|
| 1 | +// REQUIRES: gpu |
| 2 | +// UNSUPPORTED: cuda |
| 3 | +// RUN: %clangxx -fsycl %s -o %t.out |
| 4 | +// RUN: %GPU_RUN_PLACEHOLDER %t.out |
| 5 | +// |
| 6 | +// The test checks functionality of the slm_gather_rgba/slm_scatter_rgba ESIMD |
| 7 | +// API. |
| 8 | + |
| 9 | +#include "../esimd_test_utils.hpp" |
| 10 | + |
| 11 | +#include <CL/sycl.hpp> |
| 12 | +#include <CL/sycl/INTEL/esimd.hpp> |
| 13 | +#include <iostream> |
| 14 | + |
| 15 | +using namespace cl::sycl; |
| 16 | + |
| 17 | +constexpr int MASKED_LANE_NUM_REV = 1; |
| 18 | +constexpr int NUM_RGBA_CHANNELS = get_num_channels_enabled( |
| 19 | + sycl::ext::intel::experimental::esimd::rgba_channel_mask::ABGR); |
| 20 | + |
| 21 | +template <typename T, unsigned VL, auto CH_MASK> struct Kernel { |
| 22 | + T *bufOut; |
| 23 | + Kernel(T *bufOut) : bufOut(bufOut) {} |
| 24 | + |
| 25 | + void operator()(sycl::nd_item<1> ndi) const SYCL_ESIMD_KERNEL { |
| 26 | + using namespace sycl::ext::intel::experimental::esimd; |
| 27 | + constexpr int numChannels = get_num_channels_enabled(CH_MASK); |
| 28 | + uint32_t i = ndi.get_global_id(0); |
| 29 | + |
| 30 | + // In this test, each group consist of one workitem. No barriers required. |
| 31 | + // Each workitem accesses contiguous block of VL elements, where |
| 32 | + // each element consists of RGBA channels. |
| 33 | + slm_init(VL * NUM_RGBA_CHANNELS * sizeof(T)); |
| 34 | + |
| 35 | + // Prepare initial values in SLM: |
| 36 | + // 0, -1, -2, -3, -4 ... |
| 37 | + // slm_scatter only supports VL = 16 or 32, so conservatively write in |
| 38 | + // chunks of 16 elements. |
| 39 | + constexpr unsigned numStores = (VL * NUM_RGBA_CHANNELS) / 16; |
| 40 | + for (int i = 0; i < numStores; i++) { |
| 41 | + simd<T, 16> vals(-i * 16, -1); |
| 42 | + simd<uint32_t, 16> fourByteOffsets(i * 16 * sizeof(T), sizeof(T)); |
| 43 | + slm_scatter<T, 16>(vals, fourByteOffsets); |
| 44 | + } |
| 45 | + |
| 46 | + // Prepare values to store into SLM in a SOA manner, e.g.: |
| 47 | + // R R R R ... G G G G ... B B B B ... A A A A ... |
| 48 | + // 0, 4, 8, 12, ... 1, 5, 9, 13, ... 2, 6, 10, 14, ... 3, 7, 11, 15 ... |
| 49 | + simd<T, VL * numChannels> valsIn; |
| 50 | + for (unsigned i = 0; i < numChannels; i++) |
| 51 | + for (unsigned j = 0; j < VL; j++) |
| 52 | + valsIn[i * VL + j] = j * numChannels + i; |
| 53 | + |
| 54 | + // Store values to SLM. In the SLM it will be transposed into AOS: |
| 55 | + // R G B A R G B A ... |
| 56 | + // 0, 1, 2, 3, 4, 5, 6, 7 ... |
| 57 | + simd<uint32_t, VL> byteOffsets(0, sizeof(T) * NUM_RGBA_CHANNELS); |
| 58 | + slm_scatter_rgba<T, VL, CH_MASK>(valsIn, byteOffsets); |
| 59 | + |
| 60 | + // Load back values from SLM. They will be transposed back to SOA. |
| 61 | + simd<uint16_t, VL> pred = 1; |
| 62 | + pred[VL - MASKED_LANE_NUM_REV] = 0; // mask out the last lane |
| 63 | + simd<T, VL *numChannels> valsOut = |
| 64 | + slm_gather_rgba<T, VL, CH_MASK>(byteOffsets, pred); |
| 65 | + |
| 66 | + // Copy results to the output USM buffer. Maximum write block size must be |
| 67 | + // at most 8 owords, so conservatively write in chunks of 8 elements. |
| 68 | + uint32_t global_offset = i * VL * NUM_RGBA_CHANNELS; |
| 69 | + for (unsigned i = 0; i < (VL * numChannels) / 8; i++) { |
| 70 | + simd<T, 8> valsToWrite = valsOut.template select<8, 1>(i * 8); |
| 71 | + valsToWrite.copy_to(bufOut + global_offset + i * 8); |
| 72 | + } |
| 73 | + } |
| 74 | +}; |
| 75 | + |
| 76 | +std::string convertMaskToStr( |
| 77 | + sycl::ext::intel::experimental::esimd::rgba_channel_mask mask) { |
| 78 | + using namespace sycl::ext::intel::experimental::esimd; |
| 79 | + switch (mask) { |
| 80 | + case rgba_channel_mask::R: |
| 81 | + return "R"; |
| 82 | + case rgba_channel_mask::GR: |
| 83 | + return "GR"; |
| 84 | + case rgba_channel_mask::ABGR: |
| 85 | + return "ABGR"; |
| 86 | + default: |
| 87 | + return ""; |
