xref: /libCEED/backends/cuda/ceed-cuda-compile.cpp (revision b13efd58b277efef1db70d6f06eaaf4d415a7642)
1 // Copyright (c) 2017-2024, Lawrence Livermore National Security, LLC and other CEED contributors.
2 // All Rights Reserved. See the top-level LICENSE and NOTICE files for details.
3 //
4 // SPDX-License-Identifier: BSD-2-Clause
5 //
6 // This file is part of CEED:  http://github.com/ceed
7 
8 #include "ceed-cuda-compile.h"
9 
10 #include <ceed.h>
11 #include <ceed/backend.h>
12 #include <ceed/jit-tools.h>
13 #include <cuda_runtime.h>
14 #include <nvrtc.h>
15 #include <stdarg.h>
16 #include <string.h>
17 
18 #include <sstream>
19 
20 #include "ceed-cuda-common.h"
21 
22 #define CeedChk_Nvrtc(ceed, x)                                                                              \
23   do {                                                                                                      \
24     nvrtcResult result = static_cast<nvrtcResult>(x);                                                       \
25     if (result != NVRTC_SUCCESS) return CeedError((ceed), CEED_ERROR_BACKEND, nvrtcGetErrorString(result)); \
26   } while (0)
27 
28 #define CeedCallNvrtc(ceed, ...)  \
29   do {                            \
30     int ierr_q_ = __VA_ARGS__;    \
31     CeedChk_Nvrtc(ceed, ierr_q_); \
32   } while (0)
33 
34 //------------------------------------------------------------------------------
35 // Compile CUDA kernel
36 //------------------------------------------------------------------------------
37 int CeedCompile_Cuda(Ceed ceed, const char *source, CUmodule *module, const CeedInt num_defines, ...) {
38   size_t                ptx_size;
39   char                 *ptx;
40   const char           *jit_defs_path, *jit_defs_source;
41   const int             num_opts            = 3;
42   CeedInt               num_jit_source_dirs = 0;
43   const char          **opts;
44   nvrtcProgram          prog;
45   struct cudaDeviceProp prop;
46   Ceed_Cuda            *ceed_data;
47 
48   cudaFree(0);  // Make sure a Context exists for nvrtc
49 
50   std::ostringstream code;
51 
52   // Get kernel specific options, such as kernel constants
53   if (num_defines > 0) {
54     va_list args;
55     va_start(args, num_defines);
56     char *name;
57     int   val;
58 
59     for (int i = 0; i < num_defines; i++) {
60       name = va_arg(args, char *);
61       val  = va_arg(args, int);
62       code << "#define " << name << " " << val << "\n";
63     }
64     va_end(args);
65   }
66 
67   // Standard libCEED definitions for CUDA backends
68   CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-jit.h", &jit_defs_path));
69   {
70     char *source;
71 
72     CeedCallBackend(CeedLoadSourceToBuffer(ceed, jit_defs_path, &source));
73     jit_defs_source = source;
74   }
75   code << jit_defs_source;
76   code << "\n\n";
77   CeedCallBackend(CeedFree(&jit_defs_path));
78   CeedCallBackend(CeedFree(&jit_defs_source));
79 
80   // Non-macro options
81   CeedCallBackend(CeedCalloc(num_opts, &opts));
82   opts[0] = "-default-device";
83   CeedCallBackend(CeedGetData(ceed, &ceed_data));
84   CeedCallCuda(ceed, cudaGetDeviceProperties(&prop, ceed_data->device_id));
85   std::string arch_arg =
86 #if CUDA_VERSION >= 11010
87       // NVRTC used to support only virtual architectures through the option
88       // -arch, since it was only emitting PTX. It will now support actual
89       // architectures as well to emit SASS.
90       // https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#dynamic-code-generation
91       "-arch=sm_"
92 #else
93       "-arch=compute_"
94 #endif
95       + std::to_string(prop.major) + std::to_string(prop.minor);
96   opts[1] = arch_arg.c_str();
97   opts[2] = "-Dint32_t=int";
98   {
99     const char **jit_source_dirs;
100 
101     CeedCallBackend(CeedGetJitSourceRoots(ceed, &num_jit_source_dirs, &jit_source_dirs));
102     CeedCallBackend(CeedRealloc(num_opts + num_jit_source_dirs, &opts));
103     for (CeedInt i = 0; i < num_jit_source_dirs; i++) {
104       std::ostringstream include_dirs_arg;
105 
106       include_dirs_arg << "-I" << jit_source_dirs[i];
107       CeedCallBackend(CeedStringAllocCopy(include_dirs_arg.str().c_str(), (char **)&opts[num_opts + i]));
108     }
109     CeedCallBackend(CeedRestoreJitSourceRoots(ceed, &jit_source_dirs));
110   }
111 
112   // Add string source argument provided in call
113   code << source;
114 
115   // Create Program
116   CeedCallNvrtc(ceed, nvrtcCreateProgram(&prog, code.str().c_str(), NULL, 0, NULL, NULL));
117 
118   // Compile kernel
119   nvrtcResult result = nvrtcCompileProgram(prog, num_opts + num_jit_source_dirs, opts);
120 
121   for (CeedInt i = 0; i < num_jit_source_dirs; i++) {
122     CeedCallBackend(CeedFree(&opts[num_opts + i]));
123   }
124   CeedCallBackend(CeedFree(&opts));
125   if (result != NVRTC_SUCCESS) {
126     char  *log;
127     size_t log_size;
128 
129     CeedDebug256(ceed, CEED_DEBUG_COLOR_ERROR, "---------- CEED JIT SOURCE FAILED TO COMPILE ----------\n");
