diff options
| author | Pablo Tello <pablo.tello@arm.com> | 2018-09-03 11:40:33 +0100 |
|---|---|---|
| committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:54:54 +0000 |
| commit | 72686fa6ee0f04d458ed2274b4d34917628ef14d (patch) | |
| tree | 7b897efdc535ef7cea8826d36fae951a3c53438e /src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp | |
| parent | 0d2b48c4a2cc82fd3312635a97117553ea4ee735 (diff) | |
| download | ComputeLibrary-72686fa6ee0f04d458ed2274b4d34917628ef14d.tar.gz | |
COMPMID-1550: Winograd integrate RSH changes.
Refactors the transforms to make use of partial specialization.
Change-Id: Idff68d22817a00a7ee9eef5351a5a9fd33147540
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/146635
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp')
| -rw-r--r-- | src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp | 21 |
1 files changed, 2 insertions, 19 deletions
diff --git a/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp b/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp index 97b2695d69..a9d5d52d15 100644 --- a/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp +++ b/src/core/NEON/kernels/convolution/winograd/transforms/input_2x2_3x3_fp32.cpp @@ -29,22 +29,7 @@ namespace winograd { -using Transform = WinogradGEMM<2, 2, 3, 3>::InputTransform<float>; - -/****************************************************************************** - * Cost methods for the input transform. - * ===================================== - */ -template <> -template <> -int Transform::ops_performed(const Tensor4DShape &input_shape) -{ - // NOTE: Cost in FLOPs rather than instructions or uops. - const int tile_M = iceildiv(input_shape.n_rows, inner_tile_rows); - const int tile_N = iceildiv(input_shape.n_cols, inner_tile_cols); - return 16 * 16 * tile_M * tile_N * input_shape.n_channels; -} -/*****************************************************************************/ +using Transform = InputTransformImpl<3, 3, 4, 4, float>; /***************************************************************************** * F(2x2, 3x3) implies the use of a 4x4 input tile. Such tiles can require a @@ -100,7 +85,6 @@ int Transform::ops_performed(const Tensor4DShape &input_shape) * Padding right in {0, 1, 2} */ template <> -template <> template <int pad_top, int pad_left, int pad_bottom, int pad_right> void Transform::process_tile( int n_channels, @@ -328,7 +312,6 @@ void Transform::process_tile( } template <> -template <> const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom][n_pad_right] = { { @@ -405,5 +388,5 @@ const Transform::TileFn Transform::tile_fns[n_pad_top][n_pad_left][n_pad_bottom] } }; -template struct WinogradGEMM<2, 2, 3, 3>::InputTransform<float>; +template class InputTransform<3, 3, 4, 4, float>; } // namespace winograd |