From 17812ba9f7cf2c8f5121c11760ac45fbbdb7aeaf Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Mon, 4 Jun 2018 19:27:13 +0100 Subject: COMPMID-817: Tuner: Port kernels to new design. Change-Id: Iaabb1153c2abe0400ec79d51a21347debe92d642 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/134062 Tested-by: Jenkins Reviewed-by: Anthony Barbier --- src/runtime/CL/tuners/BifrostTuner.cpp | 150 ++++++++++++++++++++++++++++++++- 1 file changed, 149 insertions(+), 1 deletion(-) (limited to 'src/runtime/CL/tuners/BifrostTuner.cpp') diff --git a/src/runtime/CL/tuners/BifrostTuner.cpp b/src/runtime/CL/tuners/BifrostTuner.cpp index c0ebd24afe..edd074ba08 100644 --- a/src/runtime/CL/tuners/BifrostTuner.cpp +++ b/src/runtime/CL/tuners/BifrostTuner.cpp @@ -124,15 +124,163 @@ void tune_direct_convolution_kernel(CLDirectConvolutionLayerKernel &k) k.set_lws_hint(lws_hint); } } + +void tune_col2im_kernel(CLCol2ImKernel &k) +{ + cl::NDRange lws_hint = k.lws_hint(); + const GPUTarget gpu_target = k.get_target(); + + // Configure the local work size for Bifrost with a value obtained + // via exhaustive autotuning over 30 representative tensor shapes. + if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX)) + { + if((k._convolved_dims.first == 7) || (k._convolved_dims.first == 14)) + { + lws_hint = cl::NDRange(1, 7, 1); + } + else + { + lws_hint = cl::NDRange(1, 8, 1); + } + } + + k.set_lws_hint(lws_hint); +} + +void tune_im2col_kernel(CLIm2ColKernel &k) +{ + cl::NDRange lws_hint = k.lws_hint(); + const GPUTarget gpu_target = k.get_target(); + + // Local work size optimized for the 11x11 AlexNet convolution on Bifrost. + if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX) && k._kernel_dims.width == 11) + { + const bool is_square_kernel = (k._kernel_dims.width == k._kernel_dims.height); + if(!is_square_kernel && k._kernel_dims.width > 1 && !k._conv_info.has_padding()) + { + lws_hint = cl::NDRange(1, 1, 1); + } + } + k.set_lws_hint(lws_hint); +} + +void tune_depthwise_im2col_kernel(CLDepthwiseIm2ColKernel &k) +{ + cl::NDRange lws_hint = k.lws_hint(); + const GPUTarget gpu_target = k.get_target(); + + // Configure the local work size for Bifrost with a value obtained + // via exhaustive autotuning for the MobileNets tensor shapes. + if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX)) + { + lws_hint = cl::NDRange(1, 2, 1); + } + + k.set_lws_hint(lws_hint); +} + +void tune_gemv_kernel(CLGEMMMatrixVectorMultiplyKernel &k) +{ + cl::NDRange lws_hint = k.lws_hint(); + const GPUTarget gpu_target = k.get_target(); + + // Configure the local work size for Bifrost with a value obtained + // via exhaustive autotuning for the MobileNets tensor shapes. + if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX)) + { + lws_hint = cl::NDRange(1, 1, 1); + } + + k.set_lws_hint(lws_hint); +} + +void tune_gemm_kernel(CLGEMMMatrixMultiplyKernel &k) +{ + cl::NDRange lws_hint = k.lws_hint(); + const GPUTarget gpu_target = k.get_target(); + + // Configure LWS hint + switch(gpu_target) + { + case GPUTarget::G71: + case GPUTarget::G72: + case GPUTarget::G51: + case GPUTarget::G51BIG: + case GPUTarget::G51LIT: + case GPUTarget::TNOX: + if(k._input1->info()->dimension(1) == 24) + { + // LWS optimized for the 11x11 AlexNet convolution on Bifrost. + lws_hint = cl::NDRange(2, 2); + } + else if(k._output->info()->dimension(1) == 196) + { + lws_hint = cl::NDRange(1, 7); + } + else + { + lws_hint = cl::NDRange(8, 8); + } + break; + default: + lws_hint = cl::NullRange; + } + + k.set_lws_hint(lws_hint); +} + +void tune_pooling_kernel(CLPoolingLayerKernel &k) +{ + cl::NDRange lws_hint = k.lws_hint(); + const GPUTarget gpu_target = k.get_target(); + + // Configure the local work size (hint) from the first two dimensions of the global work size. + // On Bifrost, this works for up to 35x35xC filters, for which the pooling_layer_3_optimized + // kernel is launched with gws=(9, 33, C). In any case, the hint will be ignored if it is + // invalid (e.g. exceeds the maximum workgroup size that the kernel can be launched with). + if(k._input->info()->data_layout() == DataLayout::NCHW) + { + if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX)) + { + cl::NDRange gws = ICLKernel::gws_from_window(k.window()); + lws_hint = cl::NDRange(gws[0], gws[1], 1); + } + } + + k.set_lws_hint(lws_hint); +} } // namespace void BifrostTuner::tune_kernel_static(ICLKernel &kernel) { - // Continue on tuning if dynamic tuning if(dynamic_cast(&kernel) != nullptr) { tune_direct_convolution_kernel(*utils::cast::polymorphic_downcast(&kernel)); } + else if(dynamic_cast(&kernel) != nullptr) + { + tune_col2im_kernel(*utils::cast::polymorphic_downcast(&kernel)); + } + else if(dynamic_cast(&kernel) != nullptr) + { + tune_im2col_kernel(*utils::cast::polymorphic_downcast(&kernel)); + } + else if(dynamic_cast(&kernel) != nullptr) + { + tune_depthwise_im2col_kernel(*utils::cast::polymorphic_downcast(&kernel)); + } + else if(dynamic_cast(&kernel) != nullptr) + { + tune_gemv_kernel(*utils::cast::polymorphic_downcast(&kernel)); + } + else if(dynamic_cast(&kernel) != nullptr) + { + tune_gemm_kernel(*utils::cast::polymorphic_downcast(&kernel)); + } + else if(dynamic_cast(&kernel) != nullptr) + { + tune_pooling_kernel(*utils::cast::polymorphic_downcast(&kernel)); + } } void BifrostTuner::tune_kernel_dynamic(ICLKernel &kernel) -- cgit v1.2.1