Skip to main content

Browse the archive

Revision 06b208f28b94b202ef14360339126d2d21a0463f authored by Dillon Sharlet on 04 March 2021, 18:11:57 UTC, committed by GitHub on 04 March 2021, 18:11:57 UTC 
Move codegen backends into anonymous namespaces in source files and don't build them if not enabled (#5776)
 
* Remove unused vertex buffer parameters.

* Offload GPU code in a lowering pass instead of via CodeGen_GPU_Host. Fixes #5650, fixes #2797, fixes #2084, now #1971 is more relevant.

* clang-format.

* clang-format sorting is case sensitive!?

* clang-tidy

* Move codegen backends into anonymous namespaces in source files.

* clang-format

* Pass type arguments correctly.

* Update OffloadGPULoops.cpp

* trigger buildbots

* trigger buildbots

* Hack around tests that rely on the IR for offloaded GPU loops.

* Fix missing include.

* Remove unused include.

* clang-tidy

* Use custom lowering pass to see code before GPU offloading

* Speculative fix for segfault

* Fix const correctness

* Fix error on unused variables in generated code.

Co-authored-by: Steven Johnson <srj@google.com>
1 parent abf0f69
Raw File
process.cpp 
#include <chrono>
#include <cstdio>

#include "nl_means.h"
#include "nl_means_auto_schedule.h"

#include "HalideBuffer.h"
#include "halide_benchmark.h"
#include "halide_image_io.h"

using namespace Halide::Runtime;
using namespace Halide::Tools;

int main(int argc, char **argv) {
    if (argc < 7) {
        printf("Usage: ./process input.png patch_size search_area sigma timing_iterations output.png\n"
               "e.g.: ./process input.png 7 7 0.12 10 output.png\n");
        return 0;
    }

    Buffer<float> input = load_and_convert_image(argv[1]);
    int patch_size = atoi(argv[2]);
    int search_area = atoi(argv[3]);
    float sigma = atof(argv[4]);
    Buffer<float> output(input.width(), input.height(), 3);
    int timing_iterations = atoi(argv[5]);

    nl_means(input, patch_size, search_area, sigma, output);

    // Timing code

    printf("Input size: %d by %d, patch size: %d, search area: %d, sigma: %f\n",
           input.width(), input.height(), patch_size, search_area, sigma);

    // Manually-tuned version
    double min_t_manual = benchmark(timing_iterations, 1, [&]() {
        nl_means(input, patch_size, search_area, sigma, output);
        output.device_sync();
    });
    printf("Manually-tuned time: %gms\n", min_t_manual * 1e3);

    // Auto-scheduled version
    double min_t_auto = benchmark(timing_iterations, 1, [&]() {
        nl_means_auto_schedule(input, patch_size, search_area, sigma, output);
        output.device_sync();
    });
    printf("Auto-scheduled time: %gms\n", min_t_auto * 1e3);

    convert_and_save_image(output, argv[6]);

    printf("Success!\n");
    return 0;
}
back to top