Implement PRELU as a binary operator instead of a one-off operator type

This improves things in a few ways: - Removes a one-off operator implementation in favor of just using a standard binary operator. - The old PRELU packed weights, the new implementation just uses broadcasting like any other binary operator. - This should be more flexible and avoids initialization overhead and memory usage. PiperOrigin-RevId: 668770049
google · Sep 6, 2024 · 66de398 · 66de398
1 parent f416ab3
commit 66de398
Show file tree

Hide file tree

Showing 156 changed files with 679 additions and 19,742 deletions.
diff --git a/BUILD.bazel b/BUILD.bazel
@@ -68,7 +68,6 @@ OPERATOR_SRCS = [
     "src/operators/global-average-pooling-nwc.c",
     "src/operators/lut-elementwise-nc.c",
     "src/operators/max-pooling-nhwc.c",
-    "src/operators/prelu-nc.c",
     "src/operators/reduce-nd.c",
     "src/operators/resize-bilinear-nchw.c",
     "src/operators/resize-bilinear-nhwc.c",
@@ -314,7 +313,6 @@ MICROKERNEL_HDRS = [
     "src/xnnpack/pad.h",
     "src/xnnpack/pavgpool.h",
     "src/xnnpack/ppmm.h",
-    "src/xnnpack/prelu.h",
     "src/xnnpack/quantization.h",
     "src/xnnpack/raddexpminusmax.h",
     "src/xnnpack/raddextexp.h",
@@ -503,7 +501,6 @@ xnnpack_cc_library(
         "src/configs/lut32norm-config.c",
         "src/configs/maxpool-config.c",
         "src/configs/pavgpool-config.c",
-        "src/configs/prelu-config.c",
         "src/configs/raddstoreexpminusmax-config.c",
         "src/configs/reduce-config.c",
         "src/configs/rmax-config.c",

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -378,7 +378,6 @@ SET(OPERATOR_SRCS
   src/operators/global-average-pooling-nwc.c
   src/operators/lut-elementwise-nc.c
   src/operators/max-pooling-nhwc.c
-  src/operators/prelu-nc.c
   src/operators/reduce-nd.c
   src/operators/resize-bilinear-nchw.c
   src/operators/resize-bilinear-nhwc.c
@@ -474,7 +473,6 @@ SET(XNNPACK_SRCS
   src/configs/lut32norm-config.c
   src/configs/maxpool-config.c
   src/configs/pavgpool-config.c
-  src/configs/prelu-config.c
   src/configs/raddstoreexpminusmax-config.c
   src/configs/reduce-config.c
   src/configs/rmax-config.c
@@ -1477,7 +1475,6 @@ IF(XNNPACK_BUILD_TESTS)
       f16-ibilinear
       f16-maxpool-minmax
       f16-pavgpool-minmax
-      f16-prelu
       f16-raddstoreexpminusmax
       f16-rmax
       f16-rsum
@@ -1491,7 +1488,6 @@ IF(XNNPACK_BUILD_TESTS)
       f32-gavgpool-minmax
       f32-ibilinear-chw
       f32-ibilinear
-      f32-prelu
       f32-raddexpminusmax
       f32-raddextexp
       f32-raddstoreexpminusmax

diff --git a/bench/prelu.cc b/bench/prelu.cc
@@ -42,34 +42,34 @@ void xnnpack_prelu_f32(benchmark::State& state, const char* net) {
   std::generate(slope.begin(), slope.end(), std::ref(f32wrng));
   std::vector<float> output(batch_size * height * width * channels);
 
