cudnn_benchmarks.textproto

# This file contains custom benchmark configs, in addition to the ones generated
# in the source code.

# CUDA 9.x + Volta performance regressions in cuDNN from 7.1.4 to 7.3.0, but then fixed in 7.3.1.
convolution_benchmark {
  label: "NHWC_128x20x20x56x160"
  input {
    dimension: [128, 56, 20, 20]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  filter {
    dimension: [160, 56, 5, 5]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  output {
    dimension: [128, 160, 16, 16]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  convolution {
    pad: [0, 0]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  fwd_algo: CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM
}

# CUDA 9.x + Volta performance regressions in cuDNN from 7.1.4 to 7.3.0, but then fixed in 7.3.1.
convolution_benchmark {
  label: "NHWC_512x27x27x64x192"
  input {
    dimension: [512, 64, 27, 27]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  filter {
    dimension: [192, 64, 5, 5]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  output {
    dimension: [512, 192, 27, 27]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  convolution {
    pad: [2, 2]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_data_algo: CONVOLUTION_BWD_DATA_ALGO_1
}

# CUDA 9.x + Volta performance regressions in cuDNN 7.3.0 -> 7.3.1.
# See nvbugs/2414937.
convolution_benchmark {
  label: "NHWC_128x32x32x128x256"
  input {
    dimension: [128, 128, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  filter {
    dimension: [256, 128, 5, 1]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  output {
    dimension: [128, 256, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  convolution {
    pad: [2, 0]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_filter_algo: CONVOLUTION_BWD_FILTER_ALGO_1
}

convolution_benchmark {
  label: "NHWC_128x32x32x128x1024"
  input {
    dimension: [128, 128, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  filter {
    dimension: [1024, 128, 1, 1]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  output {
    dimension: [128, 1024, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  convolution {
    pad: [0, 0]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_filter_algo: CONVOLUTION_BWD_FILTER_ALGO_0
}

convolution_benchmark {
  label: "NHWC_128x32x32x1024x1024"
  input {
    dimension: [128, 1024, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  filter {
    dimension: [1024, 1024, 1, 1]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  output {
    dimension: [128, 1024, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  convolution {
    pad: [0, 0]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_filter_algo: CONVOLUTION_BWD_FILTER_ALGO_0
}

convolution_benchmark {
  label: "NHWC_128x32x32x1024x768"
  input {
    dimension: [128, 1024, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  filter {
    dimension: [768, 1024, 1, 1]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  output {
    dimension: [128, 768, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  convolution {
    pad: [0, 0]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_filter_algo: CONVOLUTION_BWD_FILTER_ALGO_0
}
# end of 7.3.0 -> 7.3.1 regressions

convolution_benchmark {
  label: "NHWC_BWD_FILTER_128x20x20x88x160"
  input {
    dimension: [128, 88, 20, 20]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  filter {
    dimension: [160, 88, 5, 5]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  output {
    dimension: [128, 160, 16, 16]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  convolution {
    pad: [0, 0]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_filter_algo: CONVOLUTION_BWD_FILTER_ALGO_1
}

convolution_benchmark {
  label: "NHWC_BWD_FILTER_128x20x20x56x160"
  input {
    dimension: [128, 56, 20, 20]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  filter {
    dimension: [160, 56, 5, 5]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  output {
    dimension: [128, 160, 16, 16]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  convolution {
    pad: [0, 0]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_filter_algo: CONVOLUTION_BWD_FILTER_ALGO_1
}

convolution_benchmark {
  label: "NHWC_BWD_FILTER_128x20x20x56x160"
  input {
    dimension: [128, 56, 20, 20]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  filter {
    dimension: [160, 56, 5, 5]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  output {
    dimension: [128, 160, 16, 16]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  convolution {
    pad: [0, 0]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_filter_algo: CONVOLUTION_BWD_FILTER_ALGO_1
}

