ceed-cuda-compile.cpp (33cc410d9e7c7c1ec2b31835f296422ec750fa8c) ceed-cuda-compile.cpp (fa619ecc52f58ebd3ff3ef012ebe7a24b3c56483)
1// Copyright (c) 2017-2025, 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 <dirent.h>
15#include <nvrtc.h>
16#include <stdarg.h>
1// Copyright (c) 2017-2025, 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 <dirent.h>
15#include <nvrtc.h>
16#include <stdarg.h>
17#include <stdio.h>
17#include <string.h>
18#include <sys/stat.h>
19#include <sys/types.h>
20
21#include <cstdlib>
22#include <fstream>
23#include <iostream>
24#include <sstream>

--- 14 unchanged lines hidden (view full) ---

39 } while (0)
40
41#define CeedCallSystem(ceed, command, message) CeedCallBackend(CeedCallSystem_Core(ceed, command, message))
42
43//------------------------------------------------------------------------------
44// Call system command and capture stdout + stderr
45//------------------------------------------------------------------------------
46static int CeedCallSystem_Core(Ceed ceed, const char *command, const char *message) {
18#include <string.h>
19#include <sys/stat.h>
20#include <sys/types.h>
21
22#include <cstdlib>
23#include <fstream>
24#include <iostream>
25#include <sstream>

--- 14 unchanged lines hidden (view full) ---

40 } while (0)
41
42#define CeedCallSystem(ceed, command, message) CeedCallBackend(CeedCallSystem_Core(ceed, command, message))
43
44//------------------------------------------------------------------------------
45// Call system command and capture stdout + stderr
46//------------------------------------------------------------------------------
47static int CeedCallSystem_Core(Ceed ceed, const char *command, const char *message) {
47 CeedDebug(ceed, "Running command:\n$ %s\n", command);
48 CeedDebug(ceed, "Running command:\n$ %s", command);
48 FILE *output_stream = popen((command + std::string(" 2>&1")).c_str(), "r");
49
49 FILE *output_stream = popen((command + std::string(" 2>&1")).c_str(), "r");
50
50 CeedCheck(output_stream != nullptr, ceed, CEED_ERROR_BACKEND, "Failed to %s with command: %s", message, command);
51 CeedCheck(output_stream != nullptr, ceed, CEED_ERROR_BACKEND, "Failed to %s\ncommand:\n$ %s", message, command);
51
52
52 char output[4 * CEED_MAX_RESOURCE_LEN];
53 char line[CEED_MAX_RESOURCE_LEN] = "";
54 std::string output = "";
53
55
54 while (fgets(output, sizeof(output), output_stream) != nullptr) {
56 while (fgets(line, sizeof(line), output_stream) != nullptr) {
57 output += line;
55 }
58 }
56 CeedDebug(ceed, "Command output:\n%s\n", output);
57
58 CeedCheck(pclose(output_stream) == 0, ceed, CEED_ERROR_BACKEND, "Failed to %s with command: %s\nand error: %s", message, command, output);
59 CeedDebug(ceed, "output:\n%s\n", output.c_str());
60 CeedCheck(pclose(output_stream) == 0, ceed, CEED_ERROR_BACKEND, "Failed to %s\ncommand:\n$ %s\nerror:\n%s", message, command, output.c_str());
59 return CEED_ERROR_SUCCESS;
60}
61
62//------------------------------------------------------------------------------
63// Compile CUDA kernel
64//------------------------------------------------------------------------------
65using std::ifstream;
66using std::ofstream;

--- 208 unchanged lines hidden (view full) ---

275 std::string command;
276
277 for (CeedInt i = 0; i < num_rust_source_dirs; i++) {
278 command = "cargo +" + std::string(rust_toolchain) + " build --release --target nvptx64-nvidia-cuda --config " + rust_dirs[i] +
279 "/.cargo/config.toml --manifest-path " + rust_dirs[i] + "/Cargo.toml";
280 CeedCallSystem(ceed, command.c_str(), "build Rust crate");
281 }
282
61 return CEED_ERROR_SUCCESS;
62}
63
64//------------------------------------------------------------------------------
65// Compile CUDA kernel
66//------------------------------------------------------------------------------
67using std::ifstream;
68using std::ofstream;

