RowMajorInterleaved32,
+  LayoutType.RowMajorInterleaved32: LayoutType.ColumnMajorInterleaved32,
+  LayoutType.ColumnMajorInterleaved64: LayoutType.RowMajorInterleaved64,
+  LayoutType.RowMajorInterleaved64: LayoutType.ColumnMajorInterleaved64,
+  LayoutType.TensorNHWC: LayoutType.TensorNHWC
+}
+
+#
+ShortLayoutTypeNames = {
+  LayoutType.ColumnMajor: 'n',
+  LayoutType.ColumnMajorInterleaved32: 'n2',
+  LayoutType.ColumnMajorInterleaved32: 'n32',
+  LayoutType.ColumnMajorInterleaved64: 'n64',
+  LayoutType.RowMajor: 't',
+  LayoutType.RowMajorInterleaved2: 't2',
+  LayoutType.RowMajorInterleaved32: 't32',
+  LayoutType.RowMajorInterleaved64: 't64',
+  LayoutType.TensorNHWC: 'nhwc',
+  LayoutType.TensorNDHWC: 'ndhwc',
+  LayoutType.TensorNCHW: 'nchw',
+  LayoutType.TensorNGHWC: 'nghwc',
+  LayoutType.TensorNC4HW4: 'nc4hw4', 
+  LayoutType.TensorC4RSK4: 'c4rsk4', 
+  LayoutType.TensorNC8HW8: 'nc8hw8',
+  LayoutType.TensorNC16HW16: 'nc16hw16',
+  LayoutType.TensorNC32HW32: 'nc32hw32',
+  LayoutType.TensorNC64HW64: 'nc64hw64',
+  LayoutType.TensorC32RSK32: 'c32rsk32',
+  LayoutType.TensorC64RSK64: 'c64rsk64', 
+  LayoutType.TensorK4RSC4: 'k4rsc4', 
+}
+
+#
+ShortComplexLayoutNames = {
+  (LayoutType.ColumnMajor, ComplexTransform.none): 'n',
+  (LayoutType.ColumnMajor, ComplexTransform.conj): 'c',
+  (LayoutType.RowMajor, ComplexTransform.none): 't',
+  (LayoutType.RowMajor, ComplexTransform.conj): 'h'
+}
+
+###################################################################################################
+#
+class OpcodeClass(enum.Enum):
+  Simt = enum_auto()
+  TensorOp = enum_auto()
+  WmmaTensorOp = enum_auto()
+
+OpcodeClassNames = {
+  OpcodeClass.Simt: 'simt',
+  OpcodeClass.TensorOp: 'tensorop',
+  OpcodeClass.WmmaTensorOp: 'wmma_tensorop',
+}
+
+OpcodeClassTag = {
+  OpcodeClass.Simt: 'cutlass::arch::OpClassSimt',
+  OpcodeClass.TensorOp: 'cutlass::arch::OpClassTensorOp',
+  OpcodeClass.WmmaTensorOp: 'cutlass::arch::OpClassWmmaTensorOp',
+}
+
+###################################################################################################
+
+#
+class OperationKind(enum.Enum):
+  Gemm = enum_auto()
+  Conv2d = enum_auto()    
+
+#
+OperationKindNames = {
+  OperationKind.Gemm: 'gemm'
+  , OperationKind.Conv2d: 'conv2d'
+}
+
+# 
+class Target(enum.Enum):
+  library = enum_auto()
+
+ArchitectureNames = {
+  50: 'maxwell',
+  60: 'pascal',
+  61: 'pascal',
+  70: 'volta',
+  75: 'turing',
+  80: 'ampere',
+}
+
+###################################################################################################
+
+#
+def SubstituteTemplate(template, values):
+  text = template
+  changed = True
+  while changed:
+    changed = False
+    for key, value in values.items():
+      regex = "\\$\\{%s\\}" % key
+      newtext = re.sub(regex, value, text)
+      if newtext != text:
+        changed = True
+      text = newtext
+  return text
+
+###################################################################################################
+
+#
+class GemmKind(enum.Enum):
+  Gemm = enum_auto()
+  Sparse = enum_auto()
+  Universal = enum_auto()
+  PlanarComplex = enum_auto()
+  PlanarComplexArray = enum_auto()
+  SplitKParallel = enum_auto()
+  GemvBatchedStrided = enum_auto()
+
+#
+GemmKindNames = {
+  GemmKind.Gemm: "gemm",
+  GemmKind.Sparse: "spgemm",
+  GemmKind.Universal: "gemm",
+  GemmKind.PlanarComplex: "gemm_planar_complex",
+  GemmKind.PlanarComplexArray: "gemm_planar_complex_array",
+  GemmKind.SplitKParallel: "gemm_split_k_parallel", 
+  GemmKind.GemvBatchedStrided: "gemv_batched_strided", 
+}
+
+#
+class EpilogueFunctor(enum.Enum):
+  LinearCombination = enum_auto()
+  LinearCombinationClamp = enum_auto()
+  BiasAddLinearCombination = enum_auto()
+  BiasAddLinearCombinationRelu = enum_auto()
+  BiasAddLinearCombinationHSwish = enum_auto()
+  BiasAddLinearCombinationClamp = enum_auto()
+  BiasAddLinearCombinationReluClamp = enum_auto()
+  BiasAddLinearCombinationHSwishClamp = enum_auto()
+
+
+#
+EpilogueFunctorTag = {
+  EpilogueFunctor.LinearCombination: 'cutlass::epilogue::thread::LinearCombination',
+  EpilogueFunctor.LinearCombinationClamp: 'cutlass::epilogue::thread::LinearCombinationClamp',
+  EpilogueFunctor.BiasAddLinearCombination: 'cutlass::epilogue::thread::BiasAddLinearCombination', 
+  EpilogueFunctor.BiasAddLinearCombinationRelu: 'cutlass::epilogue::thread::BiasAddLinearCombinationRelu', 
+  EpilogueFunctor.BiasAddLinearCombinationHSwish: 'cutlass::epilogue::thread::BiasAddLinearCombinationHSwish', 
+  EpilogueFunctor.BiasAddLinearCombinationClamp: 'cutlass::epilogue::thread::BiasAddLinearCombinationClamp', 
+  EpilogueFunctor.BiasAddLinearCombinationReluClamp: 'cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp', 
+  EpilogueFunctor.BiasAddLinearCombinationHSwishClamp: 'cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp', 
+}
+
+#
+ShortEpilogueNames = {
+  EpilogueFunctor.BiasAddLinearCombinationHSwishClamp: 'hswish', 
+  EpilogueFunctor.BiasAddLinearCombinationReluClamp: 'relu', 
+  EpilogueFunctor.BiasAddLinearCombinationClamp: 'identity', 
+  EpilogueFunctor.BiasAddLinearCombinationHSwish: 'hswish', 
+  EpilogueFunctor.BiasAddLinearCombinationRelu: 'relu', 
+  EpilogueFunctor.BiasAddLinearCombination: 'identity', 
+}
+
+
+
+
+
+
+#
+class SwizzlingFunctor(enum.Enum):
+  Identity1 = enum_auto()
+  Identity2 = enum_auto()
+  Identity4 = enum_auto()
+  Identity8 = enum_auto()
+  ConvFpropNCxHWx = enum_auto()
+  ConvFpropNHWC = enum_auto()
+  ConvDgradNCxHWx = enum_auto()
+
+#
+SwizzlingFunctorTag = {
+  SwizzlingFunctor.Identity1: 'cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>',
