COMPMID-2599: Implement a new and generic depthwise convolution on OpenCL (Fp32/FP16-NHWC)
Part 1
Change-Id: I5e1d27a7006199e9229e455a1df9bfc2ed4e8341
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1898
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 4f017b7..2f748de 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -221,6 +221,7 @@
{ "depthwise_convolution_3x3_f16", "depthwise_convolution.cl" },
{ "depthwise_convolution_3x3_nhwc", "depthwise_convolution.cl" },
{ "depthwise_convolution_3x3_nhwc_stride1", "depthwise_convolution.cl" },
+ { "dwc_MxN_native_fp_nhwc", "depthwise_convolution.cl" },
{ "dwc_3x3_native_qasymm8_nchw", "depthwise_convolution_quantized.cl" },
{ "dwc_3x3_native_qasymm8_dot8_nchw", "depthwise_convolution_quantized.cl" },
{ "dwc_3x3_reshaped_qasymm8_nhwc", "depthwise_convolution_quantized.cl" },
diff --git a/src/core/CL/cl_kernels/depthwise_convolution.cl b/src/core/CL/cl_kernels/depthwise_convolution.cl
index fb4a0fc..1b2f5cc 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution.cl
@@ -21,7 +21,6 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-
#include "helpers.h"
#include "activation_float_helpers.h"
@@ -1479,7 +1478,7 @@
//3x3 Convolution of elements starting in 0th row
pixels0 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src.stride_x, src.stride_y, 0, weights_addr, weights_stride_y);
//3x3 Convolution of elements starting in 2nd row
- pixels1 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src.stride_x, src.stride_y, 2, weights_addr, weights_stride_y);
+ pixels1 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src.stride_x, src.stride_y, 2, weights_addr, weights_stride_y);
#endif /* DILATION_X==1 && DILATION_Y==1 */
#ifdef HAS_BIAS
@@ -1492,6 +1491,153 @@
}
#endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(DEPTH_MULTIPLIER) && defined(DST_CHANNELS) && defined(IS_F16)
+#if defined(SRC_DIM1) && defined(SRC_DIM2) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(N0) && defined(DATA_TYPE) && defined(DILATION_X) && defined(DILATION_Y) && defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(CONV_PAD_LEFT) && defined(CONV_PAD_TOP)
+/** This function computes the depthwise convolution for NHWC data layout. This kernel assumes that the weights tensor is NOT reshaped
+ *
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
+ * @note The number of elements processed must be passed at compile time using -DN0 (e.g. -DN0=2)
+ * @note The depth multiplier must be passed at compile time using -DDEPTH_MULTIPLIER (e.g. -DDEPTH_MULTIPLIER=1)
+ * @note The first dimension of the input tensor must be passed at compile time using -DSRC_DIM1 (e.g. -DSRC_DIM1=112)
+ * @note The second dimension of the input tensor must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM2=80)
+ * @note The kernel width must be passed at compile time using -DKERNEL_WIDTH (e.g. -DKERNEL_WIDTH=5)
+ * @note The kernel height must be passed at compile time using -DKERNEL_HEIGHT (e.g. -DKERNEL_HEIGHT=5)
+ * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1)
+ * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1)
+ * @note The convolution stride along the width must be passed at compile time using -DCONV_STRIDE_X (e.g. -DCONV_STRIDE_Y=X)
+ * @note The convolution stride along the height must be passed at compile time using -DCONV_STRIDE_Y (e.g. -DCONV_STRIDE_Y=1)
+ * @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu
+ * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively
+ *
+ * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32
+ * @param[in] src_stride_x Stride of the source tensor 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 tensor 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 tensor in Z dimension (in bytes)
+ * @param[in] src_step_z src_stride_y * number of elements along Z processed per workitem(in bytes)
+ * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes)
+ * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as src_ptr