+  const size_t input_shape[4] = {batch_size, height, width, channels};
+  const size_t slope_shape[1] = {channels};
+
   xnn_status status = xnn_initialize(nullptr /* allocator */);
   if (status != xnn_status_success) {
     state.SkipWithError("failed to initialize XNNPACK");
     return;
   }
 
   xnn_operator_t prelu_op = nullptr;
-  status = xnn_create_prelu_nc_f32(
-    channels, /*slope_channels=*/channels, /*input_stride=*/channels , /*output_stride=*/channels,
-    slope.data(),
-    0 /* flags */, nullptr, nullptr, &prelu_op);
+  status = xnn_create_prelu_nd_f32(-std::numeric_limits<float>::infinity(),
+                                   std::numeric_limits<float>::infinity(),
+                                   0 /* flags */, &prelu_op);
   if (status != xnn_status_success) {
     state.SkipWithError("failed to create FP32 PReLU operator");
     return;
   }
 
-  status = xnn_reshape_prelu_nc_f32(
-    prelu_op,
-    batch_size * height * width,
-    /*threadpool=*/nullptr);
+  status =
+      xnn_reshape_prelu_nd_f32(prelu_op, 4, &input_shape[0], 1, &slope_shape[0],
+                               /*threadpool=*/nullptr);
   if (status != xnn_status_success) {
     state.SkipWithError("failed to reshape FP32 PReLU operator");
     return;
   }
 
-  status = xnn_setup_prelu_nc_f32(
-    prelu_op,
-    input.data(), output.data());
+  status = xnn_setup_prelu_nd_f32(prelu_op, input.data(), slope.data(),
+                                  output.data());
   if (status != xnn_status_success) {
     state.SkipWithError("failed to setup FP32 PReLU operator");
     return;

diff --git a/cmake/gen/avx512f_microkernels.cmake b/cmake/gen/avx512f_microkernels.cmake
@@ -19,7 +19,6 @@ SET(PROD_AVX512F_MICROKERNEL_SRCS
   src/f32-gemm/gen/f32-gemm-7x16-minmax-avx512f-broadcast.c
   src/f32-igemm/gen/f32-igemm-1x16-minmax-avx512f-broadcast.c
   src/f32-igemm/gen/f32-igemm-7x16-minmax-avx512f-broadcast.c
-  src/f32-prelu/gen/f32-prelu-avx512f-2x16.c
   src/f32-rdsum/gen/f32-rdsum-7p7x-minmax-avx512f-c64.c
   src/f32-rminmax/gen/f32-rmax-avx512f-u64-acc4.c
   src/f32-rminmax/gen/f32-rminmax-avx512f-u64-acc4.c
@@ -34,7 +33,10 @@ SET(PROD_AVX512F_MICROKERNEL_SRCS
   src/f32-vbinary/gen/f32-vminc-avx512f-u32.c
   src/f32-vbinary/gen/f32-vmul-minmax-avx512f-u32.c
   src/f32-vbinary/gen/f32-vmulc-minmax-avx512f-u32.c
+  src/f32-vbinary/gen/f32-vprelu-avx512f-u32.c
+  src/f32-vbinary/gen/f32-vpreluc-avx512f-u32.c
   src/f32-vbinary/gen/f32-vrdivc-minmax-avx512f-u32.c
+  src/f32-vbinary/gen/f32-vrpreluc-avx512f-u32.c
   src/f32-vbinary/gen/f32-vrsubc-minmax-avx512f-u32.c
   src/f32-vbinary/gen/f32-vsqrdiff-avx512f-u32.c
   src/f32-vbinary/gen/f32-vsqrdiffc-avx512f-u32.c
@@ -96,7 +98,6 @@ SET(NON_PROD_AVX512F_MICROKERNEL_SRCS
   src/f32-igemm/gen/f32-igemm-5x16-minmax-avx512f-broadcast.c
   src/f32-igemm/gen/f32-igemm-6x16-minmax-avx512f-broadcast.c
   src/f32-igemm/gen/f32-igemm-8x16-minmax-avx512f-broadcast.c
-  src/f32-prelu/gen/f32-prelu-avx512f-2x32.c
   src/f32-raddexpminusmax/gen/f32-raddexpminusmax-avx512f-p5-scalef-u64-acc2.c
   src/f32-raddexpminusmax/gen/f32-raddexpminusmax-avx512f-p5-scalef-u64-acc4.c
   src/f32-raddexpminusmax/gen/f32-raddexpminusmax-avx512f-p5-scalef-u64.c
@@ -173,12 +174,9 @@ SET(NON_PROD_AVX512F_MICROKERNEL_SRCS
   src/f32-vbinary/gen/f32-vmul-minmax-avx512f-u16.c
   src/f32-vbinary/gen/f32-vmulc-minmax-avx512f-u16.c
   src/f32-vbinary/gen/f32-vprelu-avx512f-u16.c
-  src/f32-vbinary/gen/f32-vprelu-avx512f-u32.c
   src/f32-vbinary/gen/f32-vpreluc-avx512f-u16.c
-  src/f32-vbinary/gen/f32-vpreluc-avx512f-u32.c
   src/f32-vbinary/gen/f32-vrdivc-minmax-avx512f-u16.c
   src/f32-vbinary/gen/f32-vrpreluc-avx512f-u16.c
-  src/f32-vbinary/gen/f32-vrpreluc-avx512f-u32.c
   src/f32-vbinary/gen/f32-vrsubc-minmax-avx512f-u16.c
   src/f32-vbinary/gen/f32-vsqrdiff-avx512f-u16.c
   src/f32-vbinary/gen/f32-vsqrdiffc-avx512f-u16.c

