diff options
| author | Michele Di Giorgio <michele.digiorgio@arm.com> | 2018-02-21 10:02:58 +0000 |
|---|---|---|
| committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:49:16 +0000 |
| commit | 3ebef32816435516f68cefba689dba7216464154 (patch) | |
| tree | 67d4f6bc80da08474862e6131d148167f27b14e3 /src/core/CL/kernels | |
| parent | 4d9379a9d3ada794f532ce8acdc8607f4faa2b21 (diff) | |
| download | ComputeLibrary-3ebef32816435516f68cefba689dba7216464154.tar.gz | |
COMPMID-949: Optimizing CLDepthwiseConvolution3x3Kernel for FP16
Change-Id: I2af6544eab17004c5b3de56557cb2cc5efecc915
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/122181
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Diffstat (limited to 'src/core/CL/kernels')
| -rw-r--r-- | src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp | 26 |
1 files changed, 23 insertions, 3 deletions
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp index d50e4d695e..9e5585cba4 100644 --- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp @@ -147,6 +147,7 @@ void CLDepthwiseConvolutionLayer3x3Kernel::configure(const ICLTensor *input, con // Configure the local work size for Bifrost with a value obtained // via exhaustive autotuning for the MobileNets tensor shapes. const GPUTarget gpu_target = get_target(); + const bool is_bifrost = gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72); // Configure kernel window unsigned int num_elems_read_per_iteration_x = 0; @@ -178,12 +179,31 @@ void CLDepthwiseConvolutionLayer3x3Kernel::configure(const ICLTensor *input, con num_elems_read_per_iteration_x = 3 + (num_elems_written_per_iteration_x - 1) * _conv_stride_x; break; } + if(is_bifrost) + { + if(_conv_stride_x == 1 && _conv_stride_y == 1) + { + kernel_name = "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16"; + num_elems_read_per_iteration_x = 8; + num_elems_written_per_iteration_x = 4; + num_elems_read_per_iteration_y = 6; + num_elems_written_per_iteration_y = 4; + } + else if(_conv_stride_x == 2 && _conv_stride_y == 2) + { + kernel_name = "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16"; + num_elems_read_per_iteration_x = 10; + num_elems_written_per_iteration_x = 4; + num_elems_read_per_iteration_y = 5; + num_elems_written_per_iteration_y = 2; + } + } } - else if(input->info()->data_type() == DataType::F32 && gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72)) + else if(input->info()->data_type() == DataType::F32 && is_bifrost) { if(_conv_stride_x == 1 && _conv_stride_y == 1) { - kernel_name = "depthwise_convolution_3x3_stridex1_stridey1_bifrost"; + kernel_name = "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32"; num_elems_read_per_iteration_x = 4; num_elems_read_per_iteration_y = 6; num_elems_written_per_iteration_x = 2; @@ -191,7 +211,7 @@ void CLDepthwiseConvolutionLayer3x3Kernel::configure(const ICLTensor *input, con } else if(_conv_stride_x == 2 && _conv_stride_y == 2) { - kernel_name = "depthwise_convolution_3x3_stridex2_stridey2_bifrost"; + kernel_name = "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32"; num_elems_read_per_iteration_x = 6; num_elems_read_per_iteration_y = 5; num_elems_written_per_iteration_x = 2; |