From afd38f0c617d6f89b2b4532c6c44f116617e2b6f Mon Sep 17 00:00:00 2001 From: Felix Thomasmathibalan Date: Wed, 27 Sep 2023 17:46:17 +0100 Subject: Apply clang-format on repository Code is formatted as per a revised clang format configuration file(not part of this delivery). Version 14.0.6 is used. Exclusion List: - files with .cl extension - files that are not strictly C/C++ (e.g. Android.bp, Sconscript ...) And the following directories - compute_kernel_writer/validation/ - tests/ - include/ - src/core/NEON/kernels/convolution/ - src/core/NEON/kernels/arm_gemm/ - src/core/NEON/kernels/arm_conv/ - data/ There will be a follow up for formatting of .cl files and the files under tests/ and compute_kernel_writer/validation/. Signed-off-by: Felix Thomasmathibalan Change-Id: Ib7eb1fcf4e7537b9feaefcfc15098a804a3fde0a Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10391 Benchmark: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Gunes Bayir --- .../fuse_batch_normalization/generic/fp16.cpp | 16 ++- .../fuse_batch_normalization/generic/fp32.cpp | 16 ++- .../fuse_batch_normalization/generic/impl.h | 120 ++++++++++------- src/cpu/kernels/fuse_batch_normalization/list.h | 15 ++- .../kernels/fuse_batch_normalization/nchw/all.cpp | 147 +++++++++++++-------- .../fuse_batch_normalization/nhwc/neon/fp16.cpp | 16 ++- .../fuse_batch_normalization/nhwc/neon/fp32.cpp | 16 ++- .../fuse_batch_normalization/nhwc/neon/impl.h | 143 +++++++++++--------- 8 files changed, 297 insertions(+), 192 deletions(-) (limited to 'src/cpu/kernels/fuse_batch_normalization') diff --git a/src/cpu/kernels/fuse_batch_normalization/generic/fp16.cpp b/src/cpu/kernels/fuse_batch_normalization/generic/fp16.cpp index a29ee762fc..2821af32ce 100644 --- a/src/cpu/kernels/fuse_batch_normalization/generic/fp16.cpp +++ b/src/cpu/kernels/fuse_batch_normalization/generic/fp16.cpp @@ -29,11 +29,19 @@ namespace arm_compute { namespace cpu { -void fused_batch_normalization_conv_f16(const ITensor *conv_weights, const ITensor *conv_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) +void fused_batch_normalization_conv_f16(const ITensor *conv_weights, + const ITensor *conv_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) { - return fused_batch_normalization_conv(conv_weights, conv_bias, fused_weights, fused_bias, - bn_mean, bn_var, bn_beta, bn_gamma, epsilon, window); + return fused_batch_normalization_conv(conv_weights, conv_bias, fused_weights, fused_bias, bn_mean, + bn_var, bn_beta, bn_gamma, epsilon, window); } } // namespace cpu } // namespace arm_compute diff --git a/src/cpu/kernels/fuse_batch_normalization/generic/fp32.cpp b/src/cpu/kernels/fuse_batch_normalization/generic/fp32.cpp index 076e97651d..3ca5b6977a 100644 --- a/src/cpu/kernels/fuse_batch_normalization/generic/fp32.cpp +++ b/src/cpu/kernels/fuse_batch_normalization/generic/fp32.cpp @@ -28,11 +28,19 @@ namespace arm_compute { namespace cpu { -void fused_batch_normalization_conv_f32(const ITensor *conv_weights, const ITensor *conv_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) +void fused_batch_normalization_conv_f32(const ITensor *conv_weights, + const ITensor *conv_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) { - return fused_batch_normalization_conv(conv_weights, conv_bias, fused_weights, fused_bias, - bn_mean, bn_var, bn_beta, bn_gamma, epsilon, window); + return fused_batch_normalization_conv(conv_weights, conv_bias, fused_weights, fused_bias, bn_mean, + bn_var, bn_beta, bn_gamma, epsilon, window); } } // namespace cpu } // namespace arm_compute diff --git a/src/cpu/kernels/fuse_batch_normalization/generic/impl.h b/src/cpu/kernels/fuse_batch_normalization/generic/impl.h index b9017600d6..6fa843263a 100644 --- a/src/cpu/kernels/fuse_batch_normalization/generic/impl.h +++ b/src/cpu/kernels/fuse_batch_normalization/generic/impl.h @@ -25,6 +25,7 @@ #define SRC_CORE_NEON_KERNELS_FUSE_BATCH_NORMALIZATION_GENERIC_IMPL_H #include "arm_compute/core/Helpers.h" + #include "src/core/NEON/wrapper/wrapper.h" namespace arm_compute @@ -32,8 +33,16 @@ namespace arm_compute namespace cpu { template -void fused_batch_normalization_conv(const ITensor *conv_weights, const ITensor *conv_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) +void fused_batch_normalization_conv(const ITensor *conv_weights, + const ITensor *conv_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 / conv_weights->info()->element_size(); @@ -53,13 +62,20 @@ void fused_batch_normalization_conv(const ITensor *conv_weights, const ITensor * Iterator conv_w_in(conv_weights, win); Iterator conv_w_out(run_in_place_weights ? conv_weights : fused_weights, win); - const auto conv_bias_in = (conv_bias != nullptr ? reinterpret_cast(conv_bias->ptr_to_element(Coordinates(0, 0))) : nullptr); - auto conv_bias_out = (run_in_place_bias ? conv_bias_in : reinterpret_cast(fused_bias->ptr_to_element(Coordinates(0, 0)))); + const auto conv_bias_in = + (conv_bias != nullptr ? reinterpret_cast(conv_bias->ptr_to_element(Coordinates(0, 0))) : nullptr); + auto conv_bias_out = + (run_in_place_bias ? conv_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; + 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{}); @@ -73,59 +89,61 @@ void fused_batch_normalization_conv(const ITensor *conv_weights, const ITensor * auto gamma = ScalarType(1.0); auto beta = ScalarType(0.0); auto conv_bias_in_scalar = ScalarType(0.0); - execute_window_loop(win, [&](const Coordinates & id) - { - var = input_var[id[3]]; - if(input_gamma != nullptr) + execute_window_loop( + win, + [&](const Coordinates &id) { - gamma = input_gamma[id[3]]; - } + var = input_var[id[3]]; + if (input_gamma != nullptr) + { + gamma = input_gamma[id[3]]; + } - if((id[0] == 0) && (id[1] == 0) && (id[2] == 0)) - { - if(input_beta != nullptr) + if ((id[0] == 0) && (id[1] == 0) && (id[2] == 0)) { - beta = input_beta[id[3]]; - beta_vec = wrapper::vdup_n(beta, ExactTagType{}); + if (input_beta != nullptr) + { + beta = input_beta[id[3]]; + beta_vec = wrapper::vdup_n(beta, ExactTagType{}); + } + + // Construct vectors + mean = input_mean[id[3]]; + mean_vec = wrapper::vdup_n(mean, ExactTagType{}); + + if (conv_bias_in != nullptr) + { + conv_bias_in_scalar = conv_bias_in[id[3]]; + } + auto conv_bias_tmp_scalar = (conv_bias_in_scalar - mean) / std::sqrt(var + ScalarType(epsilon)); + conv_bias_out[id[3]] = (conv_bias_tmp_scalar * gamma) + beta; } - // Construct vectors - mean = input_mean[id[3]]; - mean_vec = wrapper::vdup_n(mean, ExactTagType{}); + int x = window_start_x; + auto conv_w_in_ptr = reinterpret_cast(conv_w_in.ptr()); + auto conv_w_out_ptr = reinterpret_cast(conv_w_out.ptr()); + var_vec = wrapper::vdup_n(var, ExactTagType{}); + gamma_vec = wrapper::vdup_n(gamma, ExactTagType{}); + rvar_vec = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)); - if(conv_bias_in != nullptr) + for (; x <= (window_end_x - window_step_x); x += window_step_x) { - conv_bias_in_scalar = conv_bias_in[id[3]]; - } - auto conv_bias_tmp_scalar = (conv_bias_in_scalar - mean) / std::sqrt(var + ScalarType(epsilon)); - conv_bias_out[id[3]] = (conv_bias_tmp_scalar * gamma) + beta; - } - - int x = window_start_x; - auto conv_w_in_ptr = reinterpret_cast(conv_w_in.ptr()); - auto conv_w_out_ptr = reinterpret_cast(conv_w_out.ptr()); - var_vec = wrapper::vdup_n(var, ExactTagType{}); - gamma_vec = wrapper::vdup_n(gamma, ExactTagType{}); - rvar_vec = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)); - - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - auto wn = wrapper::vloadq(conv_w_in_ptr + x); - wn = wrapper::vmul(wn, rvar_vec); - wn = wrapper::vmul(wn, gamma_vec); + auto wn = wrapper::vloadq(conv_w_in_ptr + x); + wn = wrapper::vmul(wn, rvar_vec); + wn = wrapper::vmul(wn, gamma_vec); - // Store results - wrapper::vstore(conv_w_out_ptr + x, wn); - } + // Store results + wrapper::vstore(conv_w_out_ptr + x, wn); + } - // Compute left-over elements - for(; x < window_end_x; ++x) - { - *(conv_w_out_ptr + x) = *(conv_w_in_ptr + x) / std::sqrt(var + ScalarType(epsilon)) * gamma; - } - }, - conv_w_in, conv_w_out); -} -} + // Compute left-over elements + for (; x < window_end_x; ++x) + { + *(conv_w_out_ptr + x) = *(conv_w_in_ptr + x) / std::sqrt(var + ScalarType(epsilon)) * gamma; + } + }, + conv_w_in, conv_w_out); } +} // namespace cpu +} // namespace arm_compute #endif //SRC_CORE_NEON_KERNELS_FUSE_BATCH_NORMALIZATION_GENERIC_IMPL_H diff --git a/src/cpu/kernels/fuse_batch_normalization/list.h b/src/cpu/kernels/fuse_batch_normalization/list.h index e25b1e5fed..a03dd74f78 100644 --- a/src/cpu/kernels/fuse_batch_normalization/list.h +++ b/src/cpu/kernels/fuse_batch_normalization/list.h @@ -30,15 +30,18 @@ namespace cpu { #define DECLARE_FUSE_BATCH_NORMALIZE_CONV_KERNEL(func_name) \ void func_name(const ITensor *conv_weights, const ITensor *conv_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) #define DECLARE_FUSE_BATCH_NORMALIZE_DWC_NCHW_CONV_KERNEL(func_name) \ void func_name(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) #define DECLARE_FUSE_BATCH_NORMALIZE_DWC_NHWC_CONV_KERNEL(func_name) \ void func_name(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) DECLARE_FUSE_BATCH_NORMALIZE_CONV_KERNEL(fused_batch_normalization_conv_f16); DECLARE_FUSE_BATCH_NORMALIZE_CONV_KERNEL(fused_batch_normalization_conv_f32); @@ -50,7 +53,7 @@ DECLARE_FUSE_BATCH_NORMALIZE_DWC_NCHW_CONV_KERNEL(fused_batch_normalization_dwc_ #undef DECLARE_FUSE_BATCH_NORMALIZE_CONV_KERNEL #undef DECLARE_FUSE_BATCH_NORMALIZE_DWC_NCHW_CONV_KERNEL #undef DECLARE_FUSE_BATCH_NORMALIZE_DWC_NHWC_CONV_KERNEL -} -} +} // namespace cpu +} // namespace arm_compute -#endif // \ No newline at end of file +#endif // diff --git a/src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp b/src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp index 1e3be8792d..c0b0dfd4dc 100644 --- a/src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp +++ b/src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp @@ -29,8 +29,16 @@ namespace arm_compute namespace cpu { template -void fused_batch_normalization_dwc_nchw(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) +void fused_batch_normalization_dwc_nchw(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(); @@ -50,13 +58,20 @@ void fused_batch_normalization_dwc_nchw(const ITensor *dwc_weights, const ITenso 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 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; + 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{}); @@ -70,74 +85,92 @@ void fused_batch_normalization_dwc_nchw(const ITensor *dwc_weights, const ITenso auto gamma = ScalarType(1.0); auto beta = ScalarType(0.0); auto dwc_bias_in_scalar = ScalarType(0.0); - execute_window_loop(win, [&](const Coordinates & id) - { - var = input_var[id[2]]; - if(input_gamma != nullptr) + execute_window_loop( + win, + [&](const Coordinates &id) { - gamma = input_gamma[id[2]]; - } - - if(id[1] == 0) - { - mean = input_mean[id[2]]; - - // Construct vectors - mean_vec = wrapper::vdup_n(mean, ExactTagType{}); - if(input_beta != nullptr) + var = input_var[id[2]]; + if (input_gamma != nullptr) { - beta = input_beta[id[2]]; - beta_vec = wrapper::vdup_n(beta, ExactTagType{}); + gamma = input_gamma[id[2]]; } - if(dwc_bias_in != nullptr) + if (id[1] == 