|
| 1 | +/* |
| 2 | +// Copyright (c) 2024 Ben Ashbaugh |
| 3 | +// |
| 4 | +// SPDX-License-Identifier: MIT |
| 5 | +*/ |
| 6 | + |
| 7 | +#include <popl/popl.hpp> |
| 8 | + |
| 9 | +#include <CL/opencl.hpp> |
| 10 | + |
| 11 | +#include <algorithm> |
| 12 | +#include <chrono> |
| 13 | +#include <cinttypes> |
| 14 | +#include <vector> |
| 15 | + |
| 16 | +#include "util.hpp" |
| 17 | + |
| 18 | +static const char kernelString[] = R"CLC( |
| 19 | +float atomic_add_f(volatile global float* addr, float val) |
| 20 | +{ |
| 21 | + #if defined(__opencl_c_ext_fp32_global_atomic_add) && !defined(EMULATE) |
| 22 | + //#pragma message("using cl_ext_float_atomics") |
| 23 | + return atomic_fetch_add_explicit((volatile global atomic_float*)addr, val, memory_order_relaxed); |
| 24 | + #elif defined(cl_nv_pragma_unroll) && !defined(EMULATE) |
| 25 | + //#pragma message("using PTX atomics") |
| 26 | + float ret; asm volatile("atom.global.add.f32 %0,[%1],%2;":"=f"(ret):"l"(addr),"f"(val):"memory"); |
| 27 | + return ret; |
| 28 | + #elif __has_builtin(__builtin_amdgcn_global_atomic_fadd_f32) && !defined(EMULATE) |
| 29 | + //#pragma message("using AMD atomics") |
| 30 | + return __builtin_amdgcn_global_atomic_fadd_f32(addr, val); |
| 31 | + #elif !defined(SLOW_EMULATE) |
| 32 | + // fallback, see: https://forums.developer.nvidia.com/t/atomicadd-float-float-atomicmul-float-float/14639/7 |
| 33 | + //#pragma message("using emulated float atomics") |
| 34 | + float old = val; while((old=atomic_xchg(addr, atomic_xchg(addr, 0.0f)+old))!=0.0f); |
| 35 | + // Note: this emulated version cannot reliably return the previous value! |
| 36 | + // This makes it unsuitable for general-purpose use, but it is sufficient |
| 37 | + // for some cases, such as reductions. |
| 38 | + return 0.0f; |
| 39 | + #else |
| 40 | + // This is the traditional fallback that uses a compare and exchange loop. |
| 41 | + // It is much slower, but it supports returning the previous value. |
| 42 | + //#pragma message("using slow emulated float atomics") |
| 43 | + volatile global int* iaddr = (volatile global int*)addr; |
| 44 | + int old; |
| 45 | + int check; |
| 46 | + do { |
| 47 | + old = atomic_or(iaddr, 0); // emulated atomic load |
| 48 | + int new = as_int(as_float(old) + val); |
| 49 | + check = atomic_cmpxchg(iaddr, old, new); |
| 50 | + } while (check != old); |
| 51 | + return as_float(old); |
| 52 | + #endif |
| 53 | +} |
| 54 | +
|
| 55 | +kernel void FloatAtomicTest(global float* dst, global float* results) |
| 56 | +{ |
| 57 | + int index = get_global_id(0); |
| 58 | + results[index] = atomic_add_f(dst, 1.0f); |
| 59 | +} |
| 60 | +)CLC"; |
| 61 | + |
| 62 | +static void PrintFloatAtomicCapabilities( |
| 63 | + cl_device_fp_atomic_capabilities_ext caps ) |
| 64 | +{ |
