diff options
Diffstat (limited to 'src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp')
| -rw-r--r-- | src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp | 181 |
1 files changed, 96 insertions, 85 deletions
diff --git a/src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp b/src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp index e49fd29115..e3dd2240ca 100644 --- a/src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp +++ b/src/cpu/kernels/CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp @@ -29,12 +29,13 @@ #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/Utils.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/core/Validate.h" #include "arm_compute/core/Window.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "src/core/NEON/NEAsymm.h" + #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" +#include "src/core/NEON/NEAsymm.h" #include <arm_neon.h> @@ -53,14 +54,14 @@ Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ARM_COMPUTE_RETURN_ERROR_ON(min > max); // Check biases if exist - if(bias != nullptr) + if (bias != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias); ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != bias->dimension(0)); } - if(dst->total_size() != 0) + if (dst->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, src); @@ -71,7 +72,10 @@ Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const } // namespace template <bool is_bounded_relu> -void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window) +void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal(const ITensor *src, + const ITensor *bias, + ITensor *dst, + const Window &window) { const int32x4_t result_offset_after_shift_s32 = vdupq_n_s32(_result_offset_after_shift); const uint8x16_t min_u8 = vdupq_n_u8(static_cast<uint8_t>(_min)); @@ -89,98 +93,102 @@ void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal(co Iterator in(src, win_collapsed); Iterator out(dst, win_collapsed); - if(bias != nullptr) + if (bias != nullptr) { Window win_biases; win_biases.set(Window::DimX, Window::Dimension(0, 1, 1)); win_biases.set(Window::DimY, Window::Dimension(0, 1, 1)); Iterator bias_i(bias, win_biases); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - // Compute 16 elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) + execute_window_loop( + win_collapsed, + [&](const Coordinates &) { - int32x4x4_t in_s32 = + // Compute 16 elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) { - { - vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0), - vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4), - vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8), - vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12) - } - }; - - const int32x4x4_t bias_s32 = + int32x4x4_t in_s32 = {{vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0), + vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4), + vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8), + vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)}}; + + const int32x4x4_t bias_s32 = { + {vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0), + vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4), + vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 8), + vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 12)}}; + + // Add the bias to GEMM's result + in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]); + in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]); + in_s32.val[2] = vaddq_s32(in_s32.val[2], bias_s32.val[2]); + in_s32.val[3] = vaddq_s32(in_s32.val[3], bias_s32.val[3]); + + vst1q_u8(out.ptr() + x, + finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, + result_offset_after_shift_s32, min_u8, max_u8, is_bounded_relu)); + } + + // Compute left-over elements + for (; x < window_end_x; ++x) { - { - vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0), - vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4), - vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 8), - vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 12) - } - }; - - // Add the bias to GEMM's result - in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]); - in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]); - in_s32.val[2] = vaddq_s32(in_s32.val[2], bias_s32.val[2]); - in_s32.val[3] = vaddq_s32(in_s32.val[3], bias_s32.val[3]); - - vst1q_u8(out.ptr() + x, finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_u8, max_u8, is_bounded_relu)); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x); - int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x); - - // Add bias - in_value += bias_value; - // Finalize and store the result - *(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift, static_cast<uint8_t>(_min), static_cast<uint8_t>(_max), is_bounded_relu); - } - }, - in, out, bias_i); + const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x); + int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x); + + // Add bias + in_value += bias_value; + // Finalize and store the result + *(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, + _result_offset_after_shift, static_cast<uint8_t>(_min), + static_cast<uint8_t>(_max), is_bounded_relu); + } + }, + in, out, bias_i); } else { - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - // Compute 16 elements per iteration - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) + execute_window_loop( + win_collapsed, + [&](const Coordinates &) { - int32x4x4_t in_s32 = + // Compute 16 elements per iteration + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) { - { - vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0), - vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4), - vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8), - vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12) - } - }; - - vst1q_u8(out.ptr() + x, finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_u8, max_u8, is_bounded_relu)); - } - - // Compute left-over elements - for(; x < window_end_x; ++x) - { - const int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x); - - // Finalize and store the result - *(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift, static_cast<uint8_t>(_min), static_cast<uint8_t>(_max), is_bounded_relu); - } - }, - in, out); + int32x4x4_t in_s32 = {{vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0), + vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4), + vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8), + vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)}}; + + vst1q_u8(out.ptr() + x, + finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, + result_offset_after_shift_s32, min_u8, max_u8, is_bounded_relu)); + } + + // Compute left-over elements + for (; x < window_end_x; ++x) + { + const int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x); + + // Finalize and store the result + *(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, + _result_offset_after_shift, static_cast<uint8_t>(_min), + static_cast<uint8_t>(_max), is_bounded_relu); + } + }, + in, out); } } -void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift, - int result_offset_after_shift, int min, int max) +void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(ITensorInfo *src, + ITensorInfo *bias, + ITensorInfo *dst, + int result_fixedpoint_multiplier, + int result_shift, + int result_offset_after_shift, + int min, + int max) { ARM_COMPUTE_UNUSED(bias); // Perform validate step @@ -202,18 +210,21 @@ void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(ITens // Check if we need to clamp the result using min and max const bool is_bounded_relu = !(min <= 0 && max >= 255); - _func = is_bounded_relu ? &CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal<true> : - &CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal<false>; + _func = is_bounded_relu ? &CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal<true> + : &CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_internal<false>; } -Status CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max) +Status CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate( + const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max) { ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, bias, dst, min, max)); return Status{}; } -void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) +void CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run_op(ITensorPack &tensors, + const Window &window, + const ThreadInfo &info) { ARM_COMPUTE_UNUSED(info); ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); @@ -233,4 +244,4 @@ const char *CpuGemmLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::name() c } } // namespace kernels } // namespace cpu -} // namespace arm_compute
\ No newline at end of file +} // namespace arm_compute |