diff options
Diffstat (limited to 'src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwCast.cpp')
| -rw-r--r-- | src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwCast.cpp | 44 |
1 files changed, 23 insertions, 21 deletions
diff --git a/src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwCast.cpp b/src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwCast.cpp index 6ecf2bac44..e8e5087633 100644 --- a/src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwCast.cpp +++ b/src/dynamic_fusion/sketch/gpu/ckw_driver/components/GpuCkwCast.cpp @@ -24,16 +24,18 @@ #include "GpuCkwCast.h" #include "arm_compute/core/Error.h" -#include "arm_compute/core/Validate.h" #include "arm_compute/core/utils/helpers/AdjustVecSize.h" +#include "arm_compute/core/Validate.h" #include "ckw/TensorTileSampler.h" + #include "src/core/helpers/WindowHelpers.h" -#include "src/dynamic_fusion/sketch/gpu/GpuKernelArgument.h" -#include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.h" +#include "src/dynamic_fusion/sketch/gpu/ckw_driver/components/utils/type_converter/Common.h" #include "src/dynamic_fusion/sketch/gpu/ckw_driver/GpuCkwKernelWriter.h" #include "src/dynamic_fusion/sketch/gpu/ckw_driver/GpuCkwScopedKernelWriter.h" #include "src/dynamic_fusion/sketch/gpu/ckw_driver/GpuCkwVariableTable.h" -#include "src/dynamic_fusion/sketch/gpu/ckw_driver/components/utils/type_converter/Common.h" +#include "src/dynamic_fusion/sketch/gpu/GpuKernelArgument.h" +#include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.h" + #include <string> using namespace ckw; @@ -84,30 +86,29 @@ inline TensorTileSampler create_sampler(GpuCkwScopedKernelWriter &writer, int32_ } } // namespace -GpuCkwCast::GpuCkwCast(ComponentId id, - const ArgumentPack<ITensorInfo> &tensors, - const Attributes &attributes) - : IGpuCkwComponentDriver{ id, tensors }, - _src{}, - _dst{}, - _attributes{ attributes } +GpuCkwCast::GpuCkwCast(ComponentId id, const ArgumentPack<ITensorInfo> &tensors, const Attributes &attributes) + : IGpuCkwComponentDriver{id, tensors}, _src{}, _dst{}, _attributes{attributes} { _src = this->tensors().get_const_tensor(TensorType::ACL_SRC_0); _dst = this->tensors().get_const_tensor(TensorType::ACL_DST_0); ARM_COMPUTE_ERROR_ON_NULLPTR(_src, _dst); } -void GpuCkwCast::write_component_code(const ComponentGroup &comp_group, GpuCkwVariableTable &vtable, GpuCkwScopedKernelWriter writer) const +void GpuCkwCast::write_component_code(const ComponentGroup &comp_group, + GpuCkwVariableTable &vtable, + GpuCkwScopedKernelWriter writer) const { const auto root_window = comp_group.get_root_component()->ckw_component_driver()->get_window(); const unsigned int n0 = root_window.x().step(); const unsigned int m0 = root_window.y().step(); - GpuCkwComponentArgument *src = vtable.declare_variable(comp_group, writer, _src, TensorStorageType::ClBufferUint8Ptr, "src"); - GpuCkwComponentArgument *dst = vtable.declare_variable(comp_group, writer, _dst, TensorStorageType::ClBufferUint8Ptr, "dst"); + GpuCkwComponentArgument *src = + vtable.declare_variable(comp_group, writer, _src, TensorStorageType::ClBufferUint8Ptr, "src"); + GpuCkwComponentArgument *dst = + vtable.declare_variable(comp_group, writer, _dst, TensorStorageType::ClBufferUint8Ptr, "dst"); // Load the source tile and prepare the sampler. - if(!src->has_tile()) + if (!src->has_tile()) { const auto sampler = create_sampler(writer, m0, n0); writer->op_load_once(src, sampler); @@ -122,7 +123,7 @@ void GpuCkwCast::write_component_code(const ComponentGroup &comp_group, GpuCkwVa const auto &sampler = src->tile_sampler(); // Prepare the output tile. - if(!dst->has_tile()) + if (!dst->has_tile()) { // Get Target datatype and convert it to ckw::DataType. ckw::DataType target_dt = dynamic_fusion::to_ckw(_attributes.data_type()); @@ -143,7 +144,7 @@ void GpuCkwCast::write_component_code(const ComponentGroup &comp_group, GpuCkwVa const size_t dst_size = data_size_from_type(_dst->data_type()); const bool cast_down = (src_size >= dst_size); - if(cast_down && is_data_type_quantized(_src->data_type())) + if (cast_down && is_data_type_quantized(_src->data_type())) { const auto &constant_x80 = writer->declare_tile("0x80", 0x80); writer->op_binary_expression(src_tile, src_tile, BinaryOp::BitwiseXOR, constant_x80); @@ -151,7 +152,7 @@ void GpuCkwCast::write_component_code(const ComponentGroup &comp_group, GpuCkwVa ckw::ConvertPolicy convert_policy = ckw::ConvertPolicy::None; - if(cast_down && (is_data_type_float(_src->data_type()) || _attributes.convert_policy() == ConvertPolicy::SATURATE)) + if (cast_down && (is_data_type_float(_src->data_type()) || _attributes.convert_policy() == ConvertPolicy::SATURATE)) { convert_policy = ckw::ConvertPolicy::Saturate; } @@ -167,9 +168,10 @@ Window GpuCkwCast::get_window() const // Collapse Dim 1 (W) and Dim 2 (H) together, leave Dim 0 (C) unchanged // This is in line with the collapsing convention used by operators like Conv2d output_shape.collapse(2U, 1U); - constexpr unsigned int vector_size_byte_opencl = 16; - const unsigned int num_elems_processed_per_iteration = adjust_vec_size(vector_size_byte_opencl / _dst->element_size(), _dst->dimension(0)); - Window win = calculate_max_window(output_shape, Steps(num_elems_processed_per_iteration)); + constexpr unsigned int vector_size_byte_opencl = 16; + const unsigned int num_elems_processed_per_iteration = + adjust_vec_size(vector_size_byte_opencl / _dst->element_size(), _dst->dimension(0)); + Window win = calculate_max_window(output_shape, Steps(num_elems_processed_per_iteration)); return win; } |