+  SwizzlingFunctor.Identity2: 'cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<2>',
+  SwizzlingFunctor.Identity4: 'cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<4>',
+  SwizzlingFunctor.Identity8: 'cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<8>',
+  SwizzlingFunctor.ConvFpropNCxHWx: 'cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle',  
+  SwizzlingFunctor.ConvFpropNHWC: 'cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle',  
+  SwizzlingFunctor.ConvDgradNCxHWx: 'cutlass::conv::threadblock::ConvolutionDgradNCxHWxThreadblockSwizzle', 
+}
+
+###################################################################################################
+
+class ConvType(enum.Enum):
+  Convolution = enum_auto()
+  BatchConvolution = enum_auto()
+  Local = enum_auto()
+  LocalShare = enum_auto()
+
+ConvTypeTag = {
+    ConvType.Convolution: 'cutlass::conv::ConvType::kConvolution', 
+    ConvType.BatchConvolution: 'cutlass::conv::ConvType::kBatchConvolution', 
+    ConvType.Local: 'cutlass::conv::ConvType::kLocal', 
+    ConvType.LocalShare : 'cutlass::conv::ConvType::kLocalShare', 
+}
+
+#
+class ConvKind(enum.Enum):
+  Fprop = enum_auto()
+  Dgrad = enum_auto()
+  Wgrad = enum_auto()
+
+#
+ConvKindTag = {
+  ConvKind.Fprop: 'cutlass::conv::Operator::kFprop',
+  ConvKind.Dgrad: 'cutlass::conv::Operator::kDgrad',
+  ConvKind.Wgrad: 'cutlass::conv::Operator::kWgrad'
+}
+
+ConvKindNames = {
+  ConvKind.Fprop: 'fprop',
+  ConvKind.Dgrad: 'dgrad',
+  ConvKind.Wgrad: 'wgrad',
+}
+
+#
+class IteratorAlgorithm(enum.Enum):
+  Analytic = enum_auto()
+  Optimized = enum_auto()
+
+#
+IteratorAlgorithmTag = {
+  IteratorAlgorithm.Analytic: 'cutlass::conv::IteratorAlgorithm::kAnalytic',
+  IteratorAlgorithm.Optimized: 'cutlass::conv::IteratorAlgorithm::kOptimized',
+}
+
+IteratorAlgorithmNames = {
+  IteratorAlgorithm.Analytic: 'analytic',
+  IteratorAlgorithm.Optimized: 'optimized',
+}
+
+#
+class StrideSupport(enum.Enum):
+  Strided = enum_auto()
+  Unity = enum_auto()
+
+#
+StrideSupportTag = {
+  StrideSupport.Strided: 'cutlass::conv::StrideSupport::kStrided',
+  StrideSupport.Unity: 'cutlass::conv::StrideSupport::kUnity',
+}
+
+StrideSupportNames = {
+  StrideSupport.Strided: '',
+  StrideSupport.Unity: 'unity_stride',
+}
+
+class ImplicitGemmMode(enum.Enum):
+  GemmNt = enum_auto()
+  GemmTn = enum_auto()
+
+ImplicitGemmModeNames = {
+  ImplicitGemmMode.GemmNt: 'gemm_nt',
+  ImplicitGemmMode.GemmTn: 'gemm_tn',
+}
+
+ImplicitGemmModeTag = {
+  ImplicitGemmMode.GemmNt: 'cutlass::conv::ImplicitGemmMode::GEMM_NT',
+  ImplicitGemmMode.GemmTn: 'cutlass::conv::ImplicitGemmMode::GEMM_TN',
+}
+
+###################################################################################################
+
+#
+class MathInstruction:
+  def __init__(self, instruction_shape, element_a, element_b, element_accumulator, opcode_class, math_operation = MathOperation.multiply_add):
+    self.instruction_shape = instruction_shape
+    self.element_a = element_a
+    self.element_b = element_b
+    self.element_accumulator = element_accumulator
+    self.opcode_class = opcode_class
+    self.math_operation = math_operation
+
+
+#
+class TileDescription:
+
+  def __init__(self, threadblock_shape, stages, warp_count, math_instruction, min_compute, max_compute):
+    self.threadblock_shape = threadblock_shape
+    self.stages = stages
+    self.warp_count = warp_count
+    self.math_instruction = math_instruction
+    self.minimum_compute_capability = min_compute
+    self.maximum_compute_capability = max_compute
+
+  def procedural_name(self):
+    return "%dx%d_%dx%d" % (self.threadblock_shape[0], self.threadblock_shape[1], self.threadblock_shape[2], self.stages)
+
+#
+class TensorDescription:
+  def __init__(self, element, layout, alignment = 1, complex_transform = ComplexTransform.none):
+    self.element = element
+    self.layout = layout
+    self.alignment = alignment
+    self.complex_transform = complex_transform
+
+###################################################################################################

dnn/scripts/cutlass_generator/list.bzl

@@ -0,0 +1,578 @@
+# Generated by dnn/scripts/cutlass_generator/gen_list.py
+
+cutlass_gen_list = [
+    "cutlass_simt_sgemm_8x32_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_8x32_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_16x32_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_16x32_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_16x64_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_16x64_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_32x32_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x32_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_32x64_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x64_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_64x32_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x32_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_16x128_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_16x128_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_32x128_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x128_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_64x64_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x64_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_128x32_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_128x32_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_64x128_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x128_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_128x64_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_128x64_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_32x256_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x256_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_64x256_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x256_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_128x128_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_128x128_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_256x32_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_256x32_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_256x64_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_256x64_8x2_nn_align1.cu",
