Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Optimize resample poly kernels #512

Merged
merged 3 commits into from
Nov 2, 2023
Merged

Optimize resample poly kernels #512

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
e9c459c
Optimize resample_poly with a multi-kernel approach
tbensonatl Nov 2, 2023
c27b45c
Remove outdated comment
tbensonatl Nov 2, 2023
0bd0658
Add comment clarifying filter index calculations
tbensonatl Nov 2, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
55 changes: 46 additions & 9 deletions examples/resample_poly_bench.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,7 +46,7 @@ using namespace matx;
// length, input signal length, up/down factors) will be adjusted to a range of interest
// and the benchmark will be run with and without the proposed kernel changes.

constexpr int NUM_WARMUP_ITERATIONS = 2;
constexpr int NUM_WARMUP_ITERATIONS = 10;

// Number of iterations per timed test. Iteration times are averaged in the report.
constexpr int NUM_ITERATIONS = 20;
Expand All @@ -61,29 +61,60 @@ void ResamplePolyBench()
matx::index_t down;
} test_cases[] = {
{ 1, 256, 384, 3125 },
{ 1, 256, 384, 175 },
{ 1, 256, 4, 5 },
{ 1, 256, 1, 4 },
{ 1, 256, 1, 16 },
{ 1, 3000, 384, 3125 },
{ 1, 3000, 384, 175 },
{ 1, 3000, 4, 5 },
{ 1, 3000, 1, 4 },
{ 1, 3000, 1, 16 },
{ 1, 31000, 384, 3125 },
{ 1, 31000, 384, 175 },
{ 1, 31000, 4, 5 },
{ 1, 31000, 1, 4 },
{ 1, 31000, 1, 16 },
{ 1, 256000, 384, 3125 },
{ 1, 256000, 384, 175 },
{ 1, 256000, 64, 37 },
{ 1, 256000, 2, 3 },
{ 1, 256000, 4, 5 },
{ 1, 256000, 7, 64 },
{ 1, 256000, 7, 128 },
{ 1, 256000, 1, 4 },
{ 1, 256000, 1, 8 },
{ 1, 256000, 1, 16 },
{ 1, 256000, 1, 64 },
{ 1, 256000, 4, 1 },
{ 1, 256000, 8, 1 },
{ 1, 256000, 16, 1 },
{ 1, 256000, 64, 1 },
{ 1, 256000, 4, 1 },
{ 1, 256000, 8, 1 },
{ 1, 256000, 16, 1 },
{ 42, 256000, 384, 3125 },
{ 42, 256000, 384, 175 },
{ 42, 256000, 2, 3 },
{ 42, 256000, 4, 5 },
{ 42, 256000, 1, 4 },
{ 42, 256000, 1, 8 },
{ 42, 256000, 1, 16 },
{ 128, 256000, 384, 3125 },
{ 128, 256000, 4, 5 },
{ 128, 256000, 1, 4 },
{ 128, 256000, 1, 16 },
{ 42, 256000, 1, 64 },
{ 1, 100000000, 384, 3125 },
{ 1, 100000000, 384, 175 },
{ 1, 100000000, 2, 3 },
{ 1, 100000000, 4, 5 },
{ 1, 100000000, 7, 64 },
{ 1, 100000000, 7, 128 },
{ 1, 100000000, 1, 2 },
{ 1, 100000000, 1, 4 },
{ 1, 100000000, 1, 8 },
{ 1, 100000000, 1, 16 },
{ 1, 100000000, 1, 192 },
{ 1, 100000000, 4, 1 },
{ 1, 100000000, 8, 1 },
{ 1, 100000000, 16, 1 },
};

cudaStream_t stream;
Expand All @@ -104,9 +135,14 @@ void ResamplePolyBench()
const index_t up_len = input_len * up;
const index_t output_len = up_len / down + ((up_len % down) ? 1 : 0);

auto input = matx::make_tensor<InType, 2>({num_batches, input_len});
auto filter = matx::make_tensor<InType, 1>({filter_len});
auto output = matx::make_tensor<InType, 2>({num_batches, output_len});
auto input = matx::make_tensor<InType, 2>({num_batches, input_len}, MATX_DEVICE_MEMORY);
auto filter = matx::make_tensor<InType, 1>({filter_len}, MATX_DEVICE_MEMORY);
auto output = matx::make_tensor<InType, 2>({num_batches, output_len}, MATX_DEVICE_MEMORY);

(input = static_cast<InType>(1.0)).run(stream);
(filter = static_cast<InType>(1.0)).run(stream);

cudaStreamSynchronize(stream);

for (int k = 0; k < NUM_WARMUP_ITERATIONS; k++) {
(output = matx::resample_poly(input, filter, up, down)).run(stream);
Expand All @@ -126,7 +162,8 @@ void ResamplePolyBench()

const double gflops = static_cast<double>(num_batches*(2*filter_len_per_phase-1)*output_len) / 1.0e9;
const double avg_elapsed_us = (static_cast<double>(elapsed_ms)/NUM_ITERATIONS)*1.0e3;
printf("Batches: %5lld FilterLen: %5lld InputLen: %7lld OutputLen: %7lld Up/Down: %4lld/%4lld Elapsed Usecs: %12.1f GFLOPS: %10.3f\n",
printf("Batches: %5" INDEX_T_FMT " FilterLen: %5" INDEX_T_FMT " InputLen: %9" INDEX_T_FMT " OutputLen: %8" INDEX_T_FMT
" Up/Down: %4" INDEX_T_FMT "/%4" INDEX_T_FMT " Elapsed Usecs: %12.1f GFLOPS: %10.3f\n",
num_batches, filter_len, input_len, output_len, up, down, avg_elapsed_us, gflops/(avg_elapsed_us/1.0e6));
}

Expand Down
Loading