From 56dd726ee074cb145612d03240b710f8adb82ddd Mon Sep 17 00:00:00 2001 From: Michele Di Giorgio Date: Thu, 27 Jul 2017 09:53:49 +0100 Subject: COMPMID-448: Implement CL Quantization/Dequantization Layer. Change-Id: Id002e23a2ac48af3d245416dc6411d9a04a1e513 Reviewed-on: http://mpd-gerrit.cambridge.arm.com/81827 Reviewed-by: Gian Marco Iodice Reviewed-by: Georgios Pinitas Tested-by: Kaizen --- src/core/CL/cl_kernels/dequantization_layer.cl | 73 ++++++++++++++++++ src/core/CL/cl_kernels/minmax_layer.cl | 101 +++++++++++++++++++++++++ src/core/CL/cl_kernels/quantization_layer.cl | 75 ++++++++++++++++++ 3 files changed, 249 insertions(+) create mode 100644 src/core/CL/cl_kernels/dequantization_layer.cl create mode 100644 src/core/CL/cl_kernels/minmax_layer.cl create mode 100644 src/core/CL/cl_kernels/quantization_layer.cl (limited to 'src/core/CL/cl_kernels') diff --git a/src/core/CL/cl_kernels/dequantization_layer.cl b/src/core/CL/cl_kernels/dequantization_layer.cl new file mode 100644 index 0000000000..21e9c873ac --- /dev/null +++ b/src/core/CL/cl_kernels/dequantization_layer.cl @@ -0,0 +1,73 @@ +/* + * Copyright (c) 2017 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "helpers.h" + +/** This performs the dequantization of 8-bit unsigned integers to floating point. + * + * @param[in] input_ptr Pointer to the source image. Supported data types: QS8/QS16/F16/F32 + * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) + * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image + * @param[out] output_ptr Pointer to the destination image. Supported data types: same as @p input_ptr + * @param[in] output_stride_x Stride of the destination image in X dimension (in bytes) + * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes) + * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image + * @param[in] min_max_ptr Pointer to the min/max vector. Minimum value in position 0, maximum value in position 1. Suppported data types: F32. + * @param[in] min_max_stride_x Stride of the min/max vector in X dimension (in bytes) + * @param[in] min_max_step_x min_max_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] min_max_offset_first_element_in_bytes The offset of the first element in the min/max vector + */ +__kernel void dequantization_layer( + TENSOR3D_DECLARATION(input), + TENSOR3D_DECLARATION(output), + VECTOR_DECLARATION(min_max)) +{ + // Get pixels pointer + Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); + Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); + Vector min_max = CONVERT_TO_VECTOR_STRUCT(min_max); + + // min_max_value.s0 = min, min_max_value.s1 = max + const float2 min_max_value = vload2(0, (__global float *)min_max.ptr); + + const float4 vmin = (float4)min_max_value.s0; + const float4 scale = (float4)((min_max_value.s1 - min_max_value.s0) / 255.0f); + + // Load data + const uchar4 data = vload4(0, (__global uchar *)input.ptr); + + // Dequantize + const float4 res = convert_float4(data) * scale + vmin; + + // Store result + vstore4(res, 0, (__global float *)output.ptr); +} diff --git a/src/core/CL/cl_kernels/minmax_layer.cl b/src/core/CL/cl_kernels/minmax_layer.cl new file mode 100644 index 0000000000..1e543b43bd --- /dev/null +++ b/src/core/CL/cl_kernels/minmax_layer.cl @@ -0,0 +1,101 @@ +/* + * Copyright (c) 2017 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "helpers.h" + +#if defined(WIDTH) && defined(HEIGHT) && defined(DEPTH) +/** This function identifies the min and maximum value of an input 3D tensor. + * + * @note The width, height and depth of the input tensor must be provided at compile time using -DWIDTH, -DHEIGHT and -DDEPTH (e.g. -DWIDTH=320, -DHEIGHT=240, -DDEPTH=3) + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 + * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source image in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image + * @param[in] dst_ptr Pointer to the min/max vector. Minimum value in position 0, maximum value in position 1. Supported data types: F32. + * @param[in] dst_stride_x