From 7e589800682074b7b14b4803f934788a71800f66 Mon Sep 17 00:00:00 2001 From: Pablo Marquez Tello Date: Thu, 14 Sep 2023 09:41:37 +0100 Subject: Fuse batch normalization changes to enable fp16 in armv8a multi_isa builds * Code guarded with __ARM_FEATURE_FP16_VECTOR_ARITHMETIC needs to be moved to an fp16.cpp file to allow compilation with -march=armv8.2-a+fp16 * fp16.cpp needs to use the template fused_batch_normalization_dwc_nhwc() that had to be moved from impl.cpp to impl.h * Removed impl.cpp * Partially resolves MLCE-1102 Change-Id: Idaaa113c71729e32e565acf5fb5694c76c36d76d Signed-off-by: Pablo Marquez Tello Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10308 Benchmark: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Viet-Hoa Do Comments-Addressed: Arm Jenkins --- Android.bp | 1 - filelist.json | 3 +- src/BUILD.bazel | 1 - src/CMakeLists.txt | 1 - .../fuse_batch_normalization/nhwc/neon/impl.cpp | 158 --------------------- .../fuse_batch_normalization/nhwc/neon/impl.h | 118 ++++++++++++++- 6 files changed, 117 insertions(+), 165 deletions(-) delete mode 100644 src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.cpp diff --git a/Android.bp b/Android.bp index da69d4fe87..d7d900ccc4 100644 --- a/Android.bp +++ b/Android.bp @@ -512,7 +512,6 @@ cc_library_static { "src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp", "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp16.cpp", "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp32.cpp", - "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.cpp", "src/cpu/kernels/gemm_matrix_add/generic/neon/fp16.cpp", "src/cpu/kernels/gemm_matrix_add/generic/neon/fp32.cpp", "src/cpu/kernels/gemm_matrix_add/generic/neon/impl.cpp", diff --git a/filelist.json b/filelist.json index d3cf984d8b..2a88aec37b 100644 --- a/filelist.json +++ b/filelist.json @@ -973,8 +973,7 @@ ], "neon": { "common": [ - "src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp", - "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.cpp" + "src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp" ], "fp16": [ "src/cpu/kernels/fuse_batch_normalization/generic/fp16.cpp", diff --git a/src/BUILD.bazel b/src/BUILD.bazel index 30bb615479..ab0ea66bdd 100644 --- a/src/BUILD.bazel +++ b/src/BUILD.bazel @@ -763,7 +763,6 @@ filegroup( "cpu/kernels/fuse_batch_normalization/nchw/all.cpp", "cpu/kernels/fuse_batch_normalization/nhwc/neon/fp16.cpp", "cpu/kernels/fuse_batch_normalization/nhwc/neon/fp32.cpp", - "cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.cpp", "cpu/kernels/gemm_matrix_add/generic/neon/fp16.cpp", "cpu/kernels/gemm_matrix_add/generic/neon/fp32.cpp", "cpu/kernels/gemm_matrix_add/generic/neon/impl.cpp", diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index 5234ddf01f..e6ef5bfc6a 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -755,7 +755,6 @@ target_sources( cpu/kernels/fuse_batch_normalization/nchw/all.cpp cpu/kernels/fuse_batch_normalization/nhwc/neon/fp16.cpp cpu/kernels/fuse_batch_normalization/nhwc/neon/fp32.cpp - cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.cpp cpu/kernels/gemm_matrix_add/generic/neon/fp16.cpp cpu/kernels/gemm_matrix_add/generic/neon/fp32.cpp cpu/kernels/gemm_matrix_add/generic/neon/impl.cpp diff --git a/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.cpp b/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.cpp deleted file mode 100644 index e33af4ebbe..0000000000 --- a/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.cpp +++ /dev/null @@ -1,158 +0,0 @@ -/* - * Copyright (c) 2018-2022 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.h" - -namespace arm_compute -{ -namespace cpu -{ -template -void fused_batch_normalization_dwc_nhwc(const ITensor *dwc_weights, const ITensor *dwc_bias, ITensor *fused_weights, ITensor *fused_bias, - const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window) -{ - using ScalarType = T; - const int size = 16 / dwc_weights->info()->element_size(); - using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t; - - const bool run_in_place_weights = (fused_weights == nullptr) || (fused_weights == dwc_weights); - const bool run_in_place_bias = (fused_bias == nullptr) || (dwc_bias != nullptr && fused_bias == dwc_bias); - - // Set build options - Window win = window; - win.set(Window::DimX, Window::Dimension(0, 1, 1)); - - const int window_step_x = size; - const