Copybara import of the project:

-- 5d5cd76 by RahulSudarMCW <[email protected]>: Replace conv-hwc yaml with table header -- ad05494 by RahulSudarMCW <[email protected]>: Remove 'conv-hwc-microkernel-tester.h' -- 6d39a08 by RahulSudarMCW <[email protected]>: Declare microkernal to use table header -- 93d6323 by RahulSudarMCW <[email protected]>: Resolve failed checks FUTURE_COPYBARA_INTEGRATE_REVIEW=#7445 from RahulSundarMCW:conv-hwc 93d6323 PiperOrigin-RevId: 695826548
google · Nov 13, 2024 · 3d3c275 · 3d3c275
1 parent 689b350
commit 3d3c275
Show file tree

Hide file tree

Showing 8 changed files with 819 additions and 5,951 deletions.
diff --git a/BUILD.bazel b/BUILD.bazel
@@ -101,6 +101,7 @@ MICROKERNEL_DEFS = [
     "src/f16-vsqrt/f16-vsqrt.h",
     "src/f16-vtanh/f16-vtanh.h",
     "src/f32-avgpool/f32-avgpool-minmax.h",
+    "src/f32-conv-hwc/f32-conv-hwc.h",
     "src/f32-dwconv/f32-dwconv-minmax-multipass.h",
     "src/f32-dwconv/f32-dwconv-minmax-unipass.h",
     "src/f32-dwconv/f32-dwconv-multipass.h",

diff --git a/scripts/generate-tests.sh b/scripts/generate-tests.sh
@@ -195,9 +195,6 @@ tools/generate-rdsum-test.py --spec test/qu8-rdsum.yaml --output test/qu8-rdsum.
 ### Tests for LUT micro-kernels
 tools/generate-lut-test.py --spec test/x8-lut.yaml --output test/x8-lut.cc &
 
-### Tests for Conv HWC layout micro-kernels
-tools/generate-conv-hwc-test.py --spec test/f32-conv-hwc.yaml --output test/f32-conv-hwc.cc &
-
 ### Tests for Conv HWC2CHW layout micro-kernels
 tools/generate-conv-hwc2chw-test.py --spec test/f16-conv-hwc2chw.yaml --output test/f16-conv-hwc2chw.cc &
 tools/generate-conv-hwc2chw-test.py --spec test/f32-conv-hwc2chw.yaml --output test/f32-conv-hwc2chw.cc &

diff --git a/src/f32-conv-hwc/f32-conv-hwc.h b/src/f32-conv-hwc/f32-conv-hwc.h
@@ -0,0 +1,51 @@
+// Copyright 2023 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#ifndef XNN_UKERNEL_WITH_PARAMS
+#define XNN_UKERNEL_WITH_PARAMS(arch_flags, ukernel, kernel_size, subsampling, padding_right, padding_left, input_channels, output_channels_tile, input_widths, datatype, params_type, init_params)  \
+  XNN_UKERNEL(arch_flags, ukernel, kernel_size, subsampling, padding_right, padding_left, input_channels, output_channels_tile, input_widths, datatype)
+#define XNN_DEFINED_UKERNEL_WITH_PARAMS
+#endif
+
+#ifndef XNN_UKERNEL
+#define XNN_UKERNEL(arch_flags, ukernel, kernel_size, subsampling, padding_right, padding_left, input_channels, output_channels_tile, input_widths, datatype)   \
+  XNN_UKERNEL_WITH_PARAMS(arch_flags, ukernel, kernel_size, subsampling, padding_right, padding_left, input_channels, output_channels_tile, input_widths, datatype, void, /*init_params=*/nullptr)
+#define XNN_DEFINED_UKERNEL
+#endif
+
+XNN_UKERNEL_WITH_PARAMS(0, xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__scalar_1x1, 3, 2, 1, 0, 3, 4, 2, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(0, xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__scalar_1x1, 3, 2, 1, 1, 3, 4, 2, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+
+#if XNN_ARCH_ARM || XNN_ARCH_ARM64
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1, 3, 2, 1, 0, 3, 4, 2, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2, 3, 2, 1, 0, 3, 4, 4, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1, 3, 2, 1, 0, 3, 8, 2, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2, 3, 2, 1, 0, 3, 8, 4, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1, 3, 2, 1, 1, 3, 4, 2, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2, 3, 2, 1, 1, 3, 4, 4, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1, 3, 2, 1, 1, 3, 8, 2, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2, 3, 2, 1, 1, 3, 8, 4, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+#endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64
+
+#if XNN_ARCH_ARM64
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__aarch64_neonfma_2x1, 3, 2, 1, 0, 3, 4, 2, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__aarch64_neonfma_2x2, 3, 2, 1, 0, 3, 4, 4, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__aarch64_neonfma_2x1, 3, 2, 1, 0, 3, 8, 2, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__aarch64_neonfma_2x2, 3, 2, 1, 0, 3, 8, 4, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__aarch64_neonfma_2x1, 3, 2, 1, 1, 3, 4, 2, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__aarch64_neonfma_2x2, 3, 2, 1, 1, 3, 4, 4, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__aarch64_neonfma_2x1, 3, 2, 1, 1, 3, 8, 2, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+XNN_UKERNEL_WITH_PARAMS(xnn_arch_arm_neon, xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__aarch64_neonfma_2x2, 3, 2, 1, 1, 3, 8, 4, float, struct xnn_f32_default_params, xnn_init_f32_minmax_scalar_params)
+#endif  // XNN_ARCH_ARM64
+
+#ifdef XNN_DEFINED_UKERNEL_WITH_PARAMS
+#undef XNN_DEFINED_UKERNEL_WITH_PARAMS
+#undef XNN_UKERNEL_WITH_PARAMS
+#endif
+
+#ifdef XNN_DEFINED_UKERNEL
+#undef XNN_DEFINED_UKERNEL
+#undef XNN_UKERNEL
+#endif
diff --git a/src/xnnpack/conv.h b/src/xnnpack/conv.h
@@ -19,7 +19,7 @@ extern "C" {
 #endif
 
 
-#define DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(fn_name) \
+#define XNN_UKERNEL(arch_flags, fn_name, kernel_size, subsampling, padding_right, padding_left, input_channels, output_channels_tile, input_widths, datatype) \
   XNN_INTERNAL void fn_name(                           \
       size_t input_height,                             \
       size_t input_width,                              \
@@ -33,28 +33,9 @@ extern "C" {
       size_t output_channels,                          \
       size_t output_height_stride,                     \
       size_t output_width_stride,                      \
-      const union xnn_f32_minmax_params params[XNN_RESTRICT XNN_MIN_ELEMENTS(1)]);
-
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__aarch64_neonfma_2x1)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__aarch64_neonfma_2x2)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__scalar_1x1)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__aarch64_neonfma_2x1)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__aarch64_neonfma_2x2)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2)
-
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__aarch64_neonfma_2x1)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__aarch64_neonfma_2x2)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__scalar_1x1)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__aarch64_neonfma_2x1)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__aarch64_neonfma_2x2)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1)
-DECLARE_F32_CONV_HWC_UKERNEL_FUNCTION(xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2)
-
+      const union xnn_f32_minmax_params params[XNN_RESTRICT XNN_MIN_ELEMENTS(1)]); 
+#include "f32-conv-hwc/f32-conv-hwc.h"
+#undef XNN_UKERNEL
 
 #define DECLARE_F32_CONV_HWC2CHW_UKERNEL_FUNCTION(fn_name) \
   XNN_INTERNAL void fn_name(                                 \