| 65 | + if (caps & CL_DEVICE_GLOBAL_FP_ATOMIC_LOAD_STORE_EXT ) printf("\t\tCL_DEVICE_GLOBAL_FP_ATOMIC_LOAD_STORE_EXT\n"); |
| 66 | + if (caps & CL_DEVICE_GLOBAL_FP_ATOMIC_ADD_EXT ) printf("\t\tCL_DEVICE_GLOBAL_FP_ATOMIC_ADD_EXT\n"); |
| 67 | + if (caps & CL_DEVICE_GLOBAL_FP_ATOMIC_MIN_MAX_EXT ) printf("\t\tCL_DEVICE_GLOBAL_FP_ATOMIC_MIN_MAX_EXT\n"); |
| 68 | + if (caps & CL_DEVICE_LOCAL_FP_ATOMIC_LOAD_STORE_EXT ) printf("\t\tCL_DEVICE_LOCAL_FP_ATOMIC_LOAD_STORE_EXT\n"); |
| 69 | + if (caps & CL_DEVICE_LOCAL_FP_ATOMIC_ADD_EXT ) printf("\t\tCL_DEVICE_LOCAL_FP_ATOMIC_ADD_EXT\n"); |
| 70 | + if (caps & CL_DEVICE_LOCAL_FP_ATOMIC_MIN_MAX_EXT ) printf("\t\tCL_DEVICE_LOCAL_FP_ATOMIC_MIN_MAX_EXT\n"); |
| 71 | + |
| 72 | + cl_device_command_buffer_capabilities_khr extra = caps & ~( |
| 73 | + CL_DEVICE_GLOBAL_FP_ATOMIC_LOAD_STORE_EXT | |
| 74 | + CL_DEVICE_GLOBAL_FP_ATOMIC_ADD_EXT | |
| 75 | + CL_DEVICE_GLOBAL_FP_ATOMIC_MIN_MAX_EXT | |
| 76 | + CL_DEVICE_LOCAL_FP_ATOMIC_LOAD_STORE_EXT | |
| 77 | + CL_DEVICE_LOCAL_FP_ATOMIC_ADD_EXT | |
| 78 | + CL_DEVICE_LOCAL_FP_ATOMIC_MIN_MAX_EXT ); |
| 79 | + if (extra) { |
| 80 | + printf("\t\t(Unknown capability: %016" PRIx64 ")\n", extra); |
| 81 | + } |
| 82 | +} |
| 83 | + |
| 84 | +int main( |
| 85 | + int argc, |
| 86 | + char** argv ) |
| 87 | +{ |
| 88 | + int platformIndex = 0; |
| 89 | + int deviceIndex = 0; |
| 90 | + |
| 91 | + size_t iterations = 16; |
| 92 | + size_t gwx = 64 * 1024; |
| 93 | + |
| 94 | + bool emulate = false; |
| 95 | + bool slowEmulate = false; |
| 96 | + bool check = false; |
| 97 | + |
| 98 | + { |
| 99 | + popl::OptionParser op("Supported Options"); |
| 100 | + op. add<popl::Value< int>>( "p", "platform", "Platform Index", platformIndex, &platformIndex); |
| 101 | + op. add<popl::Value< int>>( "d", "device", "Device Index", deviceIndex, &deviceIndex); |
| 102 | + op. add<popl::Value< size_t>>( "i", "iterations", "Iterations", iterations, &iterations); |
| 103 | + op. add<popl::Value< size_t>>( "", "gwx", "Global Work Size X AKA Number of Atomics", gwx, &gwx); |
| 104 | + op. add<popl::Switch>( "e", "emulate", "Unconditionally Emulate Float Atomics", &emulate); |
| 105 | + op. add<popl::Switch>( "s", "slow-emulate", "Unconditionally Emulate Float Atomics (slowly and safely)", &slowEmulate); |
| 106 | + op. add<popl::Switch>( "c", "check", "Check Intermediate Results", &check); |
| 107 | + |
| 108 | + bool printUsage = false; |
| 109 | + try { |
| 110 | + op.parse(argc, argv); |
| 111 | + } catch (std::exception& e) { |
| 112 | + fprintf(stderr, "Error: %s\n\n", e.what()); |
| 113 | + printUsage = true; |
| 114 | + } |
| 115 | + |
| 116 | + if (printUsage || !op.unknown_options().empty() || !op.non_option_args().empty()) { |
| 117 | + fprintf(stderr, |
| 118 | + "Usage: floatatomics [options]\n" |
| 119 | + "%s", op.help().c_str()); |
| 120 | + return -1; |
