diff options
Diffstat (limited to 'src/dynamic_fusion/sketch/gpu/template_writer/cl')
4 files changed, 499 insertions, 39 deletions
diff --git a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDepthwiseConv2d.cpp b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDepthwiseConv2d.cpp new file mode 100644 index 0000000000..389bd5c65f --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDepthwiseConv2d.cpp @@ -0,0 +1,378 @@ +/* + * Copyright (c) 2022 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 "ClTemplateDepthwiseConv2d.h" + +#include "src/core/helpers/WindowHelpers.h" +#include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +ClTemplateDepthwiseConv2d::ClTemplateDepthwiseConv2d(ComponentId id, + const ArgumentPack<ITensorInfo> &tensors, + const Attributes &attributes, + const Settings &settings) + : IGpuTemplateComponentWriter{ id, tensors }, + _src{}, + _weight{}, + _bias{}, + _dst{}, + _attributes{ attributes }, + _settings{ settings } +{ + _src = this->tensors().get_const_tensor(TensorType::ACL_SRC_0); + _weight = this->tensors().get_const_tensor(TensorType::ACL_SRC_1); + if(this->tensors().get_const_tensor(TensorType::ACL_SRC_2)) + { + _bias = this->tensors().get_const_tensor(TensorType::ACL_SRC_2); + } + _dst = this->tensors().get_const_tensor(TensorType::ACL_DST_0); + ARM_COMPUTE_ERROR_ON_NULLPTR(_src, _weight, _dst); +} + +std::string ClTemplateDepthwiseConv2d::get_name() const +{ + return "depthwise_conv2d"; +} + +std::string ClTemplateDepthwiseConv2d::get_component_code(const ComponentGroup &comp_group) const +{ + ARM_COMPUTE_UNUSED(comp_group); + + constexpr int height_idx = 2; // Data Layout is NHWC + + std::string code = R"_( +//------------------ START KERNEL {{meta_kernel_id}} --------------------- +// IN_0(src) {{src}} +// IN_1(wei) {{weight}} +)_"; + + if(_bias != nullptr && _bias->has_valid_id()) + { + code += R"_( +// IN_1(bia) {{bias}} +)_"; + } + + code += R"_( +// OUT(dst, accum) {{dst}} + +TILE({{ACC_DATA_TYPE}}, M0, N0, {{dst}}); +TILE(uint, M0, 1, g_dst_indirect_y); + +{ +#define _IWEI_WIDTH {{WEI_WIDTH}} +#define _IWEI_HEIGHT {{WEI_HEIGHT}} +#define _IDST_WIDTH {{arg_dst}}_w +#define _IDST_HEIGHT {{arg_dst}}_h +#define _IM0_A M0_A +#define _IN0_A N0_A +#define _IM0_B _IWEI_WIDTH +#define _IN0_B N0 +#define _IBOUNDARY_CHECK (!((_IWEI_WIDTH == 1 && _IWEI_HEIGHT == 1 && {{PAD_LEFT}} == 0 && {{PAD_TOP}} == 0 && M0 == 1))) +)_"; + + code += R"_( + const int yo = g_ind_2 % {{arg_dst}}_h; + const int bout = g_ind_2 / {{arg_dst}}_h; +)_"; + + code += R"_( + + int xi = g_ind_1 * {{STRIDE_X}}; + int yi = yo * {{STRIDE_Y}}; + xi -= {{PAD_LEFT}}; + yi -= {{PAD_TOP}}; + + LOOP_UNROLLING(int, i, 0, 1, M0, + { + {{dst}}[i].v = 0; + }) +)_"; + + if(_weight->dimension(height_idx) < 5) + { + code += R"_( + LOOP_UNROLLING(int, yk, 0, 1, _IWEI_HEIGHT, +)_"; + } + else + { + code += R"_( + for(int yk = 0; yk < _IWEI_HEIGHT; ++yk) +)_"; + } + + code += R"_( + { + TILE({{SRC_DATA_TYPE}}, _IM0_A, _IN0_A, a); + + LOOP_UNROLLING(int, i, 0, 1, _IM0_A, + { + a[i].v = 0; + }) + + T_LOAD_NHWC_WITH_DILATION({{SRC_DATA_TYPE}}, 1, _IM0_A, _IN0_A, {{SRC_TENSOR_TYPE}}, {{src}}, bout, yi + yk * {{DILATION_Y}}, xi, (g_ind_0 / {{DEPTH_MULTIPLIER}}), {{src}}_w, {{src}}_h, {{DILATION_X}}, 1, _IBOUNDARY_CHECK, a); + + TILE({{WEI_DATA_TYPE}}, _IM0_B, _IN0_B, b); + + T_LOAD({{WEI_DATA_TYPE}}, _IM0_B, _IN0_B, {{WEI_TENSOR_TYPE}}, {{weight}}, g_ind_0, yk * _IM0_B, 1, {{weight}}_stride_y, b); + + LOOP_UNROLLING(int, m0, 0, 1, M0, + { + LOOP_UNROLLING(int, xk, 0, 1, _IWEI_WIDTH, + { +)_"; + + if(!