commit ab5b1a279284bed350d3bb75f3d9d3aec6edca0e
author Sang-Hoon Park <sang-hoon.park@arm.com> Tue Oct 15 09:29:13 2019 +0100
committer Sang-Hoon Park <sang-hoon.park@arm.com> Mon Oct 21 08:21:47 2019 +0000
tree f77d077dea99d2969d5065d25b9141ff70e0ea90
parent 422da26b44d517f38011a91888a1c5da4acfeed4

COMPMID-2744 [CL] add support for 9x9 quantized direct convolution

Change-Id: I858ce5b9a530f8568e154f5d724d267e142ef9b2
Signed-off-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2091
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>

src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp

diff --git a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
index 10119d8..7b74a5a 100644
--- a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
+++ b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
@@ -56,14 +56,23 @@
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(width_idx) != weights->dimension(height_idx), "Weights should have same width and height");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(width_idx) != 1 && weights->dimension(width_idx) != 3 && weights->dimension(width_idx) != 5 && weights->dimension(width_idx) != 9,
                                     "Kernel sizes other than 1x1, 3x3, 5x5 or 9x9 are not supported");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(width_idx) == 9 && input->data_type() == DataType::QASYMM8, "Kernel sizes of 9x9 is not supported for quantized types");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(channel_idx) != input->dimension(channel_idx),
                                     "Weights feature map dimension should match the respective input's one");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->num_dimensions() > 4, "Weights can be at most 4 dimensional");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(width_idx) == 1) && std::get<0>(conv_info.stride()) > 3, "Strides larger than 3 not supported for 1x1 convolution.");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(width_idx) == 3 || weights->dimension(width_idx) == 5) && std::get<0>(conv_info.stride()) > 2,
                                     "Strides larger than 2 not supported for 3x3 convolution.");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(width_idx) == 9) && data_layout == DataLayout::NCHW, "Only NHWC layout is supported for 9x9 convolution.");
+
+    const auto data_type = input->data_type();
+
+    if(weights->dimension(width_idx) == 9)
+    {
+        const auto supported_data_layout = is_data_type_quantized(data_type) ? DataLayout::NCHW : DataLayout::NHWC;
+        const auto error_message         = std::string("Only " + string_from_data_layout(supported_data_layout) + " layout is supported for 9x9 convolution with " + string_from_data_type(
+                                                           data_type) + " type");
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG((supported_data_layout != data_layout), error_message.c_str());
+    }
 
     if(biases != nullptr)
     {
@@ -226,6 +235,19 @@
                         ARM_COMPUTE_ERROR("Invalid convolution stride X");
                 }
                 break;
+            case 9:
+                switch(conv_stride_x)
+                {
+                    case 1:
+                        num_elems_read_per_iteration_x = 16;
+                        break;
+                    case 2:
+                        num_elems_read_per_iteration_x = 24;
+                        break;
+                    default:
+                        ARM_COMPUTE_ERROR("Invalid convolution stride X");
+                }
+                break;
             default:
                 ARM_COMPUTE_ERROR("Invalid direct convolution size");
         }
@@ -487,7 +509,7 @@
         }
         build_options.add_option(std::string("-DDATA_TYPE_PROMOTED=" + get_cl_type_from_data_type(data_type)));
         // Create kernel
-        _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(is_quantized_asymm ? "direct_convolution_1x1_3x3_5x5_quantized" : kernel_name.str(),
+        _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(is_quantized_asymm ? "direct_convolution_quantized" : kernel_name.str(),
                                                                                build_options.options()));
     }