+    "cutlass_simt_sgemm_8x32_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_8x32_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_16x32_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_16x32_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_16x64_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_16x64_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_32x32_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x32_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_32x64_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x64_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_64x32_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x32_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_16x128_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_16x128_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_32x128_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x128_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_64x64_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x64_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_128x32_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_128x32_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_64x128_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x128_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_128x64_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_128x64_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_32x256_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x256_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_64x256_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x256_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_128x128_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_128x128_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_256x32_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_256x32_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_256x64_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_256x64_8x2_nt_align1.cu",
+    "cutlass_simt_sgemm_8x32_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_8x32_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_16x32_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_16x32_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_16x64_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_16x64_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_32x32_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x32_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_32x64_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x64_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_64x32_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x32_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_16x128_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_16x128_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_32x128_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x128_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_64x64_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x64_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_128x32_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_128x32_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_64x128_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x128_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_128x64_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_128x64_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_32x256_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x256_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_64x256_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x256_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_128x128_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_128x128_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_256x32_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_256x32_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_256x64_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_256x64_8x2_tn_align1.cu",
+    "cutlass_simt_sgemm_8x32_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_8x32_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_16x32_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_16x32_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_16x64_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_16x64_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_32x32_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x32_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_32x64_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x64_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_64x32_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x32_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_16x128_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_16x128_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_32x128_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x128_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