Skip to content
This repository has been archived by the owner on Jan 12, 2019. It is now read-only.

Commit

Permalink
improve generated code for convolutional model
Browse files Browse the repository at this point in the history
  • Loading branch information
Dave Goldman committed Feb 19, 2015
1 parent 16bb506 commit ca01815
Showing 1 changed file with 17 additions and 60 deletions.
77 changes: 17 additions & 60 deletions models/expiry/modelc_bf4dd6c8.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,6 @@
#include "compile.h"
#if COMPILE_DMZ

#define EIGEN_NO_DEBUG 1 // turn off range checking and anything else that could slow us down!
#define USE_OPTIMIZED_3x3_CONVOLUTION_bf4dd6c8 0

#include "modelc_bf4dd6c8.hpp"
Expand Down Expand Up @@ -12509,14 +12508,12 @@ DMZ_INTERNAL ModelCSingleConvolved_bf4dd6c8_1 convc_bf4dd6c8_1(const ModelCConvI
uint16_t limited_input_col;
uint16_t limited_kernel_cols;
uint16_t first_kernel_col;
Eigen::Matrix<float, 5, 5> input_submatrix_5_5;
Eigen::Matrix<float, 5, Eigen::Dynamic> input_submatrix_5_D;
Eigen::Matrix<float, Eigen::Dynamic, 5> input_submatrix_D_5;
Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic> input_submatrix_D_D;
Eigen::Matrix<float, 5, 5, Eigen::RowMajor> input_submatrix_5_5;
Eigen::Matrix<float, 5, Eigen::Dynamic, Eigen::RowMajor, 5, 5> input_submatrix_5_D;
Eigen::Matrix<float, Eigen::Dynamic, 5, Eigen::RowMajor, 5, 5> input_submatrix_D_5;
Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor, 5, 5> input_submatrix_D_D;
ModelCSingleKernel_bf4dd6c8_1 zero_kernel = ModelCSingleKernel_bf4dd6c8_1::Zero();
ModelCSingleKernel_bf4dd6c8_1 padded_submatrix;
ModelCSingleKernel_bf4dd6c8_1 elemwise_mult;
float sum;

for(output_row = 0; output_row < 5 - 1; output_row++) {
limited_input_row = 0;
Expand All @@ -12531,9 +12528,7 @@ DMZ_INTERNAL ModelCSingleConvolved_bf4dd6c8_1 convc_bf4dd6c8_1(const ModelCConvI
padded_submatrix = zero_kernel;
padded_submatrix.block(first_kernel_row, first_kernel_col, limited_kernel_rows, limited_kernel_cols) = input_submatrix_D_D;

elemwise_mult = kernel.cwiseProduct(padded_submatrix);
sum = elemwise_mult.sum();
output(output_row, output_col) = sum;
output(output_row, output_col) = kernel.cwiseProduct(padded_submatrix).sum();
}


Expand All @@ -12545,9 +12540,7 @@ DMZ_INTERNAL ModelCSingleConvolved_bf4dd6c8_1 convc_bf4dd6c8_1(const ModelCConvI
padded_submatrix = zero_kernel;
padded_submatrix.block(first_kernel_row, 0, limited_kernel_rows, limited_kernel_cols) = input_submatrix_D_D;

elemwise_mult = kernel.cwiseProduct(padded_submatrix);
sum = elemwise_mult.sum();
output(output_row, output_col) = sum;
output(output_row, output_col) = kernel.cwiseProduct(padded_submatrix).sum();
}


Expand All @@ -12558,9 +12551,7 @@ DMZ_INTERNAL ModelCSingleConvolved_bf4dd6c8_1 convc_bf4dd6c8_1(const ModelCConvI
padded_submatrix = zero_kernel;
padded_submatrix.block(first_kernel_row, 0, limited_kernel_rows, 5) = input_submatrix_D_5;

elemwise_mult = kernel.cwiseProduct(padded_submatrix);
sum = elemwise_mult.sum();
output(output_row, output_col) = sum;
output(output_row, output_col) = kernel.cwiseProduct(padded_submatrix).sum();
}

}
Expand All @@ -12578,9 +12569,7 @@ DMZ_INTERNAL ModelCSingleConvolved_bf4dd6c8_1 convc_bf4dd6c8_1(const ModelCConvI
padded_submatrix = zero_kernel;
padded_submatrix.block(first_kernel_row, first_kernel_col, limited_kernel_rows, limited_kernel_cols) = input_submatrix_D_D;

elemwise_mult = kernel.cwiseProduct(padded_submatrix);
sum = elemwise_mult.sum();
output(output_row, output_col) = sum;
output(output_row, output_col) = kernel.cwiseProduct(padded_submatrix).sum();
}


Expand All @@ -12592,9 +12581,7 @@ DMZ_INTERNAL ModelCSingleConvolved_bf4dd6c8_1 convc_bf4dd6c8_1(const ModelCConvI
padded_submatrix = zero_kernel;
padded_submatrix.block(first_kernel_row, 0, limited_kernel_rows, limited_kernel_cols) = input_submatrix_D_D;

elemwise_mult = kernel.cwiseProduct(padded_submatrix);
sum = elemwise_mult.sum();
output(output_row, output_col) = sum;
output(output_row, output_col) = kernel.cwiseProduct(padded_submatrix).sum();
}


