From f1fc9d740590bfcf92778926b4d9725a5361a0b6 Mon Sep 17 00:00:00 2001 From: =?utf8?q?Prasanna=20R/SNAP=20/SRI-Bangalore/Engineer/=EC=82=BC?= =?utf8?q?=EC=84=B1=EC=A0=84=EC=9E=90?= Date: Wed, 2 Jan 2019 07:47:25 +0530 Subject: [PATCH] Add QASYMM8 support for PReLU op (#4079) This patch adds QASYMM8 support for PReLU op. Issue: #3884 Signed-off-by: prasannar --- libs/ARMComputeEx/src/core/CL/CLKernelLibrary.cpp | 5 ++ .../src/core/CL/cl_kernels/prelu_quantized.cl | 88 ++++++++++++++++++++++ .../src/core/CL/kernels/CLPReLUKernel.cpp | 29 +++++-- 3 files changed, 117 insertions(+), 5 deletions(-) create mode 100644 libs/ARMComputeEx/src/core/CL/cl_kernels/prelu_quantized.cl diff --git a/libs/ARMComputeEx/src/core/CL/CLKernelLibrary.cpp b/libs/ARMComputeEx/src/core/CL/CLKernelLibrary.cpp index b6099d4..8d69b34 100644 --- a/libs/ARMComputeEx/src/core/CL/CLKernelLibrary.cpp +++ b/libs/ARMComputeEx/src/core/CL/CLKernelLibrary.cpp @@ -242,6 +242,7 @@ const std::map CLKernelLibraryEx::_kernel_program_map {"pooling_layer_MxN_quantized_nhwc", "pooling_layer_quantized.cl"}, {"pooling_layer_MxN_quantized_nchw", "pooling_layer_quantized.cl"}, {"prelu", "prelu.cl"}, + {"prelu_qasymm8", "prelu_quantized.cl"}, {"quantization_layer", "quantization_layer.cl"}, {"reduce_min_max", "reduce_operation.cl"}, {"reduce_sum_mean", "reduce_operation.cl"}, @@ -403,6 +404,10 @@ const std::map CLKernelLibraryEx::_program_source_map #include "./cl_kernels/prelu.clembed" }, { + "prelu_quantized.cl", +#include "./cl_kernels/prelu_quantized.clembed" + }, + { "reduce_operation.cl", #include "./cl_kernels/reduce_operation.clembed" }, diff --git a/libs/ARMComputeEx/src/core/CL/cl_kernels/prelu_quantized.cl b/libs/ARMComputeEx/src/core/CL/cl_kernels/prelu_quantized.cl new file mode 100644 index 0000000..7e97b7e --- /dev/null +++ b/libs/ARMComputeEx/src/core/CL/cl_kernels/prelu_quantized.cl @@ -0,0 +1,88 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * Copyright (c) 2016-2018 ARM Limited. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include "helpers.h" +#define SUB(x, y) (x) - (y) + +#if defined(OFF_IN1) && defined(OFF_IN2) && defined(OFF_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(VEC_SIZE) + +#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE) +#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE) +#define VEC_UCHAR VEC_DATA_TYPE(uchar, VEC_SIZE) +#define CONVERT_RTE(x, type) (convert_##type##_rte((x))) +#define CONVERT_DOWN(x, type) CONVERT_RTE(x, type) + +/** Returns result of prelu function implemented as below: + * f(input) = alpha * input for input < 0, f(input) = input for input >= 0. + * + * @attention Data type can be passed using the -DDATA_TYPE_IN compile flag, e.g. -DDATA_TYPE_IN=uchar + * @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @note Can only take uchar data types. + * + * @param[in] input1_ptr Pointer to the source image. Supported Data types : QASYMM8 + * @param[in] input1_stride_x Stride of the source image in X dimension (in bytes) + * @param[in] input1_step_x input1_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] input1_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] input1_step_y input1_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] input1_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] input1_step_z input1_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] input1_offset_first_element_in_bytes The offset of the first element in the source image + * + * @param[in] alpha_ptr Pointer to the source image. Supported Data types : QASYMM8 + * @param[in] alpha_stride_x Stride of the source image in X dimension (in bytes) + * @param[in] alpha_step_x input2_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] alpha_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] alpha_step_y