Optimize depthwise convolution on OpenCL

The optimization concerns the case where the depth multiplier is > 1.
The depth multiplier for loop has been removed from the OpenCL kernel
and the GWS has been mapped to the output shape. In this way, we can
still perform a tile with N0 columns and improve the performance of
depthwise conv over 80% when depth multiplier is > 1.

Resolves COMPMID-5568

Change-Id: I604e287d4eeb31c54b9cc6c3072a698cd0e3e136
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8184
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 19 insertions, 5 deletions

diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index e821726d0e..8546471fdd 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -54,19 +54,22 @@ bool export_weights_to_cl_image_heuristic(const ITensorInfo *weights, unsigned i
     const size_t kernel_w = weights->tensor_shape()[idx_w];
     const size_t kernel_h = weights->tensor_shape()[idx_h];
 
-    if((kernel_w == 1) && (kernel_h == 1))
+    if(gpu_target == GPUTarget::G71 || get_arch_from_target(gpu_target) == GPUTarget::MIDGARD)
     {
         return false;
     }
 
-    if(depth_multiplier > 1)
+    if((kernel_w == 1) && (kernel_h == 1))
     {
         return false;
     }
 
-    if(gpu_target == GPUTarget::G71 || get_arch_from_target(gpu_target) == GPUTarget::MIDGARD)
+    if(depth_multiplier > 1)
     {
-        return false;
+        if((depth_multiplier % 4) != 0)
+        {
+            return false;
+        }
     }
 
     return true;
@@ -110,7 +113,18 @@ void initialize_dwc_native_compute_info(DWCComputeKernelInfo &dwc_compute_info,
     }
     else
     {
-        dwc_compute_info.n0 = 1;
+        if((depth_multiplier % 4) == 0)
+        {
+            dwc_compute_info.n0 = 4;
+        }
+        else if((depth_multiplier % 2) == 0)
+        {
+            dwc_compute_info.n0 = 2;
+        }
+        else
+        {
+            dwc_compute_info.n0 = 1;
+        }
     }
 
     dwc_compute_info.n0 = adjust_vec_size(dwc_compute_info.n0, weights->dimension(0));