armv8.4 i8mm optimization for convolution gemm int8 (#4034)

cmake/ncnn_add_layer.cmake

@@ -239,25 +239,25 @@ macro(ncnn_add_layer class)
             ncnn_add_arch_opt_source(${class} asimdfhm "-march=armv8.2-a+fp16+fp16fml")
         endif()
         if(NCNN_ARM84BF16)
-            ncnn_add_arch_opt_source(${class} bf16 "-march=armv8.4-a+bf16")
+            ncnn_add_arch_opt_source(${class} bf16 "-march=armv8.4-a+fp16+dotprod+bf16")
         endif()
         if(NCNN_ARM84I8MM)
-            ncnn_add_arch_opt_source(${class} i8mm "-march=armv8.4-a+i8mm")
+            ncnn_add_arch_opt_source(${class} i8mm "-march=armv8.4-a+fp16+dotprod+i8mm")
         endif()
         if(NCNN_ARM86SVE)
-            ncnn_add_arch_opt_source(${class} sve "-march=armv8.6-a+sve")
+            ncnn_add_arch_opt_source(${class} sve "-march=armv8.6-a+fp16+dotprod+sve")
         endif()
         if(NCNN_ARM86SVE2)
-            ncnn_add_arch_opt_source(${class} sve2 "-march=armv8.6-a+sve2")
+            ncnn_add_arch_opt_source(${class} sve2 "-march=armv8.6-a+fp16+dotprod+sve2")
         endif()
         if(NCNN_ARM86SVEBF16)
-            ncnn_add_arch_opt_source(${class} svebf16 "-march=armv8.6-a+sve+bf16")
+            ncnn_add_arch_opt_source(${class} svebf16 "-march=armv8.6-a+fp16+dotprod+sve+bf16")
         endif()
         if(NCNN_ARM86SVEI8MM)
-            ncnn_add_arch_opt_source(${class} svei8mm "-march=armv8.6-a+sve+i8mm")
+            ncnn_add_arch_opt_source(${class} svei8mm "-march=armv8.6-a+fp16+dotprod+sve+i8mm")
         endif()
         if(NCNN_ARM86SVEF32MM)
-            ncnn_add_arch_opt_source(${class} svef32mm "-march=armv8.6-a+sve+f32mm")
+            ncnn_add_arch_opt_source(${class} svef32mm "-march=armv8.6-a+fp16+dotprod+sve+f32mm")
         endif()
     endif()
 

src/CMakeLists.txt

@@ -414,7 +414,7 @@ endif()
 
 if(((IOS AND CMAKE_OSX_ARCHITECTURES MATCHES "arm64") OR (APPLE AND CMAKE_OSX_ARCHITECTURES MATCHES "arm64") OR (CMAKE_SYSTEM_PROCESSOR MATCHES "^(arm64|aarch64)")))
     if(NOT NCNN_RUNTIME_CPU AND NCNN_ARM86SVE)
-        set(ARM_MARCH_FLAG "-march=armv8.6-a+sve")
+        set(ARM_MARCH_FLAG "-march=armv8.6-a+fp16+dotprod+sve")
         if(NCNN_ARM86SVE2)
             set(ARM_MARCH_FLAG "${ARM_MARCH_FLAG}+sve2")
         endif()
@@ -428,7 +428,7 @@ if(((IOS AND CMAKE_OSX_ARCHITECTURES MATCHES "arm64") OR (APPLE AND CMAKE_OSX_AR
             set(ARM_MARCH_FLAG "${ARM_MARCH_FLAG}+f32mm")
         endif()
     elseif(NOT NCNN_RUNTIME_CPU AND (NCNN_ARM84BF16 OR NCNN_ARM84I8MM))
-        set(ARM_MARCH_FLAG "-march=armv8.4-a")
+        set(ARM_MARCH_FLAG "-march=armv8.4-a+fp16+dotprod")
         if(NCNN_ARM84BF16)
             set(ARM_MARCH_FLAG "${ARM_MARCH_FLAG}+bf16")
         endif()

src/layer/arm/convolution_arm_i8mm.cpp

@@ -0,0 +1,69 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#include "cpu.h"
+#include "mat.h"
+
+namespace ncnn {
+
+#include "convolution_sgemm_int8.h"
+#include "convolution_sgemm_pack1to4_int8.h"
+#include "convolution_sgemm_pack8to1_int8.h"
+#include "convolution_sgemm_pack8to4_int8.h"
+
+// pack1
+void im2col_sgemm_int8_neon_i8mm(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
+{
+    im2col_sgemm_int8_neon(bottom_im2col, top_blob, kernel, opt);
+}
+
+void convolution_im2col_sgemm_transform_kernel_int8_neon_i8mm(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
+{
+    convolution_im2col_sgemm_transform_kernel_int8_neon(kernel, kernel_tm, inch, outch, kernel_w, kernel_h);
+}
+
+// pack1to4
+void im2col_sgemm_pack1to4_int8_neon_i8mm(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
+{
+    im2col_sgemm_pack1to4_int8_neon(bottom_im2col, top_blob, kernel, opt);
+}
+
+void convolution_im2col_sgemm_transform_kernel_pack1to4_int8_neon_i8mm(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
+{
+    convolution_im2col_sgemm_transform_kernel_pack1to4_int8_neon(kernel, kernel_tm, inch, outch, kernel_w, kernel_h);
+}
+
+// pack8to1
+void im2col_sgemm_pack8to1_int8_neon_i8mm(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
+{
+    im2col_sgemm_pack8to1_int8_neon(bottom_im2col, top_blob, kernel, opt);
+}
+
+void convolution_im2col_sgemm_transform_kernel_pack8to1_int8_neon_i8mm(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
+{
+    convolution_im2col_sgemm_transform_kernel_pack8to1_int8_neon(kernel, kernel_tm, inch, outch, kernel_w, kernel_h);
+}
+
+// pack8to4
+void im2col_sgemm_pack8to4_int8_neon_i8mm(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
+{
+    im2col_sgemm_pack8to4_int8_neon(bottom_im2col, top_blob, kernel, opt);
+}
+
+void convolution_im2col_sgemm_transform_kernel_pack8to4_int8_neon_i8mm(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
+{
+    convolution_im2col_sgemm_transform_kernel_pack8to4_int8_neon(kernel, kernel_tm, inch, outch, kernel_w, kernel_h);
+}
+
+} // namespace ncnn

tests/test_convolution.cpp

@@ -413,7 +413,8 @@ static int test_convolution_1()
            || test_convolution_int8(4, 8, 16, 24, 3, 1, 1, 1, 1)
            || test_convolution_int8(4, 20, 16, 24, 3, 1, 1, 1, 0)
            || test_convolution_int8(6, 7, 64, 64, 3, 1, 2, 0, 1)
-           || test_convolution_int8(25, 33, 16, 15, 3, 1, 1, 1, 0);
+           || test_convolution_int8(25, 33, 16, 15, 3, 1, 1, 1, 0)
+           || test_convolution_int8(7, 7, 15, 12, 3, 1, 1, 1, 0);
 }
 #endif // NCNN_INT8