_64x64_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x64_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_128x32_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_128x32_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_64x128_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x128_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_128x64_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_128x64_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_32x256_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_32x256_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_64x256_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_64x256_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_128x128_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_128x128_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_256x32_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_256x32_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_256x64_8x2_tt_align1.cu",
+    "cutlass_simt_sgemm_split_k_parallel_256x64_8x2_tt_align1.cu",
+    "cutlass_simt_sgemv_batched_strided_1x128_32_tt_align4x4.cu",
+    "cutlass_simt_sgemv_batched_strided_1x128_16_tt_align4x2.cu",
+    "cutlass_simt_sgemv_batched_strided_1x128_8_tt_align4x1.cu",
+    "cutlass_simt_sgemv_batched_strided_1x128_16_tt_align2x4.cu",
+    "cutlass_simt_sgemv_batched_strided_1x128_8_tt_align2x2.cu",
+    "cutlass_simt_sgemv_batched_strided_1x128_4_tt_align2x1.cu",
+    "cutlass_simt_sgemv_batched_strided_1x128_8_tt_align1x4.cu",
+    "cutlass_simt_sgemv_batched_strided_1x128_4_tt_align1x2.cu",
+    "cutlass_simt_sgemv_batched_strided_1x128_2_tt_align1x1.cu",
+    "cutlass_simt_sgemv_batched_strided_1x64_64_tt_align4x4.cu",
+    "cutlass_simt_sgemv_batched_strided_1x64_32_tt_align4x2.cu",
+    "cutlass_simt_sgemv_batched_strided_1x64_16_tt_align4x1.cu",
+    "cutlass_simt_sgemv_batched_strided_1x64_32_tt_align2x4.cu",
+    "cutlass_simt_sgemv_batched_strided_1x64_16_tt_align2x2.cu",
+    "cutlass_simt_sgemv_batched_strided_1x64_8_tt_align2x1.cu",
+    "cutlass_simt_sgemv_batched_strided_1x64_16_tt_align1x4.cu",
+    "cutlass_simt_sgemv_batched_strided_1x64_8_tt_align1x2.cu",
+    "cutlass_simt_sgemv_batched_strided_1x64_4_tt_align1x1.cu",
+    "cutlass_simt_sgemv_batched_strided_1x32_128_tt_align4x4.cu",
+    "cutlass_simt_sgemv_batched_strided_1x32_64_tt_align4x2.cu",
+    "cutlass_simt_sgemv_batched_strided_1x32_32_tt_align4x1.cu",
+    "cutlass_simt_sgemv_batched_strided_1x32_64_tt_align2x4.cu",
+    "cutlass_simt_sgemv_batched_strided_1x32_32_tt_align2x2.cu",
+    "cutlass_simt_sgemv_batched_strided_1x32_16_tt_align2x1.cu",
+    "cutlass_simt_sgemv_batched_strided_1x32_32_tt_align1x4.cu",
+    "cutlass_simt_sgemv_batched_strided_1x32_16_tt_align1x2.cu",
+    "cutlass_simt_sgemv_batched_strided_1x32_8_tt_align1x1.cu",
+    "cutlass_simt_s8_idgrad_identity_s8_64x128x32_64x32x32_2_nc4hw4_k4rsc4.cu",
+    "cutlass_simt_s8_idgrad_identity_s8_32x128x32_32x64x32_2_nc4hw4_k4rsc4.cu",
+    "cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x64x16_2_nc4hw4_k4rsc4.cu",
+    "cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x128x16_1_nc4hw4_k4rsc4.cu",
+    "cutlass_simt_s8_idgrad_identity_s8_16x64x8_16x64x8_2_nc4hw4_k4rsc4.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu",
+    "cutlass_simt_u4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_u4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_s4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu",
+    "cutlass_simt_f32_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_simt_f32_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu",
+    "cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu",
+    "cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu",
+    "cutlass_tensorop_s4_i8832fprop_identity_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu",
+    "cutlass_tensorop_s4_i8832fprop_relu_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu",
+    "cutlass_tensorop_s4_i8832fprop_hswish_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu",
+    "cutlass_tensorop_s4_i8832fprop_identity_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu",
+    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu",
+    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu",
+    "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu",
+    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu",
+    "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu",
+    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu",
+    "cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu",
+    "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu",
+    "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu",
+    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu",
+]

dnn/scripts/cutlass_generator/manifest.py

@@ -0,0 +1,351 @@
+#
+# \file generator.py
+#
+# \brief Generates the CUTLASS Library's instances
+#
+
+import enum
+import os.path
+import shutil
+
+from library import *
+from gemm_operation import *
+from conv2d_operation import *  
+
+###################################################################################################
+
+class EmitOperationKindLibrary:
+  def __init__(self, generated_path, kind, args):
+    self.generated_path = generated_path
+    self.kind = kind
+    self.args = args
+
+    self.emitters = {
+      OperationKind.Gemm: EmitGemmConfigurationLibrary
+      , OperationKind.Conv2d: EmitConv2dConfigurationLibrary
+    }
+
+    self.configurations = [];
+
+    self.header_template ="""
+/*
+ Generated by manifest.py - Do not edit.
