diff options
author | Michalis Spyrou <michalis.spyrou@arm.com> | 2019-08-22 16:52:00 +0100 |
---|---|---|
committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2019-08-30 13:39:54 +0000 |
commit | 3f632f3f16e29ebeb7065b30008060fd4bfd09f1 (patch) | |
tree | ce15e27cf559bd1828b302e1257c187fece0b477 /src/core/NEON/kernels/NEDequantizationLayerKernel.cpp | |
parent | 351bd137e48c5276963274ac741b172483e98d21 (diff) | |
download | ComputeLibrary-3f632f3f16e29ebeb7065b30008060fd4bfd09f1.tar.gz |
COMPMID-2418: CLDequantizationLayer support for QASYMM8_PER_CHANNEL
Add support for QASYMM8_PER_CHANNEL in CLDequantiazationLayer.
Added tests for NHWC and also updated NEON code to work with NHWC
data layout.
Cleaned up the reference implementation.
Change-Id: Ic1d51f16f7f625503fffdbbb66f6487aa588f08c
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1828
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/NEDequantizationLayerKernel.cpp')
-rw-r--r-- | src/core/NEON/kernels/NEDequantizationLayerKernel.cpp | 64 |
1 files changed, 62 insertions, 2 deletions
diff --git a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp index d880c80d82..49de3ec8b3 100644 --- a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp +++ b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp @@ -160,7 +160,7 @@ void run_dequantization_qasymm8(const ITensor *input, ITensor *output, const Win } template <typename T> -void run_dequantization_qasymm8_per_channel(const ITensor *input, ITensor *output, const Window &window) +void run_dequantization_qasymm8_per_channel_nchw(const ITensor *input, ITensor *output, const Window &window) { const std::vector<float> scale = input->info()->quantization_info().scale(); const std::vector<int32_t> offset = input->info()->quantization_info().offset(); @@ -202,6 +202,66 @@ void run_dequantization_qasymm8_per_channel(const ITensor *input, ITensor *outpu } template <typename T> +void run_dequantization_qasymm8_per_channel_nhwc(const ITensor *input, ITensor *output, const Window &window) +{ + const std::vector<float> scale = input->info()->quantization_info().scale(); + const std::vector<int32_t> offset = input->info()->quantization_info().offset(); + + const int window_step_x = 16; + const auto window_start_x = static_cast<int>(window.x().start()); + const auto window_end_x = static_cast<int>(window.x().end()); + + // Reset first dimension to handle tail calculations manually + Window win(window); + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + // Create iterators + Iterator in(input, win); + Iterator out(output, win); + + execute_window_loop(win, [&](const Coordinates & id) + { + const auto in_ptr = reinterpret_cast<const uint8_t *>(in.ptr()); + const auto out_ptr = reinterpret_cast<T *>(out.ptr()); + + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const float32x4x4_t vscale = + { + { + scale[x + 0], scale[x + 1], scale[x + 2], scale[x + 3], + scale[x + 4], scale[x + 5], scale[x + 6], scale[x + 7], + scale[x + 8], scale[x + 9], scale[x + 10], scale[x + 11], + scale[x + 12], scale[x + 13], scale[x + 14], scale[x + 15] + } + }; + const int32x4x4_t voffset = + { + { + offset[x + 0], offset[x + 1], offset[x + 2], offset[x + 3], + offset[x + 4], offset[x + 5], offset[x + 6], offset[x + 7], + offset[x + 8], offset[x + 9], offset[x + 10], offset[x + 11], + offset[x + 12], offset[x + 13], offset[x + 14], offset[x + 15] + } + }; + const auto vin = wrapper::vloadq(in_ptr + x); + const auto vdeq = vdequantize(vin, vscale, voffset); + + store_result<T>(reinterpret_cast<T *>(out_ptr + x), vdeq); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + uint8_t val = *(in_ptr + x); + *(out_ptr + x) = static_cast<T>(dequantize(val, scale[x], offset[x])); + } + }, + in, out); +} + +template <typename T> void run_dequantization_qsymm8(const ITensor *input, ITensor *output, const Window &window) { const UniformQuantizationInfo &qinfo = input->info()->quantization_info().uniform(); @@ -294,7 +354,7 @@ void run_dequantization_core(const ITensor *input, ITensor *output, const Window run_dequantization_qasymm8<T>(input, output, window); break; case DataType::QASYMM8_PER_CHANNEL: - run_dequantization_qasymm8_per_channel<T>(input, output, window); + input->info()->data_layout() == DataLayout::NHWC ? run_dequantization_qasymm8_per_channel_nhwc<T>(input, output, window) : run_dequantization_qasymm8_per_channel_nchw<T>(input, output, window); break; case DataType::QSYMM8: run_dequantization_qsymm8<T>(input, output, window); |