--- 208 unchanged lines hidden (view full) ---

277 std::string command;
278
279 for (CeedInt i = 0; i < num_rust_source_dirs; i++) {
280 command = "cargo +" + std::string(rust_toolchain) + " build --release --target nvptx64-nvidia-cuda --config " + rust_dirs[i] +
281 "/.cargo/config.toml --manifest-path " + rust_dirs[i] + "/Cargo.toml";
282 CeedCallSystem(ceed, command.c_str(), "build Rust crate");
283 }
284
285 // Get Clang version
286 bool use_llvm_version = ceed_data->use_llvm_version;
287 int llvm_version = ceed_data->llvm_version;
288
289 if (llvm_version == 0) {
290 command = "$(find $(rustup run " + std::string(rust_toolchain) + " rustc --print sysroot) -name llvm-link) --version";
291 CeedDebug(ceed, "Attempting to detect Rust LLVM version.\ncommand:\n$ %s", command.c_str());
292 FILE *output_stream = popen((command + std::string(" 2>&1")).c_str(), "r");
293
294 CeedCheck(output_stream != nullptr, ceed, CEED_ERROR_BACKEND, "Failed to detect Rust LLVM version");
295
296 char line[CEED_MAX_RESOURCE_LEN] = "";
297 std::string output = "";
298
299 while (fgets(line, sizeof(line), output_stream) != nullptr) {
300 output += line;
301 }
302 CeedDebug(ceed, "output:\n%s", output.c_str());
303 CeedCheck(pclose(output_stream) == 0, ceed, CEED_ERROR_BACKEND, "Failed to detect Rust LLVM version\ncommand:\n$ %s\nerror:\n%s",
304 command.c_str(), output.c_str());
305
306 const char *version_substring = strstr(output.c_str(), "LLVM version ");
307
308 version_substring += 13;
309
310 char *next_dot = strchr((char *)version_substring, '.');
311
312 next_dot[0] = '\0';
313 ceed_data->llvm_version = llvm_version = std::stoi(version_substring);
314 CeedDebug(ceed, "Rust LLVM version number: %d\n", llvm_version);
315
316 command = std::string("clang++-") + std::to_string(llvm_version);
317 output_stream = popen((command + std::string(" 2>&1")).c_str(), "r");
318 ceed_data->use_llvm_version = use_llvm_version = pclose(output_stream) == 0;
319 }
320
283 // Compile wrapper kernel
321 // Compile wrapper kernel
284 command = "clang++ -flto=thin --cuda-gpu-arch=sm_" + std::to_string(prop.major) + std::to_string(prop.minor) +
285 " --cuda-device-only -emit-llvm -S temp/kernel_" + std::to_string(build_id) + "_0_source.cu -o temp/kernel_" +
286 std::to_string(build_id) + "_1_wrapped.ll ";
322 command = "clang++" + (use_llvm_version ? (std::string("-") + std::to_string(llvm_version)) : "") + " -flto=thin --cuda-gpu-arch=sm_" +
323 std::to_string(prop.major) + std::to_string(prop.minor) + " --cuda-device-only -emit-llvm -S temp/kernel_" + std::to_string(build_id) +
324 "_0_source.cu -o temp/kernel_" + std::to_string(build_id) + "_1_wrapped.ll ";
287 command += opts[4];
288 CeedCallSystem(ceed, command.c_str(), "JiT kernel source");
289 CeedCallSystem(ceed, ("chmod 0777 temp/kernel_" + std::to_string(build_id) + "_1_wrapped.ll").c_str(), "update JiT file permissions");
290
325 command += opts[4];
326 CeedCallSystem(ceed, command.c_str(), "JiT kernel source");
327 CeedCallSystem(ceed, ("chmod 0777 temp/kernel_" + std::to_string(build_id) + "_1_wrapped.ll").c_str(), "update JiT file permissions");
328
291 // the find command finds the rust-installed llvm-link tool and runs it
329 // Find Rust's llvm-link tool and runs it
292 command = "$(find $(rustup run " + std::string(rust_toolchain) + " rustc --print sysroot) -name llvm-link) temp/kernel_" +
293 std::to_string(build_id) +