+*/
+
+#include "cutlass/cutlass.h"
+#include "cutlass/library/library.h"
+#include "cutlass/library/manifest.h"
+
+namespace cutlass {
+namespace library {
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+"""
+    self.entry_template = """
+
+//
+// Entry point to construct operations
+//
+void initialize_all_${operation_name}_operations(Manifest &manifest) {
+"""
+    self.configuration_prototype_template = "void initialize_${configuration_name}(Manifest &manifest);\n"
+    self.configuration_template ="  initialize_${configuration_name}(manifest);\n"
+
+    self.epilogue_template ="""
+
+}
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+} // namespace library
+} // namespace cutlass
+
+"""
+
+  #
+  def __enter__(self):
+    self.operation_path = os.path.join(self.generated_path, OperationKindNames[self.kind])
+    os.mkdir(self.operation_path)
+
+    self.top_level_path = os.path.join(self.operation_path, "all_%s_operations.cu" % OperationKindNames[self.kind])
+
+    self.top_level_file = open(self.top_level_path, "w")
+    self.top_level_file.write(self.header_template)
+
+    self.source_files = [self.top_level_path,]
+
+    return self
+
+  #
+  def emit(self, configuration_name, operations):
+
+    with self.emitters[self.kind](self.operation_path, configuration_name) as configuration_emitter:
+      for operation in operations:
+        configuration_emitter.emit(operation)
+      
+      self.source_files.append(configuration_emitter.configuration_path)
+
+    self.configurations.append(configuration_name)
+    self.top_level_file.write(SubstituteTemplate(self.configuration_prototype_template, {'configuration_name': configuration_name} ))
+
+  #
+  def __exit__(self, exception_type, exception_value, traceback):
+    self.top_level_file.write(SubstituteTemplate(self.entry_template, {'operation_name': OperationKindNames[self.kind]}))
+
+    for configuration_name in self.configurations:
+      self.top_level_file.write(SubstituteTemplate(self.configuration_template, {'configuration_name': configuration_name}))
+
+    self.top_level_file.write(self.epilogue_template)
+    self.top_level_file.close()
+
+###################################################################################################
+###################################################################################################
+
+class Options:
+  def __init__(self):
+    pass
+
+###################################################################################################
+
+#
+class Manifest:
+
+  #
+  def __init__(self, args):
+    self.operations = {}
+    self.args = args
+
+    architectures = args.architectures.split(';') if len(args.architectures) else ['50',]
+    self.compute_capabilities = [int(x) for x in architectures]
+    
+    self.selected_kernels = []
+    
+    if args.operations == 'all':
+      self.operations_enabled = []
+    else:
+
+      operations_list = [
+        OperationKind.Gemm
+        , OperationKind.Conv2d     
+      ] 
+
+      self.operations_enabled = [x for x in operations_list if OperationKindNames[x] in args.operations.split(',')]
+
+    if args.kernels == 'all':
+      self.kernel_names = []
+    else:
+      self.kernel_names = [x for x in args.kernels.split(',') if x != '']
+
+    self.ignore_kernel_names = [x for x in args.ignore_kernels.split(',') if x != '']
+
+    if args.kernel_filter_file is None:
+        self.kernel_filter_list = []
+    else:
+        self.kernel_filter_list = self.get_kernel_filters(args.kernel_filter_file)
+
+
+    self.operation_count = 0
+    self.operations_by_name = {}
+    self.top_level_prologue = '''
+
+#include "cutlass/library/library.h"
+#include "cutlass/library/manifest.h"
+
+namespace cutlass {
+namespace library {
+
+${prototypes}
+
+void initialize_all(Manifest &manifest) {
+
+'''
+    self.top_level_reserve = '  manifest.reserve(${operation_count});\n\n'
+    self.top_level_epilogue = '''
+}
+
+} // namespace library
+} // namespace cutlass
+
+'''
+
+
+  def get_kernel_filters (self, kernelListFile):
+    if os.path.isfile(kernelListFile):
+        with open(kernelListFile, 'r') as fileReader:
+            lines = [line.rstrip() for line in fileReader if not line.startswith("#")]
+        
+        lines = [re.compile(line) for line in lines if line]
+        return lines
+    else:
+        return []
+
+
+
+  def filter_out_kernels(self, kernel_name, kernel_filter_list):
+
+    for kernel_filter_re in kernel_filter_list:
+        if kernel_filter_re.search(kernel_name) is not None:
+            return True
+        
+    return False
+
+    
+  #
+  def _filter_string_matches(self, filter_string, haystack):
+    ''' Returns true if all substrings appear in the haystack in order'''
+    substrings = filter_string.split('*')
+    for sub in substrings:
+      idx = haystack.find(sub)
+      if idx < 0:
+        return False
+      haystack = haystack[idx + len(sub):]
+    return True
+
+  #
+  def filter(self, operation):
+    ''' Filtering operations based on various criteria'''
+
+    # filter based on compute capability
+    enabled = False
+    for cc in self.compute_capabilities:
+      if cc >= operation.tile_description.minimum_compute_capability and \