diff --git a/cmake/gen/avx512fp16_microkernels.cmake b/cmake/gen/avx512fp16_microkernels.cmake
@@ -26,7 +26,10 @@ SET(PROD_AVX512FP16_MICROKERNEL_SRCS
   src/f16-vbinary/gen/f16-vminc-avx512fp16-u64.c
   src/f16-vbinary/gen/f16-vmul-minmax-avx512fp16-u64.c
   src/f16-vbinary/gen/f16-vmulc-minmax-avx512fp16-u64.c
+  src/f16-vbinary/gen/f16-vprelu-avx512fp16-u64.c
+  src/f16-vbinary/gen/f16-vpreluc-avx512fp16-u64.c
   src/f16-vbinary/gen/f16-vrdivc-minmax-avx512fp16-u64.c
+  src/f16-vbinary/gen/f16-vrpreluc-avx512fp16-u64.c
   src/f16-vbinary/gen/f16-vrsubc-minmax-avx512fp16-u64.c
   src/f16-vbinary/gen/f16-vsqrdiff-avx512fp16-u64.c
   src/f16-vbinary/gen/f16-vsqrdiffc-avx512fp16-u64.c
@@ -83,12 +86,9 @@ SET(NON_PROD_AVX512FP16_MICROKERNEL_SRCS
   src/f16-vbinary/gen/f16-vmul-minmax-avx512fp16-u32.c
   src/f16-vbinary/gen/f16-vmulc-minmax-avx512fp16-u32.c
   src/f16-vbinary/gen/f16-vprelu-avx512fp16-u32.c
-  src/f16-vbinary/gen/f16-vprelu-avx512fp16-u64.c
   src/f16-vbinary/gen/f16-vpreluc-avx512fp16-u32.c
-  src/f16-vbinary/gen/f16-vpreluc-avx512fp16-u64.c
   src/f16-vbinary/gen/f16-vrdivc-minmax-avx512fp16-u32.c
   src/f16-vbinary/gen/f16-vrpreluc-avx512fp16-u32.c
-  src/f16-vbinary/gen/f16-vrpreluc-avx512fp16-u64.c
   src/f16-vbinary/gen/f16-vrsubc-minmax-avx512fp16-u32.c
   src/f16-vbinary/gen/f16-vsqrdiff-avx512fp16-u32.c
   src/f16-vbinary/gen/f16-vsqrdiffc-avx512fp16-u32.c

