Optimize Transposed Convolution for CL backend (FP32/16)
This patch optimizes transposed convolution for CL backend by rewriting it in a single kernel instead of three (flip_kernel + upsample + conv). The new kernel skips the upsampling step which reduces the input space of convolution by stride_x * stride_y, resulting in significant performance improvement. It also skips the kernel flipping by traversing the weights accordingly, thus reduces the memory footprint.
Resolves: COMPMID-5676
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Change-Id: I8a333212dc7c5f7f0597aa58b0d56d44814baa14
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8588
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/gpu/cl/ClKernelLibrary.cpp b/src/gpu/cl/ClKernelLibrary.cpp
index 0f08f5d..4e03639 100644
--- a/src/gpu/cl/ClKernelLibrary.cpp
+++ b/src/gpu/cl/ClKernelLibrary.cpp
@@ -448,6 +448,7 @@
{ "space_to_batch_nhwc", "nhwc/space_to_batch.cl" },
{ "space_to_batch_static_nhwc", "nhwc/space_to_batch.cl" },
{ "space_to_depth_nhwc", "nhwc/space_to_depth.cl" },
+ { "transposed_convolution_nhwc", "nhwc/transposed_convolution.cl" },
{ "upsample_layer_nhwc", "nhwc/upsample_layer.cl" },
{ "winograd_filter_transform_4x1_3x1_nhwc", "nhwc/winograd_filter_transform.cl" },
{ "winograd_filter_transform_1x4_1x3_nhwc", "nhwc/winograd_filter_transform.cl" },
@@ -943,6 +944,10 @@
#include "./cl_kernels/nhwc/space_to_depth.clembed"
},
{
+ "nhwc/transposed_convolution.cl",
+#include "./cl_kernels/nhwc/transposed_convolution.clembed"
+ },
+ {
"nhwc/winograd_filter_transform.cl",
#include "./cl_kernels/nhwc/winograd_filter_transform.clembed"
},