130     CeedDebug(ceed, "Source:\n%s\n", code.str().c_str());
131     CeedDebug256(ceed, CEED_DEBUG_COLOR_ERROR, "---------- CEED JIT SOURCE FAILED TO COMPILE ----------\n");
132     CeedCallNvrtc(ceed, nvrtcGetProgramLogSize(prog, &log_size));
133     CeedCallBackend(CeedMalloc(log_size, &log));
134     CeedCallNvrtc(ceed, nvrtcGetProgramLog(prog, log));
135     return CeedError(ceed, CEED_ERROR_BACKEND, "%s\n%s", nvrtcGetErrorString(result), log);
136   }
137 
138 #if CUDA_VERSION >= 11010
139   CeedCallNvrtc(ceed, nvrtcGetCUBINSize(prog, &ptx_size));
140   CeedCallBackend(CeedMalloc(ptx_size, &ptx));
141   CeedCallNvrtc(ceed, nvrtcGetCUBIN(prog, ptx));
142 #else
143   CeedCallNvrtc(ceed, nvrtcGetPTXSize(prog, &ptx_size));
144   CeedCallBackend(CeedMalloc(ptx_size, &ptx));
145   CeedCallNvrtc(ceed, nvrtcGetPTX(prog, ptx));
146 #endif
147   CeedCallNvrtc(ceed, nvrtcDestroyProgram(&prog));
148 
149   CeedCallCuda(ceed, cuModuleLoadData(module, ptx));
150   CeedCallBackend(CeedFree(&ptx));
151   return CEED_ERROR_SUCCESS;
152 }
153 
154 //------------------------------------------------------------------------------
155 // Get CUDA kernel
156 //------------------------------------------------------------------------------
157 int CeedGetKernel_Cuda(Ceed ceed, CUmodule module, const char *name, CUfunction *kernel) {
158   CeedCallCuda(ceed, cuModuleGetFunction(kernel, module, name));
159   return CEED_ERROR_SUCCESS;
160 }
161 
162 //------------------------------------------------------------------------------
163 // Run CUDA kernel with block size selected automatically based on the kernel
164 //     (which may use enough registers to require a smaller block size than the
165 //      hardware is capable)
166 //------------------------------------------------------------------------------
167 int CeedRunKernelAutoblockCuda(Ceed ceed, CUfunction kernel, size_t points, void **args) {
168   int min_grid_size, max_block_size;
169 
170   CeedCallCuda(ceed, cuOccupancyMaxPotentialBlockSize(&min_grid_size, &max_block_size, kernel, NULL, 0, 0x10000));
171   CeedCallBackend(CeedRunKernel_Cuda(ceed, kernel, CeedDivUpInt(points, max_block_size), max_block_size, args));
172   return CEED_ERROR_SUCCESS;
173 }
174 
175 //------------------------------------------------------------------------------
176 // Run CUDA kernel
177 //------------------------------------------------------------------------------
178 int CeedRunKernel_Cuda(Ceed ceed, CUfunction kernel, const int grid_size, const int block_size, void **args) {
179   CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, kernel, grid_size, block_size, 1, 1, 0, args));
180   return CEED_ERROR_SUCCESS;
181 }
182 
183 //------------------------------------------------------------------------------
184 // Run CUDA kernel for spatial dimension
185 //------------------------------------------------------------------------------
186 int CeedRunKernelDim_Cuda(Ceed ceed, CUfunction kernel, const int grid_size, const int block_size_x, const int block_size_y, const int block_size_z,
187                           void **args) {
188   CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, kernel, grid_size, block_size_x, block_size_y, block_size_z, 0, args));
189   return CEED_ERROR_SUCCESS;
190 }
191 
192 //------------------------------------------------------------------------------
193 // Run CUDA kernel for spatial dimension with shared memory
194 //------------------------------------------------------------------------------
195 int CeedRunKernelDimShared_Cuda(Ceed ceed, CUfunction kernel, const int grid_size, const int block_size_x, const int block_size_y,
196                                 const int block_size_z, const int shared_mem_size, void **args) {
197 #if CUDA_VERSION >= 9000
198   cuFuncSetAttribute(kernel, CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES, shared_mem_size);
199 #endif
200   CUresult result = cuLaunchKernel(kernel, grid_size, 1, 1, block_size_x, block_size_y, block_size_z, shared_mem_size, NULL, args, NULL);
201 
202   if (result == CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES) {
203     int max_threads_per_block, shared_size_bytes, num_regs;
204 
205     cuFuncGetAttribute(&max_threads_per_block, CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, kernel);
206     cuFuncGetAttribute(&shared_size_bytes, CU_FUNC_ATTRIBUTE_SHARED_SIZE_BYTES, kernel);
207     cuFuncGetAttribute(&num_regs, CU_FUNC_ATTRIBUTE_NUM_REGS, kernel);
208     return CeedError(ceed, CEED_ERROR_BACKEND,
209                      "CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES: max_threads_per_block %d on block size (%d,%d,%d), shared_size %d, num_regs %d",
210                      max_threads_per_block, block_size_x, block_size_y, block_size_z, shared_size_bytes, num_regs);
211   } else CeedChk_Cu(ceed, result);
212   return CEED_ERROR_SUCCESS;
213 }
214 
215 //------------------------------------------------------------------------------
216