COMPMID-817: Tuner: Port kernels to new design.

Change-Id: Iaabb1153c2abe0400ec79d51a21347debe92d642
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/134062
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

6 files changed, 7 insertions, 91 deletions

diff --git a/src/core/CL/kernels/CLCol2ImKernel.cpp b/src/core/CL/kernels/CLCol2ImKernel.cpp
index e15da7258a..4e444206f1 100644
--- a/src/core/CL/kernels/CLCol2ImKernel.cpp
+++ b/src/core/CL/kernels/CLCol2ImKernel.cpp
@@ -110,21 +110,6 @@ void CLCol2ImKernel::configure(const ICLTensor *input, ICLTensor *output, std::p
 
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("col2im", build_opts.options()));
 
-    // Configure the local work size for Bifrost with a value obtained
-    // via exhaustive autotuning over 30 representative tensor shapes.
-    const GPUTarget gpu_target = get_target();
-    if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX))
-    {
-        if((_convolved_dims.first == 7) || (_convolved_dims.first == 14))
-        {
-            _lws_hint = cl::NDRange(1, 7, 1);
-        }
-        else
-        {
-            _lws_hint = cl::NDRange(1, 8, 1);
-        }
-    }
-
     // Configure kernel window
     auto win_config = validate_and_configure_window(input->info(), output->info(), _convolved_dims);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
diff --git a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
index 41ff2202ca..c89b16eedc 100644
--- a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
@@ -90,15 +90,6 @@ void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *outpu
 
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_im2col", build_opts.options()));
 
-    // Configure the local work size for Bifrost with a value obtained
-    // via exhaustive autotuning for the MobileNets tensor shapes.
-    const GPUTarget gpu_target = get_target();
-
-    if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX))
-    {
-        _lws_hint = cl::NDRange(1, 2, 1);
-    }
-
     // Configure  kernel window
     Window win = calculate_max_window(*output->info(), Steps());
     // CLDepthwiseIm2ColKernel doesn't need padding so update_window_and_padding() can be skipped
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
index 7a9760b778..fc52f4e124 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
@@ -194,51 +194,9 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen
     _output         = output;
     _slide_matrix_b = _input1->info()->num_dimensions() >= _input0->info()->num_dimensions();
 
-    const DataType data_type = input0->info()->data_type();
-    const int      fp_pos    = input0->info()->fixed_point_position();
-
-    // Get target architecture
-    GPUTarget gpu_target = get_target();
-
-    // Configure LWS hint
-    switch(gpu_target)
-    {
-        case GPUTarget::MIDGARD:
-        case GPUTarget::T600:
-        case GPUTarget::T700:
-        case GPUTarget::T800:
-            if(output->info()->dimension(1) == 196)
-            {
-                _lws_hint = cl::NDRange(1, 7);
-            }
-            else
-            {
-                _lws_hint = cl::NDRange(8, 8);
-            }
-            break;
-        case GPUTarget::G71:
-        case GPUTarget::G72:
-        case GPUTarget::G51:
-        case GPUTarget::G51BIG:
-        case GPUTarget::G51LIT:
-        case GPUTarget::TNOX:
-            if(input1->info()->dimension(1) == 24)
-            {
-                // LWS optimized for the 11x11 AlexNet convolution on Bifrost.
-                _lws_hint = cl::NDRange(2, 2);
-            }
-            else if(output->info()->dimension(1) == 196)
-            {
-                _lws_hint = cl::NDRange(1, 7);
-            }
-            else
-            {
-                _lws_hint = cl::NDRange(8, 8);
-            }
-            break;
-        default:
-            _lws_hint = cl::NullRange;
-    }
+    const DataType  data_type  = input0->info()->data_type();
+    const int       fp_pos     = input0->info()->fixed_point_position();
+    const GPUTarget gpu_target = get_target();
 
