diff options
Diffstat (limited to 'src/gpu')
| -rw-r--r-- | src/gpu/cl/kernels/ClDirectConv2dKernel.cpp | 1 | ||||
| -rw-r--r-- | src/gpu/cl/operators/ClConv2d.cpp | 9 |
2 files changed, 3 insertions, 7 deletions
diff --git a/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp b/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp index 7107def8ff..5af7aa9662 100644 --- a/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp +++ b/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp @@ -453,6 +453,7 @@ void ClDirectConv2dKernel::configure(const CLCompileContext &compile_context, IT build_options.add_option("-DM0=" + support::cpp11::to_string(m0)); build_options.add_option("-DK0=" + support::cpp11::to_string(k0)); build_options.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0)); + build_options.add_option_if((src->dimension(channel_idx) % k0) != 0, "-DLEFTOVER_LOOP"); build_options.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(act_info.activation()))); if(is_data_type_quantized(data_type)) diff --git a/src/gpu/cl/operators/ClConv2d.cpp b/src/gpu/cl/operators/ClConv2d.cpp index 92b22e758d..2b9a51df22 100644 --- a/src/gpu/cl/operators/ClConv2d.cpp +++ b/src/gpu/cl/operators/ClConv2d.cpp @@ -258,20 +258,15 @@ ConvolutionMethod ClConv2d::get_convolution_method(const ITensorInfo *src, const const bool is_large_kernel_sz = (weights->dimension(idx_w) >= kernel_sz_direct_conv_thr) && (weights->dimension(idx_h) >= kernel_sz_direct_conv_thr); const bool is_ifm_ge_16 = src->dimension(idx_c) >= 16; const bool is_ifm_gt_ofm = weights->dimension(0U) * weights->dimension(1U) * weights->dimension(2U) > weights->dimension(3U); - const bool is_ofm_le_4 = weights->dimension(3U) <= 4; // Run Winograd if valid and IFM >= 16 if(is_wino_valid && is_ifm_ge_16) { return ConvolutionMethod::WINOGRAD; } - // Run Direct for Large kernel size - if(is_large_kernel_sz && is_ifm_gt_ofm && is_direct_valid) - { - return ConvolutionMethod::DIRECT; - } - if(is_ofm_le_4 && is_ifm_gt_ofm && is_direct_valid) + // Run Direct for Large kernel size + if(is_large_kernel_sz && is_ifm_ge_16 && is_direct_valid && is_ifm_gt_ofm) { return ConvolutionMethod::DIRECT; } |