# These (NCHW|NHWC)_BWD_DATA_128_[89]_32_32 tests show that NCHW is faster than NHWC on Volta, CUDA 9.0, cuDNN 7.4.x.
convolution_benchmark {
  label: "NCHW_BWD_DATA_128_9_32_32_160_ALGO_0"
  input {
    dimension: [128, 9, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  filter {
    dimension: [160, 9, 2, 2]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  output {
    dimension: [128, 160, 16, 16]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  convolution {
    pad: [0, 0]
    filter_stride: [2, 2]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_data_algo: CONVOLUTION_BWD_DATA_ALGO_0
}

convolution_benchmark {
  label: "NCHW_BWD_DATA_128_9_32_32_160_ALGO_1"
  input {
    dimension: [128, 9, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  filter {
    dimension: [160, 9, 2, 2]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  output {
    dimension: [128, 160, 16, 16]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  convolution {
    pad: [0, 0]
    filter_stride: [2, 2]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_data_algo: CONVOLUTION_BWD_DATA_ALGO_1
}

convolution_benchmark {
  label: "NCHW_BWD_DATA_128_8_32_32_160_ALGO_1"
  input {
    dimension: [128, 8, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  filter {
    dimension: [160, 8, 2, 2]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  output {
    dimension: [128, 160, 16, 16]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  convolution {
    pad: [0, 0]
    filter_stride: [2, 2]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_data_algo: CONVOLUTION_BWD_DATA_ALGO_1
}

convolution_benchmark {
  label: "NCHW_BWD_DATA_128_8_32_32_40_ALGO_0"
  input {
    dimension: [128, 8, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  filter {
    dimension: [40, 8, 2, 2]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  output {
    dimension: [128, 40, 16, 16]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  convolution {
    pad: [0, 0]
    filter_stride: [2, 2]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_data_algo: CONVOLUTION_BWD_DATA_ALGO_0
}

convolution_benchmark {
  label: "NCHW_BWD_DATA_128_8_32_32_40_ALGO_1"
  input {
    dimension: [128, 8, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  filter {
    dimension: [40, 8, 2, 2]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  output {
    dimension: [128, 40, 16, 16]
    data_type: DATA_HALF
    format: TENSOR_NCHW
  }
  convolution {
    pad: [0, 0]
    filter_stride: [2, 2]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_data_algo: CONVOLUTION_BWD_DATA_ALGO_1
}

convolution_benchmark {
  label: "NHWC_BWD_DATA_128_9_32_32_160_ALGO_1"
  input {
    dimension: [128, 9, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  filter {
    dimension: [160, 9, 2, 2]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  output {
    dimension: [128, 160, 16, 16]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  convolution {
    pad: [0, 0]
    filter_stride: [2, 2]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_data_algo: CONVOLUTION_BWD_DATA_ALGO_1
}

convolution_benchmark {
  label: "NHWC_BWD_DATA_128_8_32_32_160_ALGO_1"
  input {
    dimension: [128, 8, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  filter {
    dimension: [160, 8, 2, 2]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  output {
    dimension: [128, 160, 16, 16]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  convolution {
    pad: [0, 0]
    filter_stride: [2, 2]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_data_algo: CONVOLUTION_BWD_DATA_ALGO_1
}

convolution_benchmark {
  label: "NHWC_BWD_DATA_128_8_32_32_40_ALGO_1"
  input {
    dimension: [128, 8, 32, 32]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  filter {
    dimension: [40, 8, 2, 2]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  output {
    dimension: [128, 40, 16, 16]
    data_type: DATA_HALF
    format: TENSOR_NHWC
  }
  convolution {
    pad: [0, 0]
    filter_stride: [2, 2]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_data_algo: CONVOLUTION_BWD_DATA_ALGO_1
}

convolution_benchmark {
  label: "NCHW_BWD_FILTER_LARGE_PADDING_SLOW"
  input {
    dimension: [512, 3072, 3, 3]
    data_type: DATA_FLOAT
    format: TENSOR_NCHW
  }
  filter {
    dimension: [1, 3072, 32, 32]
    data_type: DATA_FLOAT
    format: TENSOR_NCHW
  }
  output {
    dimension: [512, 1, 30, 30]
    data_type: DATA_FLOAT
    format: TENSOR_NCHW
  }
  convolution {
    pad: [29, 29]
    compute_mode: DATA_FLOAT
    math_type: TENSOR_OP_MATH
    mode: CROSS_CORRELATION
  }
  bwd_filter_algo: CONVOLUTION_BWD_FILTER_ALGO_0
}