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 --- .../NEON/kernels/NEReductionOperationKernel.cpp | 1955 ++++++++++---------- 1 file changed, 1024 insertions(+), 931 deletions(-) (limited to 'src/core/NEON/kernels/NEReductionOperationKernel.cpp') diff --git a/src/core/NEON/kernels/NEReductionOperationKernel.cpp b/src/core/NEON/kernels/NEReductionOperationKernel.cpp index 19955af493..455d604b3b 100644 --- a/src/core/NEON/kernels/NEReductionOperationKernel.cpp +++ b/src/core/NEON/kernels/NEReductionOperationKernel.cpp @@ -28,16 +28,17 @@ #include "arm_compute/core/ITensor.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Utils.h" -#include "arm_compute/core/Validate.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/core/Validate.h" + #include "src/core/CPP/Validate.h" -#include "src/core/NEON/INEKernel.h" -#include "src/core/NEON/NEMath.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" +#include "src/core/NEON/INEKernel.h" +#include "src/core/NEON/NEMath.h" +#include "src/core/NEON/wrapper/wrapper.h" #include "support/SaturateCast.h" -#include "src/core/NEON/wrapper/wrapper.h" #include namespace arm_compute @@ -48,7 +49,7 @@ namespace template void combine_and_store(int16x8_t t1, int16x8_t t2, Iterator &output, int offset = 0) { - if(std::is_same::value) + if (std::is_same::value) { auto res = wrapper::vcombine(wrapper::vqmovun(t1), wrapper::vqmovun(t2)); wrapper::vstore(output.ptr() + offset, res); @@ -63,8 +64,8 @@ void combine_and_store(int16x8_t t1, int16x8_t t2, Iterator &output, int offset template uint32x4x4_t calculate_index(uint32_t idx, T a, T b, uint32x4x4_t c, ReductionOperation op, int axis) { - uint32x4_t mask{ 0 }; - if(op == ReductionOperation::ARG_IDX_MIN) + uint32x4_t mask{0}; + if (op == ReductionOperation::ARG_IDX_MIN) { mask = wrapper::vcgt(b, a); } @@ -73,12 +74,12 @@ uint32x4x4_t calculate_index(uint32_t idx, T a, T b, uint32x4x4_t c, ReductionOp mask = wrapper::vclt(b, a); } - uint32x4_t vec_idx = { idx, idx + 1, idx + 2, idx + 3 }; - if(axis != 0) + uint32x4_t vec_idx = {idx, idx + 1, idx + 2, idx + 3}; + if (axis != 0) { vec_idx = wrapper::vdup_n(idx, wrapper::traits::vector_128_tag{}); } - uint32x4x4_t res = { { wrapper::vbsl(mask, vec_idx, c.val[0]), 0, 0, 0 } }; + uint32x4x4_t res = {{wrapper::vbsl(mask, vec_idx, c.val[0]), 0, 0, 0}}; return res; } @@ -86,9 +87,9 @@ uint32x4x4_t calculate_index(uint32_t idx, T a, T b, uint32x4x4_t c, ReductionOp template uint32x4x4_t calculate_index_quantized(uint32_t idx, T a, T b, uint32x4x4_t c, ReductionOperation op, int axis) { - uint32x4x4_t mask{ { 0 } }; - uint8x16_t mask_u8{ 0 }; - if(op == ReductionOperation::ARG_IDX_MIN) + uint32x4x4_t mask{{0}}; + uint8x16_t mask_u8{0}; + if (op == ReductionOperation::ARG_IDX_MIN) { mask_u8 = wrapper::vcgt(b, a); } @@ -96,44 +97,43 @@ uint32x4x4_t calculate_index_quantized(uint32_t idx, T a, T b, uint32x4x4_t c, R { mask_u8 = wrapper::vclt(b, a); } - auto wide_u16_1 = wrapper::vorr(vshll_n_u8(wrapper::vgetlow(mask_u8), 8), wrapper::vmovl(wrapper::vgetlow(mask_u8))); - auto wide_u16_2 = wrapper::vorr(vshll_n_u8(wrapper::vgethigh(mask_u8), 8), wrapper::vmovl(wrapper::vgethigh(mask_u8))); - mask.val[0] = wrapper::vorr(vshll_n_u16(wrapper::vgetlow(wide_u16_1), 16), wrapper::vmovl(wrapper::vgetlow(wide_u16_1))); - mask.val[1] = wrapper::vorr(vshll_n_u16(wrapper::vgethigh(wide_u16_1), 16), wrapper::vmovl(wrapper::vgethigh(wide_u16_1))); - mask.val[2] = wrapper::vorr(vshll_n_u16(wrapper::vgetlow(wide_u16_2), 16), wrapper::vmovl(wrapper::vgetlow(wide_u16_2))); - mask.val[3] = wrapper::vorr(vshll_n_u16(wrapper::vgethigh(wide_u16_2), 16), wrapper::vmovl(wrapper::vgethigh(wide_u16_2))); - - uint32x4x4_t vec_idx = { { { idx + 0, idx + 1, idx + 2, idx + 3 }, - { idx + 4, idx + 5, idx + 6, idx + 7 }, - { idx + 8, idx + 9, idx + 10, idx + 11 }, - { idx + 12, idx + 13, idx + 14, idx + 15 } - } - }; - if(axis != 0) + auto wide_u16_1 = + wrapper::vorr(vshll_n_u8(wrapper::vgetlow(mask_u8), 8), wrapper::vmovl(wrapper::vgetlow(mask_u8))); + auto wide_u16_2 = + wrapper::vorr(vshll_n_u8(wrapper::vgethigh(mask_u8), 8), wrapper::vmovl(wrapper::vgethigh(mask_u8))); + mask.val[0] = + wrapper::vorr(vshll_n_u16(wrapper::vgetlow(wide_u16_1), 16), wrapper::vmovl(wrapper::vgetlow(wide_u16_1))); + mask.val[1] = + wrapper::vorr(vshll_n_u16(wrapper::vgethigh(wide_u16_1), 16), wrapper::vmovl(wrapper::vgethigh(wide_u16_1))); + mask.val[2] = + wrapper::vorr(vshll_n_u16(wrapper::vgetlow(wide_u16_2), 16), wrapper::vmovl(wrapper::vgetlow(wide_u16_2))); + mask.val[3] = + wrapper::vorr(vshll_n_u16(wrapper::vgethigh(wide_u16_2), 16), wrapper::vmovl(wrapper::vgethigh(wide_u16_2))); + + uint32x4x4_t vec_idx = {{{idx + 0, idx + 1, idx + 2, idx + 3}, + {idx + 4, idx + 5, idx + 6, idx + 7}, + {idx + 8, idx + 9, idx + 10, idx + 11}, + {idx + 12, idx + 13, idx + 14, idx + 15}}}; + if (axis != 0) { vec_idx.val[0] = wrapper::vdup_n(idx, wrapper::traits::vector_128_tag{}); vec_idx.val[1] = wrapper::vdup_n(idx, wrapper::traits::vector_128_tag{}); vec_idx.val[2] = wrapper::vdup_n(idx, wrapper::traits::vector_128_tag{}); vec_idx.val[3] = wrapper::vdup_n(idx, wrapper::traits::vector_128_tag{}); } - uint32x4x4_t res = - { - { - vbslq_u32(mask.val[0], vec_idx.val[0], c.val[0]), - vbslq_u32(mask.val[1], vec_idx.val[1], c.val[1]), - vbslq_u32(mask.val[2], vec_idx.val[2], c.val[2]), - vbslq_u32(mask.val[3], vec_idx.val[3], c.val[3]) - } - }; + uint32x4x4_t res = { + {vbslq_u32(mask.val[0], vec_idx.val[0], c.val[0]), vbslq_u32(mask.val[1], vec_idx.val[1], c.val[1]), + vbslq_u32(mask.val[2], vec_idx.val[2], c.val[2]), vbslq_u32(mask.val[3], vec_idx.val[3], c.val[3])}}; return res; } // Helper function to calculate the minimum value of the input vector. All the elements in the output vector contain the min value. template -inline typename std::enable_if < std::is_same::value || std::is_same::value, - typename std::conditional::value, float32x2_t, int32x2_t>::type >::type - calculate_min(T in) +inline typename std::enable_if< + std::is_same::value || std::is_same::value, + typename std::conditional::value, float32x2_t, int32x2_t>::type>::type +calculate_min(T in) { auto pmin = wrapper::vpmin(wrapper::vgethigh(in), wrapper::vgetlow(in)); return wrapper::vpmin(pmin, pmin); @@ -141,9 +141,10 @@ inline typename std::enable_if < std::is_same::value || std::is_ // Helper function to calculate the minimum value of the input vector. All the elements in the output vector contain the min value. template -inline typename std::enable_if < std::is_same::value || std::is_same::value, - typename std::conditional::value, uint8x8_t, int8x8_t>::type >::type - calculate_min(T in) +inline typename std::enable_if< + std::is_same::value || std::is_same::value, + typename std::conditional::value, uint8x8_t, int8x8_t>::type>::type +calculate_min(T in) { auto pmin = wrapper::vpmin(wrapper::vgethigh(in), wrapper::vgetlow(in)); pmin = wrapper::vpmin(pmin, pmin); @@ -153,9 +154,10 @@ inline typename std::enable_if < std::is_same::value || std::is_s // Helper function to calculate the maximum value of the input vector. All the elements in the output vector contain the max value. template -inline typename std::enable_if < std::is_same::value || std::is_same::value, - typename std::conditional::value, float32x2_t, int32x2_t>::type >::type - calculate_max(T in) +inline typename std::enable_if< + std::is_same::value || std::is_same::value, + typename std::conditional::value, float32x2_t, int32x2_t>::type>::type +calculate_max(T in) { auto pmax = wrapper::vpmax(wrapper::vgethigh(in), wrapper::vgetlow(in)); return wrapper::vpmax(pmax, pmax); @@ -163,9 +165,10 @@ inline typename std::enable_if < std::is_same::value || std::is_ // Helper function to calculate the maximum value of the input vector. All the elements in the output vector contain the max value. template -inline typename std::enable_if < std::is_same::value || std::is_same::value, - typename std::conditional::value, uint8x8_t, int8x8_t>::type >::type - calculate_max(T in) +inline typename std::enable_if< + std::is_same::value || std::is_same::value, + typename std::conditional::value, uint8x8_t, int8x8_t>::type>::type +calculate_max(T in) { auto pmax = wrapper::vpmax(wrapper::vgethigh(in), wrapper::vgetlow(in)); pmax = wrapper::vpmax(pmax, pmax); @@ -176,10 +179,10 @@ inline typename std::enable_if < std::is_same::value || std::is_s template uint32_t calculate_vector_index(uint32x4x4_t vec_res_idx, T vec_res_value, ReductionOperation op) { - uint32x4_t res_idx_mask{ 0 }; + uint32x4_t res_idx_mask{0}; uint32x4_t mask_ones = vdupq_n_u32(0xFFFFFFFF); - if(op == ReductionOperation::ARG_IDX_MIN) + if (op == ReductionOperation::ARG_IDX_MIN) { auto pmin = calculate_min(vec_res_value); auto mask = wrapper::vceq(vec_res_value, wrapper::vcombine(pmin, pmin)); @@ -203,10 +206,10 @@ uint32_t calculate_vector_index(uint32x4x4_t vec_res_idx, T vec_res_value, Reduc template uint32_t calculate_vector_index_quantized(uint32x4x4_t vec_res_idx, T vec_res_value, ReductionOperation op) { - uint32x4x4_t res_idx_mask{ { 0 } }; + uint32x4x4_t res_idx_mask{{0}}; uint32x4_t mask_ones = vdupq_n_u32(0xFFFFFFFF); - uint8x16_t mask_u8{ 0 }; - if(op == ReductionOperation::ARG_IDX_MIN) + uint8x16_t mask_u8{0}; + if (op == ReductionOperation::ARG_IDX_MIN) { auto pmin = calculate_min(vec_res_value); mask_u8 = wrapper::vceq(vec_res_value, wrapper::vcombine(pmin, pmin)); @@ -218,12 +221,18 @@ uint32_t calculate_vector_index_quantized(uint32x4x4_t vec_res_idx, T vec_res_va } // Widen vectors - auto wide_u16_1 = wrapper::vorr(vshll_n_u8(wrapper::vgetlow(mask_u8), 8), wrapper::vmovl(wrapper::vgetlow(mask_u8))); - auto wide_u16_2 = wrapper::vorr(vshll_n_u8(wrapper::vgethigh(mask_u8), 8), wrapper::vmovl(wrapper::vgethigh(mask_u8))); - auto wide_u32_1 = wrapper::vorr(vshll_n_u16(wrapper::vgetlow(wide_u16_1), 16), wrapper::vmovl(wrapper::vgetlow(wide_u16_1))); - auto wide_u32_2 = wrapper::vorr(vshll_n_u16(wrapper::vgethigh(wide_u16_1), 16), wrapper::vmovl(wrapper::vgethigh(wide_u16_1))); - auto wide_u32_3 = wrapper::vorr(vshll_n_u16(wrapper::vgetlow(wide_u16_2), 16), wrapper::vmovl(wrapper::vgetlow(wide_u16_2))); - auto wide_u32_4 = wrapper::vorr(vshll_n_u16(wrapper::vgethigh(wide_u16_2), 16), wrapper::vmovl(wrapper::vgethigh(wide_u16_2))); + auto wide_u16_1 = + wrapper::vorr(vshll_n_u8(wrapper::vgetlow(mask_u8), 8), wrapper::vmovl(wrapper::vgetlow(mask_u8))); + auto wide_u16_2 = + wrapper::vorr(vshll_n_u8(wrapper::vgethigh(mask_u8), 8), wrapper::vmovl(wrapper::vgethigh(mask_u8))); + auto wide_u32_1 = + wrapper::vorr(vshll_n_u16(wrapper::vgetlow(wide_u16_1), 16), wrapper::vmovl(wrapper::vgetlow(wide_u16_1))); + auto wide_u32_2 = + wrapper::vorr(vshll_n_u16(wrapper::vgethigh(wide_u16_1), 16), wrapper::vmovl(wrapper::vgethigh(wide_u16_1))); + auto wide_u32_3 = + wrapper::vorr(vshll_n_u16(wrapper::vgetlow(wide_u16_2), 16), wrapper::vmovl(wrapper::vgetlow(wide_u16_2))); + auto wide_u32_4 = + wrapper::vorr(vshll_n_u16(wrapper::vgethigh(wide_u16_2), 16), wrapper::vmovl(wrapper::vgethigh(wide_u16_2))); res_idx_mask.val[0] = wrapper::vand(vec_res_idx.val[0], wide_u32_1); res_idx_mask.val[1] = wrapper::vand(vec_res_idx.val[1], wide_u32_2); res_idx_mask.val[2] = wrapper::vand(vec_res_idx.val[2], wide_u32_3); @@ -241,19 +250,19 @@ uint32_t calculate_vector_index_quantized(uint32x4x4_t vec_res_idx, T vec_res_va pmin = wrapper::vpmin(pmin, pmin); res = std::min(wrapper::vgetlane(pmin, 0), res); iter++; - } - while(iter < 4); + } while (iter < 4); return (res - 0xFFFFFFFF); } #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC template <> -uint32x4x4_t calculate_index(uint32_t idx, float16x8_t a, float16x8_t b, uint32x4x4_t c, ReductionOperation op, int axis) +uint32x4x4_t +calculate_index(uint32_t idx, float16x8_t a, float16x8_t b, uint32x4x4_t c, ReductionOperation op, int axis) { - uint32x4x2_t mask{ 0 }; - uint16x8_t mask_u16{ 0 }; - if(op == ReductionOperation::ARG_IDX_MIN) + uint32x4x2_t mask{0}; + uint16x8_t mask_u16{0}; + if (op == ReductionOperation::ARG_IDX_MIN) { mask_u16 = wrapper::vcgt(b, a); } @@ -263,19 +272,14 @@ uint32x4x4_t calculate_index(uint32_t idx, float16x8_t a, float16x8_t b, uint32x } mask.val[0] = wrapper::vmovl(wrapper::vgetlow(mask_u16)); mask.val[1] = wrapper::vmovl(wrapper::vgethigh(mask_u16)); - uint32x4x2_t vec_idx = { { { idx + 0, idx + 1, idx + 2, idx + 3 }, - { idx + 4, idx + 5, idx + 6, idx + 7 } - } - }; - if(axis != 0) + uint32x4x2_t vec_idx = {{{idx + 0, idx + 1, idx + 2, idx + 3}, {idx + 4, idx + 5, idx + 6, idx + 7}}}; + if (axis != 0) { vec_idx.val[0] = wrapper::vdup_n(idx, wrapper::traits::vector_128_tag{}); vec_idx.val[1] = wrapper::vdup_n(idx, wrapper::traits::vector_128_tag{}); } - uint32x4x4_t res = { wrapper::vbsl(mask.val[0], vec_idx.val[0], c.val[0]), - wrapper::vbsl(mask.val[1], vec_idx.val[1], c.val[1]), - 0, 0 - }; + uint32x4x4_t res = {wrapper::vbsl(mask.val[0], vec_idx.val[0], c.val[0]), + wrapper::vbsl(mask.val[1], vec_idx.val[1], c.val[1]), 0, 0}; return res; } @@ -298,10 +302,10 @@ inline float16x4_t calculate_max(float16x8_t in) template <> uint32_t calculate_vector_index(uint32x4x4_t vec_res_idx, float16x8_t vec_res_value, ReductionOperation op) { - uint32x4x2_t res_idx_mask{ 0 }; + uint32x4x2_t res_idx_mask{0}; uint32x4_t mask_ones = vdupq_n_u32(0xFFFFFFFF); uint16x8_t mask_u16; - if(op == ReductionOperation::ARG_IDX_MIN) + if (op == ReductionOperation::ARG_IDX_MIN) { auto pmin = calculate_min(vec_res_value); mask_u16 = wrapper::vceq(vec_res_value, wrapper::vcombine(pmin, pmin)); @@ -313,8 +317,10 @@ uint32_t calculate_vector_index(uint32x4x4_t vec_res_idx, float16x8_t vec_res_va } // Widen vectors - auto wide_u32_1 = wrapper::vorr(vshll_n_u16(wrapper::vgetlow(mask_u16), 8), wrapper::vmovl(wrapper::vgetlow(mask_u16))); - auto wide_u32_2 = wrapper::vorr(vshll_n_u16(wrapper::vgethigh(mask_u16), 8), wrapper::vmovl(wrapper::vgethigh(mask_u16))); + auto wide_u32_1 = + wrapper::vorr(vshll_n_u16(wrapper::vgetlow(mask_u16), 8), wrapper::vmovl(wrapper::vgetlow(mask_u16))); + auto wide_u32_2 = + wrapper::vorr(vshll_n_u16(wrapper::vgethigh(mask_u16), 8), wrapper::vmovl(wrapper::vgethigh(mask_u16))); res_idx_mask.val[0] = wrapper::vand(vec_res_idx.val[0], wide_u32_1); res_idx_mask.val[1] = wrapper::vand(vec_res_idx.val[1], wide_u32_2); res_idx_mask.val[0] = wrapper::vadd(res_idx_mask.val[0], mask_ones); @@ -328,8 +334,7 @@ uint32_t calculate_vector_index(uint32x4x4_t vec_res_idx, float16x8_t vec_res_va pmin = wrapper::vpmin(pmin, pmin); res = std::min(wrapper::vgetlane(pmin, 0), res); iter++; - } - while(iter < 2); + } while (iter < 2); return (res - 0xFFFFFFFF); } @@ -388,7 +393,8 @@ struct RedOpX /** SIMD vector tag type. */ using ExactTagType = typename wrapper::traits::neon_vector::tag_type; - inline void operator()(const Window &in_window, Window &out_window, const ITensor *in, ITensor *out, const ReductionOperation op) + inline void operator()( + const Window &in_window, Window &out_window, const ITensor *in, ITensor *out, const ReductionOperation op) { const size_t input_dim_0 = in->info()->dimension(0); const int window_step_x = 16 / sizeof(T); @@ -402,211 +408,217 @@ struct RedOpX Iterator output(out, out_window); execute_window_loop( - in_win_no_pad, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast(input.ptr()); - - auto init_res_value = static_cast(0.f); - switch(op) + in_win_no_pad, + [&](const Coordinates &) { - case ReductionOperation::ARG_IDX_MAX: - case ReductionOperation::ARG_IDX_MIN: - case ReductionOperation::MIN: - case ReductionOperation::MAX: - { - init_res_value = static_cast(*input_ptr); - break; - } - case ReductionOperation::PROD: - { - init_res_value = static_cast(1.f); - break; - } - default: - break; - } - auto vec_res_value = wrapper::vdup_n(init_res_value, ExactTagType{}); - uint32x4x4_t vec_res_idx{ { 0 } }; + const auto input_ptr = reinterpret_cast(input.ptr()); - // Compute window_step_x elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - const auto vec_elements = wrapper::vloadq(input_ptr + x); - switch(op) + auto init_res_value = static_cast(0.f); + switch (op) { - case ReductionOperation::SUM_SQUARE: - vec_res_value = wrapper::vadd(wrapper::vmul(vec_elements, vec_elements), vec_res_value); - break; - case ReductionOperation::MEAN_SUM: - case ReductionOperation::SUM: - vec_res_value = wrapper::vadd(vec_elements, vec_res_value); - break; - case ReductionOperation::PROD: - vec_res_value = wrapper::vmul(vec_elements, vec_res_value); - break; - case ReductionOperation::ARG_IDX_MIN: - { - auto temp_vec_res_value = wrapper::vmin(vec_elements, vec_res_value); - vec_res_idx = calculate_index(x, temp_vec_res_value, vec_res_value, vec_res_idx, op, 0); - vec_res_value = temp_vec_res_value; - break; - } case ReductionOperation::ARG_IDX_MAX: - { - auto temp_vec_res_value = wrapper::vmax(vec_elements, vec_res_value); - vec_res_idx = calculate_index(x, temp_vec_res_value, vec_res_value, vec_res_idx, op, 0); - vec_res_value = temp_vec_res_value; - break; - } + case ReductionOperation::ARG_IDX_MIN: case ReductionOperation::MIN: + case ReductionOperation::MAX: { - vec_res_value = wrapper::vmin(vec_elements, vec_res_value); + init_res_value = static_cast(*input_ptr); break; } - case ReductionOperation::MAX: + case ReductionOperation::PROD: { - vec_res_value = wrapper::vmax(vec_elements, vec_res_value); + init_res_value = static_cast(1.f); break; } default: - ARM_COMPUTE_ERROR("Not supported"); + break; } - } + auto vec_res_value = wrapper::vdup_n(init_res_value, ExactTagType{}); + uint32x4x4_t vec_res_idx{{0}}; - switch(op) - { - case ReductionOperation::SUM: - case ReductionOperation::MEAN_SUM: - case ReductionOperation::SUM_SQUARE: + // Compute window_step_x elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) { -#ifdef ARM_COMPUTE_DEBUG_ENABLED - auto res = static_cast(0.f); - for(int i = 0; i < S; ++i) + const auto vec_elements = wrapper::vloadq(input_ptr + x); + switch (op) { - res += wrapper::vgetlane(vec_res_value, i); + case ReductionOperation::SUM_SQUARE: + vec_res_value = wrapper::vadd(wrapper::vmul(vec_elements, vec_elements), vec_res_value); + break; + case ReductionOperation::MEAN_SUM: + case ReductionOperation::SUM: + vec_res_value = wrapper::vadd(vec_elements, vec_res_value); + break; + case ReductionOperation::PROD: + vec_res_value = wrapper::vmul(vec_elements, vec_res_value); + break; + case ReductionOperation::ARG_IDX_MIN: + { + auto temp_vec_res_value = wrapper::vmin(vec_elements, vec_res_value); + vec_res_idx = calculate_index(x, temp_vec_res_value, vec_res_value, + vec_res_idx, op, 0); + vec_res_value = temp_vec_res_value; + break; + } + case ReductionOperation::ARG_IDX_MAX: + { + auto temp_vec_res_value = wrapper::vmax(vec_elements, vec_res_value); + vec_res_idx = calculate_index(x, temp_vec_res_value, vec_res_value, + vec_res_idx, op, 0); + vec_res_value = temp_vec_res_value; + break; + } + case ReductionOperation::MIN: + { + vec_res_value = wrapper::vmin(vec_elements, vec_res_value); + break; + } + case ReductionOperation::MAX: + { + vec_res_value = wrapper::vmax(vec_elements, vec_res_value); + break; + } + default: + ARM_COMPUTE_ERROR("Not supported"); } -#else // ARM_COMPUTE_DEBUG_ENABLED - auto carry_res = wrapper::vpadd(wrapper::vgethigh(vec_res_value), wrapper::vgetlow(vec_res_value)); - for(int i = 0; i < S / 4; ++i) + } + + switch (op) + { + case ReductionOperation::SUM: + case ReductionOperation::MEAN_SUM: + case ReductionOperation::SUM_SQUARE: { - carry_res = wrapper::vpadd(carry_res, carry_res); - } - auto res = wrapper::vgetlane(carry_res, 0); +#ifdef ARM_COMPUTE_DEBUG_ENABLED + auto res = static_cast(0.f); + for (int i = 0; i < S; ++i) + { + res += wrapper::vgetlane(vec_res_value, i); + } +#else // ARM_COMPUTE_DEBUG_ENABLED + auto carry_res = + wrapper::vpadd(wrapper::vgethigh(vec_res_value), wrapper::vgetlow(vec_res_value)); + for (int i = 0; i < S / 4; ++i) + { + carry_res = wrapper::vpadd(carry_res, carry_res); + } + auto res = wrapper::vgetlane(carry_res, 0); #endif // ARM_COMPUTE_DEBUG_ENABLED - if(op == ReductionOperation::SUM_SQUARE) - { - // Compute left-over elements - for(; x < window_end_x; ++x) + if (op == ReductionOperation::SUM_SQUARE) { - res += (*(input_ptr + x)) * (*(input_ptr + x)); + // Compute left-over elements + for (; x < window_end_x; ++x) + { + res += (*(input_ptr + x)) * (*(input_ptr + x)); + } + } + else + { + // Compute left-over elements + for (; x < window_end_x; ++x) + { + res += *(input_ptr + x); + } + } + + if (op == ReductionOperation::MEAN_SUM) + { + res /= input_dim_0; } + + *(reinterpret_cast(output.ptr())) = res; + break; } - else + case ReductionOperation::PROD: { + auto carry_res = + wrapper::vmul(wrapper::vgethigh(vec_res_value), wrapper::vgetlow(vec_res_value)); + T res = 1; + for (int i = 0; i < S / 2; ++i) + { + res *= wrapper::vgetlane(carry_res, i); + } + // Compute left-over elements - for(; x < window_end_x; ++x) + for (; x < window_end_x; ++x) { - res += *(input_ptr + x); + res *= *(input_ptr + x); } - } - if(op == ReductionOperation::MEAN_SUM) - { - res /= input_dim_0; + *(reinterpret_cast(output.ptr())) = res; + break; } - - *(reinterpret_cast(output.ptr())) = res; - break; - } - case ReductionOperation::PROD: - { - auto carry_res = wrapper::vmul(wrapper::vgethigh(vec_res_value), wrapper::vgetlow(vec_res_value)); - T res = 1; - for(int i = 0; i < S / 2; ++i) + case ReductionOperation::ARG_IDX_MIN: { - res *= wrapper::vgetlane(carry_res, i); - } + auto idx = calculate_vector_index(vec_res_idx, vec_res_value, op); + auto res = static_cast(wrapper::vgetlane(calculate_min(vec_res_value), 0)); - // Compute left-over elements - for(; x < window_end_x; ++x) - { - res *= *(input_ptr + x); + // Compute left-over elements + for (; x < window_end_x; ++x) + { + if (*(input_ptr + x) < res) + { + idx = x; + res = *(input_ptr + x); + } + } + *(reinterpret_cast(output.ptr())) = idx; + break; } - - *(reinterpret_cast(output.ptr())) = res; - break; - } - case ReductionOperation::ARG_IDX_MIN: - { - auto idx = calculate_vector_index(vec_res_idx, vec_res_value, op); - auto res = static_cast(wrapper::vgetlane(calculate_min(vec_res_value), 0)); - - // Compute left-over elements - for(; x < window_end_x; ++x) + case ReductionOperation::ARG_IDX_MAX: { - if(*(input_ptr + x) < res) + auto idx = calculate_vector_index(vec_res_idx, vec_res_value, op); + auto res = static_cast(wrapper::vgetlane(calculate_max(vec_res_value), 0)); + + // Compute left-over elements + for (; x < window_end_x; ++x) { - idx = x; - res = *(input_ptr + x); + if (*(input_ptr + x) > res) + { + idx = x; + res = *(input_ptr + x); + } } + *(reinterpret_cast(output.ptr())) = idx; + break; } - *(reinterpret_cast(output.ptr())) = idx; - break; - } - case ReductionOperation::ARG_IDX_MAX: - { - auto idx = calculate_vector_index(vec_res_idx, vec_res_value, op); - auto res = static_cast(wrapper::vgetlane(calculate_max(vec_res_value), 0)); - - // Compute left-over elements - for(; x < window_end_x; ++x) + case ReductionOperation::MIN: { - if(*(input_ptr + x) > res) + auto res = static_cast(wrapper::vgetlane(calculate_min(vec_res_value), 0)); + + // Compute left-over elements + for (; x < window_end_x; ++x) { - idx = x; - res = *(input_ptr + x); + res = *(input_ptr + x) < res ? *(input_ptr + x) : res; } + *(reinterpret_cast(output.ptr())) = res; + break; } - *(reinterpret_cast(output.ptr())) = idx; - break; - } - case ReductionOperation::MIN: - { - auto res = static_cast(wrapper::vgetlane(calculate_min(vec_res_value), 0)); - - // Compute left-over elements - for(; x < window_end_x; ++x) + case ReductionOperation::MAX: { - res = *(input_ptr + x) < res ? *(input_ptr + x) : res; - } - *(reinterpret_cast(output.ptr())) = res; - break; - } - case ReductionOperation::MAX: - { - auto res = static_cast(wrapper::vgetlane(calculate_max(vec_res_value), 0)); + auto res = static_cast(wrapper::vgetlane(calculate_max(vec_res_value), 0)); - // Compute left-over elements - for(; x < window_end_x; ++x) - { - res = *(input_ptr + x) > res ? *(input_ptr + x) : res; + // Compute left-over elements + for (; x < window_end_x; ++x) + { + res = *(input_ptr + x) > res ? *(input_ptr + x) : res; + } + *(reinterpret_cast(output.ptr())) = res; + break; } - *(reinterpret_cast(output.ptr())) = res; - break; + default: + ARM_COMPUTE_ERROR("Not supported"); } - default: - ARM_COMPUTE_ERROR("Not supported"); - } - }, - input, output); + }, + input, output); } }; template struct RedOpX_quantized { - inline void operator()(const Window &in_window, Window &out_window, const ITensor *in, ITensor *out, const ReductionOperation op) + inline void operator()( + const Window &in_window, Window &out_window, const ITensor *in, ITensor *out, const ReductionOperation op) { using PromotedType = typename wrapper::traits::promote::type>::type; @@ -637,246 +649,257 @@ struct RedOpX_quantized const float B = out_offset - (in_scale * in_offset) / (out_scale); execute_window_loop( - in_win_no_pad, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast(input.ptr()); + in_win_no_pad, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast(input.ptr()); + + auto vec_res_value1 = + wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); + auto vec_res_value2 = + wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); + auto vec_res_value3 = + wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); + auto vec_res_value4 = + wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); + + auto vec_res_value1_f = vdupq_n_f32(static_cast(1.f)); + auto vec_res_value2_f = vdupq_n_f32(static_cast(1.f)); + auto vec_res_value3_f = vdupq_n_f32(static_cast(1.f)); + auto vec_res_value4_f = vdupq_n_f32(static_cast(1.f)); + + typename wrapper::traits::neon_vector::type vec_res_value = {0}; + + if (op == ReductionOperation::ARG_IDX_MAX || op == ReductionOperation::ARG_IDX_MIN || + op == ReductionOperation::MIN || op == ReductionOperation::MAX) + { + vec_res_value = wrapper::vdup_n(*input_ptr, wrapper::traits::vector_128_tag{}); + } - auto vec_res_value1 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); - auto vec_res_value2 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); - auto vec_res_value3 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); - auto vec_res_value4 = wrapper::vdup_n(static_cast(0.f), wrapper::traits::vector_128_tag{}); + uint32x4x4_t vec_res_idx{{0}}; + // Compute window_step_x elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vec_elements = wrapper::vloadq(input_ptr + x); + switch (op) + { + case ReductionOperation::SUM: + case ReductionOperation::MEAN_SUM: + { + const auto temp16x8t_1 = wrapper::vmovl(wrapper::vgetlow(vec_elements)); + const auto temp16x8t_2 = wrapper::vmovl(wrapper::vgethigh(vec_elements)); - auto vec_res_value1_f = vdupq_n_f32(static_cast(1.f)); - auto vec_res_value2_f = vdupq_n_f32(static_cast(1.f)); - auto vec_res_value3_f = vdupq_n_f32(static_cast(1.f)); - auto vec_res_value4_f = vdupq_n_f32(static_cast(1.f)); + const auto temp32x4t_1 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_1)); + const auto temp32x4t_2 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_1)); + const auto temp32x4t_3 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_2)); + const auto temp32x4t_4 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_2)); - typename wrapper::traits::neon_vector::type vec_res_value = { 0 }; + vec_res_value1 = wrapper::vadd(temp32x4t_1, vec_res_value1); + vec_res_value2 = wrapper::vadd(temp32x4t_2, vec_res_value2); + vec_res_value3 = wrapper::vadd(temp32x4t_3, vec_res_value3); + vec_res_value4 = wrapper::vadd(temp32x4t_4, vec_res_value4); + break; + } + case ReductionOperation::PROD: + { + const auto offset32x4f_4 = vdupq_n_f32(iq_info.offset); + const auto scale32x4f_4 = vdupq_n_f32(iq_info.scale); - if(op == ReductionOperation::ARG_IDX_MAX || op == ReductionOperation::ARG_IDX_MIN || op == ReductionOperation::MIN || op == ReductionOperation::MAX) - { - vec_res_value = wrapper::vdup_n(*input_ptr, wrapper::traits::vector_128_tag{}); - } + const auto temp16x8t_1 = wrapper::vmovl(wrapper::vgetlow(vec_elements)); + const auto temp16x8t_2 = wrapper::vmovl(wrapper::vgethigh(vec_elements)); - uint32x4x4_t vec_res_idx{ { 0 } }; - // Compute window_step_x elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - const auto vec_elements = wrapper::vloadq(input_ptr + x); - switch(op) - { - case ReductionOperation::SUM: - case ReductionOperation::MEAN_SUM: - { - const auto temp16x8t_1 = wrapper::vmovl(wrapper::vgetlow(vec_elements)); - const auto temp16x8t_2 = wrapper::vmovl(wrapper::vgethigh(vec_elements)); - - const auto temp32x4t_1 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_1)); - const auto temp32x4t_2 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_1)); - const auto temp32x4t_3 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_2)); - const auto temp32x4t_4 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_2)); - - vec_res_value1 = wrapper::vadd(temp32x4t_1, vec_res_value1); - vec_res_value2 = wrapper::vadd(temp32x4t_2, vec_res_value2); - vec_res_value3 = wrapper::vadd(temp32x4t_3, vec_res_value3); - vec_res_value4 = wrapper::vadd(temp32x4t_4, vec_res_value4); - break; - } - case ReductionOperation::PROD: - { - const auto offset32x4f_4 = vdupq_n_f32(iq_info.offset); - const auto scale32x4f_4 = vdupq_n_f32(iq_info.scale); - - const auto temp16x8t_1 = wrapper::vmovl(wrapper::vgetlow(vec_elements)); - const auto temp16x8t_2 = wrapper::vmovl(wrapper::vgethigh(vec_elements)); - - const auto temp32x4t_1 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_1)); - const auto temp32x4t_2 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_1)); - const auto temp32x4t_3 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_2)); - const auto temp32x4t_4 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_2)); - - auto temp32x4f_1 = wrapper::vcvt(temp32x4t_1); - auto temp32x4f_2 = wrapper::vcvt(temp32x4t_2); - auto temp32x4f_3 = wrapper::vcvt(temp32x4t_3); - auto temp32x4f_4 = wrapper::vcvt(temp32x4t_4); - - //de-quantize vec_elements - temp32x4f_1 = vmulq_f32(vsubq_f32(temp32x4f_1, offset32x4f_4), scale32x4f_4); - temp32x4f_2 = vmulq_f32(vsubq_f32(temp32x4f_2, offset32x4f_4), scale32x4f_4); - temp32x4f_3 = vmulq_f32(vsubq_f32(temp32x4f_3, offset32x4f_4), scale32x4f_4); - temp32x4f_4 = vmulq_f32(vsubq_f32(temp32x4f_4, offset32x4f_4), scale32x4f_4); - - vec_res_value1_f = vmulq_f32(temp32x4f_1, vec_res_value1_f); - vec_res_value2_f = vmulq_f32(temp32x4f_2, vec_res_value2_f); - vec_res_value3_f = vmulq_f32(temp32x4f_3, vec_res_value3_f); - vec_res_value4_f = vmulq_f32(temp32x4f_4, vec_res_value4_f); - break; + const auto temp32x4t_1 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_1)); + const auto temp32x4t_2 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_1)); + const auto temp32x4t_3 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_2)); + const auto temp32x4t_4 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_2)); + + auto temp32x4f_1 = wrapper::vcvt(temp32x4t_1); + auto temp32x4f_2 = wrapper::vcvt(temp32x4t_2); + auto temp32x4f_3 = wrapper::vcvt(temp32x4t_3); + auto temp32x4f_4 = wrapper::vcvt(temp32x4t_4); + + //de-quantize vec_elements + temp32x4f_1 = vmulq_f32(vsubq_f32(temp32x4f_1, offset32x4f_4), scale32x4f_4); + temp32x4f_2 = vmulq_f32(vsubq_f32(temp32x4f_2, offset32x4f_4), scale32x4f_4); + temp32x4f_3 = vmulq_f32(vsubq_f32(temp32x4f_3, offset32x4f_4), scale32x4f_4); + temp32x4f_4 = vmulq_f32(vsubq_f32(temp32x4f_4, offset32x4f_4), scale32x4f_4); + + vec_res_value1_f = vmulq_f32(temp32x4f_1, vec_res_value1_f); + vec_res_value2_f = vmulq_f32(temp32x4f_2, vec_res_value2_f); + vec_res_value3_f = vmulq_f32(temp32x4f_3, vec_res_value3_f); + vec_res_value4_f = vmulq_f32(temp32x4f_4, vec_res_value4_f); + break; + } + case ReductionOperation::ARG_IDX_MIN: + { + auto temp_vec_res_value = wrapper::vmin(vec_elements, vec_res_value); + vec_res_idx = calculate_index_quantized( + x, temp_vec_res_value, vec_res_value, vec_res_idx, op, 0); + vec_res_value = temp_vec_res_value; + break; + } + case ReductionOperation::ARG_IDX_MAX: + { + auto temp_vec_res_value = wrapper::vmax(vec_elements, vec_res_value); + vec_res_idx = calculate_index_quantized( + x, temp_vec_res_value, vec_res_value, vec_res_idx, op, 0); + vec_res_value = temp_vec_res_value; + break; + } + case ReductionOperation::MIN: + { + vec_res_value = wrapper::vmin(vec_elements, vec_res_value); + break; + } + case ReductionOperation::MAX: + { + vec_res_value = wrapper::vmax(vec_elements, vec_res_value); + break; + } + default: + ARM_COMPUTE_ERROR("Not supported"); } + } + + switch (op) + { case ReductionOperation::ARG_IDX_MIN: { - auto temp_vec_res_value = wrapper::vmin(vec_elements, vec_res_value); - vec_res_idx = calculate_index_quantized(x, temp_vec_res_value, vec_res_value, vec_res_idx, op, 0); - vec_res_value = temp_vec_res_value; + auto idx = + calculate_vector_index_quantized(vec_res_idx, vec_res_value, op); + auto res = static_cast(wrapper::vgetlane(calculate_min(vec_res_value), 0)); + + // Compute left-over elements + for (; x < window_end_x; ++x) + { + if (*(input_ptr + x) < res) + { + idx = x; + res = *(input_ptr + x); + } + } + *(reinterpret_cast(output.ptr())) = idx; break; } case ReductionOperation::ARG_IDX_MAX: { - auto temp_vec_res_value = wrapper::vmax(vec_elements, vec_res_value); - vec_res_idx = calculate_index_quantized(x, temp_vec_res_value, vec_res_value, vec_res_idx, op, 0); - vec_res_value = temp_vec_res_value; + auto idx = + calculate_vector_index_quantized(vec_res_idx, vec_res_value, op); + auto res = static_cast(wrapper::vgetlane(calculate_max(vec_res_value), 0)); + + // Compute left-over elements + for (; x < window_end_x; ++x) + { + if (*(input_ptr + x) > res) + { + idx = x; + res = *(input_ptr + x); + } + } + *(reinterpret_cast(output.ptr())) = idx; break; } case ReductionOperation::MIN: { - vec_res_value = wrapper::vmin(vec_elements, vec_res_value); + auto res = static_cast(wrapper::vgetlane(calculate_min(vec_res_value), 0)); + + // Compute left-over elements + for (; x < window_end_x; ++x) + { + res = *(input_ptr + x) < res ? *(input_ptr + x) : res; + } + *(reinterpret_cast(output.ptr())) = res; break; } case ReductionOperation::MAX: { - vec_res_value = wrapper::vmax(vec_elements, vec_res_value); - break; - } - default: - ARM_COMPUTE_ERROR("Not supported"); - } - } - - switch(op) - { - case ReductionOperation::ARG_IDX_MIN: - { - auto idx = calculate_vector_index_quantized(vec_res_idx, vec_res_value, op); - auto res = static_cast(wrapper::vgetlane(calculate_min(vec_res_value), 0)); - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - if(*(input_ptr + x) < res) - { - idx = x; - res = *(input_ptr + x); - } - } - *(reinterpret_cast(output.ptr())) = idx; - break; - } - case ReductionOperation::ARG_IDX_MAX: - { - auto idx = calculate_vector_index_quantized(vec_res_idx, vec_res_value, op); - auto res = static_cast(wrapper::vgetlane(calculate_max(vec_res_value), 0)); + auto res = static_cast(wrapper::vgetlane(calculate_max(vec_res_value), 0)); - // Compute left-over elements - for(; x < window_end_x; ++x) - { - if(*(input_ptr + x) > res) + // Compute left-over elements + for (; x < window_end_x; ++x) { - idx = x; - res = *(input_ptr + x); + res = *(input_ptr + x) > res ? *(input_ptr + x) : res; } + *(reinterpret_cast(output.ptr())) = res; + break; } - *(reinterpret_cast(output.ptr())) = idx; - break; - } - case ReductionOperation::MIN: - { - auto res = static_cast(wrapper::vgetlane(calculate_min(vec_res_value), 0)); - - // Compute left-over elements - for(; x < window_end_x; ++x) + case ReductionOperation::PROD: { - res = *(input_ptr + x) < res ? *(input_ptr + x) : res; - } - *(reinterpret_cast(output.ptr())) = res; - break; - } - case ReductionOperation::MAX: - { - auto res = static_cast(wrapper::vgetlane(calculate_max(vec_res_value), 0)); + auto carry_res = wrapper::vmul(vec_res_value1_f, vec_res_value2_f); + carry_res = wrapper::vmul(carry_res, vec_res_value3_f); + carry_res = wrapper::vmul(carry_res, vec_res_value4_f); - // Compute left-over elements - for(; x < window_end_x; ++x) - { - res = *(input_ptr + x) > res ? *(input_ptr + x) : res; - } - *(reinterpret_cast(output.ptr())) = res; - break; - } - case ReductionOperation::PROD: - { - auto carry_res = wrapper::vmul(vec_res_value1_f, vec_res_value2_f); - carry_res = wrapper::vmul(carry_res, vec_res_value3_f); - carry_res = wrapper::vmul(carry_res, vec_res_value4_f); + float res = wrapper::vgetlane(carry_res, 0); + res *= wrapper::vgetlane(carry_res, 1); + res *= wrapper::vgetlane(carry_res, 2); + res *= wrapper::vgetlane(carry_res, 3); - float res = wrapper::vgetlane(carry_res, 0); - res *= wrapper::vgetlane(carry_res, 1); - res *= wrapper::vgetlane(carry_res, 2); - res *= wrapper::vgetlane(carry_res, 3); + // Compute left-over elements + for (; x < window_end_x; ++x) + { + //de-quantize input + if (std::is_same::value) + { + res *= dequantize_qasymm8(*(input_ptr + x), iq_info); + } + else + { + res *= dequantize_qasymm8_signed(*(input_ptr + x), iq_info); + } + } - // Compute left-over elements - for(; x < window_end_x; ++x) - { - //de-quantize input - if(std::is_same::value) + //re-quantize result + if (std::is_same::value) { - res *= dequantize_qasymm8(*(input_ptr + x), iq_info); + res = quantize_qasymm8(res, iq_info); } else { - res *= dequantize_qasymm8_signed(*(input_ptr + x), iq_info); + res = quantize_qasymm8_signed(res, iq_info); } - } - //re-quantize result - if(std::is_same::value) - { - res = quantize_qasymm8(res, iq_info); + *reinterpret_cast(output.ptr()) = static_cast(res); + break; } - else + case ReductionOperation::SUM: + case ReductionOperation::MEAN_SUM: { - res = quantize_qasymm8_signed(res, iq_info); - } + auto carry_res = wrapper::vadd(vec_res_value1, vec_res_value2); + carry_res = wrapper::vadd(carry_res, vec_res_value3); + carry_res = wrapper::vadd(carry_res, vec_res_value4); - *reinterpret_cast(output.ptr()) = static_cast(res); - break; - } - case ReductionOperation::SUM: - case ReductionOperation::MEAN_SUM: - { - auto carry_res = wrapper::vadd(vec_res_value1, vec_res_value2); - carry_res = wrapper::vadd(carry_res, vec_res_value3); - carry_res = wrapper::vadd(carry_res, vec_res_value4); + auto carry_paddition = + wrapper::vpadd(wrapper::vgethigh(carry_res), wrapper::vgetlow(carry_res)); + carry_paddition = wrapper::vpadd(carry_paddition, carry_paddition); + auto res = static_cast(wrapper::vgetlane(carry_paddition, 0)); - auto carry_paddition = wrapper::vpadd(wrapper::vgethigh(carry_res), wrapper::vgetlow(carry_res)); - carry_paddition = wrapper::vpadd(carry_paddition, carry_paddition); - auto res = static_cast(wrapper::vgetlane(carry_paddition, 0)); + // Compute left-over elements + for (; x < window_end_x; ++x) + { + res += *(input_ptr + x); + } - // Compute left-over elements - for(; x < window_end_x; ++x) - { - res += *(input_ptr + x); - } + if (op == ReductionOperation::MEAN_SUM) + { + const int32_t resFinal = A * (static_cast(res)) + B; - if(op == ReductionOperation::MEAN_SUM) - { - const int32_t resFinal = A * (static_cast(res)) + B; + *reinterpret_cast(output.ptr()) = utils::cast::saturate_cast(resFinal); + } + else + { + // Subtract accumulated offsets + res -= (in_info.dimension(0) - 1) * iq_info.offset; + *reinterpret_cast(output.ptr()) = utils::cast::saturate_cast(res); + } - *reinterpret_cast(output.ptr()) = utils::cast::saturate_cast(resFinal); - } - else - { - // Subtract accumulated offsets - res -= (in_info.dimension(0) - 1) * iq_info.offset; - *reinterpret_cast(output.ptr()) = utils::cast::saturate_cast(res); + break; } - - break; + default: + ARM_COMPUTE_ERROR("Not supported"); } - default: - ARM_COMPUTE_ERROR("Not supported"); - } - }, - input, output); + }, + input, output); } }; @@ -887,7 +910,12 @@ struct RedOpYZW using ExactTagType = typename wrapper::traits::neon_vector::tag_type; using neon_vector = typename wrapper::traits::neon_vector::type; - inline void operator()(const Window &in_window, Window &out_window, const ITensor *in, ITensor *out, int axis, const ReductionOperation op) + inline void operator()(const Window &in_window, + Window &out_window, + const ITensor *in, + ITensor *out, + int axis, + const ReductionOperation op) { const TensorInfo in_info = *(in->info()); const int window_step_x = 16 / sizeof(T); @@ -900,203 +928,210 @@ struct RedOpYZW Window in_win_no_pad = in_window; in_win_no_pad.set(Window::DimX, Window::Dimension(window_start_x_tmp, window_end_x_tmp, in_window.shape().x())); Window out_win_no_pad = out_window; - out_win_no_pad.set(Window::DimX, Window::Dimension(window_start_x_tmp, window_end_x_tmp, out_window.shape().x())); + out_win_no_pad.set(Window::DimX, + Window::Dimension(window_start_x_tmp, window_end_x_tmp, out_window.shape().x())); Iterator input(in, in_win_no_pad); Iterator output(out, out_win_no_pad); execute_window_loop( - in_win_no_pad, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast(input.ptr()); - - // Compute window_step_x elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) + in_win_no_pad, + [&](const Coordinates &) { - neon_vector vec_res_value = { 0 }; - switch(op) - { - case ReductionOperation::ARG_IDX_MAX: - case ReductionOperation::ARG_IDX_MIN: - case ReductionOperation::MIN: - case ReductionOperation::MAX: - { - vec_res_value = wrapper::vloadq(input_ptr + x); - break; - } - case ReductionOperation::PROD: - { - vec_res_value = wrapper::vdup_n(static_cast(1.f), ExactTagType{}); - break; - } - default: - { - vec_res_value = wrapper::vdup_n(static_cast(0.f), ExactTagType{}); - break; - } - } - uint32x4x4_t vec_res_idx{ { 0 } }; + const auto input_ptr = reinterpret_cast(input.ptr()); - for(unsigned int dim = 0; dim < in_info.dimension(axis); ++dim) + // Compute window_step_x elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) { - const T *in_ptr = reinterpret_cast(input.ptr() + x * sizeof(T) + in_info.strides_in_bytes()[axis] * dim); - const auto vec_elements = wrapper::vloadq(in_ptr); - switch(op) + neon_vector vec_res_value = {0}; + switch (op) { - case ReductionOperation::SUM: - case ReductionOperation::MEAN_SUM: - vec_res_value = wrapper::vadd(vec_elements, vec_res_value); - break; - case ReductionOperation::SUM_SQUARE: - vec_res_value = wrapper::vadd(wrapper::vmul(vec_elements, vec_elements), vec_res_value); - break; - case ReductionOperation::PROD: - vec_res_value = wrapper::vmul(vec_elements, vec_res_value); - break; - case ReductionOperation::ARG_IDX_MIN: - { - auto temp_vec_res_value = wrapper::vmin(vec_elements, vec_res_value); - vec_res_idx = calculate_index(dim, temp_vec_res_value, vec_res_value, vec_res_idx, op, axis); - vec_res_value = temp_vec_res_value; - break; - } case ReductionOperation::ARG_IDX_MAX: + case ReductionOperation::ARG_IDX_MIN: + case