input2_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] alpha_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] alpha_step_z input2_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] alpha_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 + */ +__kernel void prelu_qasymm8( + TENSOR3D_DECLARATION(input), + TENSOR3D_DECLARATION(alpha), + TENSOR3D_DECLARATION(output)) +{ + // Get pixels pointer + Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); + Tensor3D alpha = CONVERT_TO_TENSOR3D_STRUCT(alpha); + Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); + + VEC_INT in_a = CONVERT(VLOAD(VEC_SIZE)(0, (__global uchar *)input.ptr), VEC_INT); + VEC_INT in_b = CONVERT(VLOAD(VEC_SIZE)(0, (__global uchar *)alpha.ptr), VEC_INT); + + in_a = SUB(in_a, (VEC_INT)((int)OFF_IN1)); + in_b = SUB(in_b, (VEC_INT)((int)OFF_IN2)); + + const VEC_FLOAT in1f32 = CONVERT(in_a, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN1); + const VEC_FLOAT in2f32 = CONVERT(in_b, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN2); + const VEC_FLOAT outf32 = in1f32 < 0 ? in1f32 * in2f32 : in1f32; + const VEC_FLOAT qresf32 = outf32 / ((VEC_FLOAT)(float)SCALE_OUT) + ((VEC_FLOAT)((float)OFF_OUT)); + const VEC_UCHAR res = CONVERT_SAT(CONVERT_DOWN(qresf32, VEC_INT), VEC_UCHAR); + + VSTORE(VEC_SIZE) + (res, 0, (__global uchar *)output.ptr); +} + +#endif // defined(OFF_IN1) && defined(OFF_IN2) && defined(OFF_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(VEC_SIZE) diff --git a/libs/ARMComputeEx/src/core/CL/kernels/CLPReLUKernel.cpp b/libs/ARMComputeEx/src/core/CL/kernels/CLPReLUKernel.cpp index 5f0cbfe..ae4b9ec 100644 --- a/libs/ARMComputeEx/src/core/CL/kernels/CLPReLUKernel.cpp +++ b/libs/ARMComputeEx/src/core/CL/kernels/CLPReLUKernel.cpp @@ -31,10 +31,10 @@ Status validate_info(const ITensorInfo *input, const ITensorInfo *alpha, const I const TensorShape &out_shape = TensorShape::broadcast_shape(input->tensor_shape(), alpha->tensor_shape()); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QASYMM8, - "Not support QASYMM8, yet"); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(alpha, 1, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32, + DataType::QASYMM8); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(alpha, 1, DataType::F16, DataType::F32, + DataType::QASYMM8); ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); @@ -63,11 +63,30 @@ void CLPReLUKernel::configure(const ICLTensor *input, const ICLTensor *alpha, IC _output = output; // Create kernel + std::string kernel_name = "prelu"; std::set build_opts; build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()))); build_opts.emplace( ("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration))); - _kernel = static_cast(CLKernelLibraryEx::get().create_kernel("prelu", build_opts)); + + if (is_data_type_quantized_asymmetric(input->info()->data_type())) + { + build_opts.emplace("-DOFF_IN1=" + + support::cpp11::to_string(input->info()->quantization_info().offset)); + build_opts.emplace("-DOFF_IN2=" + + support::cpp11::to_string(alpha->info()->quantization_info().offset)); + build_opts.emplace("-DOFF_OUT=" + + support::cpp11::to_string(output->info()->quantization_info().offset)); + build_opts.emplace("-DSCALE_IN1=" + + support::cpp11::to_string(input->info()->quantization_info().scale)); + build_opts.emplace("-DSCALE_IN2=" + + support::cpp11::to_string(alpha->info()->quantization_info().scale)); + build_opts.emplace("-DSCALE_OUT=" + + support::cpp11::to_string(output->info()->quantization_info().scale)); + kernel_name += "_qasymm8"; + } + _kernel = + static_cast(CLKernelLibraryEx::get().create_kernel(kernel_name, build_opts)); const std::pair broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*input->info(), *alpha->info()); -- 2.7.4