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>

10 files changed, 77 insertions, 26 deletions

diff --git a/src/gpu/cl/kernels/ClBatchConcatenateKernel.cpp b/src/gpu/cl/kernels/ClBatchConcatenateKernel.cpp
index 8c2af5ffb6..62040df63e 100644
--- a/src/gpu/cl/kernels/ClBatchConcatenateKernel.cpp
+++ b/src/gpu/cl/kernels/ClBatchConcatenateKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021 Arm Limited.
+ * Copyright (c) 2019-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -91,8 +91,13 @@ void ClBatchConcatenateKernel::configure(const CLCompileContext &compile_context
         build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
     }
 
+    std::string kernel_name = "concatenate";
+
+    // 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", build_opts.options());
+    _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
 
     // Configure kernel window
     auto win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration));
diff --git a/src/gpu/cl/kernels/ClDepthConcatenateKernel.cpp b/src/gpu/cl/kernels/ClDepthConcatenateKernel.cpp
index d716f1e430..9704294d62 100644
--- a/src/gpu/cl/kernels/ClDepthConcatenateKernel.cpp
+++ b/src/gpu/cl/kernels/ClDepthConcatenateKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2021 Arm Limited.
+ * Copyright (c) 2017-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -91,8 +91,13 @@ void ClDepthConcatenateKernel::configure(const CLCompileContext &compile_context
         build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
     }
 
+    std::string kernel_name = "concatenate";
+
+    // 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", build_opts.options());
+    _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
 
     // Configure kernel window
     auto win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration));
diff --git a/src/gpu/cl/kernels/ClHeightConcatenateKernel.cpp b/src/gpu/cl/kernels/ClHeightConcatenateKernel.cpp
index 688308098a..95eca69bde 100644
--- a/src/gpu/cl/kernels/ClHeightConcatenateKernel.cpp
+++ b/src/gpu/cl/kernels/ClHeightConcatenateKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021 Arm Limited.
+ * Copyright (c) 2019-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -89,7 +89,6 @@ void ClHeightConcatenateKernel::configure(const CLCompileContext &compile_contex
     build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(src->element_size()));
     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
     build_opts.add_option("-DHEIGHT_OFFSET=" + support::cpp11::to_string(_height_offset));
-    build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(src->dimension(2)));
     build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration));
 
     if(is_data_type_quantized_asymmetric(src->data_type()) && src->quantization_info() != dst->quantization_info())
@@ -102,9 +101,15 @@ void ClHeightConcatenateKernel::configure(const CLCompileContext &compile_contex
         build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale));
         build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
     }
+    _depth = src->dimension(2);
+
+    std::string kernel_name = "concatenate_height";
+
+    // 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_height", build_opts.options());
+    _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
     // Configure kernel window
 
     // The window needs to be based on src as we copy all the heights of src
@@ -125,6 +130,7 @@ void ClHeightConcatenateKernel::run_op(ITensorPack &tensors, const Window &windo
     unsigned int idx = 0;
     add_4D_tensor_argument(idx, src, window);
     add_4D_tensor_argument(idx, dst, window);
+    _kernel.setArg<cl_uint>(idx++, _depth);
     enqueue(queue, *this, window, lws_hint());
 }
 } // namespace kernels
diff --git a/src/gpu/cl/kernels/ClHeightConcatenateKernel.h b/src/gpu/cl/kernels/ClHeightConcatenateKernel.h
index 1e544d3025..d3c077fc22 100644
--- a/src/gpu/cl/kernels/ClHeightConcatenateKernel.h
+++ b/src/gpu/cl/kernels/ClHeightConcatenateKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -64,6 +64,7 @@ public:
 
 private:
     unsigned int _height_offset;
+    int32_t      _depth{ 0 };
 };
 } // namespace kernels
 } // namespace opencl
diff --git a/src/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.cpp b/src/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.cpp
index 6e7b7f6e14..b04a80a1e9 100644
--- a/src/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.cpp
+++ b/src/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2021 Arm Limited.
+ * Copyright (c) 2018-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -89,9 +89,6 @@ void ClWidthConcatenate2TensorsKernel::configure(const CLCompileContext &compile
     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("-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));
 
@@ -111,8 +108,16 @@ void ClWidthConcatenate2TensorsKernel::configure(const CLCompileContext &compile
         build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
     }
 
