From 578a9fc6c06ebbd6e2650372029e339a4cbcacca Mon Sep 17 00:00:00 2001 From: Michele Di Giorgio Date: Fri, 23 Aug 2019 11:49:04 +0100 Subject: COMPMID-2317: Implement CLROIAlignLayer Change-Id: Iaa61b7a3528d3f82339d2ff8a2d77e77a1c68603 Signed-off-by: Michele Di Giorgio Reviewed-on: https://review.mlplatform.org/c/1821 Reviewed-by: Pablo Marquez Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins --- src/core/CL/CLKernelLibrary.cpp | 7 +- .../CL/cl_kernels/roi_align_layer_quantized.cl | 225 +++++++++++++++++++++ src/core/CL/kernels/CLROIAlignLayerKernel.cpp | 39 +++- 3 files changed, 265 insertions(+), 6 deletions(-) create mode 100644 src/core/CL/cl_kernels/roi_align_layer_quantized.cl (limited to 'src') diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp index b938a1813a..4b3b37c3da 100644 --- a/src/core/CL/CLKernelLibrary.cpp +++ b/src/core/CL/CLKernelLibrary.cpp @@ -447,6 +447,7 @@ const std::map CLKernelLibrary::_kernel_program_map = { "RGBA8888_to_RGB888_bt709", "color_convert.cl" }, { "RGBA8888_to_YUV444_bt709", "color_convert.cl" }, { "roi_align_layer", "roi_align_layer.cl" }, + { "roi_align_layer_quantized", "roi_align_layer_quantized.cl" }, { "roi_pooling_layer", "roi_pooling_layer.cl" }, { "scale_nearest_neighbour_nchw", "scale.cl" }, { "scale_nearest_neighbour_nhwc", "scale.cl" }, @@ -933,6 +934,10 @@ const std::map CLKernelLibrary::_program_source_map = { "roi_align_layer.cl", #include "./cl_kernels/roi_align_layer.clembed" + }, + { + "roi_align_layer_quantized.cl", +#include "./cl_kernels/roi_align_layer_quantized.clembed" }, { "roi_pooling_layer.cl", @@ -1251,4 +1256,4 @@ std::string CLKernelLibrary::get_device_version() cl_uint CLKernelLibrary::get_num_compute_units() { return _device.getInfo(); -} \ No newline at end of file +} diff --git a/src/core/CL/cl_kernels/roi_align_layer_quantized.cl b/src/core/CL/cl_kernels/roi_align_layer_quantized.cl new file mode 100644 index 0000000000..f9360e98f1 --- /dev/null +++ b/src/core/CL/cl_kernels/roi_align_layer_quantized.cl @@ -0,0 +1,225 @@ +/* + * Copyright (c) 2019 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 specifies the value to shift the result of roi_dims / pooled_dims before ceiling. +// It is close to the epsilon machine (for a floating point system, x and x+EPS are the same number). +#define EPS_GRID 0.00001f + +#if defined(DATA_TYPE) && defined(POOLED_DIM_X) && defined(POOLED_DIM_Y) && defined(MAX_DIM_X) && defined(MAX_DIM_Y) && defined(MAX_DIM_Z) && defined(SPATIAL_SCALE) && defined(OFFSET_IN) && defined(OFFSET_OUT) && defined(SCALE_IN) && defined(SCALE_OUT) && defined(OFFSET_ROIS) && defined(SCALE_ROIS) // Check for compile time constants + +#define CONVERT_RTE(x, type) (convert_##type##_rte((x))) +#define CONVERT_DOWN(x, type) CONVERT_RTE(x, type) +inline float dequantize_qasymm8(uchar input, float offset, float scale) +{ + return ((float)input - offset) * scale; +} + +inline uchar quantize_qasymm8(float input, float offset, float scale) +{ + float out_f32 = input / scale + offset; + uchar res_u8 = CONVERT_SAT(CONVERT_DOWN(out_f32, int), uchar); + return res_u8; +} + +inline float4 dequantize_qasymm16(ushort4 input, float offset, float scale) +{ + float4 in_f32 = (CONVERT(input, float4) - (float4)(offset)) * (float4)(scale); + return in_f32; +} + +/** Performs a roi align on a single output pixel. + * + * @param[in] input Pointer to input Tensor3D struct. + * @param[in] region_start_x Start x index projected onto the input tensor. + * @param[in] region_end_x End x index projected onto the input tensor. + * @param[in] region_start_y Start y index projected onto the input tensor. + * @param[in] region_end_y