diff options
Diffstat (limited to 'src/core/cpu/kernels/CpuPoolingKernel.cpp')
| -rw-r--r-- | src/core/cpu/kernels/CpuPoolingKernel.cpp | 2351 |
1 files changed, 145 insertions, 2206 deletions
diff --git a/src/core/cpu/kernels/CpuPoolingKernel.cpp b/src/core/cpu/kernels/CpuPoolingKernel.cpp index a29aef4986..21afad2c3f 100644 --- a/src/core/cpu/kernels/CpuPoolingKernel.cpp +++ b/src/core/cpu/kernels/CpuPoolingKernel.cpp @@ -33,6 +33,8 @@ #include "src/core/NEON/NEAsymm.h" #include "src/core/NEON/NEFixedPoint.h" #include "src/core/NEON/NEMath.h" +#include "src/core/common/Registrars.h" +#include "src/core/cpu/kernels/pooling/neon/list.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" #include "support/ToolchainSupport.h" @@ -46,91 +48,138 @@ namespace cpu { namespace kernels { -using namespace misc::shape_calculator; - namespace { -template <typename T> -inline typename std::enable_if<std::is_same<T, int8_t>::value, int8_t>::type -quantize(float val, const UniformQuantizationInfo &info) -{ - return quantize_qasymm8_signed(val, info); -} +using namespace misc::shape_calculator; -template <typename T> -inline typename std::enable_if<std::is_same<T, uint8_t>::value, uint8_t>::type -quantize(float val, const UniformQuantizationInfo &info) +struct PoolingSelectorData { - return quantize_qasymm8(val, info); -} + DataType dt; + DataLayout dl; + int pool_stride_x; + Size2D pool_size; +}; -inline float calculate_avg_scale(bool exclude_padding, DataLayout data_layout, const Coordinates &id, const int pool_size_x, const int pool_size_y, const int upper_bound_w, const int upper_bound_h, - const int pad_x, const int pad_y, const int stride_x, const int stride_y) +using PoolingSelectorPtr = std::add_pointer<bool(const PoolingSelectorData &data)>::type; +using PoolingKernelPtr = std::add_pointer<void(const ITensor *, ITensor *, ITensor *, PoolingLayerInfo &, const Window &, const Window &)>::type; +struct PoolingKernel { - const unsigned int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const unsigned int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + const char *name; + const PoolingSelectorPtr is_selected; + PoolingKernelPtr ukernel; +}; - int start_x = id[idx_width] * stride_x - pad_x; - int start_y = id[idx_height] * stride_y - pad_y; - - const int end_x = std::min(start_x + pool_size_x, upper_bound_w); - const int end_y = std::min(start_y + pool_size_y, upper_bound_h); - if(exclude_padding) - { - start_x = std::max(0, start_x); - start_y = std::max(0, start_y); - } - return 1.f / ((end_y - start_y) * (end_x - start_x)); -} - -template <typename T, typename TVec> -inline void scale_vector_q16x8(bool exclude_padding, TVec &v, const Coordinates &id, int id_offset, int step, - const int pool_size, const int upper_bound_w, const int upper_bound_h, - const int pad_x, const int pad_y, const int stride_x, const int stride_y) +static const PoolingKernel available_kernels[] = { - int start_x = (id.x() + id_offset) * stride_x - pad_x; - int start_y = id.y() * stride_y - pad_y; - const int end_y = std::min(start_y + pool_size, upper_bound_h); - if(exclude_padding) { - start_y = std::max(0, start_y); - } - - std::array<T, 8> elems = + "poolingMxN_qasymm8_neon_nhwc", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NHWC) && (data.dt == DataType::QASYMM8)); }, + REGISTER_QASYMM8_NEON(arm_compute::cpu::poolingMxN_qasymm8_neon_nhwc) + }, { - { - wrapper::vgetlane(v, 0), - wrapper::vgetlane(v, 1), - wrapper::vgetlane(v, 2), - wrapper::vgetlane(v, 3), - wrapper::vgetlane(v, 4), - wrapper::vgetlane(v, 5), - wrapper::vgetlane(v, 6), - wrapper::vgetlane(v, 7), - } - }; + "poolingMxN_qasymm8_signed_neon_nhwc", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NHWC) && (data.dt == DataType::QASYMM8_SIGNED)); }, + REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::poolingMxN_qasymm8_signed_neon_nhwc) + }, +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) + { + "poolingMxN_fp16_neon_nhwc", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NHWC) && (data.dt == DataType::F16)); }, + REGISTER_FP16_NEON(arm_compute::cpu::poolingMxN_fp16_neon_nhwc) + }, +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */ + { + "poolingMxN_fp32_neon_nhwc", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NHWC) && (data.dt == DataType::F32)); }, + REGISTER_FP32_NEON(arm_compute::cpu::poolingMxN_fp32_neon_nhwc) + }, +#if defined(ENABLE_NCHW_KERNELS) + { + "pooling2_qasymm8_neon_nchw", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::QASYMM8) && (data.pool_size.x() == data.pool_size.y()) && (data.pool_size.x() == 2) && (data.pool_stride_x < 3)); }, + REGISTER_QASYMM8_NEON(arm_compute::cpu::pooling2_quantized_neon_nchw<uint8_t>) + }, + { + "pooling3_qasymm8_neon_nchw", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::QASYMM8) && (data.pool_size.x() == data.pool_size.y()) && (data.pool_size.x() == 3) && (data.pool_stride_x < 3)); }, + REGISTER_QASYMM8_NEON(arm_compute::cpu::pooling3_quantized_neon_nchw<uint8_t>) + }, + { + "poolingMxN_qasymm8_neon_nchw", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::QASYMM8)); }, + REGISTER_QASYMM8_NEON(arm_compute::cpu::poolingMxN_quantized_neon_nchw<uint8_t>) + }, + { + "pooling2_qasymm8_signed_neon_nchw", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::QASYMM8_SIGNED) && (data.pool_size.x() == data.pool_size.y()) && (data.pool_size.x() == 2) && (data.pool_stride_x < 3)); }, + REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::pooling2_quantized_neon_nchw<int8_t>) + }, + { + "pooling3_qasymm8_signed_neon_nchw", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::QASYMM8_SIGNED) && (data.pool_size.x() == data.pool_size.y()) && (data.pool_size.x() == 3) && (data.pool_stride_x < 3)); }, + REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::pooling3_quantized_neon_nchw<int8_t>) + }, + { + "poolingMxN_qasymm8_signed_neon_nchw", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::QASYMM8_SIGNED)); }, + REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::poolingMxN_quantized_neon_nchw<int8_t>) + }, +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) + { + "pooling2_fp16_neon_nchw", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::F16) && (data.pool_size.x() == data.pool_size.y()) && (data.pool_size.x() == 2)); }, + REGISTER_FP16_NEON(arm_compute::cpu::pooling2_fp16_neon_nchw) + }, + { + "pooling3_fp16_neon_nchw", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::F16) && (data.pool_size.x() == data.pool_size.y()) && (data.pool_size.x() == 3)); }, + REGISTER_FP16_NEON(arm_compute::cpu::pooling3_fp16_neon_nchw) + }, + { + "poolingMxN_fp16_neon_nchw", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::F16)); }, + REGISTER_FP16_NEON(arm_compute::cpu::poolingMxN_fp16_neon_nchw) + }, +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */ + { + "pooling2_fp32_neon_nchw", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::F32) && (data.pool_size.x() == data.pool_size.y()) && (data.pool_size.x() == 2)); }, + REGISTER_FP32_NEON(arm_compute::cpu::pooling2_fp32_neon_nchw) + }, + { + "pooling3_fp32_neon_nchw", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::F32) && (data.pool_size.x() == data.pool_size.y()) && (data.pool_size.x() == 3)); }, + REGISTER_FP32_NEON(arm_compute::cpu::pooling3_fp32_neon_nchw) + }, + { + "pooling7_fp32_neon_nchw", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::F32) && (data.pool_size.x() == data.pool_size.y()) && (data.pool_size.x() == 7)); }, + REGISTER_FP32_NEON(arm_compute::cpu::pooling7_fp32_neon_nchw) + }, + { + "poolingMxN_fp32_neon_nchw", + [](const PoolingSelectorData & data) { return ((data.dl == DataLayout::NCHW) && (data.dt == DataType::F32)); }, + REGISTER_FP32_NEON(arm_compute::cpu::poolingMxN_fp32_neon_nchw) + }, +#endif /* defined(ENABLE_NCHW_KERNELS) */ +}; - for(auto &el : elems) +/** Micro-kernel selector + * + * @param[in] data Selection data passed to help pick the appropriate micro-kernel + * + * @return A matching micro-kernel else nullptr + */ +const PoolingKernel *get_implementation(DataType dt, DataLayout dl, int pool_stride_x, Size2D pool_size) +{ + for(const auto &uk : available_kernels) { - int c_start_x = start_x; - const int end_x = std::min(c_start_x + pool_size, upper_bound_w); - if(exclude_padding) + if(uk.is_selected({ dt, dl, pool_stride_x, pool_size })) { - c_start_x = std::max(0, c_start_x); + return &uk; } - float scale = 1.f / ((end_y - start_y) * (end_x - c_start_x)); - el *= scale; - start_x += step * stride_x; } - - v = wrapper::vsetlane(elems[0], v, 0); - v = wrapper::vsetlane(elems[1], v, 1); - v = wrapper::vsetlane(elems[2], v, 2); - v = wrapper::vsetlane(elems[3], v, 3); - v = wrapper::vsetlane(elems[4], v, 4); - v = wrapper::vsetlane(elems[5], v, 5); - v = wrapper::vsetlane(elems[6], v, 6); - v = wrapper::vsetlane(elems[7], v, 7); + return nullptr; } Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &pool_info, @@ -172,6 +221,9 @@ Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const } } + const auto *uk = get_implementation(src->data_type(), src->data_layout(), pool_stride_x, pool_size); + ARM_COMPUTE_RETURN_ERROR_ON(uk == nullptr || uk->ukernel == nullptr); + return Status{}; } @@ -321,116 +373,6 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src, ITenso Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; return std::make_pair(err, win); } - -template <typename T> -inline T vcvtq_q32_f32(float32x4_t values); - -template <> -inline uint32x4_t vcvtq_q32_f32(float32x4_t values) -{ - return vcvtq_u32_f32(values); -} - -template <> -inline int32x4_t vcvtq_q32_f32(float32x4_t values) -{ - return vcvtq_s32_f32(values); -} - -template <typename T> -inline float32x4_t vcvtq_f32_q32(T values); - -template <> -inline float32x4_t vcvtq_f32_q32(uint32x4_t values) -{ - return vcvtq_f32_u32(values); -} - -template <> -inline float32x4_t vcvtq_f32_q32(int32x4_t values) -{ - return vcvtq_f32_s32(values); -} - -template <typename Tout> -inline Tout vrequantize_pooling_with_scale(const float32x4x4_t &acc, const float quant_rescale, const float scale_pooling, const