matmul first approach - Joejiong/buddy-mlir GitHub Wiki

./bin/buddy-opt ../examples/ConvOpt/pointwise_conv2d_nhwc_filter_hwcf.mlir -pointwise-conv-vectorization --print-ir-before-all
// -----// IR Dump Before {anonymous}::PointwiseConvVectorizationPass //----- //
module  {
  func @conv_2d_1x1(%arg0: tensor<1x4x5x2xf32>, %arg1: tensor<1x1x2x7xf32>) -> tensor<1x4x5x7xf32> {
    %0 = linalg.init_tensor [1, 4, 5, 7] : tensor<1x4x5x7xf32>
    %1 = linalg.conv_2d_nhwc_hwcf {dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %arg1 : tensor<1x4x5x2xf32>, tensor<1x1x2x7xf32>) outs(%0 : tensor<1x4x5x7xf32>) -> tensor<1x4x5x7xf32>
    return %1 : tensor<1x4x5x7xf32>
  }
}


module  {
  func @conv_2d_1x1(%arg0: tensor<1x4x5x2xf32>, %arg1: tensor<1x1x2x7xf32>) -> tensor<1x4x5x7xf32> {
    %0 = linalg.init_tensor [1, 4, 5, 7] : tensor<1x4x5x7xf32>
    %1 = tensor.collapse_shape %arg0 [[0, 1, 2], [3]] : tensor<1x4x5x2xf32> into tensor<20x2xf32>
    %2 = tensor.collapse_shape %arg1 [[0, 1, 2], [3]] : tensor<1x1x2x7xf32> into tensor<2x7xf32>
    %3 = tensor.collapse_shape %0 [[0, 1, 2], [3]] : tensor<1x4x5x7xf32> into tensor<20x7xf32>
    %4 = linalg.matmul ins(%1, %2 : tensor<20x2xf32>, tensor<2x7xf32>) outs(%3 : tensor<20x7xf32>) -> tensor<20x7xf32>
    %5 = tensor.expand_shape %4 [[0, 1, 2], [3]] : tensor<20x7xf32> into tensor<1x4x5x7xf32>
    return %5 : tensor<1x4x5x7xf32>
  }
}

> ./bin/pointwise-conv-2d-nhwc-hwcf-benchmark                           pointwise_benchmark [39d7e0c] modified
2021-12-20T18:42:07+08:00
Running ./bin/pointwise-conv-2d-nhwc-hwcf-benchmark
Run on (52 X 2500 MHz CPU s)
CPU Caches:
  L1 Data 32 KiB (x26)
  L1 Instruction 32 KiB (x26)
  L2 Unified 1024 KiB (x26)
  L3 Unified 36608 KiB (x1)
Load Average: 8.09, 11.43, 12.63
***WARNING*** Library was built as DEBUG. Timings may be affected.
-------------------------------------------------------------------------------------
Benchmark                                           Time             CPU   Iterations
-------------------------------------------------------------------------------------
BM_PointwiseConv2DNhwcHwcf/100                  0.124 ms        0.124 ms         5655
BM_PointwiseConv2DNhwcHwcfReturn/100            0.104 ms        0.104 ms         6728
BM_PointwiseConv2DNhwcHwcfReturnOrigin/100      0.150 ms        0.150 ms         4665
inputMemRef: [ 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 ] of shape: [ 1 4 5 2 ]
filterMemRef: [ 3 3 3 3 3 3 3 3 3 3 3 3 3 3 ] of shape: [ 1 1 2 7 ]
outputMemRef: [ 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 ] of shape: [ 1 4 5 7 ]
inputMemReturn: [ 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 ] of shape: [ 1 4 5 2 ]
filterMemReturn: [ 3 3 3 3 3 3 3 3 3 3 3 3 3 3 ] of shape: [ 1 1 2 7 ]
outputMemReturn: [ 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 ] of shape: [ 1 4 5 7 ]
inputMemReturn: [ 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 ] of shape: [ 1 4 5 2 ]
filterMemReturn: [ 3 3 3 3 3 3 3 3 3 3 3 3 3 3 ] of shape: [ 1 1 2 7 ]
outputMemReturnOrigin: [ 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 12 ] of shape: [ 1 4 5 7 ]