COMPMID-661 Add Bifrost lws heuristics for several depthwise_convolution kernels #49

Change-Id: Ibfa1c1cc9fc8501b22a18ecd519758f4aeb301eb
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110880
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

3 files changed, 57 insertions, 20 deletions

diff --git a/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp
index 63586b0f0f..be8fae2885 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp
@@ -93,9 +93,47 @@ void CLDepthwiseConvolution3x3Kernel::configure(const ICLTensor *input, const IC
     build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(_conv_stride_x));
     build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
 
+    // Create kernel
     std::string kernel_name = is_data_type_quantized_asymmetric(_input->info()->data_type()) ? "depthwise_convolution_3x3_quantized" : "depthwise_convolution_3x3";
     _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
+    // Set static arguments
+    if(is_data_type_quantized_asymmetric(_input->info()->data_type()))
+    {
+        float multiplier        = _input->info()->quantization_info().scale * _weights->info()->quantization_info().scale / _output->info()->quantization_info().scale;
+        int   output_multiplier = 0;
+        int   output_shift      = 0;
+        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+
+        unsigned int idx = 3 * num_arguments_per_3D_tensor() + ((_biases != nullptr) ? num_arguments_per_1D_tensor() : 0);
+
+        _kernel.setArg(idx++, -_input->info()->quantization_info().offset);
+        _kernel.setArg(idx++, -_weights->info()->quantization_info().offset);
+        _kernel.setArg(idx++, _output->info()->quantization_info().offset);
+        _kernel.setArg(idx++, output_multiplier);
+        _kernel.setArg(idx++, output_shift);
+    }
+
+    // Configure the local work size for Bifrost with a value obtained
+    // via exhaustive autotuning for the MobileNets tensor shapes.
+    const GPUTarget gpu_target = get_arch_from_target(get_target());
+    if(gpu_target == GPUTarget::BIFROST)
+    {
+        const size_t width = input->info()->dimension(0);
+        if(width >= 56) // 56 or 112
+        {
+            _lws_hint = cl::NDRange(8, 5, 2);
+        }
+        else if(width >= 14) // 14 or 28
+        {
+            _lws_hint = cl::NDRange(1, 5, 2);
+        }
+        else // 7
+        {
+            _lws_hint = cl::NDRange(1, 1, 2);
+        }
+    }
+
     // Configure kernel window
     const unsigned int num_elems_processed_per_iteration = 2;
     const unsigned int num_elems_written_per_iteration   = 2;
@@ -113,23 +151,6 @@ void CLDepthwiseConvolution3x3Kernel::configure(const ICLTensor *input, const IC
     output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
 
     ICLKernel::configure(win);
-
-    // Set static arguments
-    if(is_data_type_quantized_asymmetric(_input->info()->data_type()))
-    {
-        float multiplier        = _input->info()->quantization_info().scale * _weights->info()->quantization_info().scale / _output->info()->quantization_info().scale;
-        int   output_multiplier = 0;
-        int   output_shift      = 0;
-        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-
-        unsigned int idx = 3 * num_arguments_per_3D_tensor() + ((_biases != nullptr) ? num_arguments_per_1D_tensor() : 0);
-
-        _kernel.setArg(idx++, -_input->info()->quantization_info().offset);
-        _kernel.setArg(idx++, -_weights->info()->quantization_info().offset);
-        _kernel.setArg(idx++, _output->info()->quantization_info().offset);
-        _kernel.setArg(idx++, output_multiplier);
-        _kernel.setArg(idx++, output_shift);
-    }
 }
 
 void CLDepthwiseConvolution3x3Kernel::run(const Window &window, cl::CommandQueue &queue)
@@ -166,7 +187,7 @@ void CLDepthwiseConvolution3x3Kernel::run(const Window &window, cl::CommandQueue
         add_3D_tensor_argument(idx, _output, slice_out);
         add_3D_tensor_argument(idx, _weights, slice_weights);
 
-        enqueue(queue, *this, slice_out);
+        enqueue(queue, *this, slice_out, _lws_hint);
     }
     while(window.slide_window_slice_3D(slice_out) && win_in.slide_window_slice_3D(slice_in));
 }
diff --git a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
index c23941426e..ad9ac0ecd6 100644
--- a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
@@ -73,6 +73,14 @@ void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *outpu
     }
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_im2col", build_opts));
 
+    // Configure the local work size for Bifrost with a value obtained
+    // via exhaustive autotuning for the MobileNets tensor shapes.
+    const GPUTarget gpu_target = get_arch_from_target(get_target());
+    if(gpu_target == GPUTarget::BIFROST)
+    {
+        _lws_hint = cl::NDRange(1, 2, 1);
+    }
+
     // Configure  kernel window
     Window win = calculate_max_window(*input->info(), Steps());
     // The CLDepthwiseIm2ColKernel doesn't need padding so update_window_and_padding() can be skipped
@@ -105,7 +113,7 @@ void CLDepthwiseIm2ColKernel::run(const Window &window, cl::CommandQueue &queue)
         unsigned int idx = 0;
         add_3D_tensor_argument(idx, _input, slice_in);
         add_3D_tensor_argument(idx, _output, slice);
-        enqueue(queue, *this, slice);
+        enqueue(queue, *this, slice, _lws_hint);
     }
     while(window.slide_window_slice_3D(slice) && window.slide_window_slice_3D(slice_in));
 }
diff --git a/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp
index 70af5d63cf..951bc144aa 100644
--- a/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp
@@ -63,6 +63,14 @@ void CLGEMMMatrixVectorMultiplyKernel::configure(const ICLTensor *input0, const
 
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_mv", build_opts));
 
+    // Configure the local work size for Bifrost with a value obtained
+    // via exhaustive autotuning for the MobileNets tensor shapes.
+    const GPUTarget gpu_target = get_arch_from_target(get_target());
+    if(gpu_target == GPUTarget::BIFROST)
+    {
+        _lws_hint = cl::NDRange(1, 1, 1);
+    }
+
     // Configure kernel window
     const unsigned int num_elems_read_per_iteration = 4;
 
@@ -119,7 +127,7 @@ void CLGEMMMatrixVectorMultiplyKernel::run(const Window &window, cl::CommandQueu
         unsigned int idx_2 = num_arguments_per_3D_tensor() + num_arguments_per_2D_tensor();
         add_3D_tensor_argument(idx_0, _input0, slice_in);
         add_1D_tensor_argument(idx_2, _output, slice_out);
-        enqueue(queue, *this, slice_in);
+        enqueue(queue, *this, slice_in, _lws_hint);
     }
     while(window.slide_window_slice_3D(slice_in) && window.slide_window_slice_3D(slice_out));
 }