int32_t new_offset); - -template <> -inline uint8x16_t vrequantize_pooling_with_scale(const float32x4x4_t &acc, const float quant_rescale, const float scale_pooling, const int32_t new_offset) -{ - const float new_scale = quant_rescale / scale_pooling; - return vquantize(acc, UniformQuantizationInfo(new_scale, new_offset)); -} - -template <> -inline int8x16_t vrequantize_pooling_with_scale(const float32x4x4_t &acc, const float quant_rescale, const float scale_pooling, const int32_t new_offset) -{ - const float new_scale = quant_rescale / scale_pooling; - return vquantize_signed(acc, UniformQuantizationInfo(new_scale, new_offset)); -} - -template <typename Tin, typename Tout> -inline Tout vrequantize_pooling(Tin vec1, Tin vec2, const UniformQuantizationInfo &requant_qinfo); - -template <> -inline uint8x16_t vrequantize_pooling(uint8x8_t vec1, uint8x8_t vec2, const UniformQuantizationInfo &requant_qinfo) -{ - const float32x4x4_t acc = - { - { - vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8((vec1))))), - vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8((vec1))))), - vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8((vec2))))), - vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8((vec2))))), - } - }; - return vquantize(acc, requant_qinfo); -} - -template <> -inline int8x16_t vrequantize_pooling(int8x8_t vec1, int8x8_t vec2, const UniformQuantizationInfo &requant_qinfo) -{ - const float32x4x4_t acc = - { - { - vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8((vec1))))), - vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8((vec1))))), - vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8((vec2))))), - vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8((vec2))))), - } - }; - return vquantize_signed(acc, requant_qinfo); -} - -template <typename T> -inline T vrequantize_pooling(T &vec, const UniformQuantizationInfo &requant_qinfo); - -template <> -inline uint8x8_t vrequantize_pooling(uint8x8_t &vec, const UniformQuantizationInfo &requant_qinfo) -{ - const float32x4x2_t acc = - { - { - vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8((vec))))), - vcvtq_f32_u32(vmovl_u16(vget_high_u16(vmovl_u8((vec))))), - } - }; - return vquantize(acc, requant_qinfo); -} - -template <> -inline int8x8_t vrequantize_pooling(int8x8_t &vec, const UniformQuantizationInfo &requant_qinfo) -{ - const float32x4x2_t acc = - { - { - vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8((vec))))), - vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8((vec))))), - } - }; - return vquantize_signed(acc, requant_qinfo); -} - } // namespace BorderSize CpuPoolingKernel::border_size() const @@ -443,7 +385,6 @@ void CpuPoolingKernel::configure(ITensorInfo *src, ITensorInfo *dst, const Pooli ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); const PadStrideInfo pad_stride_info = pool_info.pad_stride_info; const bool is_global_pooling = pool_info.is_global_pooling; - const int pool_stride_x = pad_stride_info.stride().first; // Get data layout const auto data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : pool_info.data_layout; @@ -471,135 +412,12 @@ void CpuPoolingKernel::configure(ITensorInfo *src, ITensorInfo *dst, const Pooli ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, pool_info, pooled_w, pooled_h, indices, pool_size)); // Set instance variables - _pool_info = pool_info; - _data_layout = src->data_layout(); - _is_square = (pool_size.x() == pool_size.y()); - - // Get data type - const DataType data_type = src->data_type(); - const bool is_nchw = _data_layout == DataLayout::NCHW; - - if(data_type == DataType::QASYMM8) - { - if(!is_nchw) - { - _func = &CpuPoolingKernel::poolingMxN_q8_nhwc<uint8_t>; - } - else - { - if(pool_size.x() == 2 && pool_stride_x < 3 && _is_square) - { - _func = &CpuPoolingKernel::pooling2_q8_nchw<uint8_t>; - } - else if(pool_size.x() == 3 && pool_stride_x < 3 && _is_square) - { - _func = &CpuPoolingKernel::pooling3_q8_nchw<uint8_t>; - } - else - { - _func = &CpuPoolingKernel::poolingMxN_q8_nchw<uint8_t>; - } - } - } - else if(data_type == DataType::QASYMM8_SIGNED) - { - if(!is_nchw) - { - _func = &CpuPoolingKernel::poolingMxN_q8_nhwc<int8_t>; - } - else - { - if(pool_size.x() == 2 && pool_stride_x < 3 && _is_square) - { - _func = &CpuPoolingKernel::pooling2_q8_nchw<int8_t>; - } - else if(pool_size.x() == 3 && pool_stride_x < 3 && _is_square) - { - _func = &CpuPoolingKernel::pooling3_q8_nchw<int8_t>; - } - else - { - _func = &CpuPoolingKernel::poolingMxN_q8_nchw<int8_t>; - } - } - } - else if(data_type == DataType::F16) - { - if(!is_nchw) - { - _func = &CpuPoolingKernel::poolingMxN_f16_nhwc; - } - else - { - if(_is_square) - { - switch(pool_size.x()) - { - case 2: - { - _func = &CpuPoolingKernel::pooling2_f16_nchw; - } - break; - case 3: - { - _func = &CpuPoolingKernel::pooling3_f16_nchw; - } - break; - default: - { - _func = &CpuPoolingKernel::poolingMxN_f16_nchw; - break; - } - } - } - else - { - _func = &CpuPoolingKernel::poolingMxN_f16_nchw; - } - } - } - else if(data_type == DataType::F32) - { - if(!is_nchw) - { - _func = &CpuPoolingKernel::poolingMxN_f32_nhwc; - } - else - { - if(_is_square) - { - switch(pool_size.x()) - { - case 2: - { - _func = &CpuPoolingKernel::pooling2_f32_nchw; - break; - } - case 3: - { - _func = &CpuPoolingKernel::pooling3_f32_nchw; - break; - } - case 7: - { - _func = &CpuPoolingKernel::pooling7_f32_nchw; - break; - } - default: - { - _func = &CpuPoolingKernel::poolingMxN_f32_nchw; - break; - } - } - } - else - { - _func = &CpuPoolingKernel::poolingMxN_f32_nchw; - } - } - } + _pool_info = pool_info; + _data_layout = src->data_layout(); + _pool_size = pool_size; + _pool_stride_x = pad_stride_info.stride().first; - if(!is_nchw) + if(_data_layout == DataLayout::NHWC) { // Configure kernel window Window win = calculate_max_window(*dst, Steps()); @@ -618,1890 +436,6 @@ void CpuPoolingKernel::configure(ITensorInfo *src, ITensorInfo *dst, const Pooli } } -template <typename T> -inline uint32_t offset_no_padding(uint32_t padded_offset, const Coordinates &id, const ITensorInfo &info, int pool_stride_x, int pool_stride_y) -{ - const int pad_left = info.padding().left; - const int pad_right = info.padding().right; - const int pad_top = info.padding().top; - const int pad_bottom = info.padding().bottom; - const int in_stride_y = static_cast<int>(info.strides_in_bytes().y()); - const int in_stride_w = static_cast<int>(info.strides_in_bytes()[3]); - const int pad_horiz = pad_left + pad_right; - const int pad_vert = pad_top + pad_bottom; - - if(info.data_layout() == DataLayout::NCHW) - { - const uint32_t offset_base = padded_offset - - sizeof(T) * pad_horiz * id.y() * pool_stride_y /* subtract padding elems per row */ - - pad_top * sizeof(T) /* top padding */ - - sizeof(T) * pad_horiz * info.tensor_shape()[1] * id.z() - pad_vert * in_stride_y * id.z() /* for each Z plane there are height*pad_right padding elems */ - - in_stride_w * id[3]; - - return offset_base; - } - else - { - const uint32_t offset_base = padded_offset - - sizeof(T) * pad_horiz * id.y() * pool_stride_x // subtract padding elems per row - - pad_top * sizeof(T) // top padding - - sizeof(T) * pad_horiz * info.tensor_shape()[1] * id.z() * pool_stride_y // for each Z plane there are width*pad_right padding elems - - in_stride_w * id[3]; - - return offset_base; - } -} - -template <typename T> -void CpuPoolingKernel::pooling2_q8_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - Iterator src(_src, window_src); - Iterator dst(_dst, window); - - /** NEON vector types */ - using q8x8_t = typename wrapper::traits::neon_vector<T, 8>::type; - using q8x16_t = typename wrapper::traits::neon_vector<T, 16>::type; - using q8x8x2_t = typename std::conditional<std::is_same<T, uint8_t>::value, uint8x8x2_t, int8x8x2_t>::type; - using q16_t = typename wrapper::traits::promote_t<T>; - using q16x4_t = typename wrapper::traits::neon_vector<q16_t, 4>::type; - using q16x8_t = typename wrapper::traits::neon_vector<q16_t, 8>::type; - using q16x8x2_t = typename wrapper::traits::neon_vector<q16_t, 16>::type; - - constexpr int pool_size = 2; - int pool_stride_x = 0; - int pool_stride_y = 0; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - const T *const src_top_ptr = reinterpret_cast<const T *>(_src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top)))); - const T *const src_bottom_ptr = reinterpret_cast<const T *>(_src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1))); - - const int scale_step_x = (pool_stride_x == 1) ? 2 : 1; - - const UniformQuantizationInfo src_qinfo = _src->info()->quantization_info().uniform(); - const UniformQuantizationInfo dst_qinfo = _dst->info()->quantization_info().uniform(); - const bool have_different_qinfo = src_qinfo != dst_qinfo; - - const float requant_scale = dst_qinfo.scale / src_qinfo.scale; - const int32_t requant_offset = dst_qinfo.offset - static_cast<int32_t>(static_cast<float>(src_qinfo.offset) / requant_scale); - const UniformQuantizationInfo requant_qinfo = UniformQuantizationInfo(requant_scale, requant_offset); - - execute_window_loop(window, [&](const Coordinates & id) - { - const auto top_data = wrapper::vloadq(src_top_ptr + src.offset()); - const auto bottom_data = wrapper::vloadq(src_bottom_ptr + src.offset()); - q8x8_t lower_res = {}; - q8x8_t upper_res = {}; - - if(pooling_type != PoolingType::MAX) - { - const q16x8x2_t top_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(top_data)), wrapper::vmovl(wrapper::vgethigh(top_data)) } }; - const q16x8x2_t bottom_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(bottom_data)), wrapper::vmovl(wrapper::vgethigh(bottom_data)) } }; - - // Add rows - const q16x8x2_t vrsum = - { - { - wrapper::vadd(top_data_q16.val[0], bottom_data_q16.val[0]), - wrapper::vadd(top_data_q16.val[1], bottom_data_q16.val[1]), - } - }; - - // Pair-wise add row data - const q16x4_t vpsum_1 = wrapper::vpadd(wrapper::vgetlow(vrsum.val[0]), wrapper::vgethigh(vrsum.val[0])); - const q16x4_t vpsum_2 = wrapper::vpadd(wrapper::vgetlow(vrsum.val[1]), wrapper::vgethigh(vrsum.val[1])); - - q16x8_t