commit ba8690b4f0f82dfb66e3da819361e4032e9fa4db
author ramelg01 <ramy.elgammal@arm.com> Fri Feb 04 11:37:00 2022 +0000
committer Ramy Elgammal <ramy.elgammal@arm.com> Tue Feb 08 22:51:40 2022 +0000
tree 1a882d2ddae128b883b4b3560a28fb53272f1707
parent d56d94dff3b6fb1cc1807df19a8ead08c7f0faae

Improve start-up time for winograd_filter_transform_*_nhwc

 - pass tensor's dimensions at runtime rather than compile time
 - Add guard macro to compile only kernel of internest

Resolves: COMPMID-5118
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: Ie42c3c07fdd817ce62e7cad354381bc22c6e9264
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7058
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

src/gpu/cl/kernels/ClWinogradFilterTransformKernel.cpp

diff --git a/src/gpu/cl/kernels/ClWinogradFilterTransformKernel.cpp b/src/gpu/cl/kernels/ClWinogradFilterTransformKernel.cpp
index 4ba6ba8..136376a 100644
--- a/src/gpu/cl/kernels/ClWinogradFilterTransformKernel.cpp
+++ b/src/gpu/cl/kernels/ClWinogradFilterTransformKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2021 Arm Limited.
+ * Copyright (c) 2018-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -108,7 +108,16 @@
 
     // Set build options
     CLBuildOptions build_opts;
-    build_opts.add_option("-DSRC_DIM_Z=" + support::cpp11::to_string(src->dimension(2)));
+
+    // For NHWC layouts pass tensor dimesions at runtime
+    if(src->data_layout() == DataLayout::NHWC)
+    {
+        _src_dim_z = src->dimension(2);
+    }
+    else
+    {
+        build_opts.add_option("-DSRC_DIM_Z=" + support::cpp11::to_string(src->dimension(2)));
+    }
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type()));
     build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL");
     build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_FILTER_TRANSFORM_VERTICAL");
@@ -117,7 +126,10 @@
 
     // Create kernel
     std::string kernel_name = "winograd_filter_transform_" + output_tile_size.to_string() + "_" + kernel_size.to_string() + "_" + lower_string(string_from_data_layout(src->data_layout()));
-    _kernel                 = create_kernel(compile_context, kernel_name, build_opts.options());
+
+    // A macro guard to compile ONLY the kernel of interest
+    build_opts.add_option("-D" + upper_string(kernel_name));
+    _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
 
     // Configure kernel window
     auto win_config = validate_and_configure_window(src, dst);
@@ -149,8 +161,12 @@
     unsigned int idx = 0;
     add_4D_tensor_argument(idx, src, window);
     add_3D_tensor_argument(idx, dst, window_out);
+    if(src->info()->data_layout() == DataLayout::NHWC)
+    {
+        _kernel.setArg<cl_uint>(idx++, _src_dim_z);
+    }
     enqueue(queue, *this, window, lws_hint());
 }
 } // namespace kernels
 } // namespace opencl
-} // namespace arm_compute
\ No newline at end of file
+} // namespace arm_compute

src/gpu/cl/kernels/ClWinogradFilterTransformKernel.h

diff --git a/src/gpu/cl/kernels/ClWinogradFilterTransformKernel.h b/src/gpu/cl/kernels/ClWinogradFilterTransformKernel.h
index fe0c3da..b213030 100644
--- a/src/gpu/cl/kernels/ClWinogradFilterTransformKernel.h
+++ b/src/gpu/cl/kernels/ClWinogradFilterTransformKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2021 Arm Limited.
+ * Copyright (c) 2018-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -70,6 +70,9 @@
 
     // Inherited methods overridden:
     void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    int32_t _src_dim_z{ 0 };
 };
 } // namespace kernels
 } // namespace opencl