diff --git a/cmake/gen/avx_microkernels.cmake b/cmake/gen/avx_microkernels.cmake
@@ -21,7 +21,6 @@ SET(PROD_AVX_MICROKERNEL_SRCS
   src/f32-gemm/gen/f32-gemm-5x16-minmax-avx-broadcast.c
   src/f32-igemm/gen/f32-igemm-1x16-minmax-avx-broadcast.c
   src/f32-igemm/gen/f32-igemm-5x16-minmax-avx-broadcast.c
-  src/f32-prelu/gen/f32-prelu-avx-2x16.c
   src/f32-qc4w-gemm/gen/f32-qc4w-gemm-1x16-minmax-avx-broadcast.c
   src/f32-qc4w-gemm/gen/f32-qc4w-gemm-3x16-minmax-avx-broadcast.c
   src/f32-qc8w-gemm/gen/f32-qc8w-gemm-1x16-minmax-avx-broadcast.c
@@ -42,7 +41,10 @@ SET(PROD_AVX_MICROKERNEL_SRCS
   src/f32-vbinary/gen/f32-vminc-avx-u16.c
   src/f32-vbinary/gen/f32-vmul-minmax-avx-u16.c
   src/f32-vbinary/gen/f32-vmulc-minmax-avx-u16.c
+  src/f32-vbinary/gen/f32-vprelu-avx-u16.c
+  src/f32-vbinary/gen/f32-vpreluc-avx-u16.c
   src/f32-vbinary/gen/f32-vrdivc-minmax-avx-u16.c
+  src/f32-vbinary/gen/f32-vrpreluc-avx-u16.c
   src/f32-vbinary/gen/f32-vrsubc-minmax-avx-u16.c
   src/f32-vbinary/gen/f32-vsqrdiff-avx-u16.c
   src/f32-vbinary/gen/f32-vsqrdiffc-avx-u16.c
@@ -165,7 +167,6 @@ SET(NON_PROD_AVX_MICROKERNEL_SRCS
   src/f32-igemm/gen/f32-igemm-6x8-minmax-avx-broadcast.c
   src/f32-igemm/gen/f32-igemm-6x16-minmax-avx-broadcast.c
   src/f32-igemm/gen/f32-igemm-7x8-minmax-avx-broadcast.c
-  src/f32-prelu/gen/f32-prelu-avx-2x8.c
   src/f32-qc4w-gemm/gen/f32-qc4w-gemm-2x16-minmax-avx-broadcast.c
   src/f32-qc4w-gemm/gen/f32-qc4w-gemm-4x16-minmax-avx-broadcast.c
   src/f32-qc4w-gemm/gen/f32-qc4w-gemm-5x16-minmax-avx-broadcast.c
@@ -214,12 +215,9 @@ SET(NON_PROD_AVX_MICROKERNEL_SRCS
   src/f32-vbinary/gen/f32-vmul-minmax-avx-u8.c
   src/f32-vbinary/gen/f32-vmulc-minmax-avx-u8.c
   src/f32-vbinary/gen/f32-vprelu-avx-u8.c
-  src/f32-vbinary/gen/f32-vprelu-avx-u16.c
   src/f32-vbinary/gen/f32-vpreluc-avx-u8.c
-  src/f32-vbinary/gen/f32-vpreluc-avx-u16.c
   src/f32-vbinary/gen/f32-vrdivc-minmax-avx-u8.c
   src/f32-vbinary/gen/f32-vrpreluc-avx-u8.c
-  src/f32-vbinary/gen/f32-vrpreluc-avx-u16.c
   src/f32-vbinary/gen/f32-vrsubc-minmax-avx-u8.c
   src/f32-vbinary/gen/f32-vsqrdiff-avx-u8.c
   src/f32-vbinary/gen/f32-vsqrdiffc-avx-u8.c

