refactor(dnn/cuda): refactor cutlass kernel generator for gemm and gemv

GitOrigin-RevId: 11d78ab227

dnn/scripts/Makefile

@@ -37,21 +37,21 @@ all: ${PARAM_DEFS} ${ELEMWISE_IMPL} ${CUDA_CONV_IMPL} $(CUDA_MATMUL_IMPL)
 ../src/cuda/elemwise_multi_type/kimpl: gen_elemwise_multi_type_kern_impls.py
 	./$^ --type cuda $@
 
-../src/cuda/conv_bias/int8/kimpl: gen_cuda_conv_bias_kern_impls.py cutlass_generator
+../src/cuda/conv_bias/int8/kimpl: gen_cuda_conv_bias_kern_impls.py cutlass_generator/generator.py
 	./gen_cuda_conv_bias_kern_impls.py --type dp4a $@
-	python3 ./cutlass_generator/generator.py --operations all --type simt $@
+	python3 ./cutlass_generator/generator.py --operations conv2d --type simt $@
 
-../src/cuda/conv_bias/int8_imma/kimpl: gen_cuda_conv_bias_kern_impls.py cutlass_generator
+../src/cuda/conv_bias/int8_imma/kimpl: gen_cuda_conv_bias_kern_impls.py cutlass_generator/generator.py
 	./gen_cuda_conv_bias_kern_impls.py --type imma $@
 	python3 ./cutlass_generator/generator.py --operations conv2d --type tensorop8816 $@
 
 ../src/cuda/batch_conv_bias/int8/kimpl: gen_cuda_batch_conv_bias_kern_impls.py 
 	./$^ --type dp4a $@
 
-../src/cuda/matrix_mul/fp32_simt/kimpl: gen_cutlass_matmul_kern_impls.py
-	./$^ $@
+../src/cuda/matrix_mul/fp32_simt/kimpl: cutlass_generator/generator.py
+	python3 ./cutlass_generator/generator.py --operations gemm --type simt $@
 
-../src/cuda/matrix_mul/fp32_simt_gemv/kimpl: gen_cutlass_gemv_batched_strided_kern_impls.py
-	./$^ $@
+../src/cuda/matrix_mul/fp32_simt_gemv/kimpl: cutlass_generator
+	python3 ./cutlass_generator/generator.py --operations gemv --type simt $@
 
 .PHONY: all

dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_32x64x32_32x64x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_64x128x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_64x32x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_64x64x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_identity_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_identity_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_identity_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_identity_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_identity_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_identity_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_identity_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_identity_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_identity_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_identity_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_identity_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_relu_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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_s4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu

@@ -40,16 +40,19 @@ using Convolution =
     cutlass::arch::OpMultiplyAdd>;
 
 
+
 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);
+  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