| 1 | //===-- Test for the parallel scan and reduction operations on the GPU ----===// |
|---|
| 2 | // |
|---|
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
|---|
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
|---|
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
|---|
| 6 | // |
|---|
| 7 | //===----------------------------------------------------------------------===// |
|---|
| 8 | |
|---|
| 9 | #include "src/__support/CPP/bit.h" |
|---|
| 10 | #include "src/__support/GPU/utils.h" |
|---|
| 11 | #include "test/IntegrationTest/test.h" |
|---|
| 12 | |
|---|
| 13 | using namespace LIBC_NAMESPACE; |
|---|
| 14 | |
|---|
| 15 | static uint32_t sum(uint32_t n) { return n * (n + 1) / 2; } |
|---|
| 16 | |
|---|
| 17 | // Tests a reduction within a convergant warp or wavefront using some known |
|---|
| 18 | // values. For example, if every element in the lane is one, then the sum should |
|---|
| 19 | // be the size of the warp or wavefront, i.e. 1 + 1 + 1 ... + 1. |
|---|
| 20 | static void test_reduce() { |
|---|
| 21 | uint64_t mask = gpu::get_lane_mask(); |
|---|
| 22 | uint32_t x = gpu::reduce(lane_mask: mask, x: 1); |
|---|
| 23 | EXPECT_EQ(x, gpu::get_lane_size()); |
|---|
| 24 | |
|---|
| 25 | uint32_t y = gpu::reduce(lane_mask: mask, x: gpu::get_lane_id()); |
|---|
| 26 | EXPECT_EQ(y, sum(gpu::get_lane_size() - 1)); |
|---|
| 27 | |
|---|
| 28 | uint32_t z = 0; |
|---|
| 29 | if (gpu::get_lane_id() % 2) |
|---|
| 30 | z = gpu::reduce(lane_mask: gpu::get_lane_mask(), x: 1); |
|---|
| 31 | gpu::sync_lane(mask); |
|---|
| 32 | |
|---|
| 33 | EXPECT_EQ(z, gpu::get_lane_id() % 2 ? gpu::get_lane_size() / 2 : 0); |
|---|
| 34 | } |
|---|
| 35 | |
|---|
| 36 | // Tests an accumulation scan within a convergent warp or wavefront using some |
|---|
| 37 | // known values. For example, if every element in the lane is one, then the scan |
|---|
| 38 | // should have each element be equivalent to its ID, i.e. 1, 1 + 1, ... |
|---|
| 39 | static void test_scan() { |
|---|
| 40 | uint64_t mask = gpu::get_lane_mask(); |
|---|
| 41 | |
|---|
| 42 | uint32_t x = gpu::scan(lane_mask: mask, x: 1); |
|---|
| 43 | EXPECT_EQ(x, gpu::get_lane_id() + 1); |
|---|
| 44 | |
|---|
| 45 | uint32_t y = gpu::scan(lane_mask: mask, x: gpu::get_lane_id()); |
|---|
| 46 | EXPECT_EQ(y, sum(gpu::get_lane_id())); |
|---|
| 47 | |
|---|
| 48 | uint32_t z = 0; |
|---|
| 49 | if (gpu::get_lane_id() % 2) |
|---|
| 50 | z = gpu::scan(lane_mask: gpu::get_lane_mask(), x: 1); |
|---|
| 51 | gpu::sync_lane(mask); |
|---|
| 52 | |
|---|
| 53 | EXPECT_EQ(z, gpu::get_lane_id() % 2 ? gpu::get_lane_id() / 2 + 1 : 0); |
|---|
| 54 | } |
|---|
| 55 | |
|---|
| 56 | TEST_MAIN(int argc, char **argv, char **envp) { |
|---|
| 57 | test_reduce(); |
|---|
| 58 | |
|---|
| 59 | test_scan(); |
|---|
| 60 | |
|---|
| 61 | return 0; |
|---|
| 62 | } |
|---|
| 63 | |
|---|