res_lower = wrapper::vcombine(vpsum_1, vpsum_2); - - // Scale lower result - scale_vector_q16x8<q16_t, q16x8_t>(exclude_padding, res_lower, id, 0, scale_step_x, - pool_size, upper_bound_w, upper_bound_h, - pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - lower_res = wrapper::vmovn(res_lower); - - // Compute upper result for stride_x == 1 - if(pool_stride_x == 1) - { - // Shifted row sum - const q16x8x2_t vrsum_shifted = - { - { - wrapper::vext_1(vrsum.val[0], vrsum.val[1]), - wrapper::vext_1(vrsum.val[1], vrsum.val[1]) - } - }; - - // Pair-wise add shifted row - q16x8_t res_upper = wrapper::vcombine( - wrapper::vpadd(wrapper::vgetlow(vrsum_shifted.val[0]), wrapper::vgethigh(vrsum_shifted.val[0])), - wrapper::vpadd(wrapper::vgetlow(vrsum_shifted.val[1]), wrapper::vgethigh(vrsum_shifted.val[1]))); - - // Scale upper result - scale_vector_q16x8<q16_t, q16x8_t>(exclude_padding, res_upper, id, 1, 2, - pool_size, upper_bound_w, upper_bound_h, - pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - upper_res = wrapper::vmovn(res_upper); - } - } - else - { - const q8x16_t max_data = wrapper::vmax(top_data, bottom_data); - lower_res = wrapper::vpmax(wrapper::vgetlow(max_data), wrapper::vgethigh(max_data)); - if(pool_stride_x == 1) - { - const q8x16_t max_data_shifted = wrapper::vext_1(max_data, max_data); - upper_res = wrapper::vpmax(wrapper::vgetlow(max_data_shifted), wrapper::vgethigh(max_data_shifted)); - } - } - - if(have_different_qinfo) - { - const auto requantized_dst = vrequantize_pooling<q8x8_t, q8x16_t>(lower_res, upper_res, requant_qinfo); - lower_res = wrapper::vgetlow(requantized_dst); - upper_res = wrapper::vgethigh(requantized_dst); - } - - // Store result - if(pool_stride_x == 1) - { - const q8x8x2_t res = { { lower_res, upper_res } }; - wrapper::vstore(reinterpret_cast<T *>(dst.ptr()), res); - } - else - { - wrapper::vstore(reinterpret_cast<T *>(dst.ptr()), lower_res); - } - }, - src, dst); -} - -void CpuPoolingKernel::pooling3_f16_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - ARM_COMPUTE_UNUSED(pooling_type); - ARM_COMPUTE_UNUSED(exclude_padding); -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - Iterator src(_src, window_src); - Iterator dst(_dst, window); - - constexpr const int pool_size = 3; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - const unsigned char *const src_top_ptr = _src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); - const unsigned char *const src_middle_ptr = _src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); - const unsigned char *const src_bottom_ptr = _src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 2)); - - execute_window_loop(window, [&](const Coordinates & id) - { - float16x4_t top_data = vld1_f16(reinterpret_cast<const float16_t *>(src_top_ptr + src.offset())); - float16x4_t middle_data = vld1_f16(reinterpret_cast<const float16_t *>(src_middle_ptr + src.offset())); - float16x4_t bottom_data = vld1_f16(reinterpret_cast<const float16_t *>(src_bottom_ptr + src.offset())); - float16x4_t res = {}; - - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - top_data = vmul_f16(top_data, top_data); - middle_data = vmul_f16(middle_data, middle_data); - bottom_data = vmul_f16(bottom_data, bottom_data); - } - - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - const float16x4_t scale_v = vdup_n_f16(scale); - // Perform pooling - const float16x4_t sum_data = vadd_f16(vadd_f16(top_data, bottom_data), middle_data); - res = vpadd_f16(vset_lane_f16(0.f, sum_data, 3), sum_data); - res = vmul_f16(vpadd_f16(res, res), scale_v); - } - else - { - const float16x4_t max_data = vmax_f16(vmax_f16(top_data, bottom_data), middle_data); - res = vpmax_f16(vset_lane_f16(-std::numeric_limits<float>::max(), max_data, 3), max_data); - res = vpmax_f16(res, res); - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - res = vinv_f16(vinvsqrt_f16(res)); - } - - *(reinterpret_cast<float16_t *>(dst.ptr())) = vget_lane_f16(res, 0); - }, - src, dst); -#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - ARM_COMPUTE_UNUSED(window_src); - ARM_COMPUTE_UNUSED(window); - ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -} - -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -template <typename T> -inline typename std::enable_if<std::is_same<T, float16_t>::value, float32x2_t>::type -f16_to_f32(float16x4_t src) -{ - float32x2_t dst = { static_cast<float>(vget_lane_f16(src, 0)), static_cast<float>(vget_lane_f16(src, 1)) }; - return dst; -} -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - -template <typename T> -inline typename std::enable_if<std::is_same<T, float>::value, float32x2_t>::type -f16_to_f32(float32x2_t src) -{ - return src; -} - -template <typename T> -void CpuPoolingKernel::pooling2_nchw_maxpool_indices(const Window &window_src, const Window &window) -{ - Iterator src(_src, window_src); - Iterator dst(_dst, window); - Iterator indices(_indices, window); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const uint8_t *const src_top_ptr = _src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); - const uint8_t *const src_bottom_ptr = _src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); - const int pad_left = _src->info()->padding().left; - const int pad_right = _src->info()->padding().right; - const int in_stride_y = static_cast<int>(_src->info()->strides_in_bytes().y()); - - execute_window_loop(window, [&](const Coordinates & id) - { - auto top_data = wrapper::vload(reinterpret_cast<const T *>(src_top_ptr + src.offset())); - auto bottom_data = wrapper::vload(reinterpret_cast<const T *>(src_bottom_ptr + src.offset())); - float32x2_t top_data_f32 = f16_to_f32<T>(top_data); - float32x2_t bottom_data_f32 = f16_to_f32<T>(bottom_data); - - // Calculate max data, compare top first, then bottom, to make sue the first max is recorded. - const float32x2_t max_data_top = vpmax_f32(top_data_f32, top_data_f32); - const float32x2_t max_data_bottom = vpmax_f32(bottom_data_f32, bottom_data_f32); - const float32x2_t max_data = vmax_f32(max_data_top, max_data_bottom); - *(reinterpret_cast<T *>(dst.ptr())) = static_cast<T>(vget_lane_f32(max_data, 0)); - - // Calculate max data indice, which will be used in max unpool. - const uint32_t offset_base = offset_no_padding<T>(src.offset(), id, *_src->info(), pool_stride_x, pool_stride_y); - const uint32_t offset_top = (uint32_t)(offset_base / sizeof(T)); - const uint32_t offset_bottom = offset_top + in_stride_y / sizeof(T) - pad_right - pad_left; - const uint32x2_t voffset_top = { offset_top, offset_top + 1u }; - const uint32x2_t voffset_bottom = { offset_bottom, offset_bottom + 1u }; - const uint32x2_t tmp_indices_top = vbsl_u32(vcge_f32(top_data_f32, vrev64_f32(top_data_f32)), voffset_top, vrev64_u32(voffset_top)); - const uint32x2_t tmp_indices_bottom = vbsl_u32(vcge_f32(bottom_data_f32, vrev64_f32(bottom_data_f32)), voffset_bottom, vrev64_u32(voffset_bottom)); - *(reinterpret_cast<int *>(indices.ptr())) = vget_lane_u32(vbsl_u32(vcge_f32(max_data_top, max_data_bottom), tmp_indices_top, tmp_indices_bottom), 0); - }, - src, dst, indices); -} - -void CpuPoolingKernel::pooling2_f16_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - ARM_COMPUTE_UNUSED(pooling_type); - ARM_COMPUTE_UNUSED(exclude_padding); -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - if(pooling_type == PoolingType::MAX && _indices) - { - pooling2_nchw_maxpool_indices<float16_t>(window_src, window); - } - else - { - Iterator src(_src, window_src); - Iterator dst(_dst, window); - constexpr int pool_size = 2; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x, pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - const unsigned char *const src_top_ptr = _src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); - const unsigned char *const src_bottom_ptr = _src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); - - execute_window_loop(window, [&](const Coordinates & id) - { - float16x4_t top_data = vld1_f16(reinterpret_cast<const float16_t *>(src_top_ptr + src.offset())); - float16x4_t bottom_data = vld1_f16(reinterpret_cast<const float16_t *>(src_bottom_ptr + src.offset())); - float16x4_t res = {}; - - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - top_data = vmul_f16(top_data, top_data); - bottom_data = vmul_f16(bottom_data, bottom_data); - } - - if(pooling_type != PoolingType::MAX) - { - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - const float16x4_t scale_v = vdup_n_f16(scale); - - const float16x4_t sum_data = vadd_f16(top_data, bottom_data); - res = vmul_f16(vpadd_f16(sum_data, sum_data), scale_v); - } - else - { - const float16x4_t max_data = vmax_f16(top_data, bottom_data); - res = vpmax_f16(max_data, max_data); - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - res = vinv_f16(vinvsqrt_f16(res)); - } - - // Store result - *(reinterpret_cast<float16_t *>(dst.ptr())) = vget_lane_f16(res, 0); - }, - src, dst); - } -#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - ARM_COMPUTE_UNUSED(window_src); - ARM_COMPUTE_UNUSED(window); - ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -} - -template <typename T> -void CpuPoolingKernel::pooling3_q8_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - Iterator src(_src, window_src); - Iterator dst(_dst, window); - - /** NEON vector types */ - using q8x8_t = typename wrapper::traits::neon_vector<T, 8>::type; - using q8x16_t = typename wrapper::traits::neon_vector<T, 16>::type; - using q8x8x2_t = typename std::conditional<std::is_same<T, uint8_t>::value, uint8x8x2_t, int8x8x2_t>::type; - using q16_t = typename wrapper::traits::promote_t<T>; - using q16x8_t = typename wrapper::traits::neon_vector<q16_t, 8>::type; - using q16x8x2_t = typename wrapper::traits::neon_vector<q16_t, 16>::type; - - constexpr int pool_size = 3; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - const UniformQuantizationInfo &src_qinfo = _src->info()->quantization_info().uniform(); - const UniformQuantizationInfo &dst_qinfo = _dst->info()->quantization_info().uniform(); - - const float requant_scale = dst_qinfo.scale / src_qinfo.scale; - const int32_t