diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/transform.hpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/transform.hpp | 67 |
1 files changed, 28 insertions, 39 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transform.hpp b/src/core/NEON/kernels/arm_gemm/transform.hpp index c80bb59941..35e61b05a4 100644 --- a/src/core/NEON/kernels/arm_gemm/transform.hpp +++ b/src/core/NEON/kernels/arm_gemm/transform.hpp @@ -35,63 +35,51 @@ * being a multiple of the block sizes. */ template <unsigned IntBy, unsigned int BlockBy, bool Transposed, size_t TOutSize, size_t TInSize> -struct TransformImpl -{ +struct TransformImpl { template <typename TOut, typename TIn> - static void Transform(TOut *out, const TIn *const in, const int stride, - const int y0, const int ymax, const int x0, const int xmax) - { + static void Transform(TOut* out, const TIn* const in, const int stride, + const int y0, const int ymax, const int x0, const int xmax) { const int n_whole_y_blocks = (ymax - y0) / IntBy; - const int y_remainders = (ymax - y0) % IntBy; - const int n_y_blocks = n_whole_y_blocks + (y_remainders ? 1 : 0); + const int y_remainders = (ymax - y0) % IntBy; + const int n_y_blocks = n_whole_y_blocks + (y_remainders ? 1 : 0); const int n_whole_x_blocks = (xmax - x0) / BlockBy; - const int x_remainders = (xmax - x0) % BlockBy; - const int n_x_blocks = n_whole_x_blocks + (x_remainders ? 1 : 0); + const int x_remainders = (xmax - x0) % BlockBy; + const int n_x_blocks = n_whole_x_blocks + (x_remainders ? 1 : 0); // "Y" loop: advance down the rows of the source IntBy rows at a time. // Set up fill_rows to show the number rows to copy from, and blank_rows // for the number of blank rows to add. - for(int y_block = 0; y_block < n_y_blocks; y_block++) - { - int fill_rows = (y_block < n_whole_y_blocks) ? IntBy : y_remainders; + for (int y_block=0 ; y_block < n_y_blocks; y_block++) { + int fill_rows = (y_block < n_whole_y_blocks) ? IntBy : y_remainders; int blank_rows = IntBy - fill_rows; int y_base = y0 + (y_block * IntBy); // So now advance along this block of rows, BlockBy columns at a time. - for(int x_block = 0; x_block < n_x_blocks; x_block++) - { - int fill_cols = (x_block < n_whole_x_blocks) ? BlockBy : x_remainders; + for (int x_block=0 ; x_block < n_x_blocks; x_block++) { + int fill_cols = (x_block < n_whole_x_blocks) ? BlockBy : x_remainders; int blank_cols = BlockBy - fill_cols; int x_base = x0 + (x_block * BlockBy); - for(int row = 0; row < fill_rows; row++) - { - for(int col = 0; col < fill_cols; col++) - { + for (int row = 0; row < fill_rows; row++) { + for (int col = 0; col < fill_cols; col++) { // In-range copy. If it's transposed, we reverse the sense of rows and columns here. - if(Transposed) - { + if (Transposed) { *out++ = static_cast<TOut>(in[(x_base + col) * stride + y_base + row]); - } - else - { + } else { *out++ = static_cast<TOut>(in[(y_base + row) * stride + x_base + col]); } } // "col" tail - row is in range but column is out of range. - for(int col = 0; col < blank_cols; col++) - { + for (int col=0; col < blank_cols; col++) { *out++ = static_cast<TOut>(0); } } // "row" tail - row is out of range so fill with zeros always. - for(int row = 0; row < blank_rows; row++) - { - for(int col = 0; col < (fill_cols + blank_cols); col++) - { + for (int row = 0; row < blank_rows; row++) { + for (int col=0; col < (fill_cols + blank_cols); col++) { *out++ = static_cast<TOut>(0); } } @@ -100,9 +88,8 @@ struct TransformImpl } template <typename T> - static inline void Transform(T *out, const T *const in, const int stride, - const int k0, const int kmax, const int x0, const int xmax) - { + static inline void Transform(T* out, const T* const in, const int stride, + const int k0, const int kmax, const int x0, const int xmax) { Transform<T, T>(out, in, stride, k0, kmax, x0, xmax); } }; @@ -110,13 +97,15 @@ struct TransformImpl /*****************************************************************************/ template <unsigned int IntBy, unsigned int BlockBy, bool Transposed, typename TOut, typename TIn> void Transform( - TOut *out, const TIn *const in, const int stride, - const int k0, const int kmax, const int x0, const int xmax) -{ - // Redirect to a specialised implementation predicated on argument size. - TransformImpl<IntBy, BlockBy, Transposed, sizeof(TOut), sizeof(TIn)>::Transform( - out, in, stride, k0, kmax, x0, xmax); + TOut* out, const TIn* const in, const int stride, + const int k0, const int kmax, const int x0, const int xmax +) { + // Redirect to a specialised implementation predicated on argument size. + TransformImpl<IntBy, BlockBy, Transposed, sizeof(TOut), sizeof(TIn)>::Transform( + out, in, stride, k0, kmax, x0, xmax + ); } /*****************************************************************************/ #include "transforms/list.hpp" + |