+        cc <= operation.tile_description.maximum_compute_capability:
+
+        enabled = True
+        break
+
+    if not enabled:
+      return False
+
+    if len(self.operations_enabled) and not operation.operation_kind in self.operations_enabled:
+      return False
+
+    # eliminate duplicates
+    if operation.procedural_name() in self.operations_by_name.keys():
+      return False
+
+    # Filter based on list of valid substrings
+    if len(self.kernel_names):
+      name = operation.procedural_name()
+      enabled = False
+
+      # compare against the include list
+      for name_substr in self.kernel_names:
+        if self._filter_string_matches(name_substr, name):
+          enabled = True
+          break
+
+      # compare against the exclude list
+      for name_substr in self.ignore_kernel_names:
+        if self._filter_string_matches(name_substr, name):
+          enabled = False
+          break
+          
+    if len(self.kernel_filter_list) > 0:
+        enabled = False
+        if self.filter_out_kernels(operation.procedural_name(), self.kernel_filter_list):
+            enabled = True
+
+
+    # todo: filter based on compute data type
+    return enabled
+  #
+
+  #
+  def append(self, operation):
+    ''' 
+      Inserts the operation.
+
+      operation_kind -> configuration_name -> []
+    '''
+
+    if self.filter(operation):
+    
+      self.selected_kernels.append(operation.procedural_name())
+
+      self.operations_by_name[operation.procedural_name()] = operation
+
+      # add the configuration
+      configuration_name = operation.configuration_name()
+
+      if operation.operation_kind not in self.operations.keys():
+        self.operations[operation.operation_kind] = {}
+
+      if configuration_name not in self.operations[operation.operation_kind].keys():
+        self.operations[operation.operation_kind][configuration_name] = []
+
+      self.operations[operation.operation_kind][configuration_name].append(operation)
+      self.operation_count += 1
+  #
+
+  #
+  def emit(self, target = GeneratorTarget.Library):
+
+    operation_emitters = {
+      GeneratorTarget.Library: EmitOperationKindLibrary 
+    }
+
+    generated_path = os.path.join(self.args.curr_build_dir, 'generated')
+
+    # create generated/
+    if os.path.exists(generated_path):
+      shutil.rmtree(generated_path)
+
+    os.mkdir(generated_path)
+
+    source_files = []
+
+    top_level_path = os.path.join(generated_path, 'initialize_all.cpp')
+    with open(top_level_path, 'w') as top_level_file:
+
+      if target == GeneratorTarget.Library:
+        source_files.append(top_level_path)
+
+      prototypes = []
+      for operation_kind, configurations in self.operations.items():
+        prototypes.append(SubstituteTemplate(
+          "void initialize_all_${operation_kind}_operations(Manifest &manifest);",
+          {'operation_kind': OperationKindNames[operation_kind]}))
+
+      top_level_file.write(SubstituteTemplate(self.top_level_prologue,
+        {'prototypes': "\n".join(prototypes)}))
+
+      top_level_file.write(SubstituteTemplate(
+        self.top_level_reserve, {'operation_count': str(self.operation_count)}))
+
+      # for each operation kind, emit initializer for all configurations
+      for operation_kind, configurations in self.operations.items():
+        
+        with operation_emitters[target](generated_path, operation_kind, self.args) as operation_kind_emitter:
+          for configuration_name, operations in configurations.items():
+            operation_kind_emitter.emit(configuration_name, operations)
+
+          source_files += operation_kind_emitter.source_files
+
+        top_level_file.write(SubstituteTemplate(
+          "  initialize_all_${operation_kind}_operations(manifest);\n",
+          {'operation_kind': OperationKindNames[operation_kind]}))
+
+      top_level_file.write(self.top_level_epilogue)
+
+    # write the manifest.cmake file containing paths from all targets
+    manifest_path = os.path.join(generated_path, "manifest.cmake")
+    with open(manifest_path, "w") as manifest_file:
+
+      target_name = 'cutlass_library_objs'
+
+      target_text = SubstituteTemplate("""cutlass_target_sources(
+  ${target_name}
+  BATCH_SOURCES ON
+  PRIVATE
+""", { 'target_name': target_name})
+
+      manifest_file.write(target_text)
+
+      for source_file in source_files:
+        manifest_file.write("    %s\n" % str(source_file.replace('\\', '/')))
+      manifest_file.write(")")
+  #
+
+###################################################################################################

dnn/src/CMakeLists.txt

@@ -113,6 +113,31 @@ if(MGE_WITH_CUDA)
     list(APPEND SOURCES ${SOURCES_})
 
     file(GLOB_RECURSE CUSOURCES cuda/*.cu)
+
+    set(CUTLASS_GEN_SCRIPT ${CMAKE_CURRENT_SOURCE_DIR}/../scripts/cutlass_generator/generator.py)
+    set(CUTLASS_GEN_DIR ${CMAKE_CURRENT_BINARY_DIR}/cuda/cutlass/generated)
+    function(gen_cutlass_kimpl op type)
+        set(CURRENT_CUTLASS_GEN_DIR ${CUTLASS_GEN_DIR}/${op}_${type})
+        file(MAKE_DIRECTORY ${CURRENT_CUTLASS_GEN_DIR})
+        execute_process(
+            COMMAND ${PYTHON3_EXECUTABLE_WITHOUT_VERSION} ${CUTLASS_GEN_SCRIPT} --operations ${op} --type ${type} ${CURRENT_CUTLASS_GEN_DIR}
