COMPMID-617 Add window validation to CLDirectConvolutionLayer

Change-Id: Ia642dc68de6a0afe697bbce392e7ee955fa8944b
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/111460
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>

1 files changed, 169 insertions, 110 deletions

diff --git a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
index aea0161a1d..df0578bc6e 100644
--- a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
+++ b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
@@ -60,6 +60,160 @@ TensorShape get_output_shape(TensorShape input_shape, TensorShape weights_shape,
 
     return output_shape;
 }
+
+Error validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != weights->dimension(1),
+                                    "Weights should have same width as length");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != 1 && weights->dimension(0) != 3 && weights->dimension(0) != 5,
+                                    "Kernel sizes other than 1x1, 3x3 or 5x5 are not supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(2) != input->dimension(2),
+                                    "Weights feature map dimension should match the respective input's one");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != weights->dimension(1),
+                                    "Only rectangular weights are supported!");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->num_dimensions() > 4,
+                                    "Weights can be at most 4 dimensional");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(0) == 1) && std::get<0>(conv_info.stride()) > 3,
+                                    "Strides larger than 3 not supported for 1x1 convolution.");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(0) == 3 || weights->dimension(0) == 5) && std::get<0>(conv_info.stride()) > 2,
+                                    "Strides larger than 2 not supported for 3x3 convolution.");
+
+    if(biases != nullptr)
+    {
+        if(is_data_type_quantized_asymmetric(input->data_type()))
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
+        }
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->dimension(0) != weights->dimension(3),
+                                        "Biases size and number of input feature maps should match");
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->num_dimensions() > 1,
+                                        "Biases should be one dimensional");
+    }
+
+    // Checks performed when output is configured
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(),
+                                                           get_output_shape(input->tensor_shape(), weights->tensor_shape(), conv_info));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+    }
+
+    return Error{};
+}
+
+std::pair<Error, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *output, const PadStrideInfo &conv_info, const GPUTarget target)
+{
+    const unsigned int kernel_size = weights->dimension(0);
+    const DataType     data_type   = input->data_type();
+
+    // Get convolved dimensions
+    TensorShape output_shape = get_output_shape(input->tensor_shape(), weights->tensor_shape(), conv_info);
+
+    // Output auto inizialitation if not yet initialized
+    // FIXME: input->clone()->set_tensor_shape(output_shape) doesn't work with subtensors for grouped direct convolutions (AlexNet).
+    auto_init_if_empty(*output, output_shape,
+                       1,
+                       input->data_type(),
+                       input->fixed_point_position(),
+                       input->quantization_info());
+
+    unsigned int conv_stride_x   = std::get<0>(conv_info.stride());
+    unsigned int conv_stride_y   = std::get<1>(conv_info.stride());
+    unsigned int conv_pad_left   = std::min(conv_info.pad_left(), kernel_size / 2);
+    unsigned int conv_pad_top    = std::min(conv_info.pad_top(), kernel_size / 2);
+    unsigned int conv_pad_right  = std::min(conv_info.pad_right(), kernel_size / 2);
+    unsigned int conv_pad_bottom = std::min(conv_info.pad_bottom(), kernel_size / 2);
+
+    unsigned int num_elems_read_per_iteration_x    = 0;
+    unsigned int num_elems_read_per_iteration_y    = 0;
+    unsigned int num_elems_written_per_iteration_x = 0;
+    unsigned int num_elems_written_per_iteration_y = 0;
+
+    Window win            = Window();
+    bool   window_changed = false;
+
+    if((target == GPUTarget::BIFROST) && (kernel_size <= 5) && (conv_stride_x == 1) && (conv_stride_y == 1) && (data_type == DataType::F32))
+    {
+        // Configure kernel window
+        win = calculate_max_window(*output);
+
+        switch(kernel_size)
+        {
+            case 1:
+            {
+                num_elems_read_per_iteration_x    = 4;
+                num_elems_read_per_iteration_y    = 4;
+                num_elems_written_per_iteration_x = 4;
+                num_elems_written_per_iteration_y = 4;
+                break;
+            }
+            case 3:
+            {
+                num_elems_read_per_iteration_x    = 6;
+                num_elems_read_per_iteration_y    = 5;
+                num_elems_written_per_iteration_x = 4;
+                num_elems_written_per_iteration_y = 3;
+                break;
+            }
+            case 5:
+            {
+                num_elems_read_per_iteration_x    = 8;
+                num_elems_read_per_iteration_y    = 6;
+                num_elems_written_per_iteration_x = 4;
+                num_elems_written_per_iteration_y = 2;
+                break;
+            }
+            default:
+            {
+                ARM_COMPUTE_ERROR("Kernel size not optimized for Bifrost");
+            }
+        }
+    }
+    else
+    {
+        bool is_stride2 = ((kernel_size != 1) && (conv_stride_x == 2));
+
+        num_elems_read_per_iteration_x    = 8 + 2 * (kernel_size / 2) + (is_stride2 ? 6 + kernel_size / 2 : 0);
+        num_elems_read_per_iteration_y    = kernel_size;
+        num_elems_written_per_iteration_x = 8;
+        num_elems_written_per_iteration_y = 1;
+    }
+
+    // Calculate right and bottom border
+    int input_width  = input->dimension(0) - kernel_size / 2 + conv_pad_right;
+    int input_height = input->dimension(1) - kernel_size / 2 + conv_pad_bottom;
+
+    // Add padding only if necessary or it would always result in a window_changed
+    if(input_width % num_elems_read_per_iteration_x > 0)
+    {
+        input_width += num_elems_read_per_iteration_x;
+    }
+    if(input_height % num_elems_read_per_iteration_y > 0)
+    {
+        input_height += num_elems_read_per_iteration_y;
+    }
+
+    // Create window and update padding
+    win = calculate_max_window(*output, Steps(num_elems_written_per_iteration_x, num_elems_written_per_iteration_y));
+
+    AccessWindowStatic    input_access(input, -conv_pad_left, -conv_pad_top, input_width, input_height);
+    AccessWindowStatic    weights_access(weights, 0, 0, kernel_size, kernel_size);
+    AccessWindowRectangle output_access(output, 0, 0, num_elems_written_per_iteration_x, num_elems_written_per_iteration_y);
+
+    window_changed = update_window_and_padding(win, input_access, weights_access, output_access);
+
+    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+
+    Error err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Error{};
+    return std::make_pair(err, win);
+}
 } // namespace
 