| 121 | + } |
| 122 | + } |
| 123 | + |
| 124 | + std::vector<cl::Platform> platforms; |
| 125 | + cl::Platform::get(&platforms); |
| 126 | + |
| 127 | + printf("Running on platform: %s\n", |
| 128 | + platforms[platformIndex].getInfo<CL_PLATFORM_NAME>().c_str() ); |
| 129 | + |
| 130 | + std::vector<cl::Device> devices; |
| 131 | + platforms[platformIndex].getDevices(CL_DEVICE_TYPE_ALL, &devices); |
| 132 | + |
| 133 | + printf("Running on device: %s\n", |
| 134 | + devices[deviceIndex].getInfo<CL_DEVICE_NAME>().c_str() ); |
| 135 | + |
| 136 | + // On some implementations, the feature test macros for float atomics are |
| 137 | + // only defined when compiling for OpenCL C 3.0 or newer. |
| 138 | + std::string buildOptions = "-cl-std=CL3.0"; |
| 139 | + if (slowEmulate) { |
| 140 | + printf("Forcing slow and safe emulation.\n"); |
| 141 | + buildOptions += " -DEMULATE -DSLOW_EMULATE"; |
| 142 | + } else if (emulate) { |
| 143 | + printf("Forcing emulation.\n"); |
| 144 | + buildOptions += " -DEMULATE"; |
| 145 | + } else if (!checkDeviceForExtension(devices[deviceIndex], CL_EXT_FLOAT_ATOMICS_EXTENSION_NAME)) { |
| 146 | + printf("Device does not support " CL_EXT_FLOAT_ATOMICS_EXTENSION_NAME ".\n"); |
| 147 | + } else { |
| 148 | + printf("Device supports " CL_EXT_FLOAT_ATOMICS_EXTENSION_NAME ".\n"); |
| 149 | + |
| 150 | + cl_device_fp_atomic_capabilities_ext spcaps = |
| 151 | + devices[deviceIndex].getInfo<CL_DEVICE_SINGLE_FP_ATOMIC_CAPABILITIES_EXT>(); |
| 152 | + printf("CL_DEVICE_SINGLE_FP_ATOMIC_CAPABILITIES_EXT:\n"); |
| 153 | + PrintFloatAtomicCapabilities(spcaps); |
| 154 | + |
| 155 | + cl_device_fp_atomic_capabilities_ext dpcaps = |
| 156 | + devices[deviceIndex].getInfo<CL_DEVICE_DOUBLE_FP_ATOMIC_CAPABILITIES_EXT>(); |
| 157 | + printf("CL_DEVICE_DOUBLE_FP_ATOMIC_CAPABILITIES_EXT:\n"); |
| 158 | + PrintFloatAtomicCapabilities(dpcaps); |
| 159 | + |
| 160 | + cl_device_fp_atomic_capabilities_ext hpcaps = |
| 161 | + devices[deviceIndex].getInfo<CL_DEVICE_HALF_FP_ATOMIC_CAPABILITIES_EXT>(); |
| 162 | + printf("CL_DEVICE_HALF_FP_ATOMIC_CAPABILITIES_EXT:\n"); |
| 163 | + PrintFloatAtomicCapabilities(hpcaps); |
| 164 | + } |
| 165 | + |
| 166 | + cl::Context context{devices[deviceIndex]}; |
| 167 | + cl::CommandQueue commandQueue{context, devices[deviceIndex]}; |
| 168 | + |
| 169 | + cl::Program program{ context, kernelString }; |
| 170 | + program.build(buildOptions); |
| 171 | + cl::Kernel kernel = cl::Kernel{ program, "FloatAtomicTest" }; |
| 172 | + |
| 173 | + cl::Buffer dst = cl::Buffer{ |
| 174 | + context, |
| 175 | + CL_MEM_READ_WRITE, |
| 176 | + sizeof(cl_float) }; |
| 177 | + cl::Buffer intermediates = cl::Buffer{ |
| 178 | + context, |
| 179 | + CL_MEM_READ_WRITE, |
| 180 | + gwx * sizeof(cl_float) }; |
| 181 | + |
| 182 | + // execution |
| 183 | + { |
