diff options
| -rw-r--r-- | src/core/GLES_COMPUTE/cs_shaders/direct_convolution1x1.cs | 24 | ||||
| -rw-r--r-- | src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp | 9 |
2 files changed, 32 insertions, 1 deletions
diff --git a/src/core/GLES_COMPUTE/cs_shaders/direct_convolution1x1.cs b/src/core/GLES_COMPUTE/cs_shaders/direct_convolution1x1.cs index 071c1858bc..1338299f8c 100644 --- a/src/core/GLES_COMPUTE/cs_shaders/direct_convolution1x1.cs +++ b/src/core/GLES_COMPUTE/cs_shaders/direct_convolution1x1.cs @@ -160,6 +160,29 @@ void main() uint z_index = gl_GlobalInvocationID.z; TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, z_index * weights_stride_w); +#ifdef WEIGHTS_OPTIMIZATION + float w1, w2; + int nums = (int(weights_depth)) / 2; + for(int d = 0; d < nums; ++d) + { + vec2 vec2_w = LOAD_UNPACK2_CURRENT_ITEM_HALF(weights_ptr, weights_iter); + + w1 = vec2_w.x; + vec4 r1[2] = CONVOLVE(src_iter, w1); + pixels[0] += r1[0]; + pixels[1] += r1[1]; + + TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); + + w2 = vec2_w.y; + vec4 r2[2] = CONVOLVE(src_iter, w2); + pixels[0] += r2[0]; + pixels[1] += r2[1]; + + TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); + TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z * uint(2)); + } +#else /* WEIGHTS_OPTIMIZATION */ float w; for(int d = 0; d < int(weights_depth); ++d) { @@ -172,6 +195,7 @@ void main() TENSOR_ITERATOR_ADVANCE_IN_BYTES(src_iter, src_attrs.stride_z); TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, weights_attrs.stride_z); } +#endif /* WEIGHTS_OPTIMIZATION */ #ifdef BIAS vec2 vec2_b; diff --git a/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp index 5c7320aa8d..962f044214 100644 --- a/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp @@ -191,6 +191,10 @@ void GCDirectConvolutionLayerKernel<kernel_size>::configure(const IGCTensor *inp case DataType::F16: num_elems_read_per_iteration_x = 8; num_elems_written_per_iteration_x = 8; + if(weights->info()->dimension(2) % 2 == 0) + { + options.emplace("#define WEIGHTS_OPTIMIZATION"); + } break; case DataType::F32: @@ -255,7 +259,10 @@ void GCDirectConvolutionLayerKernel<kernel_size>::configure(const IGCTensor *inp switch(weights->info()->data_type()) { case DataType::F16: - weights_access = AccessWindowStatic(weights->info(), 0, 0, kernel_size + 1, kernel_size); + if((weights->info()->dimension(2) % 2 != 0) || (kernel_size != 1)) + { + weights_access = AccessWindowStatic(weights->info(), 0, 0, kernel_size + 1, kernel_size); + } if(_bias != nullptr) { bias_access = AccessWindowStatic(_bias->info(), 0, 0, _bias->info()->dimension(0) + 1, 1); |