+            RESULT_VARIABLE gen_cutlass_result
+            OUTPUT_FILE ${CURRENT_CUTLASS_GEN_DIR}/gen_cutlass.log
+            ERROR_FILE ${CURRENT_CUTLASS_GEN_DIR}/gen_cutlass.log
+        )
+        if (NOT gen_cutlass_result EQUAL 0)
+            message(FATAL_ERROR "Error generating library instances. See ${CURRENT_CUTLASS_GEN_DIR}/gen_cutlass.log")
+        endif()
+    endfunction()
+    gen_cutlass_kimpl(gemm simt)
+    gen_cutlass_kimpl(gemv simt)
+    gen_cutlass_kimpl(deconv simt)
+    gen_cutlass_kimpl(conv2d simt)
+    gen_cutlass_kimpl(conv2d tensorop8816)
+    gen_cutlass_kimpl(conv2d tensorop8832)
+    file(GLOB_RECURSE CUTLASS_SOURCES ${CUTLASS_GEN_DIR}/*.cu)
+    list(APPEND SOURCES ${CUTLASS_SOURCES})
+
     list(APPEND SOURCES ${CUSOURCES})
 endif()
 

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorCxRSKx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    int32_t, 
-    cutlass::layout::TensorNCxHWx<64>, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 128, 128>,
-    cutlass::gemm::GemmShape<64, 64, 128>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp<
-      cutlass::int4b_t,
-      16,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorCxRSKx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    int32_t, 
-    cutlass::layout::TensorNCxHWx<64>, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<256, 128, 128>,
-    cutlass::gemm::GemmShape<64, 64, 128>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp<
-      cutlass::int4b_t,
-      16,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorCxRSKx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    int32_t, 
-    cutlass::layout::TensorNCxHWx<64>, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 128, 128>,
-    cutlass::gemm::GemmShape<64, 64, 128>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::int4b_t,
-      16,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorCxRSKx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    int32_t, 
-    cutlass::layout::TensorNCxHWx<64>, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<256, 128, 128>,
-    cutlass::gemm::GemmShape<64, 64, 128>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::int4b_t,
-      16,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorCxRSKx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    int32_t, 
-    cutlass::layout::TensorNCxHWx<64>, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 128, 128>,
-    cutlass::gemm::GemmShape<64, 64, 128>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::int4b_t,
-      16,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::int4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorCxRSKx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    int32_t, 
-    cutlass::layout::TensorNCxHWx<64>, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<256, 128, 128>,
-    cutlass::gemm::GemmShape<64, 64, 128>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::int4b_t,
-      16,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    false, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorCxRSKx<64>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    int32_t, 
-    cutlass::layout::TensorNCxHWx<64>, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 128, 128>,
-    cutlass::gemm::GemmShape<64, 64, 128>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::uint4b_t,
-      16,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorCxRSKx<64>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    int32_t, 
-    cutlass::layout::TensorNCxHWx<64>, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<256, 128, 128>,
-    cutlass::gemm::GemmShape<64, 64, 128>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
-      cutlass::uint4b_t,
-      16,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorCxRSKx<64>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    int32_t, 
-    cutlass::layout::TensorNCxHWx<64>, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 128, 128>,
-    cutlass::gemm::GemmShape<64, 64, 128>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::uint4b_t,
-      16,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 32, 64>,
-    cutlass::gemm::GemmShape<64, 32, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<16>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    16, 
-    16, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<32>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorNCxHWx<8>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNHWC,
-    int32_t, 
-    cutlass::layout::TensorNHWC, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<128, 64, 64>,
-    cutlass::gemm::GemmShape<64, 64, 64>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::uint4b_t,
-      8,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle,     
-    2,
-    8, 
-    8, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_TN>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    cutlass::int4b_t, 
-    cutlass::layout::TensorCxRSKx<64>,
-    cutlass::uint4b_t, 
-    cutlass::layout::TensorNCxHWx<64>,
-    int32_t, 
-    cutlass::layout::TensorNCxHWx<64>, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassTensorOp,
