Handle async cuda copy
This commit is contained in:
parent
8a879a4e8c
commit
c689e8b21b
|
@ -2234,6 +2234,7 @@ ValueCategory MLIRScanner::VisitCallExpr(clang::CallExpr *expr) {
|
||||||
}
|
}
|
||||||
if (sr->getDecl()->getIdentifier() &&
|
if (sr->getDecl()->getIdentifier() &&
|
||||||
(sr->getDecl()->getName() == "cudaMemcpy" ||
|
(sr->getDecl()->getName() == "cudaMemcpy" ||
|
||||||
|
sr->getDecl()->getName() == "cudaMemcpyAsync" ||
|
||||||
sr->getDecl()->getName() == "cudaMemcpyToSymbol" ||
|
sr->getDecl()->getName() == "cudaMemcpyToSymbol" ||
|
||||||
sr->getDecl()->getName() == "memcpy" ||
|
sr->getDecl()->getName() == "memcpy" ||
|
||||||
sr->getDecl()->getName() == "__builtin_memcpy")) {
|
sr->getDecl()->getName() == "__builtin_memcpy")) {
|
||||||
|
@ -2586,13 +2587,15 @@ ValueCategory MLIRScanner::VisitCallExpr(clang::CallExpr *expr) {
|
||||||
"calloc",
|
"calloc",
|
||||||
"free",
|
"free",
|
||||||
"fgets",
|
"fgets",
|
||||||
"__erno_location",
|
"__errno_location",
|
||||||
"__assert_fail",
|
"__assert_fail",
|
||||||
"cudaEventElapsedTime",
|
"cudaEventElapsedTime",
|
||||||
"cudaEventSynchronize",
|
"cudaEventSynchronize",
|
||||||
"cudaDeviceGetAttribute",
|
"cudaDeviceGetAttribute",
|
||||||
"cudaFuncGetAttributes",
|
"cudaFuncGetAttributes",
|
||||||
"cudaGetDevice",
|
"cudaGetDevice",
|
||||||
|
"cudaGetDeviceCount",
|
||||||
|
"clock_gettime",
|
||||||
"cudaOccupancyMaxActiveBlocksPerMultiprocessor",
|
"cudaOccupancyMaxActiveBlocksPerMultiprocessor",
|
||||||
"cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags",
|
"cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags",
|
||||||
"cudaEventRecord"};
|
"cudaEventRecord"};
|
||||||
|
|
Loading…
Reference in New Issue