|
| 1 | +//==-------- usm_gather_scatter_rgba.cpp - DPC++ ESIMD on-device test -----==// |
| 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 | +// REQUIRES: gpu |
| 9 | +// UNSUPPORTED: cuda |
| 10 | +// RUN: %clangxx -fsycl %s -o %t.out |
| 11 | +// RUN: %GPU_RUN_PLACEHOLDER %t.out |
| 12 | +// |
| 13 | +// The test checks functionality of the gather_rgba/scatter_rgba USM-based ESIMD |
| 14 | +// intrinsics. |
| 15 | + |
| 16 | +#include "esimd_test_utils.hpp" |
| 17 | + |
| 18 | +#include <CL/sycl.hpp> |
| 19 | +#include <CL/sycl/INTEL/esimd.hpp> |
| 20 | +#include <iostream> |
| 21 | + |
| 22 | +using namespace cl::sycl; |
| 23 | + |
| 24 | +constexpr int MASKED_LANE_NUM_REV = 1; |
| 25 | +constexpr int NUM_RGBA_CHANNELS = get_num_channels_enabled( |
| 26 | + sycl::ext::intel::experimental::esimd::rgba_channel_mask::ABGR); |
| 27 | + |
| 28 | +template <typename T, unsigned VL, unsigned STRIDE, auto CH_MASK> |
| 29 | +struct Kernel { |
| 30 | + T *bufIn; |
| 31 | + T *bufOut; |
| 32 | + Kernel(T *bufIn, T *bufOut) : bufIn(bufIn), bufOut(bufOut) {} |
| 33 | + |
| 34 | + void operator()(id<1> i) const SYCL_ESIMD_KERNEL { |
| 35 | + using namespace sycl::ext::intel::experimental::esimd; |
| 36 | + constexpr int numChannels = get_num_channels_enabled(CH_MASK); |
| 37 | + |
| 38 | + // every workitem accesses contiguous block of VL * STRIDE elements, |
| 39 | + // where each element consists of RGBA channels. |
| 40 | + uint32_t global_offset = i * VL * STRIDE * NUM_RGBA_CHANNELS; |
| 41 | + |
| 42 | + simd<uint32_t, VL> byteOffsets(0, STRIDE * sizeof(T) * NUM_RGBA_CHANNELS); |
| 43 | + simd<T, VL *numChannels> v = |
| 44 | + gather_rgba<T, VL, CH_MASK>(bufIn + global_offset, byteOffsets); |
| 45 | + v += i; |
| 46 | + |
| 47 | + simd<uint16_t, VL> pred = 1; |
| 48 | + pred[VL - MASKED_LANE_NUM_REV] = 0; // mask out the last lane |
| 49 | + scatter_rgba<T, VL, CH_MASK>(bufOut + global_offset, v, byteOffsets, pred); |
| 50 | + } |
| 51 | +}; |
| 52 | + |
| 53 | +std::string convertMaskToStr( |
| 54 | + sycl::ext::intel::experimental::esimd::rgba_channel_mask mask) { |
| 55 | + using namespace sycl::ext::intel::experimental::esimd; |
| 56 | + switch (mask) { |
| 57 | + case rgba_channel_mask::R: |
| 58 | + return "R"; |
| 59 | + case rgba_channel_mask::GR: |
| 60 | + return "GR"; |
| 61 | + case rgba_channel_mask::ABGR: |
| 62 | + return "ABGR"; |
| 63 | + default: |
| 64 | + return ""; |
| 65 | + } |
| 66 | + return ""; |
| 67 | +} |
| 68 | + |
| 69 | +template <typename T, unsigned VL, unsigned STRIDE, auto CH_MASK> |
| 70 | +bool test(queue q) { |
| 71 | + size_t numWorkItems = 2; |
| 72 | + size_t size = VL * STRIDE * NUM_RGBA_CHANNELS * numWorkItems; |
| 73 | + using namespace sycl::ext::intel::experimental::esimd; |
| 74 | + constexpr int numChannels = get_num_channels_enabled(CH_MASK); |
| 75 | + |
| 76 | + std::cout << "Testing T=" << typeid(T).name() << " VL=" << VL |
| 77 | + << " STRIDE=" << STRIDE << " MASK=" << convertMaskToStr(CH_MASK) |
| 78 | + << "...\n"; |
| 79 | + |
| 80 | + auto dev = q.get_device(); |
| 81 | + auto ctxt = q.get_context(); |
| 82 | + T *A = static_cast<T *>(malloc_shared(size * sizeof(T), dev, ctxt)); |
| 83 | + T *B = static_cast<T *>(malloc_shared(size * sizeof(T), dev, ctxt)); |
| 84 | + T *gold = new T[size]; |
| 85 | + |
| 86 | + for (int i = 0; i < size; ++i) { |
| 87 | + A[i] = (T)i; |
| 88 | + B[i] = (T)-i; |
| 89 | + gold[i] = (T)-i; |
| 90 | + } |
| 91 | + |
| 92 | + // Fill out the array with gold values. The kernel only writes the elements |
