xref: /libCEED/backends/cuda/ceed-cuda-compile.cpp (revision a171b6ef1192a8f5f3031330339aaab2ded4c8fb)
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   const char           *opts[num_opts];
43   nvrtcProgram          prog;
44   struct cudaDeviceProp prop;
45   Ceed_Cuda            *ceed_data;
46 
47   cudaFree(0);  // Make sure a Context exists for nvrtc
48 
49   std::ostringstream code;
50 
51   // Get kernel specific options, such as kernel constants
52   if (num_defines > 0) {
53     va_list args;
54     va_start(args, num_defines);
55     char *name;
56     int   val;
57 
58     for (int i = 0; i < num_defines; i++) {
59       name = va_arg(args, char *);
60       val  = va_arg(args, int);
61       code << "#define " << name << " " << val << "\n";
62     }
63     va_end(args);
64   }
65 
66   // Standard libCEED definitions for CUDA backends
67   CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-jit.h", &jit_defs_path));
68   {
69     char *source;
70 
71     CeedCallBackend(CeedLoadSourceToBuffer(ceed, jit_defs_path, &source));
72     jit_defs_source = source;
73   }
74   code << jit_defs_source;
75   code << "\n\n";
76   CeedCallBackend(CeedFree(&jit_defs_path));
77   CeedCallBackend(CeedFree(&jit_defs_source));
78 
79   // Non-macro options
80   opts[0] = "-default-device";
81   CeedCallBackend(CeedGetData(ceed, &ceed_data));
82   CeedCallCuda(ceed, cudaGetDeviceProperties(&prop, ceed_data->device_id));
83   std::string arch_arg = "-arch=compute_" + std::to_string(prop.major) + std::to_string(prop.minor);
84   opts[1]              = arch_arg.c_str();
85   opts[2]              = "-Dint32_t=int";
86 
87   // Add string source argument provided in call
88   code << source;
89 
90   // Create Program
91   CeedCallNvrtc(ceed, nvrtcCreateProgram(&prog, code.str().c_str(), NULL, 0, NULL, NULL));
92 
93   // Compile kernel
94   nvrtcResult result = nvrtcCompileProgram(prog, num_opts, opts);
95 
96   if (result != NVRTC_SUCCESS) {
97     char  *log;
98     size_t log_size;
99 
100     CeedDebug256(ceed, CEED_DEBUG_COLOR_ERROR, "---------- CEED JIT SOURCE FAILED TO COMPILE ----------\n");
101     CeedDebug(ceed, "Source:\n%s\n", code.str().c_str());
102     CeedDebug256(ceed, CEED_DEBUG_COLOR_ERROR, "---------- CEED JIT SOURCE FAILED TO COMPILE ----------\n");
103     CeedCallNvrtc(ceed, nvrtcGetProgramLogSize(prog, &log_size));
104     CeedCallBackend(CeedMalloc(log_size, &log));
105     CeedCallNvrtc(ceed, nvrtcGetProgramLog(prog, log));
106     return CeedError(ceed, CEED_ERROR_BACKEND, "%s\n%s", nvrtcGetErrorString(result), log);
107   }
108 
109   CeedCallNvrtc(ceed, nvrtcGetPTXSize(prog, &ptx_size));
110   CeedCallBackend(CeedMalloc(ptx_size, &ptx));
111   CeedCallNvrtc(ceed, nvrtcGetPTX(prog, ptx));
112   CeedCallNvrtc(ceed, nvrtcDestroyProgram(&prog));
113 
114   CeedCallCuda(ceed, cuModuleLoadData(module, ptx));
115   CeedCallBackend(CeedFree(&ptx));
116   return CEED_ERROR_SUCCESS;
117 }
118 
119 //------------------------------------------------------------------------------
120 // Get CUDA kernel
121 //------------------------------------------------------------------------------
122 int CeedGetKernel_Cuda(Ceed ceed, CUmodule module, const char *name, CUfunction *kernel) {
123   CeedCallCuda(ceed, cuModuleGetFunction(kernel, module, name));
124   return CEED_ERROR_SUCCESS;
125 }
126 
127 //------------------------------------------------------------------------------
128 // Run CUDA kernel with block size selected automatically based on the kernel
129 //     (which may use enough registers to require a smaller block size than the
130 //      hardware is capable)
131 //------------------------------------------------------------------------------
132 int CeedRunKernelAutoblockCuda(Ceed ceed, CUfunction kernel, size_t points, void **args) {
133   int min_grid_size, max_block_size;
134 
135   CeedCallCuda(ceed, cuOccupancyMaxPotentialBlockSize(&min_grid_size, &max_block_size, kernel, NULL, 0, 0x10000));
136   CeedCallBackend(CeedRunKernel_Cuda(ceed, kernel, CeedDivUpInt(points, max_block_size), max_block_size, args));
137   return CEED_ERROR_SUCCESS;
138 }
139 
140 //------------------------------------------------------------------------------
141 // Run CUDA kernel
142 //------------------------------------------------------------------------------
143 int CeedRunKernel_Cuda(Ceed ceed, CUfunction kernel, const int grid_size, const int block_size, void **args) {
144   CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, kernel, grid_size, block_size, 1, 1, 0, args));
145   return CEED_ERROR_SUCCESS;
146 }
147 
148 //------------------------------------------------------------------------------
149 // Run CUDA kernel for spatial dimension
150 //------------------------------------------------------------------------------
151 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,
152                           void **args) {
153   CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, kernel, grid_size, block_size_x, block_size_y, block_size_z, 0, args));
154   return CEED_ERROR_SUCCESS;
155 }
156 
157 //------------------------------------------------------------------------------
158 // Run CUDA kernel for spatial dimension with shared memory
159 //------------------------------------------------------------------------------
160 int CeedRunKernelDimShared_Cuda(Ceed ceed, CUfunction kernel, const int grid_size, const int block_size_x, const int block_size_y,
161                                 const int block_size_z, const int shared_mem_size, void **args) {
162 #if CUDA_VERSION >= 9000
163   cuFuncSetAttribute(kernel, CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES, shared_mem_size);
164 #endif
165   CUresult result = cuLaunchKernel(kernel, grid_size, 1, 1, block_size_x, block_size_y, block_size_z, shared_mem_size, NULL, args, NULL);
166 
167   if (result == CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES) {
168     int max_threads_per_block, shared_size_bytes, num_regs;
169 
170     cuFuncGetAttribute(&max_threads_per_block, CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, kernel);
171     cuFuncGetAttribute(&shared_size_bytes, CU_FUNC_ATTRIBUTE_SHARED_SIZE_BYTES, kernel);
172     cuFuncGetAttribute(&num_regs, CU_FUNC_ATTRIBUTE_NUM_REGS, kernel);
173     return CeedError(ceed, CEED_ERROR_BACKEND,
174                      "CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES: max_threads_per_block %d on block size (%d,%d,%d), shared_size %d, num_regs %d",
175                      max_threads_per_block, block_size_x, block_size_y, block_size_z, shared_size_bytes, num_regs);
176   } else CeedChk_Cu(ceed, result);
177   return CEED_ERROR_SUCCESS;
178 }
179 
180 //------------------------------------------------------------------------------
181