294 "_1_wrapped.ll --ignore-non-bitcode --internalize --only-needed -S -o "
295 "temp/kernel_" +
296 std::to_string(build_id) + "_2_linked.ll ";
297
298 // Searches for .a files in rust directoy
330 command = "$(find $(rustup run " + std::string(rust_toolchain) + " rustc --print sysroot) -name llvm-link) temp/kernel_" +
331 std::to_string(build_id) +
332 "_1_wrapped.ll --ignore-non-bitcode --internalize --only-needed -S -o "
333 "temp/kernel_" +
334 std::to_string(build_id) + "_2_linked.ll ";
335
336 // Searches for .a files in rust directoy
299 // Note: this is necessary because rust crate names may not match the folder they are in
337 // Note: this is necessary because Rust crate names may not match the folder they are in
300 for (CeedInt i = 0; i < num_rust_source_dirs; i++) {
301 std::string dir = rust_dirs[i] + "/target/nvptx64-nvidia-cuda/release";
302 DIR *dp = opendir(dir.c_str());
303
304 CeedCheck(dp != nullptr, ceed, CEED_ERROR_BACKEND, "Could not open directory: %s", dir.c_str());
305 struct dirent *entry;
306
338 for (CeedInt i = 0; i < num_rust_source_dirs; i++) {
339 std::string dir = rust_dirs[i] + "/target/nvptx64-nvidia-cuda/release";
340 DIR *dp = opendir(dir.c_str());
341
342 CeedCheck(dp != nullptr, ceed, CEED_ERROR_BACKEND, "Could not open directory: %s", dir.c_str());
343 struct dirent *entry;
344
307 // finds files ending in .a
345 // Find files ending in .a
308 while ((entry = readdir(dp)) != nullptr) {
309 std::string filename(entry->d_name);
310
311 if (filename.size() >= 2 && filename.substr(filename.size() - 2) == ".a") {
312 command += dir + "/" + filename + " ";
313 }
314 }
315 closedir(dp);
316 // TODO: when libCEED switches to c++17, switch to std::filesystem for the loop above
317 }
318
319 // Link, optimize, and compile final CUDA kernel
346 while ((entry = readdir(dp)) != nullptr) {
347 std::string filename(entry->d_name);
348
349 if (filename.size() >= 2 && filename.substr(filename.size() - 2) == ".a") {
350 command += dir + "/" + filename + " ";
351 }
352 }
353 closedir(dp);
354 // TODO: when libCEED switches to c++17, switch to std::filesystem for the loop above
355 }
356
357 // Link, optimize, and compile final CUDA kernel
320 // note that the find command is used to find the rust-installed llvm tool
321 CeedCallSystem(ceed, command.c_str(), "link C and Rust source");
322 CeedCallSystem(
323 ceed,
324 ("$(find $(rustup run " + std::string(rust_toolchain) + " rustc --print sysroot) -name opt) --passes internalize,inline temp/kernel_" +
325 std::to_string(build_id) + "_2_linked.ll -o temp/kernel_" + std::to_string(build_id) + "_3_opt.bc")
326 .c_str(),
327 "optimize linked C and Rust source");
328 CeedCallSystem(ceed, ("chmod 0777 temp/kernel_" + std::to_string(build_id) + "_2_linked.ll").c_str(), "update JiT file permissions");

--- 147 unchanged lines hidden --- 		358 CeedCallSystem(ceed, command.c_str(), "link C and Rust source");
359 CeedCallSystem(
360 ceed,
361 ("$(find $(rustup run " + std::string(rust_toolchain) + " rustc --print sysroot) -name opt) --passes internalize,inline temp/kernel_" +
362 std::to_string(build_id) + "_2_linked.ll -o temp/kernel_" + std::to_string(build_id) + "_3_opt.bc")
363 .c_str(),
364 "optimize linked C and Rust source");
365 CeedCallSystem(ceed, ("chmod 0777 temp/kernel_" + std::to_string(build_id) + "_2_linked.ll").c_str(), "update JiT file permissions");

--- 147 unchanged lines hidden ---