requant_offset = dst_qinfo.offset - static_cast<int32_t>(static_cast<float>(src_qinfo.offset) / requant_scale); - const UniformQuantizationInfo requant_qinfo = UniformQuantizationInfo(requant_scale, requant_offset); - - const T *const src_top_ptr = reinterpret_cast<const T *>(_src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top)))); - const T *const src_middle_ptr = reinterpret_cast<const T *>(_src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1))); - const T *const src_bottom_ptr = reinterpret_cast<const T *>(_src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 2))); - - execute_window_loop(window, [&](const Coordinates & id) - { - const auto top_data = wrapper::vloadq(src_top_ptr + src.offset()); - const auto middle_data = wrapper::vloadq(src_middle_ptr + src.offset()); - const auto bottom_data = wrapper::vloadq(src_bottom_ptr + src.offset()); - q8x8_t fres = {}; - q8x16_t fqres = {}; - - if(pooling_type == PoolingType::AVG) - { - // Convert data to u16 - const q16x8x2_t top_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(top_data)), wrapper::vmovl(wrapper::vgethigh(top_data)) } }; - const q16x8x2_t middle_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(middle_data)), wrapper::vmovl(wrapper::vgethigh(middle_data)) } }; - const q16x8x2_t bottom_data_q16 = { { wrapper::vmovl(wrapper::vgetlow(bottom_data)), wrapper::vmovl(wrapper::vgethigh(bottom_data)) } }; - - // Calculate row sums - const q16x8x2_t vrsum = - { - { - wrapper::vadd(wrapper::vadd(top_data_q16.val[0], bottom_data_q16.val[0]), middle_data_q16.val[0]), - wrapper::vadd(wrapper::vadd(top_data_q16.val[1], bottom_data_q16.val[1]), middle_data_q16.val[1]), - } - }; - const q16x8x2_t vrsum_shifted_1 = - { - { - wrapper::vext_1(vrsum.val[0], vrsum.val[1]), - wrapper::vext_1(vrsum.val[1], vrsum.val[1]) - } - }; - const q16x8x2_t vrsum_shifted_2 = - { - { - wrapper::vext_2(vrsum.val[0], vrsum.val[1]), - wrapper::vext_2(vrsum.val[1], vrsum.val[1]) - } - }; - // Calculate final sum - q16x8x2_t final_sum = - { - { - wrapper::vadd(wrapper::vadd(vrsum.val[0], vrsum_shifted_1.val[0]), vrsum_shifted_2.val[0]), - wrapper::vadd(wrapper::vadd(vrsum.val[1], vrsum_shifted_1.val[1]), vrsum_shifted_2.val[1]), - } - }; - if(pool_stride_x == 2) - { - q16x8_t res = - { - wrapper::vgetlane(final_sum.val[0], 0), - wrapper::vgetlane(final_sum.val[0], 2), - wrapper::vgetlane(final_sum.val[0], 4), - wrapper::vgetlane(final_sum.val[0], 6), - wrapper::vgetlane(final_sum.val[1], 0), - wrapper::vgetlane(final_sum.val[1], 2), - wrapper::vgetlane(final_sum.val[1], 4), - wrapper::vgetlane(final_sum.val[1], 6), - }; - - scale_vector_q16x8<q16_t, q16x8_t>(exclude_padding, res, id, 0, 1, - pool_size, upper_bound_w, upper_bound_h, - pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - fres = wrapper::vmovn(res); - } - else - { - // Scale lower result - scale_vector_q16x8<q16_t, q16x8_t>(exclude_padding, final_sum.val[0], id, 0, 1, - pool_size, upper_bound_w, upper_bound_h, - pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - // Scale lower result - scale_vector_q16x8<q16_t, q16x8_t>(exclude_padding, final_sum.val[1], id, 8, 1, - pool_size, upper_bound_w, upper_bound_h, - pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - fqres = wrapper::vcombine(wrapper::vmovn(final_sum.val[0]), wrapper::vmovn(final_sum.val[1])); - } - } - else - { - const q8x16_t max_data = wrapper::vmax(wrapper::vmax(top_data, bottom_data), middle_data); - const q8x16_t max_data_shift1 = wrapper::vext_1(max_data, max_data); - const q8x16_t max_data_shift2 = wrapper::vext_2(max_data, max_data); - const q8x16_t final_max = wrapper::vmax(wrapper::vmax(max_data, max_data_shift1), max_data_shift2); - - if(pool_stride_x == 2) - { - const q8x8x2_t table = { { wrapper::vgetlow(final_max), wrapper::vgethigh(final_max) } }; - static const q8x8_t lookup_val = { 0, 2, 4, 6, 8, 10, 12, 14 }; - fres = wrapper::vtbl(table, lookup_val); - } - else - { - fqres = final_max; - } - } - - // Store result - if(pool_stride_x == 1) - { - if(src_qinfo != dst_qinfo) - { - fqres = vrequantize_pooling<q8x8_t, q8x16_t>(wrapper::vgetlow(fqres), wrapper::vgethigh(fqres), requant_qinfo); - } - wrapper::vstore(reinterpret_cast<T *>(dst.ptr()), fqres); - } - else - { - if(src_qinfo != dst_qinfo) - { - fres = vrequantize_pooling<q8x8_t>(fres, requant_qinfo); - } - wrapper::vstore(reinterpret_cast<T *>(dst.ptr()), fres); - } - }, - src, dst); -} - -void CpuPoolingKernel::poolingMxN_f16_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - ARM_COMPUTE_UNUSED(pooling_type); - ARM_COMPUTE_UNUSED(exclude_padding); -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - Iterator src(_src, window_src); - Iterator dst(_dst, window); - - const int pool_size_x = _pool_info.is_global_pooling ? _src->info()->tensor_shape().x() : _pool_info.pool_size.width; - const int pool_size_y = _pool_info.is_global_pooling ? _src->info()->tensor_shape().y() : _pool_info.pool_size.height; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - execute_window_loop(window, [&](const Coordinates & id) - { - float16_t res = 0.0f; - float16x8_t vres = vdupq_n_f16(0.0f); - - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - - // Perform pooling - - for(int y = 0; y < pool_size_y; ++y) - { - int x = 0; - for(; x <= (pool_size_x - 8); x += 8) - { - const float16x8_t data = vld1q_f16(reinterpret_cast<const float16_t *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().y()))); - - // Get power of 2 in case of l2 pooling and accumulate - if(pooling_type == PoolingType::L2) - { - vres = vaddq_f16(vres, vmulq_f16(data, data)); - } - else - { - vres = vaddq_f16(vres, data); - } - } - - // Leftover for loop - for(; x < pool_size_x; ++x) - { - float16_t data = *(reinterpret_cast<const float16_t *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().x()) - + (y - pool_pad_top) * static_cast<int>(_src->info()->strides_in_bytes().y()))); - - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - data *= data; - } - - res += data; - } - } - - // Reduction - float16x4_t tmp = vpadd_f16(vget_high_f16(vres), vget_low_f16(vres)); - res += vget_lane_f16(tmp, 0); - res += vget_lane_f16(tmp, 1); - res += vget_lane_f16(tmp, 2); - res += vget_lane_f16(tmp, 3); - - // Divide by scale - res *= scale; - } - else - { - float16x8_t vres = vdupq_n_f16(std::numeric_limits<float>::lowest()); - res = std::numeric_limits<float>::lowest(); - - for(int y = 0; y < pool_size_y; ++y) - { - int x = 0; - for(; x <= (pool_size_x - 8); x += 8) - { - const float16x8_t data = vld1q_f16(reinterpret_cast<const float16_t *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().y()))); - vres = vmaxq_f16(vres, data); - } - - // Leftover for loop - for(; x < pool_size_x; ++x) - { - const float16_t data = *(reinterpret_cast<const float16_t *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().x()) - + (y - pool_pad_top) * static_cast<int>(_src->info()->strides_in_bytes().y()))); - res = std::max(res, data); - } - } - - float16x4_t tmp = vpmax_f16(vget_high_f16(vres), vget_low_f16(vres)); - res = std::max(res, vget_lane_f16(tmp, 0)); - res = std::max(res, vget_lane_f16(tmp, 1)); - res = std::max(res, vget_lane_f16(tmp, 2)); - res = std::max(res, vget_lane_f16(tmp, 3)); - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - res = std::sqrt(res); - } - - // Store result - *(reinterpret_cast<float16_t *>(dst.ptr())) = res; - }, - src, dst); - -#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - ARM_COMPUTE_UNUSED(window_src); - ARM_COMPUTE_UNUSED(window); - ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -} - -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -void CpuPoolingKernel::pooling2_f16_nhwc_maxpool_indices(const Window &window_src, const Window &window) -{ - const int window_start_x = window.x().start(); - const int window_end_x = window.x().end(); - const int window_step_x = 8; - - Window window_out = window; - window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator src(_src, window_src); - Iterator dst(_dst, window_out); - Iterator indices(_indices, window_out); - - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - - const int pad_right = _src->info()->padding().right; - const int in_stride_y = static_cast<int>(_src->info()->strides_in_bytes().y()); - const int in_stride_z = static_cast<int>(_src->info()->strides_in_bytes().z()); - - execute_window_loop(window_out, [&](const Coordinates & id) - { - const int idx_width = id.y() * pool_stride_x; - const int idx_height = id.z() * pool_stride_y; - const int pool_limit_y = pool_pad_top - idx_height; - const int pool_limit_x = pool_pad_left - idx_width; - - const int pool_start_y = std::max(0, window_src.z().start() + pool_limit_y); - const int pool_start_x = std::max(0, window_src.y().start() + pool_limit_x); - const int in_x0_offset = (pool_start_x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z()); - const int in_x1_offset = (pool_start_x + 1 - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z()); - const int in_x2_offset = (pool_start_x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z()); - const int in_x3_offset = (pool_start_x + 1 - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z()); - - int x_off = window_start_x; - for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) - { - const auto in_x0_ptr = reinterpret_cast<const float16_t *>(src.ptr() + in_x0_offset) + x_off; - const auto in_x1_ptr = reinterpret_cast<const float16_t *>(src.ptr() + in_x1_offset) + x_off; - const auto in_x2_ptr = reinterpret_cast<const float16_t *>(src.ptr() + in_x2_offset) + x_off; - const auto in_x3_ptr = reinterpret_cast<const float16_t *>(src.ptr() + in_x3_offset) + x_off; - const auto v_x0 = vld1q_f16(in_x0_ptr); - const auto v_x1 = vld1q_f16(in_x1_ptr); - const auto v_x2 = vld1q_f16(in_x2_ptr); - const auto v_x3 = vld1q_f16(in_x3_ptr); - float16x8_t vres = vmaxq_f16(vmaxq_f16(v_x2, v_x3), vmaxq_f16(v_x0, v_x1)); - // Store result - vst1q_f16(reinterpret_cast<float16_t *>(dst.ptr()) + x_off, vres); - - const uint32_t offset_base = offset_no_padding<float16_t>(src.offset(), id, *_src->info(), pool_stride_x, pool_stride_y); - const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float16_t) + x_off; - const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float16_t) - pad_right; - const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float16_t) - pad_right * _src->info()->tensor_shape()[1]; - const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float16_t) - pad_right; - const uint32x4_t voffset_x0_0 = { offset_x0, offset_x0 + 1, offset_x0 + 2, offset_x0 + 3 }; - const uint32x4_t voffset_x0_1 = { offset_x0 + 4, offset_x0 + 5, offset_x0 + 6, offset_x0 + 7 }; - const uint16x8_t voffset_x0 = vcombine_u16(vmovn_u32(voffset_x0_0), vmovn_u32(voffset_x0_1)); - const uint32x4_t voffset_x1_0 = { offset_x1, offset_x1 + 1, offset_x1 + 2, offset_x1 + 3 }; - const uint32x4_t voffset_x1_1 = { offset_x1 + 4, offset_x1 + 5, offset_x1 + 6, offset_x1 + 7 }; - const uint16x8_t voffset_x1 = vcombine_u16(vmovn_u32(voffset_x1_0), vmovn_u32(voffset_x1_1)); - const uint32x4_t voffset_x2_0 = { offset_x2, offset_x2 + 1, offset_x2 + 2, offset_x2 + 3 }; - const uint32x4_t voffset_x2_1 = { offset_x2 + 4, offset_x2 + 5, offset_x2 + 6, offset_x2 + 7 }; - const uint16x8_t voffset_x2 = vcombine_u16(vmovn_u32(voffset_x2_0), vmovn_u32(voffset_x2_1)); - const uint32x4_t voffset_x3_0 = { offset_x3, offset_x3 + 1, offset_x3 + 2, offset_x3 + 3 }; - const uint32x4_t voffset_x3_1 = { offset_x3 + 4, offset_x3 + 5, offset_x3 + 6, offset_x3 + 7 }; - const uint16x8_t voffset_x3 = vcombine_u16(vmovn_u32(voffset_x3_0), vmovn_u32(voffset_x3_1)); - const uint16x8_t tmp_indices0 = vbslq_u16(vcgeq_f16(v_x0, v_x1), voffset_x0, voffset_x1); - const uint16x8_t tmp_indices1 = vbslq_u16(vcgeq_f16(v_x2, v_x3), voffset_x2, voffset_x3); - const uint16x8_t tmp_indices2 = vbslq_u16(vcgeq_f16(vmaxq_f16(v_x0, v_x1), vmaxq_f16(v_x2, v_x3)), tmp_indices0, tmp_indices1); - const uint32x4_t tmp_indeces3_0 = vmovl_u16(vget_low_u16(tmp_indices2)); - const uint32x4_t tmp_indeces3_1 = vmovl_u16(vget_high_u16(tmp_indices2)); - // Store indicies - vst1q_u32(reinterpret_cast<uint32_t *>(indices.ptr()) + x_off, tmp_indeces3_0); - vst1q_u32(reinterpret_cast<uint32_t *>(indices.ptr() + 16) + x_off, tmp_indeces3_1); - } - - // Left-overs loop - for(; x_off < window_end_x; ++x_off) - { - const auto x0 = *(reinterpret_cast<const float16_t *>(src.ptr() + in_x0_offset) + x_off); - const auto x1 = *(reinterpret_cast<const float16_t *>(src.ptr() + in_x1_offset) + x_off); - const auto x2 = *(reinterpret_cast<const float16_t *>(src.ptr() + in_x2_offset) + x_off); - const auto x3 = *(reinterpret_cast<const float16_t *>(src.ptr() + in_x3_offset) + x_off); - float16_t res = std::max(std::max(x2, x3), std::max(x0, x1)); - - // Store result - *(reinterpret_cast<float16_t *>(dst.ptr()) + x_off) = res; - - const uint32_t offset_base = offset_no_padding<float16_t>(src.offset(), id, *_src->info(), pool_stride_x, pool_stride_y); - const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float16_t) + x_off; - const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float16_t) - pad_right; - const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float16_t) - pad_right * _src->info()->tensor_shape()[1]; - const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float16_t) - pad_right; - const uint32_t tmp_idx0 = (x0 >= x1) ? offset_x0 : offset_x1; - const uint32_t tmp_idx1 = (x2 >= x3) ? offset_x2 : offset_x3; - const uint32_t tmp_idx2 = (std::max(x0, x1) >= std::max(x2, x3)) ? tmp_idx0 : tmp_idx1; - - // Store indices - *(reinterpret_cast<uint32_t *>(indices.ptr()) + x_off) = tmp_idx2; - } - }, - src, dst, indices); -} -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - -void CpuPoolingKernel::poolingMxN_f16_nhwc(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - ARM_COMPUTE_UNUSED(pooling_type); - ARM_COMPUTE_UNUSED(exclude_padding); -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - if(_pool_info.pool_size == Size2D(2, 2) && pooling_type == PoolingType::MAX && _indices) - { - pooling2_f16_nhwc_maxpool_indices(window_src, window); - } - const int window_start_x = window.x().start(); - const int window_end_x = window.x().end(); - const int window_step_x = 8; - - Window window_out = window; - window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator src(_src, window_src); - Iterator dst(_dst, window_out); - - const int pool_size_x = _pool_info.is_global_pooling ? _src->info()->tensor_shape().y() : _pool_info.pool_size.width; - const int pool_size_y = _pool_info.is_global_pooling ? _src->info()->tensor_shape().z() : _pool_info.pool_size.height; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _src->info()->dimension(2) + (exclude_padding ? 0 : pool_pad_bottom); - - float16x8_t vres; - - execute_window_loop(window_out, [&](const Coordinates & id) - { - const int idx_width = id.y() * pool_stride_x; - const int idx_height = id.z() * pool_stride_y; - const int pool_limit_y = pool_pad_top - idx_height; - const int pool_limit_x = pool_pad_left - idx_width; - - const int pool_start_y = std::max(0, window_src.z().start() + pool_limit_y); - const int pool_end_y = std::min(pool_size_y, window_src.z().end() + pool_limit_y); - const int pool_start_x = std::max(0, window_src.y().start() + pool_limit_x); - const int pool_end_x = std::min(pool_size_x, window_src.y().end() + pool_limit_x); - - int x_off = window_start_x; - for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) - { - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, - pool_stride_y); - const float16x8_t scale_v = vdupq_n_f16(scale); - - // Perform pooling - vres = vdupq_n_f16(0.0f); - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float16x8_t data = vld1q_f16(reinterpret_cast<const float16_t *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z())) + x_off); - - // Get power of 2 in case of l2 pooling and accumulate - if(pooling_type == PoolingType::L2) - { - vres = vaddq_f16(vres, vmulq_f16(data, data)); - } - else - { - vres = vaddq_f16(vres, data); - } - } - } - // Divide by scale - vres = vmulq_f16(vres, scale_v); - } - else - { - vres = vdupq_n_f16(std::numeric_limits<float>::lowest()); - - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float16x8_t data = vld1q_f16(reinterpret_cast<const float16_t *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z())) + x_off); - vres = vmaxq_f16(vres, data); - } - } - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - float16x8_t sqrt_reciprocal = vrsqrteq_f16(vres); - vres = vmulq_f16(vres, vmulq_f16(vrsqrtsq_f16(vmulq_f16(vres, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal)); - } - - // Store result - vst1q_f16(reinterpret_cast<float16_t *>(dst.ptr()) + x_off, vres); - } - - // Left-overs loop - for(; x_off < window_end_x; ++x_off) - { - float16_t res = 0.0f; - - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - const float16_t scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, - pool_stride_y); - - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float data = *(reinterpret_cast<const float16_t *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z())) + x_off); - - // Get power of 2 in case of l2 pooling and accumulate - if(pooling_type == PoolingType::L2) - { - res += data * data; - } - else - { - res += data; - } - } - } - - // Divide by scale - res *= scale; - } - else - { - res = std::numeric_limits<float>::lowest(); - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float16_t data = *(reinterpret_cast<const float16_t *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z())) + x_off); - res = std::max(res, data); - } - } - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - res = std::sqrt(res); - } - - // Store result - *(reinterpret_cast<float16_t *>(dst.ptr()) + x_off) = res; - } - }, - src, dst); - -#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - ARM_COMPUTE_UNUSED(window_src); - ARM_COMPUTE_UNUSED(window); - ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -} - -void CpuPoolingKernel::poolingMxN_f32_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - Iterator src(_src, window_src); - Iterator dst(_dst, window); - - const int pool_size_x = _pool_info.is_global_pooling ? _src->info()->tensor_shape().x() : _pool_info.pool_size.width; - const int pool_size_y = _pool_info.is_global_pooling ? _src->info()->tensor_shape().y() : _pool_info.pool_size.height; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - execute_window_loop(window, [&](const Coordinates & id) - { - float res = 0.0f; - - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - - // Perform pooling - float32x4_t vres = vdupq_n_f32(0.0f); - - for(int y = 0; y < pool_size_y; ++y) - { - int x = 0; - for(; x <= (pool_size_x - 4); x += 4) - { - const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().y()))); - - // Get power of 2 in case of l2 pooling and accumulate - if(pooling_type == PoolingType::L2) - { - vres = vmlaq_f32(vres, data, data); - } - else - { - vres = vaddq_f32(vres, data); - } - } - - // Leftover for loop - for(; x < pool_size_x; ++x) - { - float data = *(reinterpret_cast<const float *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().y()))); - - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - data *= data; - } - - res += data; - } - } - -#if defined(__aarch64__) - // Reduction operation available on 64 bit architectures only - res += vaddvq_f32(vres); -#else // __aarch64__ - // Reduction - float32x2_t tmp = vpadd_f32(vget_high_f32(vres), vget_low_f32(vres)); - tmp = vpadd_f32(tmp, tmp); - - res += vget_lane_f32(tmp, 0); -#endif // __aarch64__ - // Divide by scale - res *= scale; - } - else - { - float32x4_t vres = vdupq_n_f32(std::numeric_limits<float>::lowest()); - res = std::numeric_limits<float>::lowest(); - - for(int y = 0; y < pool_size_y; ++y) - { - int x = 0; - for(; x <= (pool_size_x - 4); x += 4) - { - const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().y()))); - vres = vmaxq_f32(vres, data); - } - - // Leftover for loop - for(; x < pool_size_x; ++x) - { - const float data = *(reinterpret_cast<const float *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().y()))); - res = std::max(res, data); - } - } -#if defined(__aarch64__) - // Reduction operation available on 64 bit architectures only - res = std::max(vmaxvq_f32(vres), res); -#else // __aarch64__ - float32x2_t tmp = vpmax_f32(vget_high_f32(vres), vget_low_f32(vres)); - tmp = vpmax_f32(tmp, tmp); - - res = std::max(res, vget_lane_f32(tmp, 0)); -#endif // __aarch64__ - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - res = std::sqrt(res); - } - - // Store result - *(reinterpret_cast<float *>(dst.ptr())) = res; - }, - src, dst); -} - -void CpuPoolingKernel::pooling2_f32_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, - bool exclude_padding) -{ - if(pooling_type == PoolingType::MAX && _indices) - { - pooling2_nchw_maxpool_indices<float>(window_src, window); - } - else - { - Iterator src(_src, window_src); - Iterator dst(_dst, window); - constexpr int pool_size = 2; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - const uint8_t *const src_top_ptr = _src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); - const uint8_t *const src_bottom_ptr = _src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); - - execute_window_loop(window, [&](const Coordinates & id) - { - const auto in_top_ptr = reinterpret_cast<const float *>(src_top_ptr + src.offset()); - const auto in_bottom_ptr = reinterpret_cast<const float *>(src_bottom_ptr + src.offset()); - float32x2_t top_data = vld1_f32(in_top_ptr); - float32x2_t bottom_data = vld1_f32(in_bottom_ptr); - float32x2_t res = {}; - float final_res = 0; - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - top_data = vmul_f32(top_data, top_data); - bottom_data = vmul_f32(bottom_data, bottom_data); - } - - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - const float32x2_t scale_v = vdup_n_f32(scale); - - // Perform pooling - const float32x2_t sum_data = vadd_f32(top_data, bottom_data); - res = vmul_f32(vpadd_f32(sum_data, sum_data), scale_v); - } - else - { - const float32x2_t max_data = vmax_f32(top_data, bottom_data); - res = vpmax_f32(max_data, max_data); - } - final_res = vget_lane_f32(res, 0); - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - final_res = sqrt(final_res); - } - - // Store result - *(reinterpret_cast<float *>(dst.ptr())) = final_res; - }, - src, dst); - } -} - -void CpuPoolingKernel::pooling3_f32_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - Iterator src(_src, window_src); - Iterator dst(_dst, window); - - constexpr const int pool_size = 3; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - const uint8_t *const src_top_ptr = _src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); - const uint8_t *const src_middle_ptr = _src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); - const uint8_t *const src_bottom_ptr = _src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 2)); - - execute_window_loop(window, [&](const Coordinates & id) - { - float32x4_t top_data = vld1q_f32(reinterpret_cast<const float *>(src_top_ptr + src.offset())); - float32x4_t middle_data = vld1q_f32(reinterpret_cast<const float *>(src_middle_ptr + src.offset())); - float32x4_t bottom_data = vld1q_f32(reinterpret_cast<const float *>(src_bottom_ptr + src.offset())); - float32x2_t res = {}; - float final_res = 0; - - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - top_data = vmulq_f32(top_data, top_data); - middle_data = vmulq_f32(middle_data, middle_data); - bottom_data = vmulq_f32(bottom_data, bottom_data); - } - - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - const float32x2_t scale_v = vdup_n_f32(scale); - - // Perform pooling - const float32x4_t sum_data = vaddq_f32(vaddq_f32(top_data, bottom_data), middle_data); - res = vpadd_f32(vget_high_f32(vsetq_lane_f32(0.f, sum_data, 3)), vget_low_f32(sum_data)); - res = vmul_f32(vpadd_f32(res, res), scale_v); - } - else - { - const float32x4_t max_data = vmaxq_f32(vmaxq_f32(top_data, bottom_data), middle_data); - res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits<float>::max(), max_data, 3)), vget_low_f32(max_data)); - res = vpmax_f32(res, res); - } - final_res = vget_lane_f32(res, 0); - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - final_res = sqrt(final_res); - } - - // Store result - *(reinterpret_cast<float *>(dst.ptr())) = final_res; - }, - src, dst); -} - -void CpuPoolingKernel::pooling7_f32_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - Iterator src(_src, window_src); - Iterator dst(_dst, window); - - constexpr const int pool_size = 7; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - std::array<const uint8_t *, pool_size> src_ptrs{ {} }; - for(int i = 0; i < pool_size; ++i) - { - src_ptrs[i] = _src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + i)); - } - - execute_window_loop(window, [&](const Coordinates & id) - { - float32x2_t res = {}; - float final_res = 0.f; - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size, pool_size, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - const float32x2_t scale_v = vdup_n_f32(scale); - - // Perform pooling - float32x4x2_t data = vld2q_f32(reinterpret_cast<const float *>(src_ptrs[0] + src.offset())); - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - data.val[0] = vmulq_f32(data.val[0], data.val[0]); - data.val[1] = vmulq_f32(data.val[1], data.val[1]); - } - float32x4_t sum_data = vaddq_f32(data.val[0], vsetq_lane_f32(0.f, data.val[1], 3)); - for(int i = 1; i < pool_size; ++i) - { - data = vld2q_f32(reinterpret_cast<const float *>(src_ptrs[i] + src.offset())); - // Get power of 2 in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - data.val[0] = vmulq_f32(data.val[0], data.val[0]); - data.val[1] = vmulq_f32(data.val[1], data.val[1]); - } - sum_data = vaddq_f32(sum_data, data.val[0]); - sum_data = vaddq_f32(sum_data, vsetq_lane_f32(0.f, data.val[1], 3)); - } - res = vpadd_f32(vget_high_f32(sum_data), vget_low_f32(sum_data)); - res = vmul_f32(vpadd_f32(res, res), scale_v); - } - else - { - float32x4x2_t max_data = vld2q_f32(reinterpret_cast<const float *>(src_ptrs[0] + src.offset())); - for(int i = 1; i < pool_size; ++i) - { - const float32x4x2_t data = vld2q_f32(reinterpret_cast<const float *>(src_ptrs[i] + src.offset())); - max_data = vmax2q_f32(max_data, data); - } - res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits<float>::max(), max_data.val[1], 3)), vget_low_f32(max_data.val[1])); - res = vpmax_f32(res, vpmax_f32(vget_high_f32(max_data.val[0]), vget_low_f32(max_data.val[0]))); - res = vpmax_f32(res, res); - } - final_res = vget_lane_f32(res, 0); - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - final_res = sqrt(final_res); - } - - // Store result - *(reinterpret_cast<float *>(dst.ptr())) = final_res; - }, - src, dst); -} - -void CpuPoolingKernel::poolingMxN_f32_nhwc(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - if(_pool_info.pool_size == Size2D(2, 2) && pooling_type == PoolingType::MAX && _indices) - { - pooling2_f32_nhwc_maxpool_indices(window_src, window); - } - else - { - const int window_start_x = window.x().start(); - const int window_end_x = window.x().end(); - const int window_step_x = 4; - - Window window_out = window; - window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator src(_src, window_src); - Iterator dst(_dst, window_out); - - const int pool_size_x = _pool_info.is_global_pooling ? _src->info()->tensor_shape().y() : _pool_info.pool_size.width; - const int pool_size_y = _pool_info.is_global_pooling ? _src->info()->tensor_shape().z() : _pool_info.pool_size.height; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _src->info()->dimension(2) + (exclude_padding ? 0 : pool_pad_bottom); - - float32x4_t vres; - - execute_window_loop(window_out, [&](const Coordinates & id) - { - const int idx_width = id.y() * pool_stride_x; - const int idx_height = id.z() * pool_stride_y; - const int pool_limit_y = pool_pad_top - idx_height; - const int pool_limit_x = pool_pad_left - idx_width; - - const int pool_start_y = std::max(0, window_src.z().start() + pool_limit_y); - const int pool_end_y = std::min(pool_size_y, window_src.z().end() + pool_limit_y); - const int pool_start_x = std::max(0, window_src.y().start() + pool_limit_x); - const int pool_end_x = std::min(pool_size_x, window_src.y().end() + pool_limit_x); - - int x_off = window_start_x; - for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) - { - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, - pool_stride_y); - const float32x4_t scale_v = vdupq_n_f32(scale); - - // Perform pooling - vres = vdupq_n_f32(0.0f); - - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z())) + x_off); - - // Get power of 2 in case of l2 pooling and accumulate - if(pooling_type == PoolingType::L2) - { - vres = vmlaq_f32(vres, data, data); - } - else - { - vres = vaddq_f32(vres, data); - } - } - } - // Divide by scale - vres = vmulq_f32(vres, scale_v); - } - else - { - vres = vdupq_n_f32(std::numeric_limits<float>::lowest()); - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z())) + x_off); - vres = vmaxq_f32(vres, data); - } - } - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - float32x4_t l2_res = { static_cast<float>(sqrt(vgetq_lane_f32(vres, 0))), - static_cast<float>(sqrt(vgetq_lane_f32(vres, 1))), - static_cast<float>(sqrt(vgetq_lane_f32(vres, 2))), - static_cast<float>(sqrt(vgetq_lane_f32(vres, 3))) - }; - vres = l2_res; - } - - // Store result - vst1q_f32(reinterpret_cast<float *>(dst.ptr()) + x_off, vres); - } - - // Left-overs loop - for(; x_off < window_end_x; ++x_off) - { - float res = 0.0f; - - if(pooling_type != PoolingType::MAX) - { - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, - pool_stride_y); - - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float data = *(reinterpret_cast<const float *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z())) + x_off); - - // Get power of 2 in case of l2 pooling and accumulate - if(pooling_type == PoolingType::L2) - { - res += data * data; - } - else - { - res += data; - } - } - } - - // Divide by scale - res *= scale; - } - else - { - res = std::numeric_limits<float>::lowest(); - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const