End y index projected onto the input tensor. + * @param[in] pz z index of the input tensor. + * + * @return An average pooled value from the region specified in the input tensor. + */ +inline DATA_TYPE roi_align_1x1(const Tensor3D *input, float region_start_x, + float bin_size_x, + float grid_size_x, + float region_end_x, + float region_start_y, + float bin_size_y, + float grid_size_y, + float region_end_y, + int pz) +{ + // Iterate through the pooling region + float sum = 0; + for(int iy = 0; iy < grid_size_y; ++iy) + { + for(int ix = 0; ix < grid_size_x; ++ix) + { + // Align the window in the middle of every bin + const float y = region_start_y + (iy + 0.5f) * bin_size_y / (float)grid_size_y; + const float x = region_start_x + (ix + 0.5f) * bin_size_x / (float)grid_size_x; + + // Interpolation in the unit square + const int y_low = (int)y; + const int x_low = (int)x; + const int y_high = y_low + 1; + const int x_high = x_low + 1; + + const float ly = y - y_low; + const float lx = x - x_low; + const float hy = 1.f - ly; + const float hx = 1.f - lx; + + const float w1 = hy * hx; + const float w2 = hy * lx; + const float w3 = ly * hx; + const float w4 = ly * lx; +#if defined(NHWC) + const DATA_TYPE data1 = *(__global DATA_TYPE *)tensor3D_offset(input, pz, x_low, y_low); + const DATA_TYPE data2 = *(__global DATA_TYPE *)tensor3D_offset(input, pz, x_high, y_low); + const DATA_TYPE data3 = *(__global DATA_TYPE *)tensor3D_offset(input, pz, x_low, y_high); + const DATA_TYPE data4 = *(__global DATA_TYPE *)tensor3D_offset(input, pz, x_high, y_high); +#else // !defined(NHWC) + const DATA_TYPE data1 = *(__global DATA_TYPE *)tensor3D_offset(input, x_low, y_low, pz); + const DATA_TYPE data2 = *(__global DATA_TYPE *)tensor3D_offset(input, x_high, y_low, pz); + const DATA_TYPE data3 = *(__global DATA_TYPE *)tensor3D_offset(input, x_low, y_high, pz); + const DATA_TYPE data4 = *(__global DATA_TYPE *)tensor3D_offset(input, x_high, y_high, pz); +#endif // defined(NHWC) + const float data1_f32 = dequantize_qasymm8(data1, OFFSET_IN, SCALE_IN); + const float data2_f32 = dequantize_qasymm8(data2, OFFSET_IN, SCALE_IN); + const float data3_f32 = dequantize_qasymm8(data3, OFFSET_IN, SCALE_IN); + const float data4_f32 = dequantize_qasymm8(data4, OFFSET_IN, SCALE_IN); + sum += w1 * data1_f32 + w2 * data2_f32 + w3 * data3_f32 + w4 * data4_f32; + } + } + + const float res_f32 = sum / (grid_size_x * grid_size_y); + return quantize_qasymm8(res_f32, OFFSET_OUT, SCALE_OUT); +} + +/** Performs a roi align function. + * + * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=uchar + * @note Datasize must be passed using -DDATA_SIZE e.g. -DDATA_SIZE=32; + * @note Input dimensions must be passed using -DMAX_DIM_X, -DMAX_DIM_Y and -DMAX_DIM_Z; + * @note Pooled region dimensions must be passed using -DPOOLED_DIM_X and -DPOOLED_DIM_Y; + * @note Spatial scale must be passed using -DSPATIAL_SCALE; + * @note Sampling ratio (i.e., the number of samples in each bin) may be passed using -DSAMPLING_RATIO. If not defined each roi + * will have a default sampling ratio of roi_dims/pooling_dims + * + * @param[in] input_ptr Pointer to the source tensor. Supported data types: QASYMM8 + * @param[in] input_stride_x Stride of the source tensor 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 tensor 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 pooled region of the source tensor as specifed by ROI + * @param[in] rois_ptr Pointer to the ROIs tensor. Layout: { batch_index, x1, y1, x2, y2 }. + * Supported data types: QASYMM16 with 0.125f scale and 0 offset + * @param[in] rois_stride_x Stride of the ROIs tensor in X dimension (in