 CLDirectConvolutionLayerKernel::CLDirectConvolutionLayerKernel()
@@ -83,6 +237,7 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
     TensorShape output_shape = get_output_shape(input->info()->tensor_shape(), weights->info()->tensor_shape(), conv_info);
 
     // Output auto inizialitation if not yet initialized
+    // FIXME: input->clone()->set_tensor_shape(output_shape) doesn't work with subtensors for grouped direct convolutions (AlexNet).
     auto_init_if_empty(*output->info(),
                        output_shape,
                        1,
@@ -91,11 +246,11 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
                        input->info()->quantization_info());
 
     // Perform validation step
-    ARM_COMPUTE_ERROR_THROW_ON(CLDirectConvolutionLayerKernel::validate(input->info(),
-                                                                        weights->info(),
-                                                                        (biases != nullptr) ? biases->info() : nullptr,
-                                                                        output->info(),
-                                                                        conv_info));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(),
+                                                  weights->info(),
+                                                  (biases != nullptr) ? biases->info() : nullptr,
+                                                  output->info(),
+                                                  conv_info));
 
     _conv_stride_x = std::get<0>(conv_info.stride());
     _conv_stride_y = std::get<1>(conv_info.stride());
@@ -126,14 +281,6 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
         kernel_name << "_f32_bifrost";
         _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name.str(), build_options.options()));
 
