Rework DepthwiseConvolution heuristic on OpenCL

Resolves COMPMID-5632

Change-Id: I2bdbe69a610ca2510fbd74d5d412842679299762
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8365
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 31 insertions, 6 deletions

diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index 8546471fdd..3eadaee0de 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -44,7 +44,7 @@ namespace
 {
 bool export_weights_to_cl_image_heuristic(const ITensorInfo *weights, unsigned int depth_multiplier, GPUTarget gpu_target)
 {
-    if(!export_weights_to_cl_image(weights))
+    if(!export_to_cl_image(weights))
     {
         return false;
     }
@@ -75,9 +75,12 @@ bool export_weights_to_cl_image_heuristic(const ITensorInfo *weights, unsigned i
     return true;
 }
 
-void initialize_dwc_native_compute_info(DWCComputeKernelInfo &dwc_compute_info, const ITensorInfo *weights, const PadStrideInfo &conv_info, const Size2D &dilation, unsigned int depth_multiplier,
+void initialize_dwc_native_compute_info(DWCComputeKernelInfo &dwc_compute_info, const ITensorInfo *input, const ITensorInfo *weights, const PadStrideInfo &conv_info, const Size2D &dilation,
+                                        unsigned int depth_multiplier,
                                         GPUTarget gpu_target)
 {
+    ARM_COMPUTE_UNUSED(input);
+
     if(!is_data_type_float(weights->data_type()))
     {
         dwc_compute_info.export_weights_to_cl_image = false;
@@ -97,6 +100,7 @@ void initialize_dwc_native_compute_info(DWCComputeKernelInfo &dwc_compute_info,
     // Floating point path
 
     // First check if we can export to cl_image.
+    dwc_compute_info.export_input_to_cl_image   = false;
     dwc_compute_info.export_weights_to_cl_image = export_weights_to_cl_image_heuristic(weights, depth_multiplier, gpu_target);
 
     // Set n0
@@ -135,7 +139,28 @@ void initialize_dwc_native_compute_info(DWCComputeKernelInfo &dwc_compute_info,
         const size_t idx_w    = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::WIDTH);
         const size_t kernel_w = weights->tensor_shape()[idx_w];
 
-        dwc_compute_info.m0 = (kernel_w >= 9) || (kernel_w == 1) ? 1 : 2;
+        if((kernel_w >= 9) || (kernel_w == 1))
+        {
+            dwc_compute_info.m0 = 1;
+        }
+        else
+        {
+            if(weights->data_type() == DataType::F16)
+            {
+                if((input->dimension(1) % 5) == 0)
+                {
+                    dwc_compute_info.m0 = 5;
+                }
+                else
+                {
+                    dwc_compute_info.m0 = 4;
+                }
+            }
+            else
+            {
+                dwc_compute_info.m0 = 2;
+            }
+        }
     }
     else
     {
@@ -237,7 +262,7 @@ void CLDepthwiseConvolutionLayer::configure(const CLCompileContext &compile_cont
     }
 
     DWCComputeKernelInfo dwc_native_compute_info;
-    initialize_dwc_native_compute_info(dwc_native_compute_info, weights_to_use->info(), conv_info, dilation, depth_multiplier, gpu_target);
+    initialize_dwc_native_compute_info(dwc_native_compute_info, input->info(), weights_to_use->info(), conv_info, dilation, depth_multiplier, gpu_target);
 
     const ConvolutionInfo conv_kernel_info{ conv_info, depth_multiplier, act_info, dilation };
 
@@ -322,7 +347,7 @@ Status CLDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITe
         ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(weights, &permuted_weights, PermutationVector(2U, 0U, 1U)));
 
         DWCComputeKernelInfo dwc_native_compute_info;
-        initialize_dwc_native_compute_info(dwc_native_compute_info, &permuted_weights, conv_info, dilation, depth_multiplier, gpu_target);
+        initialize_dwc_native_compute_info(dwc_native_compute_info, input, &permuted_weights, conv_info, dilation, depth_multiplier, gpu_target);
 
         ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayerNativeKernel::validate(&permuted_input, &permuted_weights, biases, &permuted_output,
                                                                                       dwc_native_compute_info, conv_kernel_info, &output_multipliers_shifts_info, &output_multipliers_shifts_info));
@@ -331,7 +356,7 @@ Status CLDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITe
     else
     {
         DWCComputeKernelInfo dwc_native_compute_info;
-        initialize_dwc_native_compute_info(dwc_native_compute_info, weights, conv_info, dilation, depth_multiplier, gpu_target);
+        initialize_dwc_native_compute_info(dwc_native_compute_info, input, weights, conv_info, dilation, depth_multiplier, gpu_target);
         ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayerNativeKernel::validate(input, weights, biases, output, dwc_native_compute_info, conv_kernel_info, &output_multipliers_shifts_info,
                                                                                       &output_multipliers_shifts_info));
     }