auto window_start_x = static_cast(window.x().start()); - const auto window_end_x = static_cast(window.x().end()); - - Iterator dwc_w_in(dwc_weights, win); - Iterator dwc_w_out(run_in_place_weights ? dwc_weights : fused_weights, win); - - const auto dwc_bias_in = (dwc_bias != nullptr ? reinterpret_cast(dwc_bias->ptr_to_element(Coordinates(0, 0))) : nullptr); - auto dwc_bias_out = (run_in_place_bias ? dwc_bias_in : reinterpret_cast(fused_bias->ptr_to_element(Coordinates(0, 0)))); - - const auto input_mean = reinterpret_cast(bn_mean->ptr_to_element(Coordinates(0, 0))); - const auto input_var = reinterpret_cast(bn_var->ptr_to_element(Coordinates(0, 0))); - const auto input_gamma = (bn_gamma != nullptr) ? reinterpret_cast(bn_gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; - const auto input_beta = (bn_beta != nullptr) ? reinterpret_cast(bn_beta->ptr_to_element(Coordinates(0, 0))) : nullptr; - - auto mean_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); - auto var_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); - auto gamma_vec = wrapper::vdup_n(ScalarType(1), ExactTagType{}); - auto beta_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); - auto rvar_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); - auto dwc_bias_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); - const auto epsilon_vec = wrapper::vdup_n(ScalarType(epsilon), ExactTagType{}); - - auto gamma = ScalarType(1.0); - auto beta = ScalarType(0.0); - auto dwc_bias_in_scalar = ScalarType(0); - - execute_window_loop(win, [&](const Coordinates & id) - { - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - var_vec = wrapper::vloadq(input_var + x); - if(input_gamma != nullptr) - { - gamma_vec = wrapper::vloadq(input_gamma + x); - } - - if((id[2] == 0) && (id[1] == 0)) - { - mean_vec = wrapper::vloadq(input_mean + x); - - // Construct vectors - if(input_beta != nullptr) - { - beta_vec = wrapper::vloadq(input_beta + x); - } - - if(dwc_bias_in != nullptr) - { - dwc_bias_vec = wrapper::vloadq(dwc_bias_in + x); - } - - auto dwc_bias_tmp_vec = wrapper::vmul(wrapper::vsub(dwc_bias_vec, mean_vec), wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec))); - dwc_bias_tmp_vec = wrapper::vadd(wrapper::vmul(dwc_bias_tmp_vec, gamma_vec), beta_vec); - wrapper::vstore(dwc_bias_out + x, dwc_bias_tmp_vec); - } - - auto dwc_w_in_ptr = reinterpret_cast(dwc_w_in.ptr()); - auto dwc_w_out_ptr = reinterpret_cast(dwc_w_out.ptr()); - - auto wn = wrapper::vloadq(dwc_w_in_ptr + x); - rvar_vec = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)); - wn = wrapper::vmul(wn, rvar_vec); - wn = wrapper::vmul(wn, gamma_vec); - - // Store results - wrapper::vstore(dwc_w_out_ptr + x, wn); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - auto var = input_var[x]; - if(input_gamma != nullptr) - { - gamma = input_gamma[x]; - } - - if(id[2] == 0 && id[1] == 0) - { - auto mean = input_mean[x]; - if(input_beta != nullptr) - { - beta = input_beta[x]; - } - if(dwc_bias_in != nullptr) - { - dwc_bias_in_scalar = dwc_bias_in[x]; - } - - auto dwc_bias_tmp_scalar = (dwc_bias_in_scalar - mean) / std::sqrt(var + ScalarType(epsilon)); - dwc_bias_out[x] = (dwc_bias_tmp_scalar * gamma) + beta; - } - - const auto dwc_w_in_ptr = reinterpret_cast(dwc_w_in.ptr()); - auto dwc_w_out_ptr = reinterpret_cast(dwc_w_out.ptr()); - - *(dwc_w_out_ptr + x) = *(dwc_w_in_ptr + x) / std::sqrt(var + ScalarType(epsilon)) * gamma; - } - }, - dwc_w_in, dwc_w_out); -} - -template void fused_batch_normalization_dwc_nhwc(const ITensor *dwc_weights, const ITensor *dwc_bias, ITensor *fused_weights, ITensor *fused_bias, - const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window); - -#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) -template void fused_batch_normalization_dwc_nhwc(const ITensor *dwc_weights, const ITensor *dwc_bias, ITensor *fused_weights, ITensor *fused_bias, - const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window); -#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ - -} // namespace cpu -} // namespace arm_compute diff --git a/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.h b/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.h index 3b813132b1..6f0386276f 100644 --- a/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.h +++ b/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021-2022 Arm Limited. + * Copyright (c) 2018-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -33,8 +33,122 @@ namespace cpu { template void fused_batch_normalization_dwc_nhwc(const ITensor *dwc_weights, const ITensor *dwc_bias, ITensor *fused_weights, ITensor *fused_bias, - const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window); + const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window) +{ + using ScalarType = T; + const int size = 16 / dwc_weights->info()->element_size(); + using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t; + + const bool run_in_place_weights = (fused_weights == nullptr) || (fused_weights == dwc_weights); + const bool run_in_place_bias = (fused_bias == nullptr) || (dwc_bias != nullptr && fused_bias == dwc_bias); + + // Set build options + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + const int window_step_x = size; + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + + Iterator dwc_w_in(dwc_weights, win); + Iterator dwc_w_out(run_in_place_weights ? dwc_weights : fused_weights, win); + + const auto dwc_bias_in = (dwc_bias != nullptr ? reinterpret_cast(dwc_bias->ptr_to_element(Coordinates(0, 0))) : nullptr); + auto dwc_bias_out = (run_in_place_bias ? dwc_bias_in : reinterpret_cast(fused_bias->ptr_to_element(Coordinates(0, 0)))); + + const auto input_mean = reinterpret_cast(bn_mean->ptr_to_element(Coordinates(0, 0))); + const auto input_var = reinterpret_cast(bn_var->ptr_to_element(Coordinates(0, 0))); + const auto input_gamma = (bn_gamma != nullptr) ? reinterpret_cast(bn_gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; + const auto input_beta = (bn_beta != nullptr) ? reinterpret_cast(bn_beta->ptr_to_element(Coordinates(0, 0))) : nullptr; + + auto mean_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); + auto var_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); + auto gamma_vec = wrapper::vdup_n(ScalarType(1), ExactTagType{}); + auto beta_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); + auto rvar_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); + auto dwc_bias_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); + const auto epsilon_vec = wrapper::vdup_n(ScalarType(epsilon), ExactTagType{}); + + auto gamma = ScalarType(1.0); + auto beta = ScalarType(0.0); + auto dwc_bias_in_scalar = ScalarType(0); + + execute_window_loop(win, [&](const Coordinates & id) + { + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + var_vec = wrapper::vloadq(input_var + x); + if(input_gamma != nullptr) + { + gamma_vec = wrapper::vloadq(input_gamma + x); + } + + if((id[2] == 0) && (id[1] == 0)) + { + mean_vec = wrapper::vloadq(input_mean + x); + + // Construct vectors + if(input_beta != nullptr) + { + beta_vec = wrapper::vloadq(input_beta + x); + } + + if(dwc_bias_in != nullptr) + { + dwc_bias_vec = wrapper::vloadq(dwc_bias_in + x); + } + + auto dwc_bias_tmp_vec = wrapper::vmul(wrapper::vsub(dwc_bias_vec, mean_vec), wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec))); + dwc_bias_tmp_vec = wrapper::vadd(wrapper::vmul(dwc_bias_tmp_vec, gamma_vec), beta_vec); + wrapper::vstore(dwc_bias_out + x, dwc_bias_tmp_vec); + } + + auto dwc_w_in_ptr = reinterpret_cast(dwc_w_in.ptr()); + auto dwc_w_out_ptr = reinterpret_cast(dwc_w_out.ptr()); + + auto wn = wrapper::vloadq(dwc_w_in_ptr + x); + rvar_vec = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)); + wn = wrapper::vmul(wn, rvar_vec); + wn = wrapper::vmul(wn, gamma_vec); + + // Store results + wrapper::vstore(dwc_w_out_ptr + x, wn); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + auto var = input_var[x]; + if(input_gamma != nullptr) + { + gamma = input_gamma[x]; + } + + if(id[2] == 0 && id[1] == 0) + { + auto mean = input_mean[x]; + if(input_beta != nullptr) + { + beta = input_beta[x]; + } + if(dwc_bias_in != nullptr) + { + dwc_bias_in_scalar = dwc_bias_in[x]; + } + + auto dwc_bias_tmp_scalar = (dwc_bias_in_scalar - mean) / std::sqrt(var + ScalarType(epsilon)); + dwc_bias_out[x] = (dwc_bias_tmp_scalar * gamma) + beta; + } + + const auto dwc_w_in_ptr = reinterpret_cast(dwc_w_in.ptr()); + auto dwc_w_out_ptr = reinterpret_cast(dwc_w_out.ptr()); + *(dwc_w_out_ptr + x) = *(dwc_w_in_ptr + x) / std::sqrt(var + ScalarType(epsilon)) * gamma; + } + }, + dwc_w_in, dwc_w_out); +} } // namespace cpu } // namespace arm_compute #endif //SRC_CORE_NEON_KERNELS_FUSE_BATCH_NORMALIZATION_IMPL_H -- cgit v1.2.1