0) { - dwc_bias_in_scalar = dwc_bias_in[id[2]]; + mean = input_mean[id[2]]; + + // Construct vectors + mean_vec = wrapper::vdup_n(mean, ExactTagType{}); + if (input_beta != nullptr) + { + beta = input_beta[id[2]]; + beta_vec = wrapper::vdup_n(beta, ExactTagType{}); + } + + if (dwc_bias_in != nullptr) + { + dwc_bias_in_scalar = dwc_bias_in[id[2]]; + } + + auto dwc_bias_tmp_scalar = (dwc_bias_in_scalar - mean) / std::sqrt(var + ScalarType(epsilon)); + dwc_bias_out[id[2]] = (dwc_bias_tmp_scalar * gamma) + beta; } - auto dwc_bias_tmp_scalar = (dwc_bias_in_scalar - mean) / std::sqrt(var + ScalarType(epsilon)); - dwc_bias_out[id[2]] = (dwc_bias_tmp_scalar * gamma) + beta; - } + int x = window_start_x; + auto dwc_w_in_ptr = reinterpret_cast(dwc_w_in.ptr()); + auto dwc_w_out_ptr = reinterpret_cast(dwc_w_out.ptr()); + var_vec = wrapper::vdup_n(var, ExactTagType{}); + gamma_vec = wrapper::vdup_n(gamma, ExactTagType{}); + rvar_vec = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)); - int x = window_start_x; - auto dwc_w_in_ptr = reinterpret_cast(dwc_w_in.ptr()); - auto dwc_w_out_ptr = reinterpret_cast(dwc_w_out.ptr()); - var_vec = wrapper::vdup_n(var, ExactTagType{}); - gamma_vec = wrapper::vdup_n(gamma, ExactTagType{}); - rvar_vec = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)); - - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - auto wn = wrapper::vloadq(dwc_w_in_ptr + x); - wn = wrapper::vmul(wn, rvar_vec); - wn = wrapper::vmul(wn, gamma_vec); + for (; x <= (window_end_x - window_step_x); x += window_step_x) + { + auto wn = wrapper::vloadq(dwc_w_in_ptr + x); + wn = wrapper::vmul(wn, rvar_vec); + wn = wrapper::vmul(wn, gamma_vec); - // Store results - wrapper::vstore(dwc_w_out_ptr + x, wn); - } + // Store results + wrapper::vstore(dwc_w_out_ptr + x, wn); + } - // Compute left-over elements - for(; x < window_end_x; ++x) - { - *(dwc_w_out_ptr + x) = *(dwc_w_in_ptr + x) / std::sqrt(var + ScalarType(epsilon)) * gamma; - } - }, - dwc_w_in, dwc_w_out); + // Compute left-over elements + for (; x < window_end_x; ++x) + { + *(dwc_w_out_ptr + x) = *(dwc_w_in_ptr + x) / std::sqrt(var + ScalarType(epsilon)) * gamma; + } + }, + dwc_w_in, dwc_w_out); } -void fused_batch_normalization_dwc_nchw_f32(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) +void fused_batch_normalization_dwc_nchw_f32(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) { - return fused_batch_normalization_dwc_nchw(dwc_weights, dwc_bias, fused_weights, fused_bias, - bn_mean, bn_var, bn_beta, bn_gamma, epsilon, window); + return fused_batch_normalization_dwc_nchw(dwc_weights, dwc_bias, fused_weights, fused_bias, bn_mean, + bn_var, bn_beta, bn_gamma, epsilon, window); } #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) -void fused_batch_normalization_dwc_nchw_f16(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) +void fused_batch_normalization_dwc_nchw_f16(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) { - return fused_batch_normalization_dwc_nchw(dwc_weights, dwc_bias, fused_weights, fused_bias, - bn_mean, bn_var, bn_beta, bn_gamma, epsilon, window); + return fused_batch_normalization_dwc_nchw(dwc_weights, dwc_bias, fused_weights, fused_bias, bn_mean, + bn_var, bn_beta, bn_gamma, epsilon, window); } #endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ diff --git a/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp16.cpp b/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp16.cpp index 275211ff38..1d88d3b494 100644 --- a/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp16.cpp +++ b/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp16.cpp @@ -30,11 +30,19 @@ namespace arm_compute { namespace cpu { -void fused_batch_normalization_dwc_nhwc_f16(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) +void