+ * @param[in] dst_stride_x Stride of the destination tensor 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_stride_y Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes)
+ * @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: F16/F32
+ * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
+ * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
+ * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: same as src_ptr
+ * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes)
+ * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
+ */
+__kernel void dwc_MxN_native_fp_nhwc(
+ TENSOR4D_DECLARATION(src),
+ TENSOR4D_DECLARATION(dst),
+ TENSOR3D_DECLARATION(weights),
+#if defined(HAS_BIAS)
+ VECTOR_DECLARATION(biases)
+#endif // defined(HAS_BIAS)
+)
+{
+ int x = get_global_id(0); // channels
+ int y = get_global_id(1); // spatial coordinate x
+#if defined(DST_DEPTH)
+ int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y
+ int b = get_global_id(2) / (int)DST_DEPTH; // batch
+#else // defined(DST_DEPTH)
+ int z = get_global_id(2); // spatial coordinate y
+#endif // defined(DST_DEPTH)
+
+ __global uchar *s_addr = src_ptr + src_offset_first_element_in_bytes +
+ x * sizeof(DATA_TYPE) * (int)N0;
+
+ __global uchar *d_addr = dst_ptr + dst_offset_first_element_in_bytes +
+ x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0 +
+ y * dst_stride_y +
+ z * dst_stride_z;
+
+ __global uchar *w_addr = weights_ptr + weights_offset_first_element_in_bytes +
+ x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0;
+
+#if defined(HAS_BIAS)
+ __global uchar *b_addr = biases_ptr + biases_offset_first_element_in_bytes +
+ x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0;
+#endif // defined(HAS_BIAS)
+
+#if defined(DST_DEPTH)
+ s_addr += b * src_stride_w;
+ d_addr += b * dst_stride_w;
+#endif // defined(DST_DEPTH)
+
+ for(int d = 0; d < (int)DEPTH_MULTIPLIER; ++d)
+ {
+ // Each work-item computes N0x1x1 elements
+ VEC_DATA_TYPE(DATA_TYPE, N0)
+ res = 0;
+
+ int x_coord = y * CONV_STRIDE_X - (int)CONV_PAD_LEFT;
+ int y_coord = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP;
+
+ for(int yk = 0; yk < KERNEL_HEIGHT; ++yk)
+ {
+ if(y_coord >= 0 && y_coord < SRC_DIM2)
+ {
+ int x_coord_tmp = x_coord;
+
+ for(int xk = 0; xk < KERNEL_WIDTH; ++xk)
+ {
+ if(x_coord_tmp >= 0 && x_coord_tmp < SRC_DIM1)
+ {
+ int s_offset = x_coord_tmp * (int)src_stride_y + y_coord * (int)src_stride_z;
+ int w_offset = xk * weights_stride_y + yk * weights_stride_z;
+
+ // Load input and weights values
+ VEC_DATA_TYPE(DATA_TYPE, N0)
+ i = VLOAD(N0)(0, (__global DATA_TYPE *)(s_addr + s_offset));
+ VEC_DATA_TYPE(DATA_TYPE, N0)
+ w = VLOAD(N0)(0, (__global DATA_TYPE *)(w_addr + w_offset));
+
+#if GPU_ARCH == GPU_ARCH_MIDGARD
+ res += i * w;
+#else // GPU_ARCH == GPU_ARCH_MIDGARD
+ res = fma(i, w, res);
+#endif // GPU_ARCH == GPU_ARCH_MIDGARD
+ }
+ x_coord_tmp += DILATION_X;
+ }
+ }
+ y_coord += DILATION_Y;
+ }
+
+#if defined(HAS_BIAS)
+ res += VLOAD(N0)(0, (__global DATA_TYPE *)(b_addr));
+#endif // defined(HAS_BIAS)
+
+ res = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, res, A_VAL, B_VAL);
+
+ VSTORE(N0)
+ (res, 0, (__global DATA_TYPE *)(d_addr));
+
+ w_addr += sizeof(DATA_TYPE);
+ d_addr += sizeof(DATA_TYPE);
+#if defined(HAS_BIAS)
+ b_addr += sizeof(DATA_TYPE);
+#endif // defined(HAS_BIAS)
+ }
+}
+#endif // defined(SRC_DIM1) && defined(SRC_DIM2) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defiend(N0) && defined(DATA_TYPE) && defined(DILATION_X) && defined(DILATION_Y) && defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(CONV_PAD_LEFT) && defined(CONV_PAD_TOP)
+
#if defined(VEC_SIZE) && defined(SRC_DIM_2) && defined(CONV_PAD_TOP) && defined(CONV_PAD_LEFT) && defined(DATA_TYPE)
#if DATA_TYPE != float || DATA_TYPE != half
@@ -1501,6 +1647,7 @@
#define VEC_FLOAT VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
#if defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y)
+
/** This function computes the depthwise convolution for NHWC data layout when the stride along the width or height is not 1.
