commit 89aa4eb56d56c81a9d53f94dffa5fa88742e986c
author ramelg01 <ramy.elgammal@arm.com> Tue Feb 08 23:01:31 2022 +0000
committer Ramy Elgammal <ramy.elgammal@arm.com> Fri Feb 11 11:01:10 2022 +0000
tree 64ac3cb37d44fcfb8cf7add9100a8f0230a51d8f
parent 2134d1bdb81e4959560d5becea06c43c083a9811

Improve start-up time for concatenation layers

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

Resolves: COMPMID-5121
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: I76b7c0cf56d803f58ebff5494c904ace2a86ef5a
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7097
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/ClWidthConcatenate4TensorsKernel.cpp

diff --git a/src/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.cpp b/src/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.cpp
index a08490c..7416377 100644
--- a/src/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.cpp
+++ b/src/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2021 Arm Limited.
+ * Copyright (c) 2018-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -93,16 +93,16 @@
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src1->data_type()));
     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
     build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_leftover));
-    build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(src1->dimension(2)));
-    build_opts.add_option("-DINPUT1_WIDTH=" + support::cpp11::to_string(src1->dimension(0)));
-    build_opts.add_option("-DINPUT2_WIDTH=" + support::cpp11::to_string(src2->dimension(0)));
-    build_opts.add_option("-DINPUT3_WIDTH=" + support::cpp11::to_string(src3->dimension(0)));
-    build_opts.add_option("-DINPUT4_WIDTH=" + support::cpp11::to_string(src4->dimension(0)));
     build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(src1->element_size()));
     build_opts.add_option("-DINPUT1_ROTATE_N=" + support::cpp11::to_string((src1->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration));
     build_opts.add_option("-DINPUT2_ROTATE_N=" + support::cpp11::to_string((src1->dimension(0) + src2->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration));
     build_opts.add_option("-DINPUT3_ROTATE_N=" + support::cpp11::to_string((src1->dimension(0) + src2->dimension(0) + src3->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration));
 
+    _depth        = src1->dimension(2);
+    _input1_width = src1->dimension(0);
+    _input2_width = src2->dimension(0);
+    _input3_width = src3->dimension(0);
+
     // If soources have different quantization info set quantization parameters needed for the re-quantization process
     const bool have_different_qinfo = helpers::tensor_info::tensors_have_different_quantization_info(dst, src1, src2, src3, src4);
     if(is_data_type_quantized_asymmetric(src1->data_type()) && have_different_qinfo)
@@ -124,9 +124,13 @@
         build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset));
         build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
     }
+    std::string kernel_name = "concatenate_width_x4";
+
+    // A macro guard to compile ONLY the kernel of interest
+    build_opts.add_option("-D" + upper_string(kernel_name));
 
     // Create kernel
-    _kernel = create_kernel(compile_context, "concatenate_width_x4", build_opts.options());
+    _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
 
     // Configure kernel window
     Window win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration));
@@ -176,6 +180,10 @@
         add_4D_tensor_argument(idx, src2, slice);
         add_4D_tensor_argument(idx, src3, slice);
         add_4D_tensor_argument(idx, dst, slice);
+        _kernel.setArg<cl_int>(idx++, _depth);
+        _kernel.setArg<cl_int>(idx++, _input1_width);
+        _kernel.setArg<cl_int>(idx++, _input2_width);
+        _kernel.setArg<cl_int>(idx++, _input3_width);
         enqueue(queue, *this, window, lws_hint());
     }
     while(window.slide_window_slice_4D(slice));