+    _depth        = src1->dimension(2);
+    _input1_width = src1->dimension(0);
+
+    std::string kernel_name = "concatenate_width_x2";
+
+    // 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_x2", 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));
@@ -150,6 +155,8 @@ void ClWidthConcatenate2TensorsKernel::run_op(ITensorPack &tensors, const Window
         add_4D_tensor_argument(idx, src0, slice);
         add_4D_tensor_argument(idx, src1, slice);
         add_4D_tensor_argument(idx, dst, slice);
+        _kernel.setArg<cl_int>(idx++, _depth);
+        _kernel.setArg<cl_int>(idx++, _input1_width);
         enqueue(queue, *this, window, lws_hint());
     }
     while(window.slide_window_slice_4D(slice));
diff --git a/src/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.h b/src/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.h
index 8065fb9f75..5c54479002 100644
--- a/src/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.h
+++ b/src/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -60,6 +60,10 @@ public:
 
     // Inherited methods overridden:
     void run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) override;
+
+private:
+    int32_t _depth{ 0 };
+    int32_t _input1_width{ 0 };
 };
 } // namespace kernels
 } // namespace opencl
diff --git a/src/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.cpp b/src/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.cpp
index a08490c565..741637795a 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 @@ void ClWidthConcatenate4TensorsKernel::configure(const CLCompileContext &compile
     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 @@ void ClWidthConcatenate4TensorsKernel::configure(const CLCompileContext &compile
         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 @@ void ClWidthConcatenate4TensorsKernel::run_op(ITensorPack &tensors, const Window
         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));
diff --git a/src/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.h b/src/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.h
index 80afb3b85d..baf8d381be 100644
--- a/src/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.h
+++ b/src/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -63,6 +63,12 @@ public:
 
     // Inherited methods overridden:
     void run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) override;
+
+private:
+    int32_t _depth{ 0 };
+    int32_t _input1_width{ 0 };
+    int32_t _input2_width{ 0 };
+    int32_t _input3_width{ 0 };
 };
 } // namespace kernels
 } // namespace opencl
diff --git a/src/gpu/cl/kernels/ClWidthConcatenateKernel.cpp b/src/gpu/cl/kernels/ClWidthConcatenateKernel.cpp
index 88b5a5e334..7ed609f08d 100644
--- a/src/gpu/cl/kernels/ClWidthConcatenateKernel.cpp
+++ b/src/gpu/cl/kernels/ClWidthConcatenateKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2021 Arm Limited.
+ * Copyright (c) 2018-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -87,7 +87,6 @@ void ClWidthConcatenateKernel::configure(const CLCompileContext &compile_context
     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(src->dimension(0) % num_elems_processed_per_iteration));
     build_opts.add_option("-DWIDTH_OFFSET=" + support::cpp11::to_string(width_offset));
-    build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(src->dimension(2)));
 
     if(is_data_type_quantized_asymmetric(src->data_type()) && src->quantization_info() != dst->quantization_info())
     {
@@ -99,9 +98,15 @@ void ClWidthConcatenateKernel::configure(const CLCompileContext &compile_context
         build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iqinfo.scale));
         build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oqinfo.scale));
     }
+    _depth                  = src->dimension(2);
+    std::string kernel_name = "concatenate_width";
+
+    // 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", build_opts.options());
+    _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
+
     // Configure kernel window
     Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration));
     ICLKernel::configure_internal(win.collapse(win, Window::DimZ));
@@ -120,6 +125,7 @@ void ClWidthConcatenateKernel::run_op(ITensorPack &tensors, const Window &window
     unsigned int idx = 0;
     add_4D_tensor_argument(idx, src, window);
     add_4D_tensor_argument(idx, dst, window);
+    _kernel.setArg<cl_uint>(idx++, _depth);
     enqueue(queue, *this, window, lws_hint());
 }
 } // namespace kernels
diff --git a/src/gpu/cl/kernels/ClWidthConcatenateKernel.h b/src/gpu/cl/kernels/ClWidthConcatenateKernel.h
index 71df077ada..3ace4400e6 100644
--- a/src/gpu/cl/kernels/ClWidthConcatenateKernel.h
+++ b/src/gpu/cl/kernels/ClWidthConcatenateKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -61,6 +61,9 @@ public:
 
     // Inherited methods overridden:
     void run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) override;
+
+private:
+    int32_t _depth{ 0 };
 };
 } // namespace kernels
 } // namespace opencl