diff options
Diffstat (limited to 'src/core/NEON/kernels/batchnormalization/impl')
5 files changed, 325 insertions, 243 deletions
diff --git a/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp index ed5254a0a4..e3d9b670b3 100644 --- a/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp +++ b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp @@ -24,8 +24,9 @@ #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensorPack.h" #include "arm_compute/core/Window.h" -#include "src/core/NEON/NEMath.h" + #include "src/core/NEON/kernels/detail/NEActivationFunctionDetail.h" +#include "src/core/NEON/NEMath.h" #include "src/core/NEON/wrapper/wrapper.h" #include <arm_neon.h> @@ -37,12 +38,26 @@ namespace arm_compute { namespace { -using BatchNomalizationPtr = void (*)(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, - float epsilon, ActivationLayerInfo &act_info, const Window &window); +using BatchNomalizationPtr = void (*)(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window); template <typename T> -void batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, - float epsilon, ActivationLayerInfo &act_info, const Window &window) +void batch_normalization(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window) { /** SIMD vector tag type. */ using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<float16_t, wrapper::traits::BitWidth::W128>; @@ -57,86 +72,99 @@ void batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const Iterator input(src, win_collapsed); Iterator output(dst, win_collapsed); - const auto input_mean = reinterpret_cast<const float16_t *>(mean->ptr_to_element(Coordinates(0, 0))); - const auto input_var = reinterpret_cast<const float16_t *>(var->ptr_to_element(Coordinates(0, 0))); - const auto input_gamma = (gamma != nullptr) ? reinterpret_cast<const float16_t *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; - const auto input_beta = (beta != nullptr) ? reinterpret_cast<const float16_t *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr; + const auto input_mean = reinterpret_cast<const float16_t *>(mean->ptr_to_element(Coordinates(0, 0))); + const auto input_var = reinterpret_cast<const float16_t *>(var->ptr_to_element(Coordinates(0, 0))); + const auto input_gamma = + (gamma != nullptr) ? reinterpret_cast<const float16_t *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; + const auto input_beta = + (beta != nullptr) ? reinterpret_cast<const float16_t *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr; T activation_functor(act_info); const auto epsilon_vec = wrapper::vdup_n(static_cast<float16_t>(epsilon), ExactTagType{}); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast<const float16_t *>(input.ptr()); - const auto output_ptr = reinterpret_cast<float16_t *>(output.ptr()); - - // Perform core calculations using vector operations - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) + execute_window_loop( + win_collapsed, + [&](const Coordinates &) { - // Conctruct vectors - const auto mean_vec = wrapper::vloadq(input_mean + x); - const auto var_vec = wrapper::vloadq(input_var + x); - const auto gamma_vec = (input_gamma != nullptr) ? wrapper::vloadq(input_gamma + x) : wrapper::vdup_n(static_cast<float16_t>(1.f), ExactTagType{}); - const auto beta_vec = (input_beta != nullptr) ? wrapper::vloadq(input_beta + x) : wrapper::vdup_n(static_cast<float16_t>(0.f), ExactTagType{}); - - // Calculate denominator - const auto denominator = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)); - - // Calculate x bar - const auto numerator = wrapper::vsub(wrapper::vloadq(input_ptr + x), mean_vec); - const auto x_bar = wrapper::vmul(numerator, denominator); - auto res = wrapper::vmla(beta_vec, x_bar, gamma_vec); - - // Perform fused activation - if(act_info.enabled()) + const auto input_ptr = reinterpret_cast<const float16_t *>(input.ptr()); + const auto output_ptr = reinterpret_cast<float16_t *>(output.ptr()); + + // Perform core calculations