File: gpu_different_blocks_threads_dimensions.cpp

package info (click to toggle)
halide 21.0.0-4
  • links: PTS, VCS
  • area: main
  • in suites: forky, sid
  • size: 55,752 kB
  • sloc: cpp: 289,334; ansic: 22,751; python: 7,486; makefile: 4,299; sh: 2,508; java: 1,549; javascript: 282; pascal: 207; xml: 127; asm: 9
file content (34 lines) | stat: -rw-r--r-- 762 bytes parent folder | download | duplicates (4)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
#include "Halide.h"

using namespace Halide;

int main(int argc, char **argv) {
    if (!get_jit_target_from_environment().has_gpu_feature()) {
        printf("[SKIP] No GPU target enabled.\n");
        return 0;
    }

    Func f, g;
    Var x, y, xi, yi;
    f(x, y) = x + y;
    g(x, y) = f(x, y);

    // At one point in time, FuseGPUThreadLoops assumed that the
    // number of blocks dimensions matched the number of threads
    // dimensions. This test checks that things still work if they're
    // mismatched.

    g.tile(x, y, xi, yi, 16, 16)
        .gpu_blocks(y)
        .gpu_threads(xi, yi);

    f.compute_at(g, x)
        .store_in(MemoryType::Heap)
        .gpu_threads(x, y);

    g.compile_jit();

    printf("Success!\n");

    return 0;
}