fused_batch_normalization_dwc_nhwc_f16(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) { - return fused_batch_normalization_dwc_nhwc(dwc_weights, dwc_bias, fused_weights, fused_bias, - bn_mean, bn_var, bn_beta, bn_gamma, epsilon, window); + return fused_batch_normalization_dwc_nhwc(dwc_weights, dwc_bias, fused_weights, fused_bias, bn_mean, + bn_var, bn_beta, bn_gamma, epsilon, window); } } // namespace cpu diff --git a/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp32.cpp b/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp32.cpp index 67169c5325..1f336bb196 100644 --- a/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp32.cpp +++ b/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/fp32.cpp @@ -29,11 +29,19 @@ namespace arm_compute { namespace cpu { -void fused_batch_normalization_dwc_nhwc_f32(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) +void fused_batch_normalization_dwc_nhwc_f32(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) { - return fused_batch_normalization_dwc_nhwc(dwc_weights, dwc_bias, fused_weights, fused_bias, - bn_mean, bn_var, bn_beta, bn_gamma, epsilon, window); + return fused_batch_normalization_dwc_nhwc(dwc_weights, dwc_bias, fused_weights, fused_bias, bn_mean, + bn_var, bn_beta, bn_gamma, epsilon, window); } } // namespace cpu 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 6f0386276f..5b74a7aef6 100644 --- a/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.h +++ b/src/cpu/kernels/fuse_batch_normalization/nhwc/neon/impl.h @@ -25,6 +25,7 @@ #define SRC_CORE_NEON_KERNELS_FUSE_BATCH_NORMALIZATION_IMPL_H #include "arm_compute/core/Helpers.h" + #include "src/core/NEON/wrapper/wrapper.h" namespace arm_compute @@ -32,8 +33,16 @@ 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) +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(); @@ -53,13 +62,20 @@ void fused_batch_normalization_dwc_nhwc(const ITensor *dwc_weights, const ITenso 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 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; + 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{}); @@ -73,81 +89,84 @@ void fused_batch_normalization_dwc_nhwc(const ITensor *dwc_weights, const ITenso 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) + execute_window_loop( + win, + [&](const Coordinates &id) { - 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)) + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) { - mean_vec = wrapper::vloadq(input_mean + x); - - // Construct vectors - if(input_beta != nullptr) + var_vec = wrapper::vloadq(input_var + x); + if (input_gamma != nullptr) { - beta_vec = wrapper::vloadq(input_beta + x); + gamma_vec = wrapper::vloadq(input_gamma + x); } - if(dwc_bias_in != nullptr) + if ((id[2] == 0) && (id[1] == 0)) { - dwc_bias_vec = wrapper::vloadq(dwc_bias_in + x); + 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_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 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); + 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]; + // Store results + wrapper::vstore(dwc_w_out_ptr + x, wn); } - if(id[2] == 0 && id[1] == 0) + // Compute left-over elements + for (; x < window_end_x; ++x) { - auto mean = input_mean[x]; - if(input_beta != nullptr) + auto var = input_var[x]; + if (input_gamma != nullptr) { - beta = input_beta[x]; + gamma = input_gamma[x]; } - if(dwc_bias_in != nullptr) + + if (id[2] == 0 && id[1] == 0) { - dwc_bias_in_scalar = dwc_bias_in[x]; + 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; } - 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()); + 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); + *(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 -- cgit v1.2.1