COMPMID-1581: Collapse windows

Change-Id: Iec56c9a96d9736a63f13b65efa33311950f20661
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/148572
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: bsgcomp <bsgcomp@arm.com>

3 files changed, 34 insertions, 27 deletions

diff --git a/src/core/CL/kernels/CLCol2ImKernel.cpp b/src/core/CL/kernels/CLCol2ImKernel.cpp
index 74bbb9b4df..d748745999 100644
--- a/src/core/CL/kernels/CLCol2ImKernel.cpp
+++ b/src/core/CL/kernels/CLCol2ImKernel.cpp
@@ -106,7 +106,7 @@ void CLCol2ImKernel::configure(const ICLTensor *input, ICLTensor *output, const
     build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(input->info()->element_size()));
     build_opts.add_option("-DWIDTH_INPUT=" + support::cpp11::to_string(input->info()->dimension(0)));
     build_opts.add_option("-DWIDTH_OUTPUT=" + support::cpp11::to_string(_convolved_dims.width));
-    build_opts.add_option_if(num_groups > 1, "-DGROUPING");
+    build_opts.add_option("-DNUM_GROUPS=" + support::cpp11::to_string(num_groups));
 
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("col2im", build_opts.options()));
 
@@ -143,22 +143,26 @@ void CLCol2ImKernel::run(const Window &window, cl::CommandQueue &queue)
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
 
+    bool is_collapsed     = false;
+    bool is_collapsed_out = false;
+
     Window out_window;
     out_window.use_tensor_dimensions(_output->info()->tensor_shape());
 
-    Window slice     = window.first_slice_window_3D();
-    Window slice_out = out_window.first_slice_window_3D();
+    Window collapsed     = window.collapse_if_possible(ICLKernel::window(), Window::DimZ, &is_collapsed);
+    Window collapsed_out = out_window.collapse_if_possible(out_window, 3, &is_collapsed_out);
 
-    unsigned int idx = 2 * num_arguments_per_3D_tensor();
-    _kernel.setArg<cl_uint>(idx++, _output->info()->strides_in_bytes()[3]);
+    ARM_COMPUTE_ERROR_ON(is_collapsed != is_collapsed_out);
 
+    Window slice     = collapsed.first_slice_window_3D();
+    Window slice_out = collapsed_out.first_slice_window_4D();
     do
     {
         // Set inputs
         unsigned int idx = 0;
         add_3D_tensor_argument(idx, _input, slice);
-        add_3D_tensor_argument(idx, _output, slice_out);
+        add_4D_tensor_argument(idx, _output, slice_out);
         enqueue(queue, *this, slice, lws_hint());
     }
-    while(window.slide_window_slice_3D(slice) && out_window.slide_window_slice_3D(slice_out));
+    while(collapsed.slide_window_slice_3D(slice) && collapsed_out.slide_window_slice_4D(slice_out));
 }
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
index a40aa2856c..de7e2b8737 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
@@ -225,8 +225,17 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input,
     _conv_pad_top  = conv_info.pad_top();
     _border_size   = BorderSize(_conv_pad_top, conv_info.pad_right(), conv_info.pad_bottom(), _conv_pad_left);
 
+    // Configure kernel window
+    std::string     kernel_name;
+    const GPUTarget gpu_target = get_target();
+
+    auto win_config = validate_and_configure_window(input->info(), weights->info(), output->info(), conv_info, depth_multiplier, gpu_target, kernel_name);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure_internal(win_config.second);
+
     // Set build options
     CLBuildOptions build_opts;
+    build_opts.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(_output->info()->tensor_shape().z()));
     build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(depth_multiplier));
     build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(_conv_stride_x));
     build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
@@ -273,15 +282,6 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input,
             }
         }
     }
-
-    // Configure kernel window
-    std::string     kernel_name;
-    const GPUTarget gpu_target = get_target();
-
-    auto win_config = validate_and_configure_window(input->info(), weights->info(), output->info(), conv_info, depth_multiplier, gpu_target, kernel_name);
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-    ICLKernel::configure_internal(win_config.second);
-
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // Set config_id for enabling LWS tuning
@@ -316,15 +316,17 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::run(const Window &window, cl::Com
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
 
+    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+
     // Create input window and adjust
-    Window win_in = window;
-    win_in.adjust(Window::DimX, -_conv_pad_left, true);
-    win_in.adjust(Window::DimY, -_conv_pad_top, true);
-    win_in.set_dimension_step(Window::DimX, window.x().step() * _conv_stride_x);
-    win_in.set_dimension_step(Window::DimY, window.y().step() * _conv_stride_y);
-
-    Window slice_in      = win_in.first_slice_window_3D();
-    Window slice_out     = window.first_slice_window_3D();
+    Window collapsed_in = collapsed;
+    collapsed_in.adjust(Window::DimX, -_conv_pad_left, true);
+    collapsed_in.adjust(Window::DimY, -_conv_pad_top, true);
+    collapsed_in.set_dimension_step(Window::DimX, collapsed_in.x().step() * _conv_stride_x);
+    collapsed_in.set_dimension_step(Window::DimY, collapsed_in.y().step() * _conv_stride_y);
+
+    Window slice_in      = collapsed_in.first_slice_window_3D();
+    Window slice_out     = collapsed.first_slice_window_3D();
     Window slice_weights = window.first_slice_window_3D();
     slice_weights.set_dimension_step(Window::DimX, 0);
     slice_weights.set_dimension_step(Window::DimY, 0);
@@ -347,5 +349,5 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::run(const Window &window, cl::Com
 
         enqueue(queue, *this, slice_out, lws_hint());
     }
-    while(window.slide_window_slice_3D(slice_out) && win_in.slide_window_slice_3D(slice_in));
+    while(collapsed.slide_window_slice_3D(slice_out) && collapsed_in.slide_window_slice_3D(slice_in));
 }
diff --git a/src/core/CL/kernels/CLFillBorderKernel.cpp b/src/core/CL/kernels/CLFillBorderKernel.cpp
index baf6bb6024..69206678d0 100644
--- a/src/core/CL/kernels/CLFillBorderKernel.cpp
+++ b/src/core/CL/kernels/CLFillBorderKernel.cpp
@@ -168,7 +168,8 @@ void CLFillBorderKernel::run(const Window &window, cl::CommandQueue &queue)
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
 
-    Window slice = window.first_slice_window_3D();
+    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+    Window slice     = collapsed.first_slice_window_3D();
 
     do
     {
@@ -176,5 +177,5 @@ void CLFillBorderKernel::run(const Window &window, cl::CommandQueue &queue)
         add_3D_tensor_argument(idx, _tensor, slice);
         enqueue(queue, *this, slice, cl::NullRange);
     }
-    while(window.slide_window_slice_3D(slice));
+    while(collapsed.slide_window_slice_3D(slice));
 }