_settings.is_fma_available()) + { + code += R"_( + {{dst}}[m0].v += a[xk + m0].v * b[xk].v; +)_"; + } + else + { + code += R"_( + {{dst}}[m0].v = fma(a[xk + m0].v, b[xk].v, {{dst}}[m0].v); +)_"; + } + + code += R"_( + }) + }) + } +)_"; + + if(_weight->dimension(height_idx) < 5) + { + code += R"_( + ) +)_"; + } + + if(_bias && _bias->has_valid_id()) + { + code += R"_( + TILE({{BIA_DATA_TYPE}}, 1, N0, {{bias}}); + + T_LOAD({{BIA_DATA_TYPE}}, 1, N0, BUFFER, {{bias}}, g_ind_0, 0, 0, 0, {{bias}}); + + T_ELTWISE_BROADCAST_ADD_X({{ACC_DATA_TYPE}}, M0, N0, {{dst}}, {{bias}}, {{dst}}); +)_"; + } + + code += R"_( + LOOP_UNROLLING(int, i, 0, 1, M0, + { + g_dst_indirect_y[i].v = (uint)min((int)(g_ind_1 + i), (int)({{arg_dst}}_w) - 1); + g_dst_indirect_y[i].v += (int)(g_ind_2 % {{arg_dst}}_h) * (int)({{arg_dst}}_w); + g_dst_indirect_y[i].v += (int)(g_ind_2 / {{arg_dst}}_h) * (int)({{arg_dst}}_w * {{arg_dst}}_h); + }) +} +//------------------ END KERNEL {{meta_kernel_id}} --------------------- +)_"; + + return code; +} + +void ClTemplateDepthwiseConv2d::declare_variables(GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const +{ + const GpuKernelArgumentInfo::Type input_type = _settings.export_input_to_cl_image() ? + GpuKernelArgumentInfo::Type::Tensor_4D_t_Image : + GpuKernelArgumentInfo::Type::Tensor_4D_t_Buffer; + + vtable.declare_variable( + _src, + GpuKernelArgumentInfo(input_type), + comp_group.is_intermediate_tensor(_src), + "src"); + + const GpuKernelArgumentInfo::Type weight_type = _settings.export_weights_to_cl_image() ? + GpuKernelArgumentInfo::Type::Tensor_4D_t_Image : + GpuKernelArgumentInfo::Type::Tensor_4D_t_Buffer; + + vtable.declare_variable( + _weight, + GpuKernelArgumentInfo(weight_type), + comp_group.is_intermediate_tensor(_weight), + "weight"); + + if(_bias != nullptr && _bias->has_valid_id()) // optional bias + { + vtable.declare_variable( + _bias, + GpuKernelArgumentInfo(GpuKernelArgumentInfo::Type::Vector), + comp_group.is_intermediate_tensor(_bias), + "bias"); + } + vtable.declare_variable( + _dst, + GpuKernelArgumentInfo(GpuKernelArgumentInfo::Type::Tensor_4D_t_Buffer), + comp_group.is_intermediate_tensor(_dst), + "dst"); +} + +TagLUT ClTemplateDepthwiseConv2d::get_tag_lut(const GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const +{ + TagLUT lut{}; + + // Arguments and global shared variables + lut["src"] = vtable.get_variable(_src); + lut["weight"] = vtable.get_variable(_weight); + + if(_bias != nullptr && _bias->has_valid_id()) // optional bias + { + lut["bias"] = vtable.get_variable(_bias); + lut["BIA_DATA_TYPE"] = get_cl_type_from_data_type(_bias->data_type()); + } + lut["dst"] = vtable.get_variable(_dst); + + const auto dst_argument = vtable.get_variable(comp_group.get_dst_tensors()[0]); + lut["arg_dst"] = dst_argument.uniq_name; + + // Local build options + lut["meta_kernel_id"] = id(); + lut["ACC_DATA_TYPE"] = _src->data_type(); + lut["SRC_DATA_TYPE"] = _src->data_type(); + lut["WEI_DATA_TYPE"] = _weight->data_type(); + + switch(vtable.get_variable(_src).kernel_argument_info.type) + { + case GpuKernelArgumentInfo::Type::Image_Export_To_ClImage2D: + case GpuKernelArgumentInfo::Type::Image_3D_Export_To_ClImage2D: + case GpuKernelArgumentInfo::Type::Tensor_4D_t_Image: + lut["SRC_TENSOR_TYPE"] = "IMAGE"; + break; + default: + lut["SRC_TENSOR_TYPE"] = "BUFFER"; + break; + } + + switch(vtable.get_variable(_weight).kernel_argument_info.type) + { + case GpuKernelArgumentInfo::Type::Image_Export_To_ClImage2D: + case GpuKernelArgumentInfo::Type::Image_3D_Export_To_ClImage2D: + case GpuKernelArgumentInfo::Type::Tensor_4D_t_Image: + lut["WEI_TENSOR_TYPE"] = "IMAGE"; + break; + default: + lut["WEI_TENSOR_TYPE"] = "BUFFER"; + break; + } + + // Data Layout is NHWC + constexpr int width_idx = 1; + constexpr int height_idx = 2; + + lut["WEI_WIDTH"] = _weight->dimension(width_idx); + lut["WEI_HEIGHT"] = _weight->dimension(height_idx); + + lut["STRIDE_X"] = _attributes.stride().x(); + lut["STRIDE_Y"] = _attributes.stride().y(); + + lut["PAD_LEFT"] = _attributes.pad().left; + lut["PAD_TOP"] = _attributes.pad().top; + + lut["DILATION_X"] = _attributes.dilation().x(); + lut["DILATION_Y"] = _attributes.dilation().y(); + + lut["DEPTH_MULTIPLIER"] = _attributes.depth_multiplier(); + + return lut; +} + +CLBuildOptions ClTemplateDepthwiseConv2d::get_build_options(const ComponentGroup &comp_group) const +{ + ARM_COMPUTE_UNUSED(comp_group); + + constexpr unsigned int width_idx = 1; // Data Layout is NHWC + + const unsigned int n0 = _settings.n0(); + const unsigned int m0 = _settings.m0(); + const unsigned int m0_a = _weight->dimension(width_idx) + m0 - 1; + const unsigned int n0_a = _attributes.depth_multiplier() > 1 ? 1 : n0; + const unsigned int partial_store_n0 = _dst->dimension(0) % n0; + + CLBuildOptions build_opts{}; + + if(_settings.fast_relaxed_math()) + { + build_opts.add_option("-cl-fast-relaxed-math"); + } + else + { + // -cl-fast-relaxed-math also sets -cl-finite-math-only and -cl-unsafe-math-optimizations + // to disable -cl-finite-math-only, we only include -cl-unsafe-math-optimizations + build_opts.add_option("-cl-unsafe-math-optimizations"); + } + + build_opts.add_option("-DN0=" + support::cpp11::to_string(n0)); + build_opts.add_option("-DM0=" + support::cpp11::to_string(m0)); + build_opts.add_option("-DN0_A=" + support::cpp11::to_string(n0_a)); + build_opts.add_option("-DM0_A=" + support::cpp11::to_string(m0_a)); + build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0)); + + return build_opts; +} + +std::string ClTemplateDepthwiseConv2d::get_config_id() const +{ + std::string config_id{}; + + config_id += support::cpp11::to_string(_src->dimension(0)); + config_id += "_"; + config_id += support::cpp11::to_string(_src->dimension(1)); + config_id += "_"; + config_id += support::cpp11::to_string(_src->dimension(2)); + config_id += "_"; + config_id += support::cpp11::to_string(_dst->dimension(0)); + config_id += "_"; + config_id += support::cpp11::to_string(_dst->dimension(1)); + config_id += "_"; + config_id += support::cpp11::to_string(_dst->dimension(2)); + config_id += "_"; + config_id += string_from_data_type(_src->data_type()); + + return