Expand All @@ -12605,9 +12592,7 @@ DMZ_INTERNAL ModelCSingleConvolved_bf4dd6c8_1 convc_bf4dd6c8_1(const ModelCConvI
padded_submatrix = zero_kernel;
padded_submatrix.block(first_kernel_row, 0, limited_kernel_rows, 5) = input_submatrix_D_5;

elemwise_mult = kernel.cwiseProduct(padded_submatrix);
sum = elemwise_mult.sum();
output(output_row, output_col) = sum;
output(output_row, output_col) = kernel.cwiseProduct(padded_submatrix).sum();
}

}
Expand All @@ -12623,31 +12608,26 @@ DMZ_INTERNAL ModelCSingleConvolved_bf4dd6c8_1 convc_bf4dd6c8_1(const ModelCConvI
padded_submatrix = zero_kernel;
padded_submatrix.block(0, first_kernel_col, 5, limited_kernel_cols) = input_submatrix_5_D;

elemwise_mult = kernel.cwiseProduct(padded_submatrix);
sum = elemwise_mult.sum();
output(output_row, output_col) = sum;
output(output_row, output_col) = kernel.cwiseProduct(padded_submatrix).sum();
}

for(output_col = 14 - 5; output_col < 14; output_col++) {
for(output_col = (11 + 5 - 1) - (5 - 1); output_col < 14; output_col++) {
limited_input_col = output_col - 5 + 1;
limited_kernel_cols = 11 - limited_input_col;

input_submatrix_5_D = input.block(limited_input_row, limited_input_col, 5, limited_kernel_cols);
padded_submatrix = zero_kernel;
padded_submatrix.block(0, 0, 5, limited_kernel_cols) = input_submatrix_5_D;

elemwise_mult = kernel.cwiseProduct(padded_submatrix);
sum = elemwise_mult.sum();
output(output_row, output_col) = sum;
output(output_row, output_col) = kernel.cwiseProduct(padded_submatrix).sum();
}

for(output_col = 5 - 1; output_col < 14 - 5; output_col++) {
for(output_col = 5 - 1; output_col < (11 + 5 - 1) - (5 - 1); output_col++) {
limited_input_col = output_col - 5 + 1;

input_submatrix_5_5 = input.block(limited_input_row, limited_input_col, 5, 5);
elemwise_mult = kernel.cwiseProduct(input_submatrix_5_5);
sum = elemwise_mult.sum();
output(output_row, output_col) = sum;

output(output_row, output_col) = kernel.cwiseProduct(input_submatrix_5_5).sum();
}
}

Expand All @@ -12667,27 +12647,6 @@ DMZ_INTERNAL ModelCSingleDownsampled_bf4dd6c8_1 downc_bf4dd6c8_1(const ModelCSin
DMZ_INTERNAL ModelCConvResult_bf4dd6c8_1 convolve_bf4dd6c8_1(const ModelCConvInput_bf4dd6c8_1& input) {
ModelCConvResult_bf4dd6c8_1 accumulated_results;

#ifndef __LP64__
// There is a mysterious bug which causes this method to sometimes
// produce bad results when running on a 32-bit processor.
// I have not been able to pin it down, but I strongly suspect some
// bug in either Eigen or Clang.
// That suspicion is founded in part on the fact that the following
// workaround actually works!

if (!input.any() && input.sum() != 0) {
// any() is true if any coefficient is non-zero.
// So !any() is true only if all coefficients are zero.
// sum() != 0 is true only if !(all coefficients are zero).
// So the overall expression will NEVER be true.
// But the compiler is not sufficiently brilliant to know that,
// and therefore this never-executed code will be generated.
// Generating this code, even without executing it, is apparently
// sufficient to work around the mysterious 32-bit bug.
std::cerr << "card.io dmz: This is a bug workaround; ignore the following:\n" << input << "\n";
}
#endif

Eigen::Map<ModelCAllKernels_bf4dd6c8_1, Eigen::Aligned> all_kernels((float *)data_359cb697);
Eigen::Map<ModelCConvB_bf4dd6c8_1, Eigen::Aligned> conv_b((float *)data_f0fed3cf);

Expand Down Expand Up @@ -12758,9 +12717,7 @@ DMZ_INTERNAL ModelCSingleConvolved_bf4dd6c8_2 convc_bf4dd6c8_2(const ModelCConvI
// Scalar handling of non-vectorized leftovers
for(uint16_t output_col = vector_processed_cols; output_col < 3; output_col++) {
ModelCSingleKernel_bf4dd6c8_2 input_submatrix = input.block<5, 5>(output_row, output_col);
ModelCSingleKernel_bf4dd6c8_2 elemwise_mult = kernel.cwiseProduct(input_submatrix);
float sum = elemwise_mult.sum();
output(output_row, output_col) = sum;
output(output_row, output_col) = kernel.cwiseProduct(input_submatrix).sum();
}
}
return output;
Expand Down

0 comments on commit ca01815

Please sign in to comment.