diff --git a/cmake/gen/f16c_microkernels.cmake b/cmake/gen/f16c_microkernels.cmake
@@ -18,7 +18,6 @@ SET(PROD_F16C_MICROKERNEL_SRCS
   src/f16-gavgpool/gen/f16-gavgpool-7p7x-minmax-f16c-c8.c
   src/f16-gavgpool/gen/f16-gavgpool-7x-minmax-f16c-c8.c
   src/f16-maxpool/f16-maxpool-9p8x-minmax-f16c-c8.c
-  src/f16-prelu/gen/f16-prelu-f16c-2x16.c
   src/f16-rminmax/f16-rmax-f16c-u32.c
   src/f16-vbinary/gen/f16-vadd-minmax-f16c-u16.c
   src/f16-vbinary/gen/f16-vaddc-minmax-f16c-u16.c
@@ -30,7 +29,10 @@ SET(PROD_F16C_MICROKERNEL_SRCS
   src/f16-vbinary/gen/f16-vminc-f16c-u16.c
   src/f16-vbinary/gen/f16-vmul-minmax-f16c-u16.c
   src/f16-vbinary/gen/f16-vmulc-minmax-f16c-u16.c
+  src/f16-vbinary/gen/f16-vprelu-f16c-u16.c
+  src/f16-vbinary/gen/f16-vpreluc-f16c-u16.c
   src/f16-vbinary/gen/f16-vrdivc-minmax-f16c-u8.c
+  src/f16-vbinary/gen/f16-vrpreluc-f16c-u16.c
   src/f16-vbinary/gen/f16-vrsubc-minmax-f16c-u16.c
   src/f16-vbinary/gen/f16-vsqrdiff-f16c-u16.c
   src/f16-vbinary/gen/f16-vsqrdiffc-f16c-u16.c
@@ -64,7 +66,6 @@ SET(NON_PROD_F16C_MICROKERNEL_SRCS
   src/f16-gavgpool/gen/f16-gavgpool-7x-minmax-f16c-c16.c
   src/f16-gavgpool/gen/f16-gavgpool-7x-minmax-f16c-c24.c
   src/f16-gavgpool/gen/f16-gavgpool-7x-minmax-f16c-c32.c
-  src/f16-prelu/gen/f16-prelu-f16c-2x8.c
   src/f16-vbinary/gen/f16-vadd-minmax-f16c-u8.c
   src/f16-vbinary/gen/f16-vaddc-minmax-f16c-u8.c
   src/f16-vbinary/gen/f16-vdiv-minmax-f16c-u16.c
@@ -76,12 +77,9 @@ SET(NON_PROD_F16C_MICROKERNEL_SRCS
   src/f16-vbinary/gen/f16-vmul-minmax-f16c-u8.c
   src/f16-vbinary/gen/f16-vmulc-minmax-f16c-u8.c
   src/f16-vbinary/gen/f16-vprelu-f16c-u8.c
-  src/f16-vbinary/gen/f16-vprelu-f16c-u16.c
   src/f16-vbinary/gen/f16-vpreluc-f16c-u8.c
-  src/f16-vbinary/gen/f16-vpreluc-f16c-u16.c
   src/f16-vbinary/gen/f16-vrdivc-minmax-f16c-u16.c
   src/f16-vbinary/gen/f16-vrpreluc-f16c-u8.c
-  src/f16-vbinary/gen/f16-vrpreluc-f16c-u16.c
   src/f16-vbinary/gen/f16-vrsubc-minmax-f16c-u8.c
   src/f16-vbinary/gen/f16-vsqrdiff-f16c-u8.c
   src/f16-vbinary/gen/f16-vsqrdiffc-f16c-u8.c