| 93 | + // that are not masked. For example, |
| 94 | + // for STRIDE=1 and MASK=R, we have the following indices written: |
| 95 | + // 0, 4, 8, 12 ... |
| 96 | + // for STRIDE=2 and MASK=RG, we have the following indices written: |
| 97 | + // 0, 1, 8, 9, 16, 17 ... |
| 98 | + // All the other elements will be equal to '-A[i]'. |
| 99 | + auto blockSize = VL * STRIDE * NUM_RGBA_CHANNELS; |
| 100 | + for (unsigned i = 0; i < size; i += NUM_RGBA_CHANNELS * STRIDE) |
| 101 | + for (unsigned j = 0; j < numChannels; j++) |
| 102 | + gold[i + j] = A[i + j] + (i / (blockSize)); |
| 103 | + |
| 104 | + // Account for masked out last lanes (with pred argument to scatter_rgba). |
| 105 | + auto maskedElementOffset = (VL - 1) * STRIDE * NUM_RGBA_CHANNELS; |
| 106 | + for (unsigned i = maskedElementOffset; i < size; i += blockSize) |
| 107 | + for (unsigned j = 0; j < numChannels; j++) |
| 108 | + gold[i + j] = -A[i + j]; |
| 109 | + |
| 110 | + try { |
| 111 | + range<1> glob_range{numWorkItems}; |
| 112 | + auto e = q.submit([&](handler &cgh) { |
| 113 | + Kernel<T, VL, STRIDE, CH_MASK> kernel(A, B); |
| 114 | + cgh.parallel_for(glob_range, kernel); |
| 115 | + }); |
| 116 | + e.wait(); |
| 117 | + } catch (cl::sycl::exception const &e) { |
| 118 | + std::cout << "SYCL exception caught: " << e.what() << '\n'; |
| 119 | + free(A, ctxt); |
| 120 | + free(B, ctxt); |
| 121 | + delete[] gold; |
| 122 | + return e.get_cl_code(); |
| 123 | + } |
| 124 | + |
| 125 | + int err_cnt = 0; |
| 126 | + for (unsigned i = 0; i < size; ++i) { |
| 127 | + if (B[i] != gold[i]) { |
| 128 | + if (++err_cnt < 35) { |
| 129 | + std::cout << "failed at index " << i << ": " << B[i] |
| 130 | + << " != " << gold[i] << " (gold)\n"; |
| 131 | + } |
| 132 | + } |
| 133 | + } |
| 134 | + |
| 135 | + if (err_cnt > 0) { |
| 136 | + std::cout << " pass rate: " |
| 137 | + << ((float)(size - err_cnt) / (float)size) * 100.0f << "% (" |
| 138 | + << (size - err_cnt) << "/" << size << ")\n"; |
| 139 | + } |
| 140 | + |
| 141 | + free(A, ctxt); |
| 142 | + free(B, ctxt); |
| 143 | + delete[] gold; |
| 144 | + |
| 145 | + std::cout << (err_cnt > 0 ? " FAILED\n" : " Passed\n"); |
| 146 | + return err_cnt > 0 ? false : true; |
| 147 | +} |
| 148 | + |
| 149 | +template <typename T, unsigned VL, unsigned STRIDE> bool test(queue q) { |
| 150 | + using namespace sycl::ext::intel::experimental::esimd; |
| 151 | + bool passed = true; |
| 152 | + passed &= test<T, VL, STRIDE, rgba_channel_mask::R>(q); |
| 153 | + passed &= test<T, VL, STRIDE, rgba_channel_mask::GR>(q); |
| 154 | + passed &= test<T, VL, STRIDE, rgba_channel_mask::ABGR>(q); |
| 155 | + return passed; |
| 156 | +} |
| 157 | + |
| 158 | +int main(void) { |
| 159 | + queue q(esimd_test::ESIMDSelector{}, esimd_test::createExceptionHandler()); |
| 160 | + |
| 161 | + auto dev = q.get_device(); |
| 162 | + std::cout << "Running on " << dev.get_info<info::device::name>() << "\n"; |
| 163 | + |
| 164 | + bool passed = true; |
| 165 | + passed &= test<int, 16, 1>(q); |
| 166 | + passed &= test<int, 16, 2>(q); |
| 167 | + passed &= test<int, 16, 4>(q); |
| 168 | + passed &= test<int, 32, 1>(q); |
| 169 | + passed &= test<int, 32, 3>(q); |
| 170 | + passed &= test<int, 32, 8>(q); |
| 171 | + passed &= test<float, 16, 1>(q); |
| 172 | + passed &= test<float, 16, 2>(q); |
| 173 | + passed &= test<float, 16, 4>(q); |
| 174 | + passed &= test<float, 32, 1>(q); |
| 175 | + passed &= test<float, 32, 3>(q); |
| 176 | + passed &= test<float, 32, 8>(q); |
| 177 | + |
| 178 | + return passed ? 0 : 1; |
| 179 | +} |
0 commit comments