Stride of the min/max vector in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the min/max vector + */ +__kernel void minmax_layer( + TENSOR3D_DECLARATION(src), + VECTOR_DECLARATION(dst)) +{ + Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); + Vector dst = CONVERT_TO_VECTOR_STRUCT(dst); + + float4 min_value = (float4)FLT_MAX; + float4 max_value = (float4) - FLT_MAX; + float2 min_max_value = (float2)(FLT_MAX, -FLT_MAX); + + for(int z = 0; z < DEPTH; ++z) + { + for(int y = 0; y < HEIGHT; ++y) + { + int x = 0; + __global float *src_addr = (__global float *)(src.ptr + y * src_stride_y + z * src_stride_z); + + for(; x <= (int)(WIDTH - 8); x += 8) + { + float8 value = *(src_addr + x); + + min_value = select(value.s0123, min_value, min_value < value.s0123); + min_value = select(value.s4567, min_value, min_value < value.s4567); + + max_value = select(value.s0123, max_value, max_value > value.s0123); + max_value = select(value.s4567, max_value, max_value > value.s4567); + } + + for(; x < WIDTH; ++x) + { + float value = *(src_addr + x); + + min_max_value.s0 = min(min_max_value.s0, value); + min_max_value.s1 = max(min_max_value.s1, value); + } + } + } + + // Perform min/max reduction + min_value.s01 = min(min_value.s01, min_value.s23); + min_value.s0 = min(min_value.s0, min_value.s1); + max_value.s01 = max(max_value.s01, max_value.s23); + max_value.s0 = max(max_value.s0, max_value.s1); + + min_max_value.s0 = min(min_max_value.s0, min_value.s0); + min_max_value.s1 = max(min_max_value.s1, max_value.s0); + + if(min_max_value.s0 == min_max_value.s1) + { + min_max_value.s0 = 0.0f; + min_max_value.s1 = 1.0f; + } + + // Store min and max + vstore2(min_max_value, 0, (__global float *)dst.ptr); +} +#endif // defined(WIDTH) && defined(HEIGHT) && defined(DEPTH) \ No newline at end of file diff --git a/src/core/CL/cl_kernels/quantization_layer.cl b/src/core/CL/cl_kernels/quantization_layer.cl new file mode 100644 index 0000000000..80ea54012f --- /dev/null +++ b/src/core/CL/cl_kernels/quantization_layer.cl @@ -0,0 +1,75 @@ +/* + * Copyright (c) 2017 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "helpers.h" + +/** This performs the quantization of floating point inputs to 8-bit unsigned integers. + * + * @param[in] input_ptr Pointer to the source image. Supported data types: F32 + * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) + * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image + * @param[out] output_ptr Pointer to the destination image. Supported data types: U8 + * @param[in] output_stride_x Stride of the destination image in X dimension (in bytes) + * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes) + * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image + * @param[in] min_max_ptr Pointer to the min/max vector. Minimum value in position 0, maximum value in position 1. Supported data types: F32. + * @param[in] min_max_stride_x Stride of the min/max vector in X dimension (in bytes) + * @param[in] min_max_step_x min_max_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] min_max_offset_first_element_in_bytes The offset of the first element in the min/max vector + */ +__kernel void quantization_layer( + TENSOR3D_DECLARATION(input), + TENSOR3D_DECLARATION(output), + VECTOR_DECLARATION(min_max)) +{ + // Get pixels pointer + Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); + Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); + + // min_max_value.s0 = min, min_max_value.s1 = max + const float2 min_max_value = vload2(0, (__global float *)(min_max_ptr + min_max_offset_first_element_in_bytes)); + + const float4 vmin = (float4)min_max_value.s0; + const float4 vrange = (float4)(min_max_value.s1 - min_max_value.s0); + + // Load data + float4 data = vload4(0, (__global float *)input.ptr); + + // Map float values to range [0.0, 1.0] + data = (data - vmin) / vrange; + + // Quantize and saturate + uchar4 res = convert_uchar4_sat(data * 256.0f); + + // Store result + vstore4(res, 0, (__global uchar *)output.ptr); +} -- cgit v1.2.1