| 88 | + } |
| 89 | + return ""; |
| 90 | +} |
| 91 | + |
| 92 | +template <typename T, unsigned VL, auto CH_MASK> bool test(queue q) { |
| 93 | + using namespace sycl::ext::intel::experimental::esimd; |
| 94 | + constexpr int numChannels = get_num_channels_enabled(CH_MASK); |
| 95 | + constexpr size_t size = VL * numChannels; |
| 96 | + |
| 97 | + std::cout << "Testing T=" << typeid(T).name() << " VL=" << VL |
| 98 | + << " MASK=" << convertMaskToStr(CH_MASK) << "...\n"; |
| 99 | + |
| 100 | + auto dev = q.get_device(); |
| 101 | + auto ctxt = q.get_context(); |
| 102 | + T *A = static_cast<T *>(malloc_shared(size * sizeof(T), dev, ctxt)); |
| 103 | + T *gold = new T[size]; |
| 104 | + |
| 105 | + for (int i = 0; i < size; ++i) { |
| 106 | + A[i] = (T)-i; |
| 107 | + } |
| 108 | + |
| 109 | + // Fill out the array with gold values. |
| 110 | + // R R R R ... G G G G ... B B B B ... A A A A ... |
| 111 | + // 0, 4, 8, 12, ... 1, 5, 9, 13, ... 2, 6, 10, 14, ... 3, 7, 11, 15 ... |
| 112 | + for (unsigned i = 0; i < numChannels; i++) |
| 113 | + for (unsigned j = 0; j < VL; j++) |
| 114 | + gold[i * VL + j] = j * numChannels + i; |
| 115 | + |
| 116 | + // Account for masked out last lanes (with pred argument to slm_gather_rgba). |
| 117 | + unsigned maskedIndex = VL - 1; |
| 118 | + for (unsigned i = 0; i < numChannels; i++, maskedIndex += VL) |
| 119 | + gold[maskedIndex] = 0; |
| 120 | + |
| 121 | + try { |
| 122 | + // We need that many workitems |
| 123 | + sycl::range<1> GlobalRange{1}; |
| 124 | + // Number of workitems in a workgroup |
| 125 | + sycl::range<1> LocalRange{1}; |
| 126 | + sycl::nd_range<1> Range{GlobalRange * LocalRange, LocalRange}; |
| 127 | + |
| 128 | + auto e = q.submit([&](handler &cgh) { |
| 129 | + Kernel<T, VL, CH_MASK> kernel(A); |
| 130 | + cgh.parallel_for(Range, kernel); |
| 131 | + }); |
| 132 | + e.wait(); |
| 133 | + } catch (sycl::exception const &e) { |
| 134 | + std::cerr << "SYCL exception caught: " << e.what() << '\n'; |
| 135 | + free(A, ctxt); |
| 136 | + delete[] gold; |
| 137 | + return static_cast<bool>(e.code()); |
| 138 | + } |
| 139 | + |
| 140 | + int err_cnt = 0; |
| 141 | + for (unsigned i = 0; i < size; ++i) { |
| 142 | + if (A[i] != gold[i]) { |
| 143 | + if (++err_cnt < 35) { |
| 144 | + std::cerr << "failed at index " << i << ": " << A[i] |
| 145 | + << " != " << gold[i] << " (gold)\n"; |
| 146 | + } |
| 147 | + } |
| 148 | + } |
| 149 | + |
| 150 | + if (err_cnt > 0) { |
| 151 | + std::cout << " pass rate: " |
| 152 | + << ((float)(size - err_cnt) / (float)size) * 100.0f << "% (" |
| 153 | + << (size - err_cnt) << "/" << size << ")\n"; |
| 154 | + } |
| 155 | + |
| 156 | + free(A, ctxt); |
| 157 | + delete[] gold; |
| 158 | + |
| 159 | + std::cout << (err_cnt > 0 ? " FAILED\n" : " Passed\n"); |
| 160 | + return err_cnt > 0 ? false : true; |
| 161 | +} |
| 162 | + |
| 163 | +template <typename T, unsigned VL> bool test(queue q) { |
| 164 | + using namespace sycl::ext::intel::experimental::esimd; |
| 165 | + bool passed = true; |
| 166 | + passed &= test<T, VL, rgba_channel_mask::R>(q); |
| 167 | + passed &= test<T, VL, rgba_channel_mask::GR>(q); |
| 168 | + passed &= test<T, VL, rgba_channel_mask::ABGR>(q); |
| 169 | + return passed; |
| 170 | +} |
| 171 | + |
| 172 | +int main(void) { |
| 173 | + queue q(esimd_test::ESIMDSelector{}, esimd_test::createExceptionHandler()); |
| 174 | + |
| 175 | + auto dev = q.get_device(); |
| 176 | + std::cout << "Running on " << dev.get_info<info::device::name>() << "\n"; |
| 177 | + |
| 178 | + bool passed = true; |
| 179 | + passed &= test<int, 8>(q); |
| 180 | + passed &= test<int, 16>(q); |
| 181 | + passed &= test<int, 32>(q); |
| 182 | + passed &= test<float, 8>(q); |
| 183 | + passed &= test<float, 16>(q); |
| 184 | + passed &= test<float, 32>(q); |
| 185 | + |
| 186 | + return passed ? 0 : 1; |
| 187 | +} |
0 commit comments