bytes) + * @param[in] rois_step_x Step of the ROIs tensor in X dimension (in bytes) + * @param[in] rois_stride_y Stride of the ROIs tensor in Y dimension (in bytes) + * @param[in] rois_step_y Step of the ROIs tensor in Y dimension (in bytes) + * @param[in] rois_offset_first_element_in_bytes The offset of the first element in the ROIs tensor + * @param[out] output_ptr Pointer to the destination tensor. Supported data types: Supported data types: same as @p input_ptr + * @param[in] output_stride_x Stride of the destination tensor 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 tensor 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 destination 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 tensor + * @param[in] input_stride_w Stride of the source tensor in W dimension (in bytes) + * @param[in] output_stride_w Stride of the destination tensor in W dimension (in bytes) + */ +__kernel void roi_align_layer_quantized( + TENSOR3D_DECLARATION(input), + IMAGE_DECLARATION(rois), + TENSOR3D_DECLARATION(output), + unsigned int input_stride_w, unsigned int output_stride_w) +{ + // Get pixels pointer + Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(input); + Image rois = CONVERT_TO_IMAGE_STRUCT_NO_STEP(rois); + Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(output); + +#if defined(NHWC) + const int px = get_global_id(1); + const int py = get_global_id(2); + const int pw = get_global_id(0); +#else // !defined(NHWC) + const int px = get_global_id(0); + const int py = get_global_id(1); + const int pw = get_global_id(2); +#endif // defined(NHWC) + + // Load roi parameters + // roi is laid out as follows { batch_index, x1, y1, x2, y2 } + const ushort roi_batch = *((__global ushort *)offset(&rois, 0, pw)); + float4 roi = dequantize_qasymm16(vload4(0, (__global ushort *)offset(&rois, 1, pw)), OFFSET_ROIS, SCALE_ROIS); + float2 roi_anchor = roi.s01 * convert_float(SPATIAL_SCALE); + float2 roi_dims = fmax((roi.s23 - roi.s01) * convert_float(SPATIAL_SCALE), 1.f); + + // Calculate pooled region start and end + float2 spatial_indx = (float2)(px, py); + float2 pooled_dims = (float2)(POOLED_DIM_X, POOLED_DIM_Y); + float2 max_spatial_dims = (float2)(MAX_DIM_X, MAX_DIM_Y); + + float2 bin_size = (float2)((roi_dims.s0 / (float)POOLED_DIM_X), (roi_dims.s1 / (float)POOLED_DIM_Y)); + float2 region_start = spatial_indx * bin_size + roi_anchor; + float2 region_end = (spatial_indx + 1) * bin_size + roi_anchor; + + region_start = clamp(region_start, 0, max_spatial_dims); + region_end = clamp(region_end, 0, max_spatial_dims); + +#if defined(SAMPLING_RATIO) + float2 roi_bin_grid = SAMPLING_RATIO; +#else // !defined(SAMPLING_RATIO) + // Note that we subtract EPS_GRID before ceiling. This is to avoid situations where 1.000001 gets ceiled to 2. + float2 roi_bin_grid = ceil(bin_size - EPS_GRID); +#endif // defined(SAMPLING_RATIO) + + // Move input and output pointer across the fourth dimension + input.ptr += roi_batch * input_stride_w; + output.ptr += pw * output_stride_w; + for(int pz = 0; pz < MAX_DIM_Z; ++pz) + { +#if defined(NHWC) + __global DATA_TYPE *_output_ptr = (__global DATA_TYPE *)tensor3D_offset(&output, pz, px, py); +#else // !defined(NHWC) + __global DATA_TYPE *_output_ptr = (__global DATA_TYPE *)tensor3D_offset(&output, px, py, pz); +#endif // defined(NHWC) + *_output_ptr = (__global DATA_TYPE)roi_align_1x1(&input, + region_start.x, + bin_size.x, + roi_bin_grid.x, + region_end.x, + region_start.y, + bin_size.y, + roi_bin_grid.y, + region_end.y, pz); + } +} +#endif // Check for compile time constants diff --git a/src/core/CL/kernels/CLROIAlignLayerKernel.cpp b/src/core/CL/kernels/CLROIAlignLayerKernel.cpp