diff options
Diffstat (limited to 'src/runtime/heuristics/indirect_conv/ClIndirectConvDefaultConfigValhall.cpp')
-rw-r--r-- | src/runtime/heuristics/indirect_conv/ClIndirectConvDefaultConfigValhall.cpp | 60 |
1 files changed, 33 insertions, 27 deletions
diff --git a/src/runtime/heuristics/indirect_conv/ClIndirectConvDefaultConfigValhall.cpp b/src/runtime/heuristics/indirect_conv/ClIndirectConvDefaultConfigValhall.cpp index 990f050112..3380d8f1b7 100644 --- a/src/runtime/heuristics/indirect_conv/ClIndirectConvDefaultConfigValhall.cpp +++ b/src/runtime/heuristics/indirect_conv/ClIndirectConvDefaultConfigValhall.cpp @@ -35,17 +35,19 @@ namespace cl_indirect_conv { using namespace arm_compute::misc::shape_calculator; -ClIndirectConvDefaultConfigValhall::ClIndirectConvDefaultConfigValhall(GPUTarget gpu) - : IClIndirectConvKernelConfig(gpu) +ClIndirectConvDefaultConfigValhall::ClIndirectConvDefaultConfigValhall(GPUTarget gpu) : IClIndirectConvKernelConfig(gpu) { } -DirectConvComputeKernelInfo ClIndirectConvDefaultConfigValhall::configure(const ITensorInfo *src, const ITensorInfo *wei, const PadStrideInfo &conv_info) +DirectConvComputeKernelInfo ClIndirectConvDefaultConfigValhall::configure(const ITensorInfo *src, + const ITensorInfo *wei, + const PadStrideInfo &conv_info) { - using ConfigurationFunctionExecutorPtr = DirectConvComputeKernelInfo (ClIndirectConvDefaultConfigValhall::*)(const ITensorInfo *src, const ITensorInfo *wei, const PadStrideInfo &conv_info); + using ConfigurationFunctionExecutorPtr = DirectConvComputeKernelInfo (ClIndirectConvDefaultConfigValhall::*)( + const ITensorInfo *src, const ITensorInfo *wei, const PadStrideInfo &conv_info); - ClIndirectConvConfigArray<ConfigurationFunctionExecutorPtr> configs_G77(&ClIndirectConvDefaultConfigValhall::configure_G77_f32, - &ClIndirectConvDefaultConfigValhall::configure_G77_f16); + ClIndirectConvConfigArray<ConfigurationFunctionExecutorPtr> configs_G77( + &ClIndirectConvDefaultConfigValhall::configure_G77_f32, &ClIndirectConvDefaultConfigValhall::configure_G77_f16); // Important note: Indirect convolution should not be used when the kernel size is 1x1 (pointwise). The reason is because the indirect buffer makes // indirect convolution less efficient than direct convolution or gemm. For this reason, the heuristic of indirect convolution has not been tuned @@ -57,22 +59,24 @@ DirectConvComputeKernelInfo ClIndirectConvDefaultConfigValhall::configure(const return (this->*func)(src, wei, conv_info); } -DirectConvComputeKernelInfo ClIndirectConvDefaultConfigValhall::configure_G77_f32(const ITensorInfo *src, const ITensorInfo *wei, const PadStrideInfo &conv_info) +DirectConvComputeKernelInfo ClIndirectConvDefaultConfigValhall::configure_G77_f32(const ITensorInfo *src, + const ITensorInfo *wei, + const PadStrideInfo &conv_info) { DirectConvComputeKernelInfo desc; - if(src->data_layout() == DataLayout::NHWC) + if (src->data_layout() == DataLayout::NHWC) { - const TensorShape dst_shape = misc::shape_calculator::compute_deep_convolution_shape(*src, *wei, conv_info); + const TensorShape dst_shape = misc::shape_calculator::compute_deep_convolution_shape(*src, *wei, conv_info); const bool export_weights_to_cl_image = export_to_cl_image(wei); - const int32_t stride_x = conv_info.stride().first; - const int32_t stride_y = conv_info.stride().second; - const int32_t ofm = dst_shape[0]; - const int32_t m = (dst_shape[1]/ stride_x) * (dst_shape[2] / stride_y); + const int32_t stride_x = conv_info.stride().first; + const int32_t stride_y = conv_info.stride().second; + const int32_t ofm = dst_shape[0]; + const int32_t m = (dst_shape[1] / stride_x) * (dst_shape[2] / stride_y); desc.export_weights_to_cl_image = export_weights_to_cl_image; - if(ofm <= 4) + if (ofm <= 4) { desc.m0 = 1; desc.n0 = 2; @@ -82,7 +86,7 @@ DirectConvComputeKernelInfo ClIndirectConvDefaultConfigValhall::configure_G77_f3 { // The 16000 threshold value has been identified as the right // one for using the biggest block size allowed on F32: 5x4x4 - if(m < 16000) + if (m < 16000) { desc.m0 = 4; desc.n0 = 4; @@ -100,31 +104,33 @@ DirectConvComputeKernelInfo ClIndirectConvDefaultConfigValhall::configure_G77_f3 return desc; } -DirectConvComputeKernelInfo ClIndirectConvDefaultConfigValhall::configure_G77_f16(const ITensorInfo *src, const ITensorInfo *wei, const PadStrideInfo &conv_info) +DirectConvComputeKernelInfo ClIndirectConvDefaultConfigValhall::configure_G77_f16(const ITensorInfo *src, + const ITensorInfo *wei, + const PadStrideInfo &conv_info) { DirectConvComputeKernelInfo desc; - if(src->data_layout() == DataLayout::NHWC) + if (src->data_layout() == DataLayout::NHWC) { - const TensorShape wei_shape = wei->tensor_shape(); - const TensorShape dst_shape = misc::shape_calculator::compute_deep_convolution_shape(*src, *wei, conv_info); + const TensorShape wei_shape = wei->tensor_shape(); + const TensorShape dst_shape = misc::shape_calculator::compute_deep_convolution_shape(*src, *wei, conv_info); const bool export_weights_to_cl_image = export_to_cl_image(wei); - const int32_t ofm = dst_shape[0]; - const int32_t m = dst_shape[1] * dst_shape[2]; - const int32_t k = wei_shape[0]; + const int32_t ofm = dst_shape[0]; + const int32_t m = dst_shape[1] * dst_shape[2]; + const int32_t k = wei_shape[0]; desc.export_weights_to_cl_image = export_weights_to_cl_image; - if(ofm <= 4) + if (ofm <= 4) { // k0 should be as larger as possible. However, we should avoid // having left-over for loops that make the implementation slower. - if((k % 16) == 0) + if ((k % 16) == 0) { desc.k0 = 16; } - else if((k % 8) == 0) + else if ((k % 8) == 0) { desc.k0 = 8; } @@ -140,11 +146,11 @@ DirectConvComputeKernelInfo ClIndirectConvDefaultConfigValhall::configure_G77_f1 { // The 16000 threshold value has been identified as the right // one for using the biggest block size allowed on F16: 8x4 - if(m >= 16000 && k < 4) + if (m >= 16000 && k < 4) { desc.m0 = 8; desc.n0 = 4; - desc.k0 = 4; // k0 is clamped to k inside the kernel when k is less than 4 + desc.k0 = 4; // k0 is clamped to k inside the kernel when k is less than 4 } else { |