diff --git a/cmake/gen/neon_microkernels.cmake b/cmake/gen/neon_microkernels.cmake
@@ -41,7 +41,6 @@ SET(PROD_NEON_MICROKERNEL_SRCS
   src/f32-maxpool/f32-maxpool-9p8x-minmax-neon-c4.c
   src/f32-pavgpool/f32-pavgpool-9p8x-minmax-neon-c4.c
   src/f32-pavgpool/f32-pavgpool-9x-minmax-neon-c4.c
-  src/f32-prelu/gen/f32-prelu-neon-2x8.c
   src/f32-qc4w-gemm/gen/f32-qc4w-gemm-1x8-minmax-neon-lane-ld64.c
   src/f32-qc4w-gemm/gen/f32-qc4w-gemm-4x8-minmax-neon-lane-ld64.c
   src/f32-qc8w-gemm/gen/f32-qc8w-gemm-1x8-minmax-neon-lane-ld64.c
@@ -62,6 +61,9 @@ SET(PROD_NEON_MICROKERNEL_SRCS
   src/f32-vbinary/gen/f32-vminc-neon-u8.c
   src/f32-vbinary/gen/f32-vmul-minmax-neon-u8.c
   src/f32-vbinary/gen/f32-vmulc-minmax-neon-u8.c
+  src/f32-vbinary/gen/f32-vprelu-neon-u8.c
+  src/f32-vbinary/gen/f32-vpreluc-neon-u8.c
+  src/f32-vbinary/gen/f32-vrpreluc-neon-u8.c
   src/f32-vbinary/gen/f32-vrsubc-minmax-neon-u8.c
   src/f32-vbinary/gen/f32-vsqrdiff-neon-u8.c
   src/f32-vbinary/gen/f32-vsqrdiffc-neon-u8.c
@@ -343,14 +345,6 @@ SET(NON_PROD_NEON_MICROKERNEL_SRCS
   src/f32-ppmm/gen/f32-ppmm-4x16-minmax-neon.c
   src/f32-ppmm/gen/f32-ppmm-8x8-minmax-neon-prfm.c
   src/f32-ppmm/gen/f32-ppmm-8x8-minmax-neon.c
-  src/f32-prelu/gen/f32-prelu-neon-1x4.c
-  src/f32-prelu/gen/f32-prelu-neon-1x8.c
-  src/f32-prelu/gen/f32-prelu-neon-1x16.c
-  src/f32-prelu/gen/f32-prelu-neon-2x4.c
-  src/f32-prelu/gen/f32-prelu-neon-2x16.c
-  src/f32-prelu/gen/f32-prelu-neon-4x4.c
-  src/f32-prelu/gen/f32-prelu-neon-4x8.c
-  src/f32-prelu/gen/f32-prelu-neon-4x16.c
   src/f32-qc4w-gemm/gen/f32-qc4w-gemm-1x8-minmax-neon-dup-ld64.c
   src/f32-qc4w-gemm/gen/f32-qc4w-gemm-4x8-minmax-neon-dup-ld64.c
   src/f32-qc4w-gemm/gen/f32-qc4w-gemm-5x8-minmax-neon-lane-ld64.c
@@ -432,11 +426,8 @@ SET(NON_PROD_NEON_MICROKERNEL_SRCS
   src/f32-vbinary/gen/f32-vmul-minmax-neon-u4.c
   src/f32-vbinary/gen/f32-vmulc-minmax-neon-u4.c
   src/f32-vbinary/gen/f32-vprelu-neon-u4.c
-  src/f32-vbinary/gen/f32-vprelu-neon-u8.c
   src/f32-vbinary/gen/f32-vpreluc-neon-u4.c
-  src/f32-vbinary/gen/f32-vpreluc-neon-u8.c
   src/f32-vbinary/gen/f32-vrpreluc-neon-u4.c
-  src/f32-vbinary/gen/f32-vrpreluc-neon-u8.c
   src/f32-vbinary/gen/f32-vrsubc-minmax-neon-u4.c
   src/f32-vbinary/gen/f32-vsqrdiff-neon-u4.c
   src/f32-vbinary/gen/f32-vsqrdiffc-neon-u4.c