index 50729f2421..134286bae1 100644 --- a/src/core/CL/kernels/CLROIAlignLayerKernel.cpp +++ b/src/core/CL/kernels/CLROIAlignLayerKernel.cpp @@ -45,11 +45,10 @@ namespace Status validate_arguments(const ITensorInfo *input, const ITensorInfo *rois, ITensorInfo *output, const ROIPoolingLayerInfo &pool_info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, rois, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, rois); ARM_COMPUTE_RETURN_ERROR_ON(rois->dimension(0) != 5); ARM_COMPUTE_RETURN_ERROR_ON(rois->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F32, DataType::F16); ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC, DataLayout::NCHW); ARM_COMPUTE_RETURN_ERROR_ON((pool_info.pooled_width() == 0) || (pool_info.pooled_height() == 0)); @@ -59,6 +58,19 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *rois, ITe ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(compute_roi_align_shape(*input, *rois, pool_info), output->tensor_shape()); } + + if(input->data_type() == DataType::QASYMM8) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(rois, 1, DataType::QASYMM16); + + const UniformQuantizationInfo rois_qinfo = rois->quantization_info().uniform(); + ARM_COMPUTE_RETURN_ERROR_ON(rois_qinfo.scale != 0.125f); + ARM_COMPUTE_RETURN_ERROR_ON(rois_qinfo.offset != 0); + } + else + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, rois); + } return Status{}; } @@ -104,9 +116,12 @@ void CLROIAlignLayerKernel::configure(const ICLTensor *input, const ICLTensor *r _rois = rois; _pool_info = pool_info; + const DataType data_type = input->info()->data_type(); + const bool is_qasymm = is_data_type_quantized_asymmetric(data_type); + // Set build options CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)); build_opts.add_option("-DDATA_SIZE=" + get_data_size_from_data_type(input->info()->data_type())); build_opts.add_option("-DMAX_DIM_X=" + support::cpp11::to_string(_input->info()->dimension(get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH)))); build_opts.add_option("-DMAX_DIM_Y=" + support::cpp11::to_string(_input->info()->dimension(get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT)))); @@ -117,9 +132,23 @@ void CLROIAlignLayerKernel::configure(const ICLTensor *input, const ICLTensor *r build_opts.add_option_if(input->info()->data_layout() == DataLayout::NHWC, "-DNHWC"); build_opts.add_option_if(pool_info.sampling_ratio() > 0, "-DSAMPLING_RATIO=" + support::cpp11::to_string(pool_info.sampling_ratio())); + if(is_qasymm) + { + const UniformQuantizationInfo iq_info = input->info()->quantization_info().uniform(); + const UniformQuantizationInfo roisq_info = rois->info()->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = output->info()->quantization_info().uniform(); + + build_opts.add_option("-DOFFSET_IN=" + float_to_string_with_full_precision(iq_info.offset)); + build_opts.add_option("-DSCALE_IN=" + float_to_string_with_full_precision(iq_info.scale)); + build_opts.add_option("-DOFFSET_ROIS=" + float_to_string_with_full_precision(roisq_info.offset)); + build_opts.add_option("-DSCALE_ROIS=" + float_to_string_with_full_precision(roisq_info.scale)); + build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset)); + build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale)); + } + // Create kernel - std::string kernel_name = "roi_align_layer"; - _kernel = static_cast(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); + const std::string kernel_name = (is_qasymm) ? "roi_align_layer_quantized" : "roi_align_layer"; + _kernel = static_cast(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); ICLKernel::configure_internal(win_config.second); } -- cgit v1.2.1