config_id; +} + +std::set<std::string> ClTemplateDepthwiseConv2d::get_headers_list() const +{ + return std::set<std::string>{ "helpers.h", "tile_helpers.h" }; +} + +Window ClTemplateDepthwiseConv2d::get_window() const +{ + ARM_COMPUTE_ERROR_ON_MSG(_dst->tensor_shape().total_size() == 0U, "Destination tensor is not initialized"); + + Window win = calculate_max_window(*_dst, Steps(_settings.n0(), _settings.m0())); + return win.collapse(win, Window::DimZ); +} + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute diff --git a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDepthwiseConv2d.h b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDepthwiseConv2d.h new file mode 100644 index 0000000000..84b689ef64 --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDepthwiseConv2d.h @@ -0,0 +1,111 @@ +/* + * Copyright (c) 2022 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. + */ +#ifndef SRC_DYNAMIC_FUSION_SKETCH_GPU_TEMPLATE_WRITER_CL_CLTEMPLATEDEPTHWISECONV2D +#define SRC_DYNAMIC_FUSION_SKETCH_GPU_TEMPLATE_WRITER_CL_CLTEMPLATEDEPTHWISECONV2D + +#include "arm_compute/dynamic_fusion/sketch/attributes/DepthwiseConv2dAttributes.h" +#include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDepthwiseConv2d.h" +#include "src/dynamic_fusion/sketch/gpu/template_writer/IGpuTemplateComponentWriter.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +class ClTemplateDepthwiseConv2d final : public IGpuTemplateComponentWriter +{ +public: + using Attributes = ClComponentDepthwiseConv2d::Attributes; + using Settings = ClComponentDepthwiseConv2d::Settings; + /** Constructor + * + * Similar to @ref ClComponentDepthwiseConv2d::validate() + * + * @param[in] id Component id + * @param[in] tensors Tensor arguments to the components + * @param[in] attributes Component attributes + * @param[in] settings Component settings + */ + ClTemplateDepthwiseConv2d(ComponentId id, + const ArgumentPack<ITensorInfo> &tensors, + const Attributes &attributes, + const Settings &settings); + /** Prevent instances of this class from being copy constructed */ + ClTemplateDepthwiseConv2d(const ClTemplateDepthwiseConv2d &depthwise_conv2d) = delete; + /** Prevent instances of this class from being copied */ + ClTemplateDepthwiseConv2d &operator=(const ClTemplateDepthwiseConv2d &depthwise_conv2d) = delete; + /** Allow instances of this class to be move constructed */ + ClTemplateDepthwiseConv2d(ClTemplateDepthwiseConv2d &&depthwise_conv2d) = default; + /** Allow instances of this class to be moved */ + ClTemplateDepthwiseConv2d &operator=(ClTemplateDepthwiseConv2d &&depthwise_conv2d) = default; + /** Generate kernel component name */ + std::string get_name() const override; + /** Generate kernel component code template + * + * @param[in] comp_group Component group of which the component is a part of + * + * @return std::string Component code + */ + std::string get_component_code(const ComponentGroup &comp_group) const override; + /** Declare all variables used by the component in the @p vtable + * + * @param[out] vtable Variable table + * @param[in] comp_group Component group of which the component is a part of + */ + void declare_variables(GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const override; + /** Generate the tag look-up table used to