diff --git a/cmake/gen/neonfp16arith_microkernels.cmake b/cmake/gen/neonfp16arith_microkernels.cmake
@@ -40,7 +40,6 @@ SET(PROD_NEONFP16ARITH_MICROKERNEL_SRCS
   src/f16-maxpool/f16-maxpool-9p8x-minmax-neonfp16arith-c8.c
   src/f16-pavgpool/f16-pavgpool-9p8x-minmax-neonfp16arith-c8.c
   src/f16-pavgpool/f16-pavgpool-9x-minmax-neonfp16arith-c8.c
-  src/f16-prelu/gen/f16-prelu-neonfp16arith-2x16.c
   src/f16-qs8-vcvt/gen/f16-qs8-vcvt-neonfp16arith-u32.c
   src/f16-raddstoreexpminusmax/gen/f16-raddstoreexpminusmax-neonfp16arith-rr2-p2-u32.c
   src/f16-raddstoreexpminusmax/gen/f16-raddstoreexpminusmax-neonfp16arith-rr2-p2-u40.c
@@ -55,6 +54,9 @@ SET(PROD_NEONFP16ARITH_MICROKERNEL_SRCS
   src/f16-vbinary/gen/f16-vminc-neonfp16arith-u16.c
   src/f16-vbinary/gen/f16-vmul-minmax-neonfp16arith-u16.c
   src/f16-vbinary/gen/f16-vmulc-minmax-neonfp16arith-u16.c
+  src/f16-vbinary/gen/f16-vprelu-neonfp16arith-u16.c
+  src/f16-vbinary/gen/f16-vpreluc-neonfp16arith-u16.c
+  src/f16-vbinary/gen/f16-vrpreluc-neonfp16arith-u16.c
   src/f16-vbinary/gen/f16-vrsubc-minmax-neonfp16arith-u16.c
   src/f16-vbinary/gen/f16-vsqrdiff-neonfp16arith-u16.c
   src/f16-vbinary/gen/f16-vsqrdiffc-neonfp16arith-u16.c
@@ -195,7 +197,6 @@ SET(NON_PROD_NEONFP16ARITH_MICROKERNEL_SRCS
   src/f16-igemm/gen/f16-igemm-4x16-minmax-neonfp16arith-ld64.c
   src/f16-igemm/gen/f16-igemm-8x8-minmax-neonfp16arith-ld64.c
   src/f16-igemm/gen/f16-igemm-8x16-minmax-neonfp16arith-ld64.c
-  src/f16-prelu/gen/f16-prelu-neonfp16arith-2x8.c
   src/f16-qs8-vcvt/gen/f16-qs8-vcvt-neonfp16arith-u8.c
   src/f16-qs8-vcvt/gen/f16-qs8-vcvt-neonfp16arith-u16.c
   src/f16-qs8-vcvt/gen/f16-qs8-vcvt-neonfp16arith-u24.c
@@ -271,11 +272,8 @@ SET(NON_PROD_NEONFP16ARITH_MICROKERNEL_SRCS
   src/f16-vbinary/gen/f16-vmul-minmax-neonfp16arith-u8.c
   src/f16-vbinary/gen/f16-vmulc-minmax-neonfp16arith-u8.c
   src/f16-vbinary/gen/f16-vprelu-neonfp16arith-u8.c
-  src/f16-vbinary/gen/f16-vprelu-neonfp16arith-u16.c
   src/f16-vbinary/gen/f16-vpreluc-neonfp16arith-u8.c
-  src/f16-vbinary/gen/f16-vpreluc-neonfp16arith-u16.c
   src/f16-vbinary/gen/f16-vrpreluc-neonfp16arith-u8.c
-  src/f16-vbinary/gen/f16-vrpreluc-neonfp16arith-u16.c
   src/f16-vbinary/gen/f16-vrsubc-minmax-neonfp16arith-u8.c
   src/f16-vbinary/gen/f16-vsqrdiff-neonfp16arith-u8.c
   src/f16-vbinary/gen/f16-vsqrdiffc-neonfp16arith-u8.c