float data = *(reinterpret_cast<const float *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z())) + x_off); - res = std::max(res, data); - } - } - } - - // Calculate square-root in case of l2 pooling - if(pooling_type == PoolingType::L2) - { - res = std::sqrt(res); - } - - // Store result - *(reinterpret_cast<float *>(dst.ptr()) + x_off) = res; - } - }, - src, dst); - } -} - -void CpuPoolingKernel::pooling2_f32_nhwc_maxpool_indices(const Window &window_src, const Window &window) -{ - const int window_start_x = window.x().start(); - const int window_end_x = window.x().end(); - const int window_step_x = 4; - - Window window_out = window; - window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator src(_src, window_src); - Iterator dst(_dst, window_out); - Iterator indices(_indices, window_out); - - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - - float32x4_t vres; - float res; - - const int pad_right = _src->info()->padding().right; - const int in_stride_y = static_cast<int>(_src->info()->strides_in_bytes().y()); - const int in_stride_z = static_cast<int>(_src->info()->strides_in_bytes().z()); - - execute_window_loop(window_out, [&](const Coordinates & id) - { - const int idx_width = id.y() * pool_stride_x; - const int idx_height = id.z() * pool_stride_y; - const int pool_limit_y = pool_pad_top - idx_height; - const int pool_limit_x = pool_pad_left - idx_width; - - const int pool_start_y = std::max(0, window_src.z().start() + pool_limit_y); - const int pool_start_x = std::max(0, window_src.y().start() + pool_limit_x); - - const int in_x0_offset = (pool_start_x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z()); - const int in_x1_offset = (pool_start_x + 1 - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z()); - const int in_x2_offset = (pool_start_x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z()); - const int in_x3_offset = (pool_start_x + 1 - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z()); - - int x_off = window_start_x; - for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) - { - const auto in_x0_ptr = reinterpret_cast<const float *>(src.ptr() + in_x0_offset); - const auto in_x1_ptr = reinterpret_cast<const float *>(src.ptr() + in_x1_offset); - const auto in_x2_ptr = reinterpret_cast<const float *>(src.ptr() + in_x2_offset); - const auto in_x3_ptr = reinterpret_cast<const float *>(src.ptr() + in_x3_offset); - const auto v_x0 = vld1q_f32(in_x0_ptr + x_off); - const auto v_x1 = vld1q_f32(in_x1_ptr + x_off); - const auto v_x2 = vld1q_f32(in_x2_ptr + x_off); - const auto v_x3 = vld1q_f32(in_x3_ptr + x_off); - vres = vmaxq_f32(vmaxq_f32(v_x2, v_x3), vmaxq_f32(v_x0, v_x1)); - // Store result - vst1q_f32(reinterpret_cast<float *>(dst.ptr()) + x_off, vres); - - const uint32_t offset_base = offset_no_padding<float>(src.offset(), id, *_src->info(), pool_stride_x, pool_stride_y); - const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float) + x_off; - const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float) - pad_right; - const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float) - pad_right * _src->info()->tensor_shape()[1]; - const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float) - pad_right; - const uint32x4_t voffset_x0 = { offset_x0, offset_x0 + 1, offset_x0 + 2, offset_x0 + 3 }; - const uint32x4_t voffset_x1 = { offset_x1, offset_x1 + 1, offset_x1 + 2, offset_x1 + 3 }; - const uint32x4_t voffset_x2 = { offset_x2, offset_x2 + 1, offset_x2 + 2, offset_x2 + 3 }; - const uint32x4_t voffset_x3 = { offset_x3, offset_x3 + 1, offset_x3 + 2, offset_x3 + 3 }; - const uint32x4_t tmp_indices0 = vbslq_u32(vcgeq_f32(v_x0, v_x1), voffset_x0, voffset_x1); - const uint32x4_t tmp_indices1 = vbslq_u32(vcgeq_f32(v_x2, v_x3), voffset_x2, voffset_x3); - const uint32x4_t tmp_indices2 = vbslq_u32(vcgeq_f32(vmaxq_f32(v_x0, v_x1), vmaxq_f32(v_x2, v_x3)), tmp_indices0, tmp_indices1); - - // Store indices - vst1q_u32(reinterpret_cast<uint32_t *>(indices.ptr()) + x_off, tmp_indices2); - } - - // Left-overs loop - for(; x_off < window_end_x; ++x_off) - { - const auto x0 = *(reinterpret_cast<const float *>(src.ptr() + in_x0_offset) + x_off); - const auto x1 = *(reinterpret_cast<const float *>(src.ptr() + in_x1_offset) + x_off); - const auto x2 = *(reinterpret_cast<const float *>(src.ptr() + in_x2_offset) + x_off); - const auto x3 = *(reinterpret_cast<const float *>(src.ptr() + in_x3_offset) + x_off); - res = std::max(std::max(x2, x3), std::max(x0, x1)); - - // Store result - *(reinterpret_cast<float *>(dst.ptr()) + x_off) = res; - - const uint32_t offset_base = offset_no_padding<float>(src.offset(), id, *_src->info(), pool_stride_x, pool_stride_y); - const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float) + x_off; - const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float) - pad_right; - const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float) - pad_right * _src->info()->tensor_shape()[1]; - const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float) - pad_right; - const uint32_t tmp_idx0 = (x0 >= x1) ? offset_x0 : offset_x1; - const uint32_t tmp_idx1 = (x2 >= x3) ? offset_x2 : offset_x3; - const uint32_t tmp_idx2 = (std::max(x0, x1) >= std::max(x2, x3)) ? tmp_idx0 : tmp_idx1; - - // Store indices - *(reinterpret_cast<uint32_t *>(indices.ptr()) + x_off) = tmp_idx2; - } - }, - src, dst, indices); -} - -template <typename T> -void CpuPoolingKernel::poolingMxN_q8_nchw(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - Iterator src(_src, window_src); - Iterator dst(_dst, window); - - /** NEON vector types */ - using q8x8_t = typename wrapper::traits::neon_vector<T, 8>::type; - using q16_t = typename wrapper::traits::promote_t<T>; - using q16x8_t = typename wrapper::traits::neon_vector<q16_t, 8>::type; - using q32_t = typename wrapper::traits::promote_t<q16_t>; - using q32x4_t = typename wrapper::traits::neon_vector<q32_t, 4>::type; - - const int pool_size_x = _pool_info.is_global_pooling ? _src->info()->tensor_shape().x() : _pool_info.pool_size.width; - const int pool_size_y = _pool_info.is_global_pooling ? _src->info()->tensor_shape().y() : _pool_info.pool_size.height; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _src->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_bottom); - - const UniformQuantizationInfo &src_qinfo = _src->info()->quantization_info().uniform(); - const UniformQuantizationInfo &dst_qinfo = _dst->info()->quantization_info().uniform(); - - execute_window_loop(window, [&](const Coordinates & id) - { - T res = std::numeric_limits<T>::min(); - - if(pooling_type != PoolingType::MAX) - { - q32x4_t vres = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{}); - q32_t sres = 0; - - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NCHW, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, pool_stride_y); - - // Perform pooling - for(int y = 0; y < pool_size_y; ++y) - { - int x = 0; - for(; x <= (pool_size_x - 8); x += 8) - { - const q8x8_t data = wrapper::vload(reinterpret_cast<const T *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().y()))); - - const q16x8_t data_q16 = wrapper::vmovl(data); - vres = wrapper::vadd(vres, wrapper::vaddl(wrapper::vgethigh(data_q16), wrapper::vgetlow(data_q16))); - } - - // Leftover for loop - for(; x < pool_size_x; ++x) - { - T data = *(reinterpret_cast<const T *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().y()))); - sres += data; - } - } - - // Reduction - const auto tmp = wrapper::vpadd(wrapper::vgethigh(vres), wrapper::vgetlow(vres)); - sres += wrapper::vgetlane(tmp, 0) + wrapper::vgetlane(tmp, 1); - - // Divide by scale - res = static_cast<T>(support::cpp11::round(sres * scale)); - } - else - { - q8x8_t vres = wrapper::vdup_n(std::numeric_limits<T>::min(), wrapper::traits::vector_64_tag{}); - - for(int y = 0; y < pool_size_y; ++y) - { - int x = 0; - for(; x <= (pool_size_x - 8); x += 8) - { - const q8x8_t data = wrapper::vload(reinterpret_cast<const T *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().y()))); - vres = wrapper::vmax(vres, data); - } - // Leftover for loop - for(; x < pool_size_x; ++x) - { - const T data = *(reinterpret_cast<const T *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().x()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().y()))); - res = std::max(res, data); - } - } - - // Reduce max - vres = wrapper::vpmax(vres, vres); - vres = wrapper::vpmax(vres, vres); - vres = wrapper::vpmax(vres, vres); - - // Get max value - res = std::max(res, wrapper::vgetlane(vres, 0)); - } - // Store result - res = (src_qinfo != dst_qinfo) ? Qasymm8QuantizationHelper<T>::quantize(Qasymm8QuantizationHelper<T>::dequantize(res, src_qinfo), dst_qinfo) : res; - *(reinterpret_cast<T *>(dst.ptr())) = res; - }, - src, dst); -} - -template <typename T> -void CpuPoolingKernel::poolingMxN_q8_nhwc(const Window &window_src, const Window &window, PoolingType pooling_type, bool exclude_padding) -{ - const int window_start_x = window.x().start(); - const int window_end_x = window.x().end(); - const int window_step_x = 16; - const int window_half_step_x = window_step_x / 2; - - Window window_out = window; - window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator src(_src, window_src); - Iterator dst(_dst, window_out); - - using q8x8_t = typename wrapper::traits::neon_vector<T, 8>::type; - using q8x16_t = typename wrapper::traits::neon_vector<T, 16>::type; - using q16_t = typename wrapper::traits::promote_t<T>; - using q16x8_t = typename wrapper::traits::neon_vector<q16_t, 8>::type; - using q32_t = typename wrapper::traits::promote_t<q16_t>; - using q32x4_t = typename wrapper::traits::neon_vector<q32_t, 4>::type; - - const int pool_size_x = _pool_info.is_global_pooling ? _src->info()->tensor_shape().y() : _pool_info.pool_size.width; - const int pool_size_y = _pool_info.is_global_pooling ? _src->info()->tensor_shape().z() : _pool_info.pool_size.height; - const int pool_pad_right = _pool_info.pad_stride_info.pad_right(); - const int pool_pad_top = _pool_info.pad_stride_info.pad_top(); - const int pool_pad_left = _pool_info.pad_stride_info.pad_left(); - const int pool_pad_bottom = _pool_info.pad_stride_info.pad_bottom(); - - int pool_stride_x = 0; - int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - const int upper_bound_w = _src->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_right); - const int upper_bound_h = _src->info()->dimension(2) + (exclude_padding ? 