-        // Configure kernel window
-        Window win = calculate_max_window(*output->info());
-
-        unsigned int num_elems_read_per_iteration_x    = 0;
-        unsigned int num_elems_read_per_iteration_y    = 0;
-        unsigned int num_elems_written_per_iteration_x = 0;
-        unsigned int num_elems_written_per_iteration_y = 0;
-
         // Through extensive experimentation with over 30 representative tensor
         // shapes, we found a small number of local work size configurations
         // that result in nearly optimal execution times. Selecting the right
@@ -155,10 +302,6 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
         {
             case 1:
             {
-                num_elems_read_per_iteration_x    = 4;
-                num_elems_read_per_iteration_y    = 4;
-                num_elems_written_per_iteration_x = 4;
-                num_elems_written_per_iteration_y = 4;
                 if(mega_ops_ < 1.f)
                 {
                     _lws_hint = cl::NDRange(1, 1, 8);
@@ -175,10 +318,6 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
             }
             case 3:
             {
-                num_elems_read_per_iteration_x    = 6;
-                num_elems_read_per_iteration_y    = 5;
-                num_elems_written_per_iteration_x = 4;
-                num_elems_written_per_iteration_y = 3;
                 if(mega_ops_ < 1.f)
                 {
                     _lws_hint = cl::NDRange(1, 1, 8);
@@ -199,10 +338,6 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
             }
             case 5:
             {
-                num_elems_read_per_iteration_x    = 8;
-                num_elems_read_per_iteration_y    = 6;
-                num_elems_written_per_iteration_x = 4;
-                num_elems_written_per_iteration_y = 2;
                 if(mega_ops_ < 2.f || mega_ops_ > 80.f)
                 {
                     _lws_hint = cl::NDRange(2, 1, 4);
@@ -218,23 +353,6 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
                 ARM_COMPUTE_ERROR("Kernel size not optimized for Bifrost");
             }
         }
-
-        // Calculate right and bottom border
-        const int input_width  = input->info()->dimension(0) - kernel_size / 2 + conv_pad_right;
-        const int input_height = input->info()->dimension(1) - kernel_size / 2 + conv_pad_bottom;
-
-        // Create window and update padding
-        win = calculate_max_window(*output->info(), Steps(num_elems_written_per_iteration_x, num_elems_written_per_iteration_y));
-
-        AccessWindowStatic    input_access(input->info(), -conv_pad_left, -conv_pad_top, input_width + num_elems_read_per_iteration_x, input_height + num_elems_read_per_iteration_y);
-        AccessWindowStatic    weights_access(weights->info(), 0, 0, kernel_size, kernel_size);
-        AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_written_per_iteration_x, num_elems_written_per_iteration_y);
-
-        update_window_and_padding(win, input_access, weights_access, output_access);
-
-        output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
-
-        ICLKernel::configure(win);
     }
     else
     {
@@ -254,34 +372,13 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
         // Create kernel
         _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(is_quantized_asymm ? "direct_convolution_1x1_3x3_5x5_quantized" : kernel_name.str(),
                                                                                build_options.options()));
-
-        // Configure kernel window
-
-        bool is_stride2 = ((kernel_size != 1) && (_conv_stride_x == 2));
-
-        const unsigned int num_elems_read_per_iteration_x    = 8 + 2 * (kernel_size / 2) + (is_stride2 ? 6 + kernel_size / 2 : 0);
-        const unsigned int num_elems_read_per_iteration_y    = kernel_size;
-        const unsigned int num_elems_written_per_iteration_x = 8;
-        const unsigned int num_elems_written_per_iteration_y = 1;
-
-        // Calculate right and bottom border
-        const int input_width  = input->info()->dimension(0) - kernel_size / 2 + conv_pad_right;
-        const int input_height = input->info()->dimension(1) - kernel_size / 2 + conv_pad_bottom;
-
-        // Create window and update padding
-        Window win = calculate_max_window(*output->info(), Steps(num_elems_written_per_iteration_x, num_elems_written_per_iteration_y));
-
-        AccessWindowStatic    input_access(input->info(), -conv_pad_left, -conv_pad_top, input_width + num_elems_read_per_iteration_x, input_height + num_elems_read_per_iteration_y);
-        AccessWindowStatic    weights_access(weights->info(), 0, 0, kernel_size, kernel_size);
-        AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_written_per_iteration_x, num_elems_written_per_iteration_y);
-
-        update_window_and_padding(win, input_access, weights_access, output_access);
-
-        output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
-
-        ICLKernel::configure(win);
     }
 
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), weights->info(), output->info(), conv_info, gpu_target);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure(win_config.second);
+
     // Set static kernel arguments
     if(is_data_type_quantized_asymmetric(data_type))
     {
@@ -322,49 +419,11 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
     _config_id += support::cpp11::to_string(output->info()->dimension(1));
 }
 
-Error CLDirectConvolutionLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info)
+Error CLDirectConvolutionLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                                               const GPUTarget target)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != weights->dimension(1),
-                                    "Weights should have same width as length");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != 1 && weights->dimension(0) != 3 && weights->dimension(0) != 5,
-                                    "Kernel sizes other than 1x1, 3x3 or 5x5 are not supported");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(2) != input->dimension(2),
-                                    "Weights feature map dimension should match the respective input's one");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != weights->dimension(1),
-                                    "Only rectangular weights are supported!");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->num_dimensions() > 4,
-                                    "Weights can be at most 4 dimensional");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(0) == 1) && std::get<0>(conv_info.stride()) > 3,
-                                    "Strides larger than 3 not supported for 1x1 convolution.");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(0) == 3 || weights->dimension(0) == 5) && std::get<0>(conv_info.stride()) > 2,
-                                    "Strides larger than 2 not supported for 3x3 convolution.");
-
-    if(biases != nullptr)
-    {
-        if(is_data_type_quantized_asymmetric(input->data_type()))
-        {
-            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
-        }
-        else
-        {
-            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
-        }
-        ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->dimension(0) != weights->dimension(3),
-                                        "Biases size and number of input feature maps should match");
-        ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->num_dimensions() > 1,
-                                        "Biases should be one dimensional");
-    }
-
-    // Checks performed when output is configured
-    if(output->total_size() != 0)
-    {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(),
-                                                           get_output_shape(input->tensor_shape(), weights->tensor_shape(), conv_info));
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
-    }
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), output->clone().get(), conv_info, target).first);
 
     return Error{};
 }