-    cutlass::arch::Sm75,
-    cutlass::gemm::GemmShape<256, 128, 128>,
-    cutlass::gemm::GemmShape<64, 64, 128>,
-    cutlass::gemm::GemmShape<8, 8, 32>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp<
-      cutlass::uint4b_t,
-      16,
-      int32_t, 
-      int32_t, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    32, 
-    32, 
-    true, 
-    cutlass::arch::OpMultiplyAddSaturate,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<128, 128, 32>,
-    cutlass::gemm::GemmShape<64, 32, 32>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwish<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<128, 32, 32>,
-    cutlass::gemm::GemmShape<64, 32, 32>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwish<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<128, 64, 32>,
-    cutlass::gemm::GemmShape<64, 32, 32>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwish<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<16, 128, 16>,
-    cutlass::gemm::GemmShape<16, 128, 16>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwish<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    1,
-    4, 
-    8, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<16, 64, 8>,
-    cutlass::gemm::GemmShape<16, 64, 8>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwish<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    4, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<32, 128, 32>,
-    cutlass::gemm::GemmShape<32, 64, 32>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwish<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<32, 32, 32>,
-    cutlass::gemm::GemmShape<32, 32, 32>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwish<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<32, 64, 32>,
-    cutlass::gemm::GemmShape<32, 64, 32>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwish<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<64, 128, 32>,
-    cutlass::gemm::GemmShape<64, 32, 32>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwish<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<64, 32, 32>,
-    cutlass::gemm::GemmShape<64, 32, 32>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwish<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<64, 64, 32>,
-    cutlass::gemm::GemmShape<64, 32, 32>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombinationHSwish<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<128, 128, 32>,
-    cutlass::gemm::GemmShape<64, 32, 32>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombination<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<128, 32, 32>,
-    cutlass::gemm::GemmShape<64, 32, 32>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombination<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<128, 64, 32>,
-    cutlass::gemm::GemmShape<64, 32, 32>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombination<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<16, 128, 16>,
-    cutlass::gemm::GemmShape<16, 128, 16>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombination<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    1,
-    4, 
-    8, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<16, 64, 8>,
-    cutlass::gemm::GemmShape<16, 64, 8>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombination<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    4, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<32, 128, 32>,
-    cutlass::gemm::GemmShape<32, 64, 32>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombination<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu

@@ -1,59 +0,0 @@
-
-#if !MEGDNN_TEGRA_X1
-// ignore warning of cutlass
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl"
-
-
-// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator
-using Convolution = 
-  typename cutlass::conv::device::Convolution<
-    int8_t, 
-    cutlass::layout::TensorNCxHWx<4>,
-    int8_t, 
-    cutlass::layout::TensorCxRSKx<4>,
-    float, 
-    cutlass::layout::TensorNCHW,
-    float, 
-    cutlass::layout::TensorNCHW, 
-    int32_t, 
-    cutlass::conv::ConvType::kConvolution, 
-    cutlass::arch::OpClassSimt,
-    cutlass::arch::Sm61,
-    cutlass::gemm::GemmShape<32, 32, 32>,
-    cutlass::gemm::GemmShape<32, 32, 32>,
-    cutlass::gemm::GemmShape<1, 1, 4>,
-    cutlass::epilogue::thread::BiasAddLinearCombination<
-      float,
-      1,
-      int32_t, 
-      float, 
-      float
-    >,
-    cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle,     
-    2,
-    4, 
-    16, 
-    false, 
-    cutlass::arch::OpMultiplyAdd,
-    cutlass::conv::ImplicitGemmMode::GEMM_NT>;
-
-
-
-template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper<Convolution>(
-  const typename Convolution::ElementSrc* d_src, 
-  const typename Convolution::ElementFilter* d_filter, 
-  const typename Convolution::ElementBias* d_bias, 
-  const typename Convolution::ElementDst* d_z, 
-  typename Convolution::ElementDst* d_dst, 
-  int* workspace, 
-  typename Convolution::ConvolutionParameter const& conv_param, 
-  typename Convolution::EpilogueOutputOp::Params const& epilogue, 
-  cudaStream_t stream, 
-  typename Convolution::ExtraParam extra_param);
-
-
-#pragma GCC diagnostic pop
-#endif