*
* @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
diff --git a/src/core/CL/cl_kernels/helpers.h b/src/core/CL/cl_kernels/helpers.h
index 756c906..f501077 100644
--- a/src/core/CL/cl_kernels/helpers.h
+++ b/src/core/CL/cl_kernels/helpers.h
@@ -68,6 +68,7 @@
#define double1 double
#define vload1(OFFSET, PTR) *(OFFSET + PTR)
+#define vstore1(DATA, OFFSET, PTR) *(OFFSET + PTR) = DATA
#define VEC_DATA_TYPE_STR(type, size) type##size
#define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size)
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
new file mode 100644
index 0000000..b34c261
--- /dev/null
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
@@ -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 "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h"
+
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
+#include "arm_compute/core/CL/ICLKernel.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#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"
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const DWCWeightsKernelInfo &dwc_weights_info,
+ const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
+{
+ ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+ ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier > 1 && dwc_weights_info.n0 != 1);
+ ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1);
+ ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().second < 1);
+ ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1));
+ const size_t idx_c = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
+ ARM_COMPUTE_UNUSED(idx_c);
+ ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_c) != (input->dimension(idx_c) * depth_multiplier));
+
+ if(biases != nullptr)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0));
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
+ ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
+ }
+
+ if(output->total_size() != 0)
+ {
+ const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
+ }
+
+ return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *output, const DWCWeightsKernelInfo &dwc_weights_info,
+ const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
+{
+ // Get convolved dimensions
+ const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation);
+
+ auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape).set_quantization_info(output->quantization_info()));
+
+ const unsigned int n0 = dwc_weights_info.n0;
+
+ // Configure kernel window
+ Window win = calculate_max_window(*output, Steps(n0));
+
+ // The following access windows are only valid in case of NHWC and because n0 must unit in case depth_multiplier > 1
+ AccessWindowHorizontal input_access(input, 0, n0);
+ AccessWindowHorizontal weights_access(weights, 0, n0);
+ AccessWindowHorizontal output_access(output, 0, n0);
+
+ bool window_changed = false;
+
+ if(bias != nullptr)
+ {
+ AccessWindowHorizontal bias_access(bias, 0, n0);
+ window_changed = update_window_and_padding(win, input_access, weights_access, bias_access, output_access);
+ }
+ else
+ {
+ window_changed = update_window_and_padding(win, input_access, weights_access, output_access);
+ }
+
+ output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+
+ Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+ return std::make_pair(err, win);
+}
+} // namespace
+
+CLDepthwiseConvolutionLayerNativeKernel::CLDepthwiseConvolutionLayerNativeKernel()
+ : _input(nullptr), _weights(nullptr), _biases(nullptr), _output(nullptr), _depth_multiplier(1)
+{
+}
+
+void CLDepthwiseConvolutionLayerNativeKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCWeightsKernelInfo &dwc_weights_info,
+ const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
+{
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
+ ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), dwc_weights_info, dwc_info, conv_info, depth_multiplier,
+ dilation));
+
+ auto win_config = validate_and_configure_window(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), dwc_weights_info, dwc_info, conv_info, depth_multiplier,
+ dilation);
+ ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+
+ _input = input;
+ _output = output;
+ _weights = weights;
+ _biases = biases;
+ _depth_multiplier = depth_multiplier;
+
+ const size_t idx_w = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
+ const size_t idx_h = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