diff --git a/cmake/gen/scalar_microkernels.cmake b/cmake/gen/scalar_microkernels.cmake
@@ -70,7 +70,6 @@ SET(PROD_SCALAR_MICROKERNEL_SRCS
   src/f32-maxpool/f32-maxpool-9p8x-minmax-scalar-c1.c
   src/f32-pavgpool/f32-pavgpool-9p8x-minmax-scalar-c1.c
   src/f32-pavgpool/f32-pavgpool-9x-minmax-scalar-c1.c
-  src/f32-prelu/gen/f32-prelu-scalar-2x4.c
   src/f32-qc4w-gemm/gen/f32-qc4w-gemm-1x4-minmax-scalar.c
   src/f32-qc4w-gemm/gen/f32-qc4w-gemm-4x4-minmax-scalar.c
   src/f32-qc8w-gemm/gen/f32-qc8w-gemm-1x4-minmax-scalar.c
@@ -99,7 +98,10 @@ SET(PROD_SCALAR_MICROKERNEL_SRCS
   src/f32-vbinary/gen/f32-vminc-scalar-u8.c
   src/f32-vbinary/gen/f32-vmul-minmax-scalar-u8.c
   src/f32-vbinary/gen/f32-vmulc-minmax-scalar-u8.c
+  src/f32-vbinary/gen/f32-vprelu-scalar-u8.c
+  src/f32-vbinary/gen/f32-vpreluc-scalar-u8.c
   src/f32-vbinary/gen/f32-vrdivc-minmax-scalar-u2.c
+  src/f32-vbinary/gen/f32-vrpreluc-scalar-u8.c
   src/f32-vbinary/gen/f32-vrsubc-minmax-scalar-u8.c
   src/f32-vbinary/gen/f32-vsqrdiff-scalar-u8.c
   src/f32-vbinary/gen/f32-vsqrdiffc-scalar-u8.c
@@ -391,7 +393,6 @@ SET(NON_PROD_SCALAR_MICROKERNEL_SRCS
   src/f32-ppmm/gen/f32-ppmm-3x3-minmax-scalar.c
   src/f32-ppmm/gen/f32-ppmm-4x2-minmax-scalar.c
   src/f32-ppmm/gen/f32-ppmm-4x4-minmax-scalar.c
-  src/f32-prelu/gen/f32-prelu-scalar-2x1.c
   src/f32-qc4w-gemm/gen/f32-qc4w-gemm-2x4-minmax-scalar.c
   src/f32-qc4w-gemm/gen/f32-qc4w-gemm-4x2-minmax-scalar.c
   src/f32-qc8w-gemm/gen/f32-qc8w-gemm-1x4-relu-scalar.c
@@ -544,11 +545,9 @@ SET(NON_PROD_SCALAR_MICROKERNEL_SRCS
   src/f32-vbinary/gen/f32-vprelu-scalar-u1.c
   src/f32-vbinary/gen/f32-vprelu-scalar-u2.c
   src/f32-vbinary/gen/f32-vprelu-scalar-u4.c
-  src/f32-vbinary/gen/f32-vprelu-scalar-u8.c
   src/f32-vbinary/gen/f32-vpreluc-scalar-u1.c
   src/f32-vbinary/gen/f32-vpreluc-scalar-u2.c
   src/f32-vbinary/gen/f32-vpreluc-scalar-u4.c
-  src/f32-vbinary/gen/f32-vpreluc-scalar-u8.c
   src/f32-vbinary/gen/f32-vrdivc-minmax-scalar-u1.c
   src/f32-vbinary/gen/f32-vrdivc-minmax-scalar-u4.c
   src/f32-vbinary/gen/f32-vrdivc-minmax-scalar-u8.c
@@ -563,7 +562,6 @@ SET(NON_PROD_SCALAR_MICROKERNEL_SRCS
   src/f32-vbinary/gen/f32-vrpreluc-scalar-u1.c
   src/f32-vbinary/gen/f32-vrpreluc-scalar-u2.c
   src/f32-vbinary/gen/f32-vrpreluc-scalar-u4.c
-  src/f32-vbinary/gen/f32-vrpreluc-scalar-u8.c
   src/f32-vbinary/gen/f32-vrsubc-minmax-scalar-u1.c
   src/f32-vbinary/gen/f32-vrsubc-minmax-scalar-u2.c
   src/f32-vbinary/gen/f32-vrsubc-minmax-scalar-u4.c