diff options
Diffstat (limited to 'src/core/NEON/kernels/NEWinogradLayerKernel.cpp')
-rw-r--r-- | src/core/NEON/kernels/NEWinogradLayerKernel.cpp | 79 |
1 files changed, 78 insertions, 1 deletions
diff --git a/src/core/NEON/kernels/NEWinogradLayerKernel.cpp b/src/core/NEON/kernels/NEWinogradLayerKernel.cpp index b9109dcff2..fe633368c0 100644 --- a/src/core/NEON/kernels/NEWinogradLayerKernel.cpp +++ b/src/core/NEON/kernels/NEWinogradLayerKernel.cpp @@ -27,9 +27,86 @@ #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensor.h" #include "arm_compute/core/TensorInfo.h" +#include "support/ToolchainSupport.h" + +#include "src/core/NEON/kernels/winograd/winograd_shim_nchw.hpp" + +using T = winograd_shim_nchw::Winograd2x2_3x3GEMM<float, float>; namespace arm_compute { +class Winograd3x3F32::Private +{ +public: + Private(const KernelShape &kernel_shape, const Tensor4DShape input_shape, const PaddingType padding_type, void *kernel_storage) + : convolver(kernel_shape, input_shape, padding_type, kernel_storage) + { + } + + T convolver; +}; + +Winograd3x3F32::~Winograd3x3F32() +{ +} + +void Winograd3x3F32::nchw2nhwc(const Tensor4DShape &input_shape, const PaddingType padding_type, void *working_space, const void *const input) +{ + _pimpl->convolver.nchw2nhwc(input_shape, padding_type, working_space, reinterpret_cast<const float *>(input)); +} + +void Winograd3x3F32::nhwc2nchw(const Tensor4DShape &input_shape, const PaddingType padding_type, void *working_space, void *const output) +{ + _pimpl->convolver.nhwc2nchw(input_shape, padding_type, working_space, reinterpret_cast<float *const>(output)); +} + +void Winograd3x3F32::transform_weights(const void *const kernel, void *transform_working_space) +{ + _pimpl->convolver.transform_weights(reinterpret_cast<const float *>(kernel), transform_working_space); +} + +void Winograd3x3F32::reshape_input(const Tensor4DShape &input_shape, const PaddingType padding_type, const void *const input, void *working_space) +{ + _pimpl->convolver.reshape_input(input_shape, padding_type, reinterpret_cast<const float *>(input), working_space); +} + +void Winograd3x3F32::reshape_output(const Tensor4DShape &input_shape, const PaddingType padding_type, void *const output) +{ +#if defined(__aarch64__) + _pimpl->convolver.reshape_output(input_shape, padding_type, reinterpret_cast<float *const>(output)); +#else /* __aarch64__ */ + ARM_COMPUTE_UNUSED(input_shape); + ARM_COMPUTE_UNUSED(padding_type); + ARM_COMPUTE_UNUSED(output); + ARM_COMPUTE_ERROR("Not implemented"); +#endif /* __aarch64__ */ +} + +std::pair<void *, void *> Winograd3x3F32::get_nhwc_ptrs(const Tensor4DShape &input_shape, const PaddingType padding_type, void *working_space) +{ + return _pimpl->convolver.get_nhwc_ptrs(input_shape, padding_type, working_space); +} + +Winograd3x3F32::Winograd3x3F32(const KernelShape &kernel_shape, const Tensor4DShape input_shape, const PaddingType padding_type, void *kernel_storage) + : _pimpl(support::cpp14::make_unique<Private>(kernel_shape, input_shape, padding_type, kernel_storage)) +{ +} + +size_t NEWinogradLayerKernel::get_kernel_storage_size(const KernelShape &shape) +{ + return T::get_kernel_storage_size(shape); +} + +size_t NEWinogradLayerKernel::get_working_space_size(const Tensor4DShape &input_shape, const KernelShape &k_shape, const PaddingType padding) +{ + return T::get_working_space_size(input_shape, k_shape, padding); +} + +size_t NEWinogradLayerKernel::get_kernel_transform_working_size(const KernelShape &shape) +{ + return T::get_kernel_transform_working_size(shape); +} + NEWinogradLayerKernel::NEWinogradLayerKernel() : _convolver(nullptr), _output(nullptr) { @@ -55,6 +132,6 @@ void NEWinogradLayerKernel::run(const Window &window, const ThreadInfo &info) const size_t num_gemms_per_thread = 16 / num_threads; const size_t first_gemm = tid * num_gemms_per_thread; const size_t last_gemm = (tid == (num_threads - 1)) ? 15 : first_gemm + num_gemms_per_thread - 1; - _convolver->execute(first_gemm, last_gemm); + _convolver->_pimpl->convolver.execute(first_gemm, last_gemm); } } // namespace arm_compute |