instantiate the component code. + * + * @param[in] vtable Variable table + * @param[in] comp_group Component group of which the component is a part of + * + * @return TagLUT Tag lookup table + */ + TagLUT get_tag_lut(const GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const override; + /** Generate the build options used in the component + * + * @param[in] comp_group Component group of which the component is a part of + * + * @return CLBuildOptions Build options + */ + CLBuildOptions get_build_options(const ComponentGroup &comp_group) const override; + /** Generate the component config id string used for tuning */ + std::string get_config_id() const override; + /** Generate the header list used in the component */ + std::set<std::string> get_headers_list() const override; + /** Generate the execution window for the component */ + Window get_window() const override; + +private: + const ITensorInfo *_src; + const ITensorInfo *_weight; + const ITensorInfo *_bias; + const ITensorInfo *_dst; + Attributes _attributes; + Settings _settings; +}; +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute +#endif /* SRC_DYNAMIC_FUSION_SKETCH_GPU_TEMPLATE_WRITER_CL_CLTEMPLATEDEPTHWISECONV2D */ diff --git a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp index 870de64eb8..7ad7dd69f0 100644 --- a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp +++ b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp @@ -86,11 +86,10 @@ std::string ClTemplateDirectConv2d::get_component_code(const ComponentGroup &com code += R"_( // OUT(dst, accum) {{dst}} -// Initialize the accumulators TILE({{ACC_DATA_TYPE}}, M0, N0, {{dst}}); +TILE(uint, M0, 1, g_dst_indirect_y); + { - // All the tensor dimensions are passed at compile time. - // In case of dynamic tensor support, the following dimensions should be passed as function argument. #define _IWEI_WIDTH {{WEI_WIDTH}} #define _IWEI_HEIGHT {{WEI_HEIGHT}} #define _ISRC_WIDTH {{src}}_w @@ -101,8 +100,6 @@ TILE({{ACC_DATA_TYPE}}, M0, N0, {{dst}}); #define _IDST_CHANNELS {{arg_dst}}_c #define _IY_MULTIPLIER (_IWEI_WIDTH * _IWEI_HEIGHT) - // .v = access the whole vector (OpenCL vector) - // .s[x] = access the vector element at position x (scalar access) TILE(int, M0, 1, xi); TILE(int, M0, 1, yi); @@ -132,7 +129,6 @@ TILE({{ACC_DATA_TYPE}}, M0, N0, {{dst}}); TILE({{SRC_DATA_TYPE}}, M0, K0, a); TILE({{WEI_DATA_TYPE}}, N0, K0, b); - // Initialize tiles LOOP_UNROLLING(int, i, 0, 1, M0, { a[i].v = {{ZERO_VALUE}}; @@ -143,32 +139,24 @@ TILE({{ACC_DATA_TYPE}}, M0, N0, {{dst}}); b[i].v = {{ZERO_VALUE}}; }) - // Load tile from the src tensor T_LOAD_NHWC_INDIRECT({{SRC_DATA_TYPE}}, M0, K0, {{SRC_TENSOR_TYPE}}, {{src}}, g_ind_2, yk, xk, ck, _ISRC_WIDTH, _ISRC_HEIGHT, {{src}}_stride_y, xi, yi, a); - // Load tile from the weights tensor T_LOAD({{WEI_DATA_TYPE}}, N0, K0, {{WEI_TENSOR_TYPE}}, {{weight}}, ck, g_ind_0 * _IY_MULTIPLIER + i, _IY_MULTIPLIER, {{weight}}_stride_y, b); - // Compute the matrix multiplication between two tiles T_MMUL({{SRC_DATA_TYPE}}, {{WEI_DATA_TYPE}}, {{ACC_DATA_TYPE}}, M0, N0, K0, NT, T, a, b, {{dst}}); ck += K0; } - - // We voluntarily use SRC_CHANNELS rather than _DSRC_CHANNELS - // This #if directive should be removed in case of dynamic tensor support )_"; if(leftover_loop) { code += R"_( - // Left-over accumulations for(; k < _ISRC_CHANNELS; ++k) { TILE({{SRC_DATA_TYPE}}, M0, 1, a); TILE({{WEI_DATA_TYPE}}, N0, 1, b); - // Initialize tiles LOOP_UNROLLING(int, i, 0, 1, M0, { a[i].v = {{ZERO_VALUE}}; @@ -179,14 +167,10 @@ TILE({{ACC_DATA_TYPE}}, M0, N0, {{dst}}); b[i].v = {{ZERO_VALUE}}; }) - // Load tile from the src tensor T_LOAD_NHWC_INDIRECT({{SRC_DATA_TYPE}}, M0, 1, {{SRC_TENSOR_TYPE}}, {{src}}, g_ind_2, yk, xk, ck, _ISRC_WIDTH, _ISRC_HEIGHT, {{src}}_stride_y, xi, yi, a); - // Load tile from the weights tensor - // The T_LOAD for the left-over elements can only use BUFFER because we load one element per iteration T_LOAD({{WEI_DATA_TYPE}}, N0, 1, BUFFER, {{weight}}, ck, g_ind_0 * _IY_MULTIPLIER + i, _IY_MULTIPLIER, {{weight}}_stride_y, b); - // Compute the matrix multiplication between two tiles T_MMUL({{SRC_DATA_TYPE}}, {{WEI_DATA_TYPE}}, {{ACC_DATA_TYPE}}, M0, N0, 1, NT, T, a, b, {{dst}}); ++ck; @@ -215,12 +199,16 @@ code += R"_( T_LOAD({{BIA_DATA_TYPE}}, 1, N0, BUFFER, {{bias}}, g_ind_0, 0, 1, 0, bias0); - // c = c + bias[broadcasted] T_ELTWISE_BROADCAST_ADD_X({{ACC_DATA_TYPE}}, M0, N0, {{dst}}, bias0, {{dst}}); )_"; } code += R"_( + LOOP_UNROLLING(int, i, 0, 1, M0, + { + g_dst_indirect_y[i].v = (uint)min(g_ind_1 + i, (int)({{arg_dst}}_w * {{arg_dst}}_h) - 1); + g_dst_indirect_y[i].v += g_ind_2 * (int)({{arg_dst}}_w * {{arg_dst}}_h); + }) } //------------------ END KERNEL {{meta_kernel_id}} --------------------- )_"; diff --git a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateStore.cpp b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateStore.cpp index 6c4b8f52f2..bffb467ebb 100644 --- a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateStore.cpp +++ b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateStore.cpp @@ -46,31 +46,14 @@ std::string ClTemplateStore::get_name() const std::string ClTemplateStore::get_component_code(const ComponentGroup &comp_group) const { ARM_COMPUTE_UNUSED(comp_group); + return R"_( //------------------ START KERNEL {{meta_kernel_id}} STORE --------------------- { -// This also follows NHWC layout -// g_ind_0 maps to global_id(0) maps to Channel -// g_ind_1 maps to global_id(1) maps to Height and Weight (Collapsed Window) -// g_ind_2 maps to global_id(2) maps to N / Batch -#define _IDST_WIDTH {{dst}}_w -#define _IDST_HEIGHT {{dst}}_h - TILE(uint, M0, 1, dst_indirect_y); - - // Calculate the destination indirect Y - LOOP_UNROLLING(int, i, 0, 1, M0, - { - dst_indirect_y[i].v = (uint)min(g_ind_1 + i, (int)(_IDST_WIDTH * _IDST_HEIGHT) - 1); - dst_indirect_y[i].v += g_ind_2 * (int)(_IDST_WIDTH * _IDST_HEIGHT); - }) - bool x_cond = PARTIAL_N0 != 0 && get_global_id(0) == 0; - T_STORE_INDIRECT_WIDTH_SELECT({{DST_DATA_TYPE}}, M0, N0, PARTIAL_N0, {{DST_TENSOR_TYPE}}, {{dst}}, g_ind_0, {{dst}}_stride_y, x_cond, {{src}}, dst_indirect_y); - -#undef _IDST_WIDTH -#undef _IDST_HEIGHT - //------------------ END KERNEL {{meta_kernel_id}} STORE --------------------- + T_STORE_INDIRECT_WIDTH_SELECT({{DST_DATA_TYPE}}, M0, N0, PARTIAL_N0, {{DST_TENSOR_TYPE}}, {{dst}}, g_ind_0, {{dst}}_stride_y, x_cond, {{src}}, g_dst_indirect_y); +//------------------ END KERNEL {{meta_kernel_id}} STORE --------------------- } )_"; |