| 184 | + kernel.setArg(0, dst); |
| 185 | + kernel.setArg(1, intermediates); |
| 186 | + |
| 187 | + commandQueue.finish(); |
| 188 | + |
| 189 | + auto start = std::chrono::system_clock::now(); |
| 190 | + for (size_t i = 0; i < iterations; i++) { |
| 191 | + cl_float zero = 0.0f; |
| 192 | + commandQueue.enqueueFillBuffer( |
| 193 | + dst, |
| 194 | + zero, |
| 195 | + 0, |
| 196 | + sizeof(zero)); |
| 197 | + commandQueue.enqueueNDRangeKernel( |
| 198 | + kernel, |
| 199 | + cl::NullRange, |
| 200 | + cl::NDRange{gwx}); |
| 201 | + } |
| 202 | + |
| 203 | + commandQueue.finish(); |
| 204 | + |
| 205 | + auto end = std::chrono::system_clock::now(); |
| 206 | + std::chrono::duration<float> elapsed_seconds = end - start; |
| 207 | + printf("Finished in %f seconds\n", elapsed_seconds.count()); |
| 208 | + } |
| 209 | + |
| 210 | + // basic validation |
| 211 | + { |
| 212 | + cl_float check = 0.0f; |
| 213 | + for (size_t i = 0; i < gwx; i++) { |
| 214 | + check += 1.0f; |
| 215 | + } |
| 216 | + |
| 217 | + cl_float result = 0.0f; |
| 218 | + commandQueue.enqueueReadBuffer( |
| 219 | + dst, |
| 220 | + CL_TRUE, |
| 221 | + 0, |
| 222 | + sizeof(result), |
| 223 | + &result); |
| 224 | + if (result != check) { |
| 225 | + printf("Error: expected %f, got %f!\n", check, result); |
| 226 | + } else { |
| 227 | + printf("Basic Validation: Success.\n"); |
| 228 | + } |
| 229 | + } |
| 230 | + |
| 231 | + // intermediate results validation |
| 232 | + if (check) { |
| 233 | + if (emulate && !slowEmulate) { |
| 234 | + printf("The emulated float atomic add does not support intermediate results.\n"); |
| 235 | + } else { |
| 236 | + std::vector<cl_float> test(gwx); |
| 237 | + commandQueue.enqueueReadBuffer( |
| 238 | + intermediates, |
| 239 | + CL_TRUE, |
| 240 | + 0, |
| 241 | + gwx * sizeof(cl_float), |
| 242 | + test.data()); |
| 243 | + |
| 244 | + std::sort(test.begin(), test.end()); |
| 245 | + |
| 246 | + size_t mismatches = 0; |
| 247 | + for (size_t i = 0; i < gwx; i++) { |
| 248 | + if (i == 0 && !(test[i] == 0.0f)) { |
| 249 | + if (mismatches < 16) { |
| 250 | + printf("Error at index %zu: expected %f, got %f!\n", i, 0.0f, test[i]); |
| 251 | + } |
| 252 | + mismatches++; |
| 253 | + } else if (i > 0 && !(test[i] > test[i-1])) { |
| 254 | + if (mismatches < 16) { |
| 255 | + printf("Error at index %zu: expected %f > %f!\n", i, test[i], test[i-1]); |
| 256 | + } |
| 257 | + mismatches++; |
| 258 | + } |
| 259 | + } |
| 260 | + |
| 261 | + if (mismatches) { |
| 262 | + printf("Intermediate Results Validation: Found %zu mismatches / %zu values!!!\n", |
| 263 | + mismatches, gwx); |
| 264 | + } else { |
| 265 | + printf("Intermediate Results Validation: Success.\n"); |
| 266 | + } |
| 267 | + } |
| 268 | + } |
| 269 | + |
| 270 | + return 0; |
| 271 | +} |
0 commit comments