diff options
| author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2021-12-01 09:26:14 +0000 |
|---|---|---|
| committer | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2021-12-01 12:38:25 +0000 |
| commit | 78baa48308cba4101b4bcb4680f2f4ca90aeefd7 (patch) | |
| tree | 62b049b4cb8cffc78559c14dcaa2dbcea2ebc871 /src/gpu/cl/kernels | |
| parent | 56d55123527b5bb84a5c3516f161dd4438cdc7d8 (diff) | |
| download | ComputeLibrary-78baa48308cba4101b4bcb4680f2f4ca90aeefd7.tar.gz | |
Improve start-up direct convolution on OpenCL
- Pass arguments at runtime
- Rework ClConv2D heuristic to select direct convolution when OFM < IFM
also for small kernel sizes
Resolves COMPMID-5000
Change-Id: I9b538e29093829bc366d24d1e904341c247fa22b
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6771
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/gpu/cl/kernels')
| -rw-r--r-- | src/gpu/cl/kernels/ClDirectConv2dKernel.cpp | 12 |
1 files changed, 3 insertions, 9 deletions
diff --git a/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp b/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp index 2d851a6982..7107def8ff 100644 --- a/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp +++ b/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp @@ -438,14 +438,8 @@ void ClDirectConv2dKernel::configure(const CLCompileContext &compile_context, IT build_options.add_option("-cl-fast-relaxed-math"); build_options.add_option("-DSRC_TENSOR_TYPE=BUFFER"); - build_options.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(src->dimension(width_idx))); - build_options.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(height_idx))); - build_options.add_option("-DSRC_CHANNELS=" + support::cpp11::to_string(src->dimension(channel_idx))); build_options.add_option("-DSRC_DATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); build_options.add_option("-DDST_TENSOR_TYPE=BUFFER"); - build_options.add_option("-DDST_WIDTH=" + support::cpp11::to_string(dst->dimension(width_idx))); - build_options.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(dst->dimension(height_idx))); - build_options.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(dst->dimension(channel_idx))); build_options.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(dst->data_type())); build_options.add_option_if_else(export_to_cl_image, "-DWEI_TENSOR_TYPE=IMAGE", "-DWEI_TENSOR_TYPE=BUFFER"); build_options.add_option("-DWEI_WIDTH=" + support::cpp11::to_string(weights->dimension(width_idx))); @@ -613,13 +607,13 @@ void ClDirectConv2dKernel::run_op(ITensorPack &tensors, const Window &window, cl } unsigned int idx = 0; - add_4D_tensor_argument(idx, src, slice); - add_4D_tensor_argument(idx, dst, slice); + add_4d_tensor_nhwc_argument(idx, src); + add_4d_tensor_nhwc_argument(idx, dst); if(export_to_cl_image) { _kernel.setArg(idx++, weights_cl_image); } - add_4D_tensor_argument(idx, weights, slice); + add_4d_tensor_nhwc_argument(idx, weights); if(biases != nullptr) { add_1D_tensor_argument(idx, biases, slice); |