0 : pool_pad_bottom); - - const float32x4_t half_scale_v = vdupq_n_f32(0.5f); - const UniformQuantizationInfo src_qinfo = _src->info()->quantization_info().uniform(); - const UniformQuantizationInfo dst_qinfo = _dst->info()->quantization_info().uniform(); - - const float quant_rescale = dst_qinfo.scale / src_qinfo.scale; - // "new_offset" doesn't have to consider the "half_scale_v" in its computation - // With a requantization performed in a single step there won't be uncertainties introduced - const int32_t new_offset = dst_qinfo.offset - static_cast<int32_t>(static_cast<float>(src_qinfo.offset) / quant_rescale); - - const float requant_scale = dst_qinfo.scale / src_qinfo.scale; - const int32_t requant_offset = dst_qinfo.offset - static_cast<int32_t>(static_cast<float>(src_qinfo.offset) / requant_scale); - const UniformQuantizationInfo requant_qinfo = UniformQuantizationInfo(requant_scale, requant_offset); - - execute_window_loop(window_out, [&](const Coordinates & id) - { - const int idx_width = id.y() * pool_stride_x; - const int idx_height = id.z() * pool_stride_y; - const int pool_limit_y = pool_pad_top - idx_height; - const int pool_limit_x = pool_pad_left - idx_width; - - const int pool_start_y = std::max(0, window_src.z().start() + pool_limit_y); - const int pool_end_y = std::min(pool_size_y, window_src.z().end() + pool_limit_y); - const int pool_start_x = std::max(0, window_src.y().start() + pool_limit_x); - const int pool_end_x = std::min(pool_size_x, window_src.y().end() + pool_limit_x); - - int x_off = window_start_x; - for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) - { - if(pooling_type != PoolingType::MAX) - { - q32x4_t vres1 = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{}); - q32x4_t vres2 = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{}); - q32x4_t vres3 = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{}); - q32x4_t vres4 = wrapper::vdup_n(static_cast<q32_t>(0.f), wrapper::traits::vector_128_tag{}); - - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, - pool_stride_y); - - // Perform pooling - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const q8x16_t data = wrapper::vloadq(reinterpret_cast<const T *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z())) + x_off); - - const q16x8_t data_q16 = wrapper::vmovl(wrapper::vgetlow(data)); - const q16x8_t data2_q16 = wrapper::vmovl(wrapper::vgethigh(data)); - vres1 = wrapper::vadd(vres1, wrapper::vmovl(wrapper::vgetlow(data_q16))); - vres2 = wrapper::vadd(vres2, wrapper::vmovl(wrapper::vgethigh(data_q16))); - vres3 = wrapper::vadd(vres3, wrapper::vmovl(wrapper::vgetlow(data2_q16))); - vres4 = wrapper::vadd(vres4, wrapper::vmovl(wrapper::vgethigh(data2_q16))); - } - } - - if(src_qinfo != dst_qinfo) - { - const float32x4x4_t vres = - { - { - vcvtq_f32_q32(vres1), - vcvtq_f32_q32(vres2), - vcvtq_f32_q32(vres3), - vcvtq_f32_q32(vres4), - } - }; - const auto requantized_dst = vrequantize_pooling_with_scale<q8x16_t>(vres, quant_rescale, scale, new_offset); - // Store result - wrapper::vstore(reinterpret_cast<T *>(dst.ptr()) + x_off, wrapper::vgetlow(requantized_dst)); - wrapper::vstore(reinterpret_cast<T *>(dst.ptr()) + x_off + 8, wrapper::vgethigh(requantized_dst)); - } - else - { - const float32x4_t scale_v = vdupq_n_f32(scale); - // Divide by scale and add 0.5f to round to nearest instead of rounding towards zero - vres1 = vcvtq_q32_f32<q32x4_t>(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres1), scale_v)); - vres2 = vcvtq_q32_f32<q32x4_t>(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres2), scale_v)); - vres3 = vcvtq_q32_f32<q32x4_t>(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres3), scale_v)); - vres4 = vcvtq_q32_f32<q32x4_t>(wrapper::vmla(half_scale_v, vcvtq_f32_q32(vres4), scale_v)); - - const q8x8_t res1 = wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(vres1), wrapper::vmovn(vres2))); - const q8x8_t res2 = wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(vres3), wrapper::vmovn(vres4))); - // Store result - wrapper::vstore(reinterpret_cast<T *>(dst.ptr()) + x_off, res1); - wrapper::vstore(reinterpret_cast<T *>(dst.ptr()) + x_off + 8, res2); - } - } - else - { - q8x16_t vres = wrapper::vdup_n(std::numeric_limits<T>::min(), wrapper::traits::vector_128_tag{}); - - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const q8x16_t data = wrapper::vloadq(reinterpret_cast<const T *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z())) + x_off); - vres = wrapper::vmax(vres, data); - } - } - - // Store result - wrapper::vstore(reinterpret_cast<T *>(dst.ptr()) + x_off, (src_qinfo != dst_qinfo) ? vrequantize_pooling<q8x8_t, q8x16_t>(wrapper::vgetlow(vres), wrapper::vgethigh(vres), - requant_qinfo) : - vres); - } - } - - if(pooling_type == PoolingType::MAX) - { - for(; x_off <= (window_end_x - window_half_step_x); x_off += window_half_step_x) - { - q8x8_t vres = wrapper::vdup_n(std::numeric_limits<T>::min(), wrapper::traits::vector_64_tag{}); - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const q8x8_t data = wrapper::vload(reinterpret_cast<const T *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z())) + x_off); - vres = wrapper::vmax(vres, data); - } - } - - // Store result - wrapper::vstore(reinterpret_cast<T *>(dst.ptr()) + x_off, - (src_qinfo != dst_qinfo) ? vrequantize_pooling<q8x8_t>(vres, requant_qinfo) : vres); - } - } - - // Left-overs loop - for(; x_off < window_end_x; ++x_off) - { - if(pooling_type != PoolingType::MAX) - { - q32_t res = static_cast<q32_t>(0.f); - - // Calculate scale - const float scale = calculate_avg_scale(exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, - pool_stride_y); - - // Perform pooling - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const T data = *(reinterpret_cast<const T *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z())) + x_off); - res += data; - } - } - - if(src_qinfo != dst_qinfo) - { - const float res_f = static_cast<float>(res); - const float new_scale = quant_rescale / scale; - const auto requantized_dst = quantize<T>(res_f, UniformQuantizationInfo(new_scale, new_offset)); - - // Store result - *(reinterpret_cast<T *>(dst.ptr()) + x_off) = requantized_dst; - } - else - { - // Divide by scale and add 0.5f to round to nearest instead of rounding towards zero - res = static_cast<T>(0.5f + static_cast<float>(res) * scale); - - // Store result - *(reinterpret_cast<T *>(dst.ptr()) + x_off) = res; - } - } - else - { - T res = std::numeric_limits<T>::min(); - - for(int y = pool_start_y; y < pool_end_y; ++y) - { - for(int x = pool_start_x; x < pool_end_x; ++x) - { - const T data = *(reinterpret_cast<const T *>(src.ptr() + (x - pool_pad_left) * static_cast<int>(_src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> - (_src->info()->strides_in_bytes().z())) + x_off); - res = std::max(res, data); - } - } - - // Store result - if(src_qinfo != dst_qinfo) - { - const float res_f = static_cast<float>(res); - *(reinterpret_cast<T *>(dst.ptr()) + x_off) = quantize<T>(res_f, requant_qinfo); - } - else - { - *(reinterpret_cast<T *>(dst.ptr()) + x_off) = res; - } - } - } - - }, - src, dst); -} - Status CpuPoolingKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &pool_info, const ITensorInfo *indices) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src); @@ -2547,23 +481,21 @@ void CpuPoolingKernel::run_op(ITensorPack &tensors, const Window &window, const ARM_COMPUTE_UNUSED(info); ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window); - ARM_COMPUTE_ERROR_ON(_func == nullptr); - _src = tensors.get_const_tensor(TensorType::ACL_SRC_0); - _dst = tensors.get_tensor(TensorType::ACL_DST_0); - _indices = tensors.get_tensor(TensorType::ACL_DST_1); + const ITensor *src = tensors.get_const_tensor(TensorType::ACL_SRC_0); + ITensor *dst = tensors.get_tensor(TensorType::ACL_DST_0); + ITensor *indices = tensors.get_tensor(TensorType::ACL_DST_1); - const unsigned int pool_stride_x = _pool_info.pad_stride_info.stride().first; - const unsigned int pool_stride_y = _pool_info.pad_stride_info.stride().second; - const unsigned int pool_size = _pool_info.pool_size.width; - const bool exclude_padding = _pool_info.exclude_padding; + const unsigned int pool_stride_x = _pool_info.pad_stride_info.stride().first; + const unsigned int pool_stride_y = _pool_info.pad_stride_info.stride().second; + const unsigned int pool_size = _pool_info.pool_size.width; Window window_src(window); if(_data_layout == DataLayout::NCHW) { // Set step for src in x and y direction for the src unsigned int window_x_inc = 0; - switch(_src->info()->data_type()) + switch(src->info()->data_type()) { case DataType::QASYMM8: case DataType::QASYMM8_SIGNED: @@ -2593,12 +525,19 @@ void CpuPoolingKernel::run_op(ITensorPack &tensors, const Window &window, const else { window_src.set(Window::DimX, Window::Dimension(0, 1, 1)); - window_src.set(Window::DimY, Window::Dimension(0, _src->info()->dimension(1), pool_stride_x)); - window_src.set(Window::DimZ, Window::Dimension(0, _src->info()->dimension(2), pool_stride_y)); + window_src.set(Window::DimY, Window::Dimension(0, src->info()->dimension(1), pool_stride_x)); + window_src.set(Window::DimZ, Window::Dimension(0, src->info()->dimension(2), pool_stride_y)); } - // Run function - (this->*_func)(window_src, window, _pool_info.pool_type, exclude_padding); + const auto *uk = get_implementation(src->info()->data_type(), src->info()->data_layout(), _pool_stride_x, _pool_size); + ARM_COMPUTE_ERROR_ON(uk == nullptr || uk->ukernel == nullptr); + + uk->ukernel(src, dst, indices, _pool_info, window_src, window); +} + +const char *CpuPoolingKernel::name() const +{ + return "CpuPoolingKernel"; } } // namespace kernels } // namespace cpu |