Rework OpenCL Depthwise Convolution

- Remove dedicated kernels for NCHW. Now we only use NHWC with permute
- Remove specialized kernels for 3x3 NHWC
- Simplify CLDepthwiseConvolutionLayer.cpp to call just the native
  implementation for both floating-point and quantized data types
- Develop two parametric opencl kernels for depthwise convolution layer NHWC
  (floating-point and quantized)
- Add support to export the weights to cl_image
- Extend test for depthwise convolution on opencl

Resolves COMPMID-4417

Change-Id: I253dd5d959a70783c82e62b1771a5e9f91621cb0
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5806
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>

1 files changed, 39 insertions, 0 deletions

diff --git a/src/core/CL/CLHelpers.cpp b/src/core/CL/CLHelpers.cpp
index 6af378c7ab..3323929742 100644
--- a/src/core/CL/CLHelpers.cpp
+++ b/src/core/CL/CLHelpers.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 #include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
 #include "arm_compute/core/CL/CLTypes.h"
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Log.h"
@@ -427,4 +428,42 @@ void set_wbsm(cl::Kernel &kernel, cl_int wbsm_hint)
     ARM_COMPUTE_ERROR_ON(err != CL_SUCCESS);
 }
 
+bool export_weights_to_cl_image(const ITensorInfo *tensor)
+{
+    if(tensor->tensor_shape()[0] % 4)
+    {
+        return false;
+    }
+
+    // If not floating point
+    if(!is_data_type_float(tensor->data_type()))
+    {
+        return false;
+    }
+
+    // Check if the cl_khr_image2d_from_buffer extension is supported on the target platform
+    if(!image2d_from_buffer_supported(CLKernelLibrary::get().get_device()))
+    {
+        return false;
+    }
+
+    // Check cl image pitch alignment
+    if(get_cl_image_pitch_alignment(CLKernelLibrary::get().get_device()) == 0)
+    {
+        return false;
+    }
+
+    const size_t image_w     = tensor->tensor_shape()[0] / 4;
+    const size_t image_h     = tensor->tensor_shape()[1] * tensor->tensor_shape()[2] * tensor->tensor_shape()[3];
+    const size_t max_image_w = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_WIDTH>();
+    const size_t max_image_h = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_HEIGHT>();
+
+    if(image_w > max_image_w || image_h > max_image_h)
+    {
+        return false;
+    }
+
+    return true;
+}
+
 } // namespace arm_compute