using vector operations + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) { - activation_functor(res); + // Conctruct vectors + const auto mean_vec = wrapper::vloadq(input_mean + x); + const auto var_vec = wrapper::vloadq(input_var + x); + const auto gamma_vec = (input_gamma != nullptr) + ? wrapper::vloadq(input_gamma + x) + : wrapper::vdup_n(static_cast<float16_t>(1.f), ExactTagType{}); + const auto beta_vec = (input_beta != nullptr) + ? wrapper::vloadq(input_beta + x) + : wrapper::vdup_n(static_cast<float16_t>(0.f), ExactTagType{}); + + // Calculate denominator + const auto denominator = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)); + + // Calculate x bar + const auto numerator = wrapper::vsub(wrapper::vloadq(input_ptr + x), mean_vec); + const auto x_bar = wrapper::vmul(numerator, denominator); + auto res = wrapper::vmla(beta_vec, x_bar, gamma_vec); + + // Perform fused activation + if (act_info.enabled()) + { + activation_functor(res); + } + + // Store results + wrapper::vstore(output_ptr + x, res); } - // Store results - wrapper::vstore(output_ptr + x, res); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - // Conctruct vectors - const float16_t gamma = (input_gamma != nullptr) ? input_gamma[x] : 1.f; - const float16_t beta = (input_beta != nullptr) ? input_beta[x] : 0.f; - - const float16_t denominator = sqrt(input_var[x] + epsilon); - const float16_t numerator = input_ptr[x] - input_mean[x]; - const float16_t x_bar = numerator / denominator; - float16_t res = beta + x_bar * gamma; - - // Perform fused activation - if(act_info.enabled()) + // Compute left-over elements + for (; x < window_end_x; ++x) { - activation_functor(res); + // Conctruct vectors + const float16_t gamma = (input_gamma != nullptr) ? input_gamma[x] : 1.f; + const float16_t beta = (input_beta != nullptr) ? input_beta[x] : 0.f; + + const float16_t denominator = sqrt(input_var[x] + epsilon); + const float16_t numerator = input_ptr[x] - input_mean[x]; + const float16_t x_bar = numerator / denominator; + float16_t res = beta + x_bar * gamma; + + // Perform fused activation + if (act_info.enabled()) + { + activation_functor(res); + } + + // Store results + *reinterpret_cast<float16_t *>(output_ptr + x) = res; } - - // Store results - *reinterpret_cast<float16_t *>(output_ptr + x) = res; - } - }, - input, output); + }, + input, output); } // Fused Batched Normalization with activation functions -static std::map<ActivationLayerInfo::ActivationFunction, BatchNomalizationPtr> fused_map = -{ - { ActivationLayerInfo::ActivationFunction::RELU, &batch_normalization<detail::relu<float16_t, 8>> }, - { ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &batch_normalization<detail::brelu<float16_t, 8>> }, - { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &batch_normalization<detail::lubrelu<float16_t, 8>> } -}; -} +static std::map<ActivationLayerInfo::ActivationFunction, BatchNomalizationPtr> fused_map = { + {ActivationLayerInfo::ActivationFunction::RELU, &batch_normalization<detail::relu<float16_t, 8>>}, + {ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &batch_normalization<detail::brelu<float16_t, 8>>}, + {ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &batch_normalization<detail::lubrelu<float16_t, 8>>}}; +} // namespace namespace cpu { -void fp16_neon_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, - float epsilon, ActivationLayerInfo &act_info, const Window &window) +void fp16_neon_batch_normalization(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window) { - if(act_info.enabled()) + if (act_info.enabled()) { fused_map[act_info.activation()](src, dst, mean, var, beta, gamma, epsilon, act_info, window); } diff --git a/src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp index d6e22e1843..4e1654ee6b 100644 --- a/src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp +++ b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp @@ -24,8 +24,9 @@ #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensorPack.h" #include "arm_compute/core/Window.h" -#include "src/core/NEON/NEMath.h" + #include "src/core/NEON/kernels/detail/NEActivationFunctionDetail.h" +#include "src/core/NEON/NEMath.h" #include "src/core/NEON/wrapper/wrapper.h" #include <arm_neon.h> @@ -36,12 +37,26 @@ namespace arm_compute { namespace { -using BatchNomalizationPtr = void (*)(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, - float epsilon, ActivationLayerInfo &act_info, const Window &window); +using BatchNomalizationPtr = void (*)(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window); template <typename T> -void batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, - float epsilon, ActivationLayerInfo &act_info, const Window &window) +void batch_normalization(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window) { /** SIMD vector tag type. */ using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<float, wrapper::traits::BitWidth::W128>; @@ -56,86 +71,99 @@ void batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const Iterator input(src, win_collapsed); Iterator output(dst, win_collapsed); - const auto input_mean = reinterpret_cast<const float *>(mean->ptr_to_element(Coordinates(0, 0))); - const auto input_var = reinterpret_cast<const float *>(var->ptr_to_element(Coordinates(0, 0))); - const auto input_gamma = (gamma != nullptr) ? reinterpret_cast<const float *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; - const auto input_beta = (beta != nullptr) ? reinterpret_cast<const float *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr; + const auto input_mean = reinterpret_cast<const float *>(mean->ptr_to_element(Coordinates(0, 0))); + const auto input_var = reinterpret_cast<const float *>(var->ptr_to_element(Coordinates(0, 0))); + const auto input_gamma = + (gamma != nullptr) ? reinterpret_cast<const float *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; + const auto input_beta = + (beta != nullptr) ? reinterpret_cast<const float *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr; T activation_functor(act_info); const auto epsilon_vec = wrapper::vdup_n(static_cast<float>(epsilon), ExactTagType{}); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast<const float *>(input.ptr()); - const auto output_ptr = reinterpret_cast<float *>(output.ptr()); - - // Perform core calculations using vector operations - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) + execute_window_loop( + win_collapsed, + [&](const Coordinates &) { - // Conctruct vectors - const auto mean_vec = wrapper::vloadq(input_mean + x); - const auto var_vec = wrapper::vloadq(input_var + x); - const auto gamma_vec = (input_gamma != nullptr) ? wrapper::vloadq(input_gamma + x) : wrapper::vdup_n(static_cast<float>(1.f), ExactTagType{}); - const auto beta_vec = (input_beta != nullptr) ? wrapper::vloadq(input_beta + x) : wrapper::vdup_n(static_cast<float>(0.f), ExactTagType{}); - - // Calculate denominator - const auto denominator = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)); - - // Calculate x bar - const auto numerator = wrapper::vsub(wrapper::vloadq(input_ptr + x), mean_vec); - const auto x_bar = wrapper::vmul(numerator, denominator); - auto res = wrapper::vmla(beta_vec, x_bar, gamma_vec); - - // Perform fused activation - if(act_info.enabled()) + const auto input_ptr = reinterpret_cast<const float *>(input.ptr()); + const auto output_ptr = reinterpret_cast<float *>(output.ptr()); + + // Perform core calculations using vector operations + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) { - activation_functor(res); + // Conctruct vectors + const auto mean_vec = wrapper::vloadq(input_mean + x); + const auto var_vec = wrapper::vloadq(input_var + x); + const auto gamma_vec = (input_gamma != nullptr) + ? wrapper::vloadq(input_gamma + x) + : wrapper::vdup_n(static_cast<float>(1.f), ExactTagType{}); + const auto beta_vec = (input_beta != nullptr) + ? wrapper::vloadq(input_beta + x) + : wrapper::vdup_n(static_cast<float>(0.f), ExactTagType{}); + + // Calculate denominator + const auto denominator = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)); + + // Calculate x bar + const auto numerator = wrapper::vsub(wrapper::vloadq(input_ptr + x), mean_vec); + const auto x_bar = wrapper::vmul(numerator, denominator); + auto res = wrapper::vmla(beta_vec, x_bar, gamma_vec); + + // Perform fused activation + if (act_info.enabled()) + { + activation_functor(res); + } + + // Store results + wrapper::vstore(output_ptr + x, res); } - // Store results - wrapper::vstore(output_ptr + x, res); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - // Conctruct vectors - const float gamma = (input_gamma != nullptr) ? input_gamma[x] : 1.f; - const float beta = (input_beta != nullptr) ? input_beta[x] : 0.f; - - const float denominator = sqrt(input_var[x] + epsilon); - const float numerator = input_ptr[x] - input_mean[x]; - const float x_bar = numerator / denominator; - float res = beta + x_bar * gamma; - - // Perform fused activation - if(act_info.enabled()) + // Compute left-over elements + for (; x < window_end_x; ++x) { - activation_functor(res); + // Conctruct vectors + const float gamma = (input_gamma != nullptr) ? input_gamma[x] : 1.f; + const float beta = (input_beta != nullptr) ? input_beta[x] : 0.f; + + const float denominator = sqrt(input_var[x] + epsilon); + const float numerator = input_ptr[x] - input_mean[x]; + const float x_bar = numerator / denominator; + float res = beta + x_bar * gamma; + + // Perform fused activation + if (act_info.enabled()) + { + activation_functor(res); + } + + // Store results + *reinterpret_cast<float *>(output_ptr + x) = res; } - - // Store results - *reinterpret_cast<float *>(output_ptr + x) = res; - } - }, - input, output); + }, + input, output); } // Fused Batched Normalization with activation functions -static std::map<ActivationLayerInfo::ActivationFunction, BatchNomalizationPtr> fused_map = -{ - { ActivationLayerInfo::ActivationFunction::RELU, &batch_normalization<detail::relu<float, 4>> }, - { ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &batch_normalization<detail::brelu<float, 4>> }, - { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &batch_normalization<detail::lubrelu<float, 4>> } -}; -} +static std::map<ActivationLayerInfo::ActivationFunction, BatchNomalizationPtr> fused_map = { + {ActivationLayerInfo::ActivationFunction::RELU, &batch_normalization<detail::relu<float, 4>>}, + {ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &batch_normalization<detail::brelu<float, 4>>}, + {ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &batch_normalization<detail::lubrelu<float, 4>>}}; +} // namespace namespace cpu { -void fp32_neon_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, - float epsilon, ActivationLayerInfo &act_info, const Window &window) +void fp32_neon_batch_normalization(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window) { - if(act_info.enabled()) + if (act_info.enabled()) { fused_map[act_info.activation()](src, dst, mean, var, beta, gamma, epsilon, act_info, window); } diff --git a/src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp b/src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp index 98cd9aa7fe..48caaa3e63 100644 --- a/src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp +++ b/src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp @@ -25,6 +25,7 @@ #include "arm_compute/core/ITensorPack.h" #include "arm_compute/core/Window.h" #include "arm_compute/function_info/ActivationLayerInfo.h" + #include "src/core/NEON/SVEMath.h" #include <cmath> @@ -37,8 +38,15 @@ namespace arm_compute { namespace cpu { -void fp16_sve_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, - float epsilon, ActivationLayerInfo &act_info, const Window &window) +void fp16_sve_batch_normalization(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window) { const auto window_start_x = static_cast<int>(window.x().start()); const auto window_end_x = static_cast<int>(window.x().end()); @@ -49,69 +57,74 @@ void fp16_sve_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mea Iterator input(src, win_collapsed); Iterator output(dst, win_collapsed); - const auto input_mean = reinterpret_cast<const float16_t *>(mean->ptr_to_element(Coordinates(0, 0))); - const auto input_var = reinterpret_cast<const float16_t *>(var->ptr_to_element(Coordinates(0, 0))); - const auto input_gamma = (gamma != nullptr) ? reinterpret_cast<const float16_t *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; - const auto input_beta = (beta != nullptr) ? reinterpret_cast<const float16_t *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr; + const auto input_mean = reinterpret_cast<const float16_t *>(mean->ptr_to_element(Coordinates(0, 0))); + const auto input_var = reinterpret_cast<const float16_t *>(var->ptr_to_element(Coordinates(0, 0))); + const auto input_gamma = + (gamma != nullptr) ? reinterpret_cast<const float16_t *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; + const auto input_beta = + (beta != nullptr) ? reinterpret_cast<const float16_t *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr; const auto epsilon_vec = svdup_n_f16(epsilon); const auto const_1 = svdup_n_f16(1.f); const auto const_0 = svdup_n_f16(0.f); const auto va = svdup_n_f16(act_info.a()); const auto vb = svdup_n_f16(act_info.b()); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast<const float16_t *>(input.ptr()); - const auto output_ptr = reinterpret_cast<float16_t *>(output.ptr()); - - // Compute S elements per iteration - int x = window_start_x; - svbool_t pg = svwhilelt_b16(x, window_end_x); - do + execute_window_loop( + win_collapsed, + [&](const Coordinates &) { - // Conctruct vectors - const auto mean_vec = svld1_f16(pg, input_mean + x); - const auto var_vec = svld1_f16(pg, input_var + x); - const auto gamma_vec = (input_gamma != nullptr) ? svld1_f16(pg, input_gamma + x) : const_1; - const auto beta_vec = (input_beta != nullptr) ? svld1_f16(pg, input_beta + x) : const_0; + const auto input_ptr = reinterpret_cast<const float16_t *>(input.ptr()); + const auto output_ptr = reinterpret_cast<float16_t *>(output.ptr()); - // Calculate denominator - const auto tmp = svadd_f16_z(pg, var_vec, epsilon_vec); - auto denominator = svrsqrte_f16(tmp); - denominator = svmul_f16_z(pg, svrsqrts_f16(svmul_f16_z(pg, tmp, denominator), denominator), denominator); - denominator = svmul_f16_z(pg, svrsqrts_f16(svmul_f16_z(pg, tmp, denominator), denominator), denominator); + // Compute S elements per iteration + int x = window_start_x; + svbool_t pg = svwhilelt_b16(x, window_end_x); + do + { + // Conctruct vectors + const auto mean_vec = svld1_f16(pg, input_mean + x); + const auto var_vec = svld1_f16(pg, input_var + x); + const auto gamma_vec = (input_gamma != nullptr) ? svld1_f16(pg, input_gamma + x) : const_1; + const auto beta_vec = (input_beta != nullptr) ? svld1_f16(pg, input_beta + x) : const_0; - // Calculate x bar - const auto numerator = svsub_f16_z(pg, svld1_f16(pg, input_ptr + x), mean_vec); - const auto x_bar = svmul_f16_z(pg, numerator, denominator); - auto res = svmla_f16_z(pg, beta_vec, x_bar, gamma_vec); + // Calculate denominator + const auto tmp = svadd_f16_z(pg, var_vec, epsilon_vec); + auto denominator = svrsqrte_f16(tmp); + denominator = + svmul_f16_z(pg, svrsqrts_f16(svmul_f16_z(pg, tmp, denominator), denominator), denominator); + denominator = + svmul_f16_z(pg, svrsqrts_f16(svmul_f16_z(pg, tmp, denominator), denominator), denominator); - // Perform fused activation - if(act_info.enabled()) - { - if(act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU) - { - res = svmax_f16_z(pg, const_0, res); - } - else if(act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) - { - res = svmin_f16_z(pg, va, svmax_f16_z(pg, const_0, res)); - } - else if(act_info.activation() == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + // Calculate x bar + const auto numerator = svsub_f16_z(pg, svld1_f16(pg, input_ptr + x), mean_vec); + const auto x_bar = svmul_f16_z(pg, numerator, denominator); + auto res = svmla_f16_z(pg, beta_vec, x_bar, gamma_vec); + + // Perform fused activation + if (act_info.enabled()) { - res = svmin_f16_z(pg, va, svmax_f16_z(pg, vb, res)); + if (act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU) + { + res = svmax_f16_z(pg, const_0, res); + } + else if (act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + { + res = svmin_f16_z(pg, va, svmax_f16_z(pg, const_0, res)); + } + else if (act_info.activation() == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + res = svmin_f16_z(pg, va, svmax_f16_z(pg, vb, res)); + } } - } - // Store results - svst1_f16(pg, output_ptr + x, res); + // Store results + svst1_f16(pg, output_ptr + x, res); - x += svcntw(); - pg = svwhilelt_b16(x, window_end_x); - } - while(svptest_any(svptrue_b16(), pg)); - }, - input, output); + x += svcntw(); + pg = svwhilelt_b16(x, window_end_x); + } while (svptest_any(svptrue_b16(), pg)); + }, + input, output); } } // namespace cpu } // namespace arm_compute diff --git a/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp b/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp index 952ab320bf..df4fbfe607 100644 --- a/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp +++ b/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp @@ -25,6 +25,7 @@ #include "arm_compute/core/ITensorPack.h" #include "arm_compute/core/Window.h" #include "arm_compute/function_info/ActivationLayerInfo.h" + #include "src/core/NEON/SVEMath.h" #include <cmath> @@ -37,8 +38,15 @@ namespace arm_compute { namespace cpu { -void fp32_sve_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, - float epsilon, ActivationLayerInfo &act_info, const Window &window) +void fp32_sve_batch_normalization(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window) { const auto window_start_x = static_cast<int>(window.x().start()); const auto window_end_x = static_cast<int>(window.x().end()); @@ -49,69 +57,74 @@ void fp32_sve_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mea Iterator input(src, win_collapsed); Iterator output(dst, win_collapsed); - const auto input_mean = reinterpret_cast<const float *>(mean->ptr_to_element(Coordinates(0, 0))); - const auto input_var = reinterpret_cast<const float *>(var->ptr_to_element(Coordinates(0, 0))); - const auto input_gamma = (gamma != nullptr) ? reinterpret_cast<const float *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; - const auto input_beta = (beta != nullptr) ? reinterpret_cast<const float *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr; + const auto input_mean = reinterpret_cast<const float *>(mean->ptr_to_element(Coordinates(0, 0))); + const auto input_var = reinterpret_cast<const float *>(var->ptr_to_element(Coordinates(0, 0))); + const auto input_gamma = + (gamma != nullptr) ? reinterpret_cast<const float *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; + const auto input_beta = + (beta != nullptr) ? reinterpret_cast<const float *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr; const auto epsilon_vec = svdup_n_f32(epsilon); const auto const_1 = svdup_n_f32(1.f); const auto const_0 = svdup_n_f32(0.f); const auto va = svdup_n_f32(act_info.a()); const auto vb = svdup_n_f32(act_info.b()); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast<const float *>(input.ptr()); - const auto output_ptr = reinterpret_cast<float *>(output.ptr()); - - // Compute S elements per iteration - int x = window_start_x; - svbool_t pg = svwhilelt_b32(x, window_end_x); - do + execute_window_loop( + win_collapsed, + [&](const Coordinates &) { - // Conctruct vectors - const auto mean_vec = svld1_f32(pg, input_mean + x); - const auto var_vec = svld1_f32(pg, input_var + x); - const auto gamma_vec = (input_gamma != nullptr) ? svld1_f32(pg, input_gamma + x) : const_1; - const auto beta_vec = (input_beta != nullptr) ? svld1_f32(pg, input_beta + x) : const_0; + const auto input_ptr = reinterpret_cast<const float *>(input.ptr()); + const auto output_ptr = reinterpret_cast<float *>(output.ptr()); - // Calculate denominator - const auto tmp = svadd_f32_z(pg, var_vec, epsilon_vec); - auto denominator = svrsqrte_f32(tmp); - denominator = svmul_f32_z(pg, svrsqrts_f32(svmul_f32_z(pg, tmp, denominator), denominator), denominator); - denominator = svmul_f32_z(pg, svrsqrts_f32(svmul_f32_z(pg, tmp, denominator), denominator), denominator); + // Compute S elements per iteration + int x = window_start_x; + svbool_t pg = svwhilelt_b32(x, window_end_x); + do + { + // Conctruct vectors + const auto mean_vec = svld1_f32(pg, input_mean + x); + const auto var_vec = svld1_f32(pg, input_var + x); + const auto gamma_vec = (input_gamma != nullptr) ? svld1_f32(pg, input_gamma + x) : const_1; + const auto beta_vec = (input_beta != nullptr) ? svld1_f32(pg, input_beta + x) : const_0; - // Calculate x bar - const auto numerator = svsub_f32_z(pg, svld1_f32(pg, input_ptr + x), mean_vec); - const auto x_bar = svmul_f32_z(pg, numerator, denominator); - auto res = svmla_f32_z(pg, beta_vec, x_bar, gamma_vec); + // Calculate denominator + const auto tmp = svadd_f32_z(pg, var_vec, epsilon_vec); + auto denominator = svrsqrte_f32(tmp); + denominator = + svmul_f32_z(pg, svrsqrts_f32(svmul_f32_z(pg, tmp, denominator), denominator), denominator); + denominator = + svmul_f32_z(pg, svrsqrts_f32(svmul_f32_z(pg, tmp, denominator), denominator), denominator); - // Perform fused activation - if(act_info.enabled()) - { - if(act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU) - { - res = svmax_f32_z(pg, const_0, res); - } - else if(act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) - { - res = svmin_f32_z(pg, va, svmax_f32_z(pg, const_0, res)); - } - else if(act_info.activation() == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + // Calculate x bar + const auto numerator = svsub_f32_z(pg, svld1_f32(pg, input_ptr + x), mean_vec); + const auto x_bar = svmul_f32_z(pg, numerator, denominator); + auto res = svmla_f32_z(pg, beta_vec, x_bar, gamma_vec); + + // Perform fused activation + if (act_info.enabled()) { - res = svmin_f32_z(pg, va, svmax_f32_z(pg, vb, res)); + if (act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU) + { + res = svmax_f32_z(pg, const_0, res); + } + else if (act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + { + res = svmin_f32_z(pg, va, svmax_f32_z(pg, const_0, res)); + } + else if (act_info.activation() == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + res = svmin_f32_z(pg, va, svmax_f32_z(pg, vb, res)); + } } - } - // Store results - svst1_f32(pg, output_ptr + x, res); + // Store results + svst1_f32(pg, output_ptr + x, res); - x += svcntw(); - pg = svwhilelt_b32(x, window_end_x); - } - while(svptest_any(svptrue_b32(), pg)); - }, - input, output); + x += svcntw(); + pg = svwhilelt_b32(x, window_end_x); + } while (svptest_any(svptrue_b32(), pg)); + }, + input, output); } } // namespace cpu } // namespace arm_compute diff --git a/src/core/NEON/kernels/batchnormalization/impl/list.h b/src/core/NEON/kernels/batchnormalization/impl/list.h index 8e0ea36f5a..cbf540bd71 100644 --- a/src/core/NEON/kernels/batchnormalization/impl/list.h +++ b/src/core/NEON/kernels/batchnormalization/impl/list.h @@ -28,9 +28,9 @@ namespace arm_compute { namespace cpu { -#define DECLARE_BATCH_NORMALIZATION_KERNEL(func_name) \ - void func_name(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, \ - float epsilon, ActivationLayerInfo &act_info, const Window &window) +#define DECLARE_BATCH_NORMALIZATION_KERNEL(func_name) \ + void func_name(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, \ + const ITensor *gamma, float epsilon, ActivationLayerInfo &act_info, const Window &window) DECLARE_BATCH_NORMALIZATION_KERNEL(fp16_neon_batch_normalization); DECLARE_BATCH_NORMALIZATION_KERNEL(fp16_sve_batch_normalization); |