COMPMID-661 Add Bifrost lws heuristics for several depthwise_convolution kernels #49
Change-Id: Ibfa1c1cc9fc8501b22a18ecd519758f4aeb301eb
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110880
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
index c239414..ad9ac0e 100644
--- a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
@@ -73,6 +73,14 @@
}
_kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_im2col", build_opts));
+ // Configure the local work size for Bifrost with a value obtained
+ // via exhaustive autotuning for the MobileNets tensor shapes.
+ const GPUTarget gpu_target = get_arch_from_target(get_target());
+ if(gpu_target == GPUTarget::BIFROST)
+ {
+ _lws_hint = cl::NDRange(1, 2, 1);
+ }
+
// Configure kernel window
Window win = calculate_max_window(*input->info(), Steps());
// The CLDepthwiseIm2ColKernel doesn't need padding so update_window_and_padding() can be skipped
@@ -105,7 +113,7 @@
unsigned int idx = 0;
add_3D_tensor_argument(idx, _input, slice_in);
add_3D_tensor_argument(idx, _output, slice);
- enqueue(queue, *this, slice);
+ enqueue(queue, *this, slice, _lws_hint);
}
while(window.slide_window_slice_3D(slice) && window.slide_window_slice_3D(slice_in));
}