+ const size_t weights_width = weights->info()->dimension(idx_w);
+ const size_t weights_height = weights->info()->dimension(idx_h);
+
+ CLBuildOptions build_opts;
+ build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
+ build_opts.add_option_if(_input->info()->tensor_shape().total_size_upper(3) > 1, "-DDST_DEPTH=" + support::cpp11::to_string(static_cast<int>(_output->info()->dimension(2))));
+ build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(_input->info()->data_type()));
+ build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(dwc_info.activation_info.activation())));
+ build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(depth_multiplier));
+ build_opts.add_option("-DN0=" + support::cpp11::to_string(dwc_weights_info.n0));
+ build_opts.add_option("-DSRC_DIM1=" + support::cpp11::to_string(_input->info()->dimension(1)));
+ build_opts.add_option("-DSRC_DIM2=" + support::cpp11::to_string(_input->info()->dimension(2)));
+ build_opts.add_option("-DKERNEL_WIDTH=" + support::cpp11::to_string(weights_width));
+ build_opts.add_option("-DKERNEL_HEIGHT=" + support::cpp11::to_string(weights_height));
+ build_opts.add_option("-DCONV_PAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
+ build_opts.add_option("-DCONV_PAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
+ build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(conv_info.stride().first));
+ build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(conv_info.stride().second));
+ build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
+ build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
+ build_opts.add_option_if(dwc_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(dwc_info.activation_info.a()));
+ build_opts.add_option_if(dwc_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(dwc_info.activation_info.b()));
+
+ std::string kernel_name("dwc_MxN_native_fp_nhwc");
+
+ ICLKernel::configure_internal(win_config.second);
+ _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
+
+ // Set config_id for enabling LWS tuning
+ _config_id = kernel_name;
+ _config_id += "_";
+ _config_id += support::cpp11::to_string(input->info()->dimension(0));
+ _config_id += "_";
+ _config_id += support::cpp11::to_string(input->info()->dimension(1));
+ _config_id += "_";
+ _config_id += support::cpp11::to_string(input->info()->dimension(2));
+ _config_id += "_";
+ _config_id += support::cpp11::to_string(output->info()->dimension(0));
+ _config_id += "_";
+ _config_id += support::cpp11::to_string(output->info()->dimension(1));
+ _config_id += "_";
+ _config_id += support::cpp11::to_string(output->info()->dimension(2));
+ _config_id += "_";
+ _config_id += string_from_data_type(input->info()->data_type());
+}
+
+Status CLDepthwiseConvolutionLayerNativeKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+ const DWCWeightsKernelInfo &dwc_weights_info, const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
+{
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation));
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(),
+ biases != nullptr ? biases->clone().get() : nullptr,
+ output->clone().get(), dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation)
+ .first);
+
+ return Status{};
+}
+
+void CLDepthwiseConvolutionLayerNativeKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+ ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+ ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
+
+ // Collapse window
+ Window window_collapsed = window.collapse(ICLKernel::window(), Window::DimZ);
+ Window slice_in = window.first_slice_window_4D();
+ Window slice_out = window_collapsed.first_slice_window_4D();
+
+ if(_depth_multiplier != 1)
+ {
+ ARM_COMPUTE_ERROR_ON(slice_out.x().step() != 1);
+ slice_out.set(Window::DimX, Window::Dimension(0, _input->info()->tensor_shape()[0], 1));
+ }
+
+ if(_biases != nullptr)
+ {
+ unsigned int idx = 2 * num_arguments_per_4D_tensor() + num_arguments_per_3D_tensor();
+ add_1D_tensor_argument(idx, _biases, slice_in);
+ }
+
+ do
+ {
+ unsigned int idx = 0;
+ add_4D_tensor_argument(idx, _input, slice_in);
+ add_4D_tensor_argument(idx, _output, slice_out);
+ add_3D_tensor_argument(idx, _weights, slice_out);
+ enqueue(queue, *this, slice_out, lws_hint());
+ }
+ while(window_collapsed.slide_window_slice_4D(slice_out) && window.slide_window_slice_4D(slice_in));
+}
+} // namespace arm_compute