     ElementsProcessed num_elements_processed{};
 
diff --git a/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp
index 1d6f388def..d8ecd501b0 100644
--- a/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp
@@ -110,14 +110,6 @@ void CLGEMMMatrixVectorMultiplyKernel::configure(const ICLTensor *input0, const
         _kernel.setArg<int>(idx++, -_input1->info()->quantization_info().offset);
     }
 
-    // Configure the local work size for Bifrost with a value obtained
-    // via exhaustive autotuning for the MobileNets tensor shapes.
-    const GPUTarget gpu_target = get_target();
-    if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX))
-    {
-        _lws_hint = cl::NDRange(1, 1, 1);
-    }
-
     // Configure kernel window
     const unsigned int num_elems_read_per_iteration = 4;
 
diff --git a/src/core/CL/kernels/CLIm2ColKernel.cpp b/src/core/CL/kernels/CLIm2ColKernel.cpp
index 378456cde6..53a4dca9a3 100644
--- a/src/core/CL/kernels/CLIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLIm2ColKernel.cpp
@@ -61,7 +61,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, b
 } // namespace
 
 CLIm2ColKernel::CLIm2ColKernel()
-    : _input(nullptr), _output(nullptr), _convolved_dims(), _num_elems_processed_per_iteration(1), _run_func(nullptr), _kernel_dims()
+    : _input(nullptr), _output(nullptr), _conv_info(), _convolved_dims(), _num_elems_processed_per_iteration(1), _run_func(nullptr), _kernel_dims()
 {
 }
 
@@ -74,6 +74,7 @@ void CLIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const
 
     _input       = input;
     _output      = output;
+    _conv_info   = conv_info;
     _kernel_dims = kernel_dims;
 
     const DataType  data_type  = input->info()->data_type();
@@ -190,10 +191,9 @@ void CLIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const
                 {
                     vector_size = kernel_dims.width;
                 }
-                // Local work size and vector size optimized for the 11x11 AlexNet convolution on Bifrost.
+                // Vector size optimized for the 11x11 AlexNet convolution on Bifrost.
                 if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX) && kernel_dims.width == 11)
                 {
-                    _lws_hint   = cl::NDRange(1, 1, 1);
                     vector_size = 8;
                 }
                 const size_t width_mod_vector_size = kernel_dims.width % vector_size;
diff --git a/src/core/CL/kernels/CLPoolingLayerKernel.cpp b/src/core/CL/kernels/CLPoolingLayerKernel.cpp
index 3091df4665..b242c5550c 100644
--- a/src/core/CL/kernels/CLPoolingLayerKernel.cpp
+++ b/src/core/CL/kernels/CLPoolingLayerKernel.cpp
@@ -208,8 +208,7 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output,
     _output    = output;
     _pool_info = pool_info;
 
-    const GPUTarget gpu_target = get_target();
-    const DataType  data_type  = input->info()->data_type();
+    const DataType data_type = input->info()->data_type();
 
     // Set build options
     CLBuildOptions build_opts;
@@ -273,20 +272,11 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output,
     ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
     ICLKernel::configure(std::get<1>(win_config));
 
-    // Configure the local work size (hint) from the first two dimensions of the global work size.
-    // On Bifrost, this works for up to 35x35xC filters, for which the pooling_layer_3_optimized
-    // kernel is launched with gws=(9, 33, C). In any case, the hint will be ignored if it is
-    // invalid (e.g. exceeds the maximum workgroup size that the kernel can be launched with).
     if(data_layout == DataLayout::NCHW)
     {
         CLPoolingConfig pooling_config     = std::get<2>(win_config);
         _num_elems_processed_per_iteration = pooling_config.first;
         _border_size                       = pooling_config.second;
-        if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX))
-        {
-            cl::NDRange gws = ICLKernel::gws_from_window(std::get<1>(win_config));
-            _lws_hint       = cl::NDRange(gws[0], gws[1], 1);
-        }
     }
     else
     {