From b3d2579b567eabd98fdb4861bf1380fefa18c9aa Mon Sep 17 00:00:00 2001 From: Moritz Pflanzer Date: Wed, 26 Jul 2017 11:49:37 +0100 Subject: COMPMID-415: Move ConvolutionLayer to new validation Change-Id: I1f40dff43142c4e2c096122bfa1ca08241ff80ff Reviewed-on: http://mpd-gerrit.cambridge.arm.com/81952 Tested-by: Kaizen Reviewed-by: Anthony Barbier --- tests/validation/TensorOperations.h | 146 ------------------------------------ 1 file changed, 146 deletions(-) (limited to 'tests/validation/TensorOperations.h') diff --git a/tests/validation/TensorOperations.h b/tests/validation/TensorOperations.h index 84aa965a9f..f4d2110387 100644 --- a/tests/validation/TensorOperations.h +++ b/tests/validation/TensorOperations.h @@ -59,100 +59,6 @@ struct is_floating_point { }; -bool is_valid_pixel(int i, int min, int max) -{ - return (i >= min && i < max); -} - -// 3D convolution for floating point type -template ::value, int>::type * = nullptr> -void convolution3d(const T *in, const T *weights, const T *bias, T *out, int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights, int8_t fixed_point_position) -{ - const int half_width_weights = width_weights / 2; - const int half_height_weights = height_weights / 2; - - // Reset accumulator - T acc = static_cast(0); - - // Compute a 2D convolution for each IFM and accumulate the result - for(int ifm = 0; ifm < depth_in; ++ifm) - { - // Compute the offset for the input slice - const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in; - - // Compute 2D convolution - for(int yk = -half_height_weights; yk <= half_height_weights; ++yk) - { - for(int xk = -half_width_weights; xk <= half_width_weights; ++xk) - { - // Check if the pixel is out-of-bound - if(is_valid_pixel(xi + xk, 0, width_in) && is_valid_pixel(yi + yk, 0, height_in)) - { - const int idx = xk + half_width_weights; - const int idy = yk + half_height_weights; - - const T i_value = in[offset_slice_in + xk + yk * width_in]; - const T w_value = weights[idx + idy * width_weights + ifm * width_weights * height_weights]; - - acc += i_value * w_value; - } - } - } - } - - // Accumulate the bias and store the result - *out = acc + (*bias); -} - -// 3D convolution for fixed point type -template ::value, int>::type * = nullptr> -void convolution3d(const T *in, const T *weights, const T *bias, T *out, int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights, - int8_t fixed_point_position) -{ - const int half_width_weights = width_weights / 2; - const int half_height_weights = height_weights / 2; - - using namespace fixed_point_arithmetic; - using promoted_type = typename fixed_point_arithmetic::traits::promote::type; - - // Reset accumulator - fixed_point acc(0, fixed_point_position); - - // Compute a 2D convolution for each IFM and accumulate the result - for(int ifm = 0; ifm < depth_in; ++ifm) - { - // Compute the offset for the input slice - const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in; - - // Compute 2D convolution - for(int yk = -half_height_weights; yk <= half_height_weights; ++yk) - { - for(int xk = -half_width_weights; xk <= half_width_weights; ++xk) - { - // Check if the pixel is out-of-bound - if(is_valid_pixel(xi + xk, 0, width_in) && is_valid_pixel(yi + yk, 0, height_in)) - { - const int idx = xk + half_width_weights; - const int idy = yk + half_height_weights; - - const fixed_point i_value(in[offset_slice_in + xk + yk * width_in], fixed_point_position, true); - const fixed_point w_value(weights[idx + idy * width_weights + ifm * width_weights * height_weights], fixed_point_position, true); - const fixed_point iw = i_value * w_value; - acc = iw + acc; - } - } - } - } - - // Get the bias - const fixed_point b(*bias, fixed_point_position, true); - - // Accumulate the bias and covert back - acc = acc + b; - fixed_point res(acc); - *out = res.raw(); -} - template ::value, int>::type * = nullptr> void vector_matrix_multiply(const T *in, const T *weights, const T *bias, T *out, int cols_weights, int rows_weights, uint8_t fixed_point_position) { @@ -999,58 +905,6 @@ void batch_normalization_layer(const Tensor &in, Tensor &out, const Tensor } } -// Convolution layer -template -void convolution_layer(const Tensor &in, const Tensor &weights, const Tensor &bias, Tensor &out, const PadStrideInfo &conv_info) -{ - const int width_in = in.shape().x(); - const int height_in = in.shape().y(); - const int depth_in = in.shape().z(); - const int width_out = out.shape().x(); - const int height_out = out.shape().y(); - const int depth_out = out.shape().z(); - const int width_weights = weights.shape().x(); - const int height_weights = weights.shape().y(); - const int depth_weights = weights.shape().z(); - const int pad_xi = std::min(static_cast(conv_info.pad().first), width_weights / 2); - const int pad_yi = std::min(static_cast(conv_info.pad().second), height_weights / 2); - const int start_xi = width_weights / 2 - pad_xi; - const int start_yi = height_weights / 2 - pad_yi; - const int end_xi = width_in - start_xi; - const int end_yi = height_in - start_yi; - const int stride_xi = conv_info.stride().first; - const int stride_yi = conv_info.stride().second; - const int num_batches = in.shape().total_size() / (width_in * height_in * depth_in); - - for(int r = 0; r < num_batches; ++r) - { - for(int yi = start_yi; yi < end_yi; yi += stride_yi) - { - for(int xi = start_xi; xi < end_xi; xi += stride_xi) - { - for(int ofm = 0; ofm < depth_out; ++ofm) - { - // Compute input and output offsets - const int offset_in = r * width_in * height_in * depth_in; - const int xo = (xi - start_xi) / stride_xi; - const int yo = (yi - start_yi) / stride_yi; - const int offset_out = xo + yo * width_out + ofm * width_out * height_out + r * width_out * height_out * depth_out; - - // Compute 3D convolution - convolution3d(in.data() + offset_in, - weights.data() + ofm * width_weights * height_weights * depth_weights, - bias.data() + ofm, - out.data() + offset_out, - xi, yi, - width_in, height_in, depth_in, - width_weights, height_weights, - static_cast(in.fixed_point_position())); - } - } - } - } -} - // Fully connected layer template void fully_connected_layer(const Tensor &in, const Tensor &weights, const Tensor &bias, Tensor &out) -- cgit v1.2.1