commit 51d7119d3e7fb9f2b387632270f99270b3cd1163
author Gian Marco Iodice <gianmarco.iodice@arm.com> Wed Feb 16 14:41:28 2022 +0000
committer Gian Marco Iodice <gianmarco.iodice@arm.com> Mon Feb 21 10:31:13 2022 +0000
tree 3f24ccf4d05347bcbb5fbf47326de674158835cf
parent 5e99318e4378b1f151fc85cd241adf7b222a088c

Fix performance regression on Arm(R) Mali(TM)-G71

Resolves COMPMID-5095

Change-Id: Ib7f6e891095372f870fc08833cca922909506dad
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7158
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

src/gpu/cl/operators/ClConv2d.cpp

diff --git a/src/gpu/cl/operators/ClConv2d.cpp b/src/gpu/cl/operators/ClConv2d.cpp
index 71f8378..23c1b8a 100644
--- a/src/gpu/cl/operators/ClConv2d.cpp
+++ b/src/gpu/cl/operators/ClConv2d.cpp
@@ -261,6 +261,7 @@
                 const bool  is_ifm_ge_16       = src->dimension(idx_c) >= 16;
                 const bool  is_ofm_lte_8       = weights->dimension(3U) <= 8;
                 const bool  workload_gte_8192  = (output_shape[0] * output_shape[1] * output_shape[2]) / 16 >= 8192;
+                const bool is_ifm_gt_ofm       = src->dimension(idx_c) > weights->dimension(3U);
 
                 // Run Winograd if valid and IFM >= 16
                 if(is_wino_valid && is_ifm_ge_16)
@@ -271,9 +272,21 @@
                 // Direct convolution case
                 if(is_direct_valid)
                 {
-                    if((is_large_kernel_sz && workload_gte_8192 && is_ifm_ge_16) || (is_ofm_lte_8 && is_ifm_ge_16))
+                    if((gpu_target == arm_compute::GPUTarget::G71 ||
+                       gpu_target == arm_compute::GPUTarget::G72 ||
+                       gpu_target == arm_compute::GPUTarget::MIDGARD))
                     {
-                        return ConvolutionMethod::DIRECT;
+                        if(is_large_kernel_sz && is_ifm_ge_16 && is_ifm_gt_ofm)
+                        {
+                            return ConvolutionMethod::DIRECT;
+                        }
+                    }
+                    else
+                    {
+                        if((is_large_kernel_sz && workload_gte_8192 && is_ifm_ge_16) || (is_ofm_lte_8 && is_ifm_ge_16))
+                        {
+                            return ConvolutionMethod::DIRECT;
+                        }
                     }
                 }