ReductionOperation::MIN: + case ReductionOperation::MAX: { - auto temp_vec_res_value = wrapper::vmax(vec_elements, vec_res_value); - vec_res_idx = calculate_index(dim, temp_vec_res_value, vec_res_value, vec_res_idx, op, axis); - vec_res_value = temp_vec_res_value; + vec_res_value = wrapper::vloadq(input_ptr + x); break; } - case ReductionOperation::MIN: + case ReductionOperation::PROD: { - vec_res_value = wrapper::vmin(vec_elements, vec_res_value); + vec_res_value = wrapper::vdup_n(static_cast(1.f), ExactTagType{}); break; } - case ReductionOperation::MAX: + default: { - vec_res_value = wrapper::vmax(vec_elements, vec_res_value); + vec_res_value = wrapper::vdup_n(static_cast(0.f), ExactTagType{}); break; } - default: - ARM_COMPUTE_ERROR("Not supported"); } - } + uint32x4x4_t vec_res_idx{{0}}; - if(op == ReductionOperation::MEAN_SUM) - { - auto vec_width_inv = wrapper::vinv(wrapper::vdup_n(static_cast(in_info.dimension(axis)), ExactTagType{})); - vec_res_value = wrapper::vmul(vec_res_value, vec_width_inv); - } - - if(op == ReductionOperation::ARG_IDX_MIN || op == ReductionOperation::ARG_IDX_MAX) - { - wrapper::vstore(reinterpret_cast(output.ptr()) + x, vec_res_idx.val[0]); -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - if(std::is_same::value) + for (unsigned int dim = 0; dim < in_info.dimension(axis); ++dim) { - wrapper::vstore(reinterpret_cast(output.ptr()) + x + 4, vec_res_idx.val[1]); + const T *in_ptr = + reinterpret_cast(input.ptr() + x * sizeof(T) + in_info.strides_in_bytes()[axis] * dim); + const auto vec_elements = wrapper::vloadq(in_ptr); + switch (op) + { + case ReductionOperation::SUM: + case ReductionOperation::MEAN_SUM: + vec_res_value = wrapper::vadd(vec_elements, vec_res_value); + break; + case ReductionOperation::SUM_SQUARE: + vec_res_value = wrapper::vadd(wrapper::vmul(vec_elements, vec_elements), vec_res_value); + break; + case ReductionOperation::PROD: + vec_res_value = wrapper::vmul(vec_elements, vec_res_value); + break; + case ReductionOperation::ARG_IDX_MIN: + { + auto temp_vec_res_value = wrapper::vmin(vec_elements, vec_res_value); + vec_res_idx = + calculate_index(dim, temp_vec_res_value, vec_res_value, vec_res_idx, op, axis); + vec_res_value = temp_vec_res_value; + break; + } + case ReductionOperation::ARG_IDX_MAX: + { + auto temp_vec_res_value = wrapper::vmax(vec_elements, vec_res_value); + vec_res_idx = + calculate_index(dim, temp_vec_res_value, vec_res_value, vec_res_idx, op, axis); + vec_res_value = temp_vec_res_value; + break; + } + case ReductionOperation::MIN: + { + vec_res_value = wrapper::vmin(vec_elements, vec_res_value); + break; + } + case ReductionOperation::MAX: + { + vec_res_value = wrapper::vmax(vec_elements, vec_res_value); + break; + } + default: + ARM_COMPUTE_ERROR("Not supported"); + } } -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - } - else - { - wrapper::vstore(reinterpret_cast(output.ptr() + x * sizeof(T)), vec_res_value); - } - } - // Compute left-over elements - for(; x < window_end_x; ++x) - { - auto res_value = 0.f; - switch(op) - { - case ReductionOperation::ARG_IDX_MAX: - case ReductionOperation::ARG_IDX_MIN: - case ReductionOperation::MIN: - case ReductionOperation::MAX: + if (op == ReductionOperation::MEAN_SUM) { - res_value = *(input_ptr + x); - break; + auto vec_width_inv = + wrapper::vinv(wrapper::vdup_n(static_cast(in_info.dimension(axis)), ExactTagType{})); + vec_res_value = wrapper::vmul(vec_res_value, vec_width_inv); } - case ReductionOperation::PROD: + + if (op == ReductionOperation::ARG_IDX_MIN || op == ReductionOperation::ARG_IDX_MAX) { - res_value = static_cast(1.f); - break; + wrapper::vstore(reinterpret_cast(output.ptr()) + x, vec_res_idx.val[0]); +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + if (std::is_same::value) + { + wrapper::vstore(reinterpret_cast(output.ptr()) + x + 4, vec_res_idx.val[1]); + } +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC } - default: + else { - res_value = static_cast(0.f); - break; + wrapper::vstore(reinterpret_cast(output.ptr() + x * sizeof(T)), vec_res_value); } } - uint32_t res_idx = 0; - for(unsigned int dim = 0; dim < in_info.dimension(axis); ++dim) + // Compute left-over elements + for (; x < window_end_x; ++x) { - const T *in_ptr = reinterpret_cast(input.ptr() + x * sizeof(T) + in_info.strides_in_bytes()[axis] * dim); - - switch(op) + auto res_value = 0.f; + switch (op) { - case ReductionOperation::SUM: - case ReductionOperation::MEAN_SUM: - res_value += *in_ptr; - break; - case ReductionOperation::SUM_SQUARE: - res_value += *in_ptr * *in_ptr; - break; - case ReductionOperation::PROD: - res_value *= *in_ptr; - break; + case ReductionOperation::ARG_IDX_MAX: case ReductionOperation::ARG_IDX_MIN: + case ReductionOperation::MIN: + case ReductionOperation::MAX: { - if(*in_ptr < res_value) - { - res_value = *in_ptr; - res_idx = dim; - } + res_value = *(input_ptr + x); break; } - case ReductionOperation::ARG_IDX_MAX: + case ReductionOperation::PROD: { - if(*in_ptr > res_value) - { - res_value = *in_ptr; - res_idx = dim; - } + res_value = static_cast(1.f); break; } - case ReductionOperation::MIN: + default: { - res_value = *in_ptr < res_value ? *in_ptr : res_value; + res_value = static_cast(0.f); break; } - case ReductionOperation::MAX: + } + + uint32_t res_idx = 0; + for (unsigned int dim = 0; dim < in_info.dimension(axis); ++dim) + { + const T *in_ptr = + reinterpret_cast(input.ptr() + x * sizeof(T) + in_info.strides_in_bytes()[axis] * dim); + + switch (op) { - res_value = *in_ptr > res_value ? *in_ptr : res_value; - break; + case ReductionOperation::SUM: + case ReductionOperation::MEAN_SUM: + res_value += *in_ptr; + break; + case ReductionOperation::SUM_SQUARE: + res_value += *in_ptr * *in_ptr; + break; + case ReductionOperation::PROD: + res_value *= *in_ptr; + break; + case ReductionOperation::ARG_IDX_MIN: + { + if (*in_ptr < res_value) + { + res_value = *in_ptr; + res_idx = dim; + } + break; + } + case ReductionOperation::ARG_IDX_MAX: + { + if (*in_ptr > res_value) + { + res_value = *in_ptr; + res_idx = dim; + } + break; + } + case ReductionOperation::MIN: + { + res_value = *in_ptr < res_value ? *in_ptr : res_value; + break; + } + case ReductionOperation::MAX: + { + res_value = *in_ptr > res_value ? *in_ptr : res_value; + break; + } + default: + ARM_COMPUTE_ERROR("Not supported"); } - default: - ARM_COMPUTE_ERROR("Not supported"); } - } - if(op == ReductionOperation::MEAN_SUM) - { - res_value /= in_info.dimension(axis); - } + if (op == ReductionOperation::MEAN_SUM) + { + res_value /= in_info.dimension(axis); + } - if(op == ReductionOperation::ARG_IDX_MIN || op == ReductionOperation::ARG_IDX_MAX) - { - *(reinterpret_cast(output.ptr()) + x) = res_idx; - } - else - { - *(reinterpret_cast(output.ptr() + x * sizeof(T))) = res_value; + if (op == ReductionOperation::ARG_IDX_MIN || op == ReductionOperation::ARG_IDX_MAX) + { + *(reinterpret_cast(output.ptr()) + x) = res_idx; + } + else + { + *(reinterpret_cast(output.ptr() + x * sizeof(T))) = res_value; + } } - } - }, - input, output); + }, + input, output); } }; @@ -1107,7 +1142,8 @@ struct RedOpYZW_complex using ExactTagType = typename wrapper::traits::neon_vector::tag_type; using neon_vector = typename wrapper::traits::neon_vector::type; - inline void operator()(const Window &in_window, Window &out_window, const ITensor *in, ITensor *out, int, const ReductionOperation) + inline void operator()( + const Window &in_window, Window &out_window, const ITensor *in, ITensor *out, int, const ReductionOperation) { ARM_COMPUTE_ERROR_ON(axis != 2); ARM_COMPUTE_ERROR_ON(op != ReductionOperation::SUM); @@ -1124,70 +1160,77 @@ struct RedOpYZW_complex Window in_win_no_pad = in_window; in_win_no_pad.set(Window::DimX, Window::Dimension(window_start_x_tmp, window_end_x_tmp, in_window.shape().x())); Window out_win_no_pad = out_window; - out_win_no_pad.set(Window::DimX, Window::Dimension(window_start_x_tmp, window_end_x_tmp, out_window.shape().x())); + out_win_no_pad.set(Window::DimX, + Window::Dimension(window_start_x_tmp, window_end_x_tmp, out_window.shape().x())); Iterator input(in, in_win_no_pad); Iterator output(out, out_win_no_pad); execute_window_loop( - in_win_no_pad, [&](const Coordinates &) - { - // Compute window_step_x elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) + in_win_no_pad, + [&](const Coordinates &) { - neon_vector vec_res_value_0 = { 0 }; - neon_vector vec_res_value_1 = { 0 }; - - vec_res_value_0 = wrapper::vdup_n(static_cast(0.f), ExactTagType{}); - vec_res_value_1 = wrapper::vdup_n(static_cast(0.f), ExactTagType{}); - - T *out_ptr = reinterpret_cast(output.ptr() + 2 * x * sizeof(T)); - for(unsigned int dim = 0; dim < in_info.dimension(axis); ++dim) + // Compute window_step_x elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) { - T *in_ptr_0 = reinterpret_cast(input.ptr() + 2 * x * sizeof(T) + stride_z * dim); - T *in_ptr_1 = reinterpret_cast(input.ptr() + 2 * x * sizeof(T) + 16 + stride_z * dim); + neon_vector vec_res_value_0 = {0}; + neon_vector vec_res_value_1 = {0}; - const auto vec_elements_0 = wrapper::vloadq(in_ptr_0); - const auto vec_elements_1 = wrapper::vloadq(in_ptr_1); + vec_res_value_0 = wrapper::vdup_n(static_cast(0.f), ExactTagType{}); + vec_res_value_1 = wrapper::vdup_n(static_cast(0.f), ExactTagType{}); - vec_res_value_0 = wrapper::vadd(vec_elements_0, vec_res_value_0); - vec_res_value_1 = wrapper::vadd(vec_elements_1, vec_res_value_1); - } + T *out_ptr = reinterpret_cast(output.ptr() + 2 * x * sizeof(T)); + for (unsigned int dim = 0; dim < in_info.dimension(axis); ++dim) + { + T *in_ptr_0 = reinterpret_cast(input.ptr() + 2 * x * sizeof(T) + stride_z * dim); + T *in_ptr_1 = reinterpret_cast(input.ptr() + 2 * x * sizeof(T) + 16 + stride_z * dim); - wrapper::vstore(out_ptr, vec_res_value_0); - wrapper::vstore(out_ptr + 4, vec_res_value_1); - } + const auto vec_elements_0 = wrapper::vloadq(in_ptr_0); + const auto vec_elements_1 = wrapper::vloadq(in_ptr_1); - // Compute left-over elements - for(; x < window_end_x; ++x) - { - auto res_value_0 = 0.f; - auto res_value_1 = 0.f; + vec_res_value_0 = wrapper::vadd(vec_elements_0, vec_res_value_0); + vec_res_value_1 = wrapper::vadd(vec_elements_1, vec_res_value_1); + } - T *out_ptr = reinterpret_cast(output.ptr() + 2 * x * sizeof(T)); - for(unsigned int dim = 0; dim < in_info.dimension(axis); ++dim) + wrapper::vstore(out_ptr, vec_res_value_0); + wrapper::vstore(out_ptr + 4, vec_res_value_1); + } + + // Compute left-over elements + for (; x < window_end_x; ++x) { - T *in_ptr = reinterpret_cast(input.ptr() + 2 * x * sizeof(T) + stride_z * dim); - res_value_0 += *in_ptr; - res_value_1 += *(in_ptr + 1); + auto res_value_0 = 0.f; + auto res_value_1 = 0.f; + + T *out_ptr = reinterpret_cast(output.ptr() + 2 * x * sizeof(T)); + for (unsigned int dim = 0; dim < in_info.dimension(axis); ++dim) + { + T *in_ptr = reinterpret_cast(input.ptr() + 2 * x * sizeof(T) + stride_z * dim); + res_value_0 += *in_ptr; + res_value_1 += *(in_ptr + 1); + } + *out_ptr = res_value_0; + *(out_ptr + 1) = res_value_1; } - *out_ptr = res_value_0; - *(out_ptr + 1) = res_value_1; - } - }, - input, output); + }, + input, output); } }; template struct RedOpYZW_quantized { - inline void operator()(const Window &in_window, Window &out_window, const ITensor *in, ITensor *out, int axis, const ReductionOperation op) + inline void operator()(const Window &in_window, + Window &out_window, + const ITensor *in, + ITensor *out, + int axis, + const ReductionOperation op) { const TensorInfo in_info = *(in->info()); const UniformQuantizationInfo iq_info = in_info.quantization_info().uniform(); - using PromotedType = typename wrapper::traits::promote::type>::type; + using PromotedType = typename wrapper::traits::promote::type>::type; const auto oq_info = out->info()->quantization_info().uniform(); @@ -1201,12 +1244,14 @@ struct RedOpYZW_quantized Window in_win_no_pad = in_window; in_win_no_pad.set(Window::DimX, Window::Dimension(window_start_x_tmp, window_end_x_tmp, in_window.shape().x())); Window out_win_no_pad = out_window; - out_win_no_pad.set(Window::DimX, Window::Dimension(window_start_x_tmp, window_end_x_tmp, out_window.shape().x())); + out_win_no_pad.set(Window::DimX, + Window::Dimension(window_start_x_tmp, window_end_x_tmp, out_window.shape().x())); Iterator input(in, in_win_no_pad); Iterator output(out, out_win_no_pad); - using vector_type = typename wrapper::traits::neon_bitvector::type; + using vector_type = + typename wrapper::traits::neon_bitvector::type; using vector_type_f = typename wrapper::traits::neon_vector::type; vector_type vec_res_value1{}; @@ -1234,362 +1279,384 @@ struct RedOpYZW_quantized const auto vec_B = wrapper::vdup_n(static_cast(B), wrapper::traits::vector_128_tag{}); execute_window_loop( - in_win_no_pad, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast(input.ptr()); - - // Compute window_step_x elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) + in_win_no_pad, + [&](const Coordinates &) { - uint32x4x4_t vec_res_idx{ { 0 } }; - vec_res_value1 = wrapper::vdup_n(static_cast(0), wrapper::traits::vector_128_tag{}); - vec_res_value2 = wrapper::vdup_n(static_cast(0), wrapper::traits::vector_128_tag{}); - vec_res_value3 = wrapper::vdup_n(static_cast(0), wrapper::traits::vector_128_tag{}); - vec_res_value4 = wrapper::vdup_n(static_cast(0), wrapper::traits::vector_128_tag{}); + const auto input_ptr = reinterpret_cast(input.ptr()); + + // Compute window_step_x elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) + { + uint32x4x4_t vec_res_idx{{0}}; + vec_res_value1 = wrapper::vdup_n(static_cast(0), wrapper::traits::vector_128_tag{}); + vec_res_value2 = wrapper::vdup_n(static_cast(0), wrapper::traits::vector_128_tag{}); + vec_res_value3 = wrapper::vdup_n(static_cast(0), wrapper::traits::vector_128_tag{}); + vec_res_value4 = wrapper::vdup_n(static_cast(0), wrapper::traits::vector_128_tag{}); - vec_res_value1_f = wrapper::vdup_n(static_cast(1), wrapper::traits::vector_128_tag{}); - vec_res_value2_f = wrapper::vdup_n(static_cast(1), wrapper::traits::vector_128_tag{}); - vec_res_value3_f = wrapper::vdup_n(static_cast(1), wrapper::traits::vector_128_tag{}); - vec_res_value4_f = wrapper::vdup_n(static_cast(1), wrapper::traits::vector_128_tag{}); + vec_res_value1_f = wrapper::vdup_n(static_cast(1), wrapper::traits::vector_128_tag{}); + vec_res_value2_f = wrapper::vdup_n(static_cast(1), wrapper::traits::vector_128_tag{}); + vec_res_value3_f = wrapper::vdup_n(static_cast(1), wrapper::traits::vector_128_tag{}); + vec_res_value4_f = wrapper::vdup_n(static_cast(1), wrapper::traits::vector_128_tag{}); - auto vec_res_value = wrapper::vloadq(input_ptr + x); + auto vec_res_value = wrapper::vloadq(input_ptr + x); - for(unsigned int index_dim = 0; index_dim < in_info.dimension(axis); ++index_dim) - { - const T *in_ptr = input_ptr + x + in_info.strides_in_bytes()[axis] * index_dim; - const auto vec_elements = wrapper::vloadq(in_ptr); - switch(op) + for (unsigned int index_dim = 0; index_dim < in_info.dimension(axis); ++index_dim) { - case ReductionOperation::SUM: - case ReductionOperation::MEAN_SUM: + const T *in_ptr = input_ptr + x + in_info.strides_in_bytes()[axis] * index_dim; + const auto vec_elements = wrapper::vloadq(in_ptr); + switch (op) { - const auto temp16x8t_1 = wrapper::vmovl(wrapper::vgetlow(vec_elements)); - const auto temp16x8t_2 = wrapper::vmovl(wrapper::vgethigh(vec_elements)); - - const auto temp32x4t_1 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_1)); - const auto temp32x4t_2 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_1)); - const auto temp32x4t_3 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_2)); - const auto temp32x4t_4 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_2)); - - vec_res_value1 = wrapper::vadd(temp32x4t_1, vec_res_value1); - vec_res_value2 = wrapper::vadd(temp32x4t_2, vec_res_value2); - vec_res_value3 = wrapper::vadd(temp32x4t_3, vec_res_value3); - vec_res_value4 = wrapper::vadd(temp32x4t_4, vec_res_value4); - break; + case ReductionOperation::SUM: + case ReductionOperation::MEAN_SUM: + { + const auto temp16x8t_1 = wrapper::vmovl(wrapper::vgetlow(vec_elements)); + const auto temp16x8t_2 = wrapper::vmovl(wrapper::vgethigh(vec_elements)); + + const auto temp32x4t_1 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_1)); + const auto temp32x4t_2 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_1)); + const auto temp32x4t_3 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_2)); + const auto temp32x4t_4 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_2)); + + vec_res_value1 = wrapper::vadd(temp32x4t_1, vec_res_value1); + vec_res_value2 = wrapper::vadd(temp32x4t_2, vec_res_value2); + vec_res_value3 = wrapper::vadd(temp32x4t_3, vec_res_value3); + vec_res_value4 = wrapper::vadd(temp32x4t_4, vec_res_value4); + break; + } + case ReductionOperation::PROD: + { + const auto offset32x4f_4 = wrapper::vdup_n(static_cast(iq_info.offset), + wrapper::traits::vector_128_tag{}); + const auto scale32x4f_4 = + wrapper::vdup_n(iq_info.scale, wrapper::traits::vector_128_tag{}); + + const auto temp16x8t_1 = wrapper::vmovl(wrapper::vgetlow(vec_elements)); + const auto temp16x8t_2 = wrapper::vmovl(wrapper::vgethigh(vec_elements)); + + const auto temp32x4t_1 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_1)); + const auto temp32x4t_2 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_1)); + const auto temp32x4t_3 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_2)); + const auto temp32x4t_4 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_2)); + + auto temp32x4f_1 = wrapper::vcvt(temp32x4t_1); + auto temp32x4f_2 = wrapper::vcvt(temp32x4t_2); + auto temp32x4f_3 = wrapper::vcvt(temp32x4t_3); + auto temp32x4f_4 = wrapper::vcvt(temp32x4t_4); + + //de-quantize vec_elements + temp32x4f_1 = wrapper::vmul(wrapper::vsub(temp32x4f_1, offset32x4f_4), scale32x4f_4); + temp32x4f_2 = wrapper::vmul(wrapper::vsub(temp32x4f_2, offset32x4f_4), scale32x4f_4); + temp32x4f_3 = wrapper::vmul(wrapper::vsub(temp32x4f_3, offset32x4f_4), scale32x4f_4); + temp32x4f_4 = wrapper::vmul(wrapper::vsub(temp32x4f_4, offset32x4f_4), scale32x4f_4); + + vec_res_value1_f = wrapper::vmul(temp32x4f_1, vec_res_value1_f); + vec_res_value2_f = wrapper::vmul(temp32x4f_2, vec_res_value2_f); + vec_res_value3_f = wrapper::vmul(temp32x4f_3, vec_res_value3_f); + vec_res_value4_f = wrapper::vmul(temp32x4f_4, vec_res_value4_f); + break; + } + case ReductionOperation::ARG_IDX_MIN: + { + auto temp_vec_res_value = wrapper::vmin(vec_elements, vec_res_value); + vec_res_idx = calculate_index_quantized(index_dim, temp_vec_res_value, vec_res_value, + vec_res_idx, op, axis); + vec_res_value = temp_vec_res_value; + break; + } + case ReductionOperation::ARG_IDX_MAX: + { + auto temp_vec_res_value = wrapper::vmax(vec_elements, vec_res_value); + vec_res_idx = calculate_index_quantized(index_dim, temp_vec_res_value, vec_res_value, + vec_res_idx, op, axis); + vec_res_value = temp_vec_res_value; + break; + } + case ReductionOperation::MIN: + { + vec_res_value = wrapper::vmin(vec_elements, vec_res_value); + break; + } + case ReductionOperation::MAX: + { + vec_res_value = wrapper::vmax(vec_elements, vec_res_value); + break; + } + default: + ARM_COMPUTE_ERROR("Not supported"); } - case ReductionOperation::PROD: - { - const auto offset32x4f_4 = wrapper::vdup_n(static_cast(iq_info.offset), wrapper::traits::vector_128_tag{}); - const auto scale32x4f_4 = wrapper::vdup_n(iq_info.scale, wrapper::traits::vector_128_tag{}); - - const auto temp16x8t_1 = wrapper::vmovl(wrapper::vgetlow(vec_elements)); - const auto temp16x8t_2 = wrapper::vmovl(wrapper::vgethigh(vec_elements)); - - const auto temp32x4t_1 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_1)); - const auto temp32x4t_2 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_1)); - const auto temp32x4t_3 = wrapper::vmovl(wrapper::vgetlow(temp16x8t_2)); - const auto temp32x4t_4 = wrapper::vmovl(wrapper::vgethigh(temp16x8t_2)); - - auto temp32x4f_1 = wrapper::vcvt(temp32x4t_1); - auto temp32x4f_2 = wrapper::vcvt(temp32x4t_2); - auto temp32x4f_3 = wrapper::vcvt(temp32x4t_3); - auto temp32x4f_4 = wrapper::vcvt(temp32x4t_4); + } - //de-quantize vec_elements - temp32x4f_1 = wrapper::vmul(wrapper::vsub(temp32x4f_1, offset32x4f_4), scale32x4f_4); - temp32x4f_2 = wrapper::vmul(wrapper::vsub(temp32x4f_2, offset32x4f_4), scale32x4f_4); - temp32x4f_3 = wrapper::vmul(wrapper::vsub(temp32x4f_3, offset32x4f_4), scale32x4f_4); - temp32x4f_4 = wrapper::vmul(wrapper::vsub(temp32x4f_4, offset32x4f_4), scale32x4f_4); - - vec_res_value1_f = wrapper::vmul(temp32x4f_1, vec_res_value1_f); - vec_res_value2_f = wrapper::vmul(temp32x4f_2, vec_res_value2_f); - vec_res_value3_f = wrapper::vmul(temp32x4f_3, vec_res_value3_f); - vec_res_value4_f = wrapper::vmul(temp32x4f_4, vec_res_value4_f); - break; - } + switch (op) + { case ReductionOperation::ARG_IDX_MIN: - { - auto temp_vec_res_value = wrapper::vmin(vec_elements, vec_res_value); - vec_res_idx = calculate_index_quantized(index_dim, temp_vec_res_value, vec_res_value, vec_res_idx, op, axis); - vec_res_value = temp_vec_res_value; - break; - } case ReductionOperation::ARG_IDX_MAX: { - auto temp_vec_res_value = wrapper::vmax(vec_elements, vec_res_value); - vec_res_idx = calculate_index_quantized(index_dim, temp_vec_res_value, vec_res_value, vec_res_idx, op, axis); - vec_res_value = temp_vec_res_value; + wrapper::vstore(reinterpret_cast(output.ptr() + 4 * x), vec_res_idx.val[0]); + wrapper::vstore(reinterpret_cast(output.ptr() + 4 * x) + 4, vec_res_idx.val[1]); + wrapper::vstore(reinterpret_cast(output.ptr() + 4 * x) + 8, vec_res_idx.val[2]); + wrapper::vstore(reinterpret_cast(output.ptr() + 4 * x) + 12, + vec_res_idx.val[3]); break; } case ReductionOperation::MIN: - { - vec_res_value = wrapper::vmin(vec_elements, vec_res_value); - break; - } case ReductionOperation::MAX: { - vec_res_value = wrapper::vmax(vec_elements, vec_res_value); + wrapper::vstore(reinterpret_cast(output.ptr() + x), vec_res_value); break; } - default: - ARM_COMPUTE_ERROR("Not supported"); - } - } - - switch(op) - { - case ReductionOperation::ARG_IDX_MIN: - case ReductionOperation::ARG_IDX_MAX: - { - wrapper::vstore(reinterpret_cast(output.ptr() + 4 * x), vec_res_idx.val[0]); - wrapper::vstore(reinterpret_cast(output.ptr() + 4 * x) + 4, vec_res_idx.val[1]); - wrapper::vstore(reinterpret_cast(output.ptr() + 4 * x) + 8, vec_res_idx.val[2]); - wrapper::vstore(reinterpret_cast(output.ptr() + 4 * x) + 12, vec_res_idx.val[3]); - break; - } - case ReductionOperation::MIN: - case ReductionOperation::MAX: - { - wrapper::vstore(reinterpret_cast(output.ptr() + x), vec_res_value); - break; - } - case ReductionOperation::SUM: - { - // Subtract offsets - auto offsets = vdupq_n_s32((in_info.dimension(axis) - 1) * iq_info.offset); + case ReductionOperation::SUM: + { + // Subtract offsets + auto offsets = vdupq_n_s32((in_info.dimension(axis) - 1) * iq_info.offset); - auto vec_res_s_value1 = wrapper::vreinterpret(vec_res_value1); - auto vec_res_s_value2 = wrapper::vreinterpret(vec_res_value2); - auto vec_res_s_value3 = wrapper::vreinterpret(vec_res_value3); - auto vec_res_s_value4 = wrapper::vreinterpret(vec_res_value4); + auto vec_res_s_value1 = wrapper::vreinterpret(vec_res_value1); + auto vec_res_s_value2 = wrapper::vreinterpret(vec_res_value2); + auto vec_res_s_value3 = wrapper::vreinterpret(vec_res_value3); + auto vec_res_s_value4 = wrapper::vreinterpret(vec_res_value4); - vec_res_s_value1 = wrapper::vsub(vec_res_s_value1, offsets); - vec_res_s_value2 = wrapper::vsub(vec_res_s_value2, offsets); - vec_res_s_value3 = wrapper::vsub(vec_res_s_value3, offsets); - vec_res_s_value4 = wrapper::vsub(vec_res_s_value4, offsets); + vec_res_s_value1 = wrapper::vsub(vec_res_s_value1, offsets); + vec_res_s_value2 = wrapper::vsub(vec_res_s_value2, offsets); + vec_res_s_value3 = wrapper::vsub(vec_res_s_value3, offsets); + vec_res_s_value4 = wrapper::vsub(vec_res_s_value4, offsets); - const auto temp16x8t_1 = wrapper::vcombine(wrapper::vqmovn(vec_res_s_value1), wrapper::vqmovn(vec_res_s_value2)); - const auto temp16x8t_2 = wrapper::vcombine(wrapper::vqmovn(vec_res_s_value3), wrapper::vqmovn(vec_res_s_value4)); + const auto temp16x8t_1 = + wrapper::vcombine(wrapper::vqmovn(vec_res_s_value1), wrapper::vqmovn(vec_res_s_value2)); + const auto temp16x8t_2 = + wrapper::vcombine(wrapper::vqmovn(vec_res_s_value3), wrapper::vqmovn(vec_res_s_value4)); - combine_and_store(temp16x8t_1, temp16x8t_2, output, x); - break; - } - case ReductionOperation::MEAN_SUM: - { - vec_res_value1_f = wrapper::vmla(vec_B, wrapper::vcvt(vec_res_value1), vec_A); - vec_res_value2_f = wrapper::vmla(vec_B, wrapper::vcvt(vec_res_value2), vec_A); - vec_res_value3_f = wrapper::vmla(vec_B, wrapper::vcvt(vec_res_value3), vec_A); - vec_res_value4_f = wrapper::vmla(vec_B, wrapper::vcvt(vec_res_value4), vec_A); + combine_and_store(temp16x8t_1, temp16x8t_2, output, x); + break; + } + case ReductionOperation::MEAN_SUM: + { + vec_res_value1_f = wrapper::vmla(vec_B, wrapper::vcvt(vec_res_value1), vec_A); + vec_res_value2_f = wrapper::vmla(vec_B, wrapper::vcvt(vec_res_value2), vec_A); + vec_res_value3_f = wrapper::vmla(vec_B, wrapper::vcvt(vec_res_value3), vec_A); + vec_res_value4_f = wrapper::vmla(vec_B, wrapper::vcvt(vec_res_value4), vec_A); #ifdef __aarch64__ - vec_res_value1 = wrapper::vcvta(vec_res_value1_f); - vec_res_value2 = wrapper::vcvta(vec_res_value2_f); - vec_res_value3 = wrapper::vcvta(vec_res_value3_f); - vec_res_value4 = wrapper::vcvta(vec_res_value4_f); + vec_res_value1 = wrapper::vcvta(vec_res_value1_f); + vec_res_value2 = wrapper::vcvta(vec_res_value2_f); + vec_res_value3 = wrapper::vcvta(vec_res_value3_f); + vec_res_value4 = wrapper::vcvta(vec_res_value4_f); #else // defined(__aarch64__) - vec_res_value1 = wrapper::vcvt(vec_res_value1_f); - vec_res_value2 = wrapper::vcvt(vec_res_value2_f); - vec_res_value3 = wrapper::vcvt(vec_res_value3_f); - vec_res_value4 = wrapper::vcvt(vec_res_value4_f); + vec_res_value1 = wrapper::vcvt(vec_res_value1_f); + vec_res_value2 = wrapper::vcvt(vec_res_value2_f); + vec_res_value3 = wrapper::vcvt(vec_res_value3_f); + vec_res_value4 = wrapper::vcvt(vec_res_value4_f); #endif // __aarch64__ - const auto temp16x8t_1 = wrapper::vcombine(wrapper::vqmovn(vec_res_value1), wrapper::vqmovn(vec_res_value2)); - const auto temp16x8t_2 = wrapper::vcombine(wrapper::vqmovn(vec_res_value3), wrapper::vqmovn(vec_res_value4)); - auto res = wrapper::vcombine(wrapper::vqmovn(temp16x8t_1), wrapper::vqmovn(temp16x8t_2)); + const auto temp16x8t_1 = + wrapper::vcombine(wrapper::vqmovn(vec_res_value1), wrapper::vqmovn(vec_res_value2)); + const auto temp16x8t_2 = + wrapper::vcombine(wrapper::vqmovn(vec_res_value3), wrapper::vqmovn(vec_res_value4)); + auto res = wrapper::vcombine(wrapper::vqmovn(temp16x8t_1), wrapper::vqmovn(temp16x8t_2)); - wrapper::vstore(reinterpret_cast(output.ptr() + x), res); - break; - } - case ReductionOperation::PROD: - { - const auto offset32x4f_4 = wrapper::vdup_n(static_cast(iq_info.offset), wrapper::traits::vector_128_tag{}); - const auto iscale32x4f_4 = vinvq_f32(vdupq_n_f32(iq_info.scale)); - - //re-quantize - vec_res_value1_f = wrapper::vadd(wrapper::vmul(vec_res_value1_f, iscale32x4f_4), offset32x4f_4); - vec_res_value2_f = wrapper::vadd(wrapper::vmul(vec_res_value2_f, iscale32x4f_4), offset32x4f_4); - vec_res_value3_f = wrapper::vadd(wrapper::vmul(vec_res_value3_f, iscale32x4f_4), offset32x4f_4); - vec_res_value4_f = wrapper::vadd(wrapper::vmul(vec_res_value4_f, iscale32x4f_4), offset32x4f_4); - - vec_res_value1 = wrapper::vcvt(vec_res_value1_f); - vec_res_value2 = wrapper::vcvt(vec_res_value2_f); - vec_res_value3 = wrapper::vcvt(vec_res_value3_f); - vec_res_value4 = wrapper::vcvt(vec_res_value4_f); - - const auto temp16x8t_1 = wrapper::vcombine(wrapper::vqmovn(vec_res_value1), wrapper::vqmovn(vec_res_value2)); - const auto temp16x8t_2 = wrapper::vcombine(wrapper::vqmovn(vec_res_value3), wrapper::vqmovn(vec_res_value4)); - auto res = wrapper::vcombine(wrapper::vqmovn(temp16x8t_1), wrapper::vqmovn(temp16x8t_2)); - - wrapper::vstore(reinterpret_cast(output.ptr() + x), res); - break; + wrapper::vstore(reinterpret_cast(output.ptr() + x), res); + break; + } + case ReductionOperation::PROD: + { + const auto offset32x4f_4 = + wrapper::vdup_n(static_cast(iq_info.offset), wrapper::traits::vector_128_tag{}); + const auto iscale32x4f_4 = vinvq_f32(vdupq_n_f32(iq_info.scale)); + + //re-quantize + vec_res_value1_f = + wrapper::vadd(wrapper::vmul(vec_res_value1_f, iscale32x4f_4), offset32x4f_4); + vec_res_value2_f = + wrapper::vadd(wrapper::vmul(vec_res_value2_f, iscale32x4f_4), offset32x4f_4); + vec_res_value3_f = + wrapper::vadd(wrapper::vmul(vec_res_value3_f, iscale32x4f_4), offset32x4f_4); + vec_res_value4_f = + wrapper::vadd(wrapper::vmul(vec_res_value4_f, iscale32x4f_4), offset32x4f_4); + + vec_res_value1 = wrapper::vcvt(vec_res_value1_f); + vec_res_value2 = wrapper::vcvt(vec_res_value2_f); + vec_res_value3 = wrapper::vcvt(vec_res_value3_f); + vec_res_value4 = wrapper::vcvt(vec_res_value4_f); + + const auto temp16x8t_1 = + wrapper::vcombine(wrapper::vqmovn(vec_res_value1), wrapper::vqmovn(vec_res_value2)); + const auto temp16x8t_2 = + wrapper::vcombine(wrapper::vqmovn(vec_res_value3), wrapper::vqmovn(vec_res_value4)); + auto res = wrapper::vcombine(wrapper::vqmovn(temp16x8t_1), wrapper::vqmovn(temp16x8t_2)); + + wrapper::vstore(reinterpret_cast(output.ptr() + x), res); + break; + } + default: + ARM_COMPUTE_ERROR("Not supported"); } - default: - ARM_COMPUTE_ERROR("Not supported"); } - } - // Compute left-over elements - for(; x < window_end_x; ++x) - { - float res_value = 0.f; - int32_t res_value_q = 0; - - switch(op) + // Compute left-over elements + for (; x < window_end_x; ++x) { - case ReductionOperation::ARG_IDX_MAX: - case ReductionOperation::ARG_IDX_MIN: - case ReductionOperation::MIN: - case ReductionOperation::MAX: - { - res_value = *(input_ptr + x); - break; - } - case ReductionOperation::PROD: - { - res_value = static_cast(1.0f); - break; - } - default: - { - res_value = static_cast(0.0f); - break; - } - } - uint32_t res_idx = 0; + float res_value = 0.f; + int32_t res_value_q = 0; - for(unsigned int dim = 0; dim < in_info.dimension(axis); ++dim) - { - const T *in_ptr = reinterpret_cast(input.ptr() + x + in_info.strides_in_bytes()[axis] * dim); - switch(op) + switch (op) { - case ReductionOperation::SUM: + case ReductionOperation::ARG_IDX_MAX: + case ReductionOperation::ARG_IDX_MIN: + case ReductionOperation::MIN: + case ReductionOperation::MAX: { - res_value += *in_ptr; + res_value = *(input_ptr + x); break; } - case ReductionOperation::MEAN_SUM: + case ReductionOperation::PROD: { - res_value_q += *in_ptr; + res_value = static_cast(1.0f); break; } - case ReductionOperation::SUM_SQUARE: + default: { - res_value += *in_ptr * *in_ptr; + res_value = static_cast(0.0f); break; } - case ReductionOperation::PROD: + } + uint32_t res_idx = 0; + + for (unsigned int dim = 0; dim < in_info.dimension(axis); ++dim) + { + const T *in_ptr = + reinterpret_cast(input.ptr() + x + in_info.strides_in_bytes()[axis] * dim); + switch (op) { - //de-quantize input - if(std::is_same::value) + case ReductionOperation::SUM: { - res_value *= dequantize_qasymm8(*in_ptr, iq_info); + res_value += *in_ptr; + break; } - else + case ReductionOperation::MEAN_SUM: { - res_value *= dequantize_qasymm8_signed(*in_ptr, iq_info); + res_value_q += *in_ptr; + break; } - break; - } - case ReductionOperation::ARG_IDX_MIN: - { - if(*in_ptr < res_value) + case ReductionOperation::SUM_SQUARE: { - res_value = *in_ptr; - res_idx = dim; + res_value += *in_ptr * *in_ptr; + break; } - break; - } - case ReductionOperation::ARG_IDX_MAX: - { - if(*in_ptr > res_value) + case ReductionOperation::PROD: { - res_value = *in_ptr; - res_idx = dim; + //de-quantize input + if (std::is_same::value) + { + res_value *= dequantize_qasymm8(*in_ptr, iq_info); + } + else + { + res_value *= dequantize_qasymm8_signed(*in_ptr, iq_info); + } + break; } - break; + case ReductionOperation::ARG_IDX_MIN: + { + if (*in_ptr < res_value) + { + res_value = *in_ptr; + res_idx = dim; + } + break; + } + case ReductionOperation::ARG_IDX_MAX: + { + if (*in_ptr > res_value) + { + res_value = *in_ptr; + res_idx = dim; + } + break; + } + case ReductionOperation::MIN: + { + res_value = *in_ptr < res_value ? *in_ptr : res_value; + break; + } + case ReductionOperation::MAX: + { + res_value = *in_ptr > res_value ? *in_ptr : res_value; + break; + } + default: + ARM_COMPUTE_ERROR("Not supported"); } - case ReductionOperation::MIN: + } + + switch (op) + { + case ReductionOperation::MEAN_SUM: { - res_value = *in_ptr < res_value ? *in_ptr : res_value; + // Apply previously calculated coefficients (with rounding on aarch64) +#ifdef __aarch64__ + const int32_t res = + arm_compute::support::cpp11::round(A * (static_cast(res_value_q)) + B); +#else // defined(__aarch64__) + const int32_t res = A * (static_cast(res_value_q)) + B; +#endif // __aarch64__ + *reinterpret_cast(output.ptr() + x) = utils::cast::saturate_cast(res); break; } - case ReductionOperation::MAX: + case ReductionOperation::SUM: { - res_value = *in_ptr > res_value ? *in_ptr : res_value; + // Subtract accumulated offsets + res_value -= (in_info.dimension(axis) - 1) * iq_info.offset; + *reinterpret_cast(output.ptr() + x) = utils::cast::saturate_cast(res_value); break; } - default: - ARM_COMPUTE_ERROR("Not supported"); - } - } - - switch(op) - { - case ReductionOperation::MEAN_SUM: - { - // Apply previously calculated coefficients (with rounding on aarch64) -#ifdef __aarch64__ - const int32_t res = arm_compute::support::cpp11::round(A * (static_cast(res_value_q)) + B); -#else // defined(__aarch64__) - const int32_t res = A * (static_cast(res_value_q)) + B; -#endif // __aarch64__ - *reinterpret_cast(output.ptr() + x) = utils::cast::saturate_cast(res); - break; - } - case ReductionOperation::SUM: - { - // Subtract accumulated offsets - res_value -= (in_info.dimension(axis) - 1) * iq_info.offset; - *reinterpret_cast(output.ptr() + x) = utils::cast::saturate_cast(res_value); - break; - } - case ReductionOperation::PROD: - { - //re-quantize result - T res = 0; - if(std::is_same::value) + case ReductionOperation::PROD: { - res = quantize_qasymm8(res_value, iq_info); + //re-quantize result + T res = 0; + if (std::is_same::value) + { + res = quantize_qasymm8(res_value, iq_info); + } + else + { + res = quantize_qasymm8_signed(res_value, iq_info); + } + *(reinterpret_cast(output.ptr() + x)) = res; + break; } - else + case ReductionOperation::ARG_IDX_MIN: + case ReductionOperation::ARG_IDX_MAX: { - res = quantize_qasymm8_signed(res_value, iq_info); + *(reinterpret_cast(output.ptr() + x * 4)) = res_idx; + break; } - *(reinterpret_cast(output.ptr() + x)) = res; - break; - } - case ReductionOperation::ARG_IDX_MIN: - case ReductionOperation::ARG_IDX_MAX: - { - *(reinterpret_cast(output.ptr() + x * 4)) = res_idx; - break; + default: + *(reinterpret_cast(output.ptr() + x)) = res_value; } - default: - *(reinterpret_cast(output.ptr() + x)) = res_value; } - } - }, - input, output); + }, + input, output); } }; -void reduce_op(const Window &window, const ITensor *input, ITensor *output, unsigned int axis, const ReductionOperation op) +void reduce_op( + const Window &window, const ITensor *input, ITensor *output, unsigned int axis, const ReductionOperation op) { const bool is_complex = (input->info()->num_channels() == 2); - if(is_complex) + if (is_complex) { - switch(axis) + switch (axis) { case 2: - switch(input->info()->data_type()) + switch (input->info()->data_type()) { case DataType::F32: - switch(op) + switch (op) { case ReductionOperation::SUM: - return Reducer>::reduceZ(window, input, output, RedOpYZW_complex(), op); + return Reducer>::reduceZ( + window, input, output, RedOpYZW_complex(), + op); default: ARM_COMPUTE_ERROR("Not supported"); } @@ -1602,19 +1669,21 @@ void reduce_op(const Window &window, const ITensor *input, ITensor *output, unsi return; } - switch(axis) + switch (axis) { case 0: { - switch(input->info()->data_type()) + switch (input->info()->data_type()) { case DataType::QASYMM8: { - return Reducer>::reduceX(window, input, output, RedOpX_quantized(), op); + return Reducer>::reduceX(window, input, output, + RedOpX_quantized(), op); } case DataType::QASYMM8_SIGNED: { - return Reducer>::reduceX(window, input, output, RedOpX_quantized(), op); + return Reducer>::reduceX(window, input, output, RedOpX_quantized(), + op); } #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: @@ -1635,19 +1704,22 @@ void reduce_op(const Window &window, const ITensor *input, ITensor *output, unsi } } case 1: - switch(input->info()->data_type()) + switch (input->info()->data_type()) { case DataType::QASYMM8: { - return Reducer>::reduceY(window, input, output, RedOpYZW_quantized(), op); + return Reducer>::reduceY(window, input, output, + RedOpYZW_quantized(), op); } case DataType::QASYMM8_SIGNED: { - return Reducer>::reduceY(window, input, output, RedOpYZW_quantized(), op); + return Reducer>::reduceY(window, input, output, + RedOpYZW_quantized(), op); } #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: - return Reducer>::reduceY(window, input, output, RedOpYZW(), op); + return Reducer>::reduceY(window, input, output, RedOpYZW(), + op); #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F32: return Reducer>::reduceY(window, input, output, RedOpYZW(), op); @@ -1657,15 +1729,18 @@ void reduce_op(const Window &window, const ITensor *input, ITensor *output, unsi ARM_COMPUTE_ERROR("Not supported"); } case 2: - switch(input->info()->data_type()) + switch (input->info()->data_type()) { case DataType::QASYMM8: - return Reducer>::reduceZ(window, input, output, RedOpYZW_quantized(), op); + return Reducer>::reduceZ(window, input, output, + RedOpYZW_quantized(), op); case DataType::QASYMM8_SIGNED: - return Reducer>::reduceZ(window, input, output, RedOpYZW_quantized(), op); + return Reducer>::reduceZ(window, input, output, + RedOpYZW_quantized(), op); #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: - return Reducer>::reduceZ(window, input, output, RedOpYZW(), op); + return Reducer>::reduceZ(window, input, output, RedOpYZW(), + op); #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F32: return Reducer>::reduceZ(window, input, output, RedOpYZW(), op); @@ -1675,15 +1750,18 @@ void reduce_op(const Window &window, const ITensor *input, ITensor *output, unsi ARM_COMPUTE_ERROR("Not supported"); } case 3: - switch(input->info()->data_type()) + switch (input->info()->data_type()) { case DataType::QASYMM8: - return Reducer>::reduceW(window, input, output, RedOpYZW_quantized(), op); + return Reducer>::reduceW(window, input, output, + RedOpYZW_quantized(), op); case DataType::QASYMM8_SIGNED: - return Reducer>::reduceW(window, input, output, RedOpYZW_quantized(), op); + return Reducer>::reduceW(window, input, output, + RedOpYZW_quantized(), op); #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: - return Reducer>::reduceW(window, input, output, RedOpYZW(), op); + return Reducer>::reduceW(window, input, output, RedOpYZW(), + op); #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F32: return Reducer>::reduceW(window, input, output, RedOpYZW(), op); @@ -1704,9 +1782,10 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, u ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); - if(input->num_channels() == 1) + if (input->num_channels() == 1) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8_SIGNED, DataType::QASYMM8, DataType::S32, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8_SIGNED, DataType::QASYMM8, + DataType::S32, DataType::F16, DataType::F32); } else { @@ -1715,13 +1794,14 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, u ARM_COMPUTE_RETURN_ERROR_ON(axis != 2); } - ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Reduction axis greater than max number of dimensions"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, + "Reduction axis greater than max number of dimensions"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis"); - if(output->total_size() != 0) + if (output->total_size() != 0) { bool is_arg_min_max = (op == ReductionOperation::ARG_IDX_MAX || op == ReductionOperation::ARG_IDX_MIN); - if(!is_arg_min_max) + if (!is_arg_min_max) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON(input->num_channels() != output->num_channels()); @@ -1731,8 +1811,9 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, u ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U32, DataType::S32); } - const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_reduced_shape(input->tensor_shape(), axis); - const TensorInfo tensor_info_reshaped = input->clone()->set_tensor_shape(output_shape); + const TensorShape output_shape = + arm_compute::misc::shape_calculator::compute_reduced_shape(input->tensor_shape(), axis); + const TensorInfo tensor_info_reshaped = input->clone()->set_tensor_shape(output_shape); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_reshaped); } @@ -1745,7 +1826,10 @@ NEReductionOperationKernel::NEReductionOperationKernel() { } -void NEReductionOperationKernel::configure(const ITensor *input, ITensor *output, unsigned int axis, ReductionOperation op) +void NEReductionOperationKernel::configure(const ITensor *input, + ITensor *output, + unsigned int axis, + ReductionOperation op) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); @@ -1761,14 +1845,23 @@ void NEReductionOperationKernel::configure(const ITensor *input, ITensor *output INEKernel::configure(win); // Calculate output shape and set if empty - const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_reduced_shape(input->info()->tensor_shape(), axis); + const TensorShape output_shape = + arm_compute::misc::shape_calculator::compute_reduced_shape(input->info()->tensor_shape(), axis); // Output auto initialization if not yet initialized const bool is_arg_min_max = (op == ReductionOperation::ARG_IDX_MIN || op == ReductionOperation::ARG_IDX_MAX); DataType output_data_type = is_arg_min_max ? DataType::S32 : input->info()->data_type(); - auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape).set_data_type(output_data_type).reset_padding().set_is_resizable(true)); + auto_init_if_empty(*output->info(), input->info() + ->clone() + ->set_tensor_shape(output_shape) + .set_data_type(output_data_type) + .reset_padding() + .set_is_resizable(true)); } -Status NEReductionOperationKernel::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, ReductionOperation op) +Status NEReductionOperationKernel::validate(const ITensorInfo *input, + const ITensorInfo *output, + unsigned int axis, + ReductionOperation op) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, axis, op)); -- cgit v1.2.1