AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

SWDEV-539145 - Support extended fine grained system memory pool (#603)

Browse Source 
* Add hipHostMalloc() new flag hipHostMallocUncached which will force to allocate pinned
host memory on extended fine grained system memory pool.
* Add hipHostAlloc() new flag hipHostAllocUncached which will force to allocate pinned
host memory on extended fine grained system memory pool.
* Add hipHostRegister() new flag hipHostRegisterUncached which will force to map
host memory onto extended fine grained system momory pool.
This commit is contained in:
Sang, Tao
2025-07-01 00:40:10 -04:00
committed by GitHub
parent 2ce45143d8
commit a7d7687b8f
5 changed files with 97 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files
hipamd/src/hip_memory.cpp
+28 -6
View File
@@ -400,8 +400,19 @@ hipError_t ihipHostMalloc(void** ptr, size_t sizeBytes, unsigned int flags)
}
unsigned int ihipFlags = CL_MEM_SVM_FINE_GRAIN_BUFFER;
if (flags & hipHostMallocUncached) {
if (IS_WINDOWS) {
return hipErrorInvalidValue;
}
if (flags & (hipHostMallocNonCoherent | hipHostMallocCoherent)) {
return hipErrorInvalidValue;
}
ihipFlags |= ROCCLR_MEM_HSA_UNCACHED;
}
if (flags == 0 ||
flags & (hipHostMallocCoherent | hipHostMallocMapped | hipHostMallocNumaUser) ||
flags & (hipHostMallocCoherent | hipHostMallocMapped | hipHostMallocNumaUser |
hipHostMallocUncached) ||
(!(flags & hipHostMallocNonCoherent) && HIP_HOST_COHERENT)) {
ihipFlags |= CL_MEM_SVM_ATOMICS;
}
@@ -1332,11 +1343,21 @@ hipError_t hipHostGetFlags(unsigned int* flagsPtr, void* hostPtr) {
}
hipError_t ihipHostRegister(void* hostPtr, size_t sizeBytes, unsigned int flags) {
if (hostPtr == nullptr || sizeBytes == 0 || flags > 15) {
if (hostPtr == nullptr || sizeBytes == 0 ||
flags & ~(hipHostRegisterPortable | hipHostRegisterMapped |
hipExtHostRegisterCoarseGrained | hipExtHostRegisterUncached)) {
return hipErrorInvalidValue;
} else {
unsigned int memFlags = CL_MEM_USE_HOST_PTR | CL_MEM_SVM_ATOMICS;
if (flags & hipExtHostRegisterUncached) {
if (IS_WINDOWS) {
return hipErrorInvalidValue;
}
memFlags |= ROCCLR_MEM_HSA_UNCACHED;
}
amd::Memory* mem = new (*hip::host_context) amd::Buffer(*hip::host_context,
CL_MEM_USE_HOST_PTR | CL_MEM_SVM_ATOMICS, sizeBytes);
memFlags, sizeBytes);
constexpr bool sysMemAlloc = false;
constexpr bool skipAlloc = false;
@@ -1358,6 +1379,7 @@ hipError_t ihipHostRegister(void* hostPtr, size_t sizeBytes, unsigned int flags)
}
}
if (mem != nullptr) {
mem->getUserData().deviceId = hip::getCurrentDevice()->deviceId();
// Save the HIP memory flags so that they can be accessed later
@@ -1416,8 +1438,8 @@ hipError_t hipHostAlloc(void** ptr, size_t sizeBytes, unsigned int flags) {
if (ptr == nullptr) {
HIP_RETURN(hipErrorInvalidValue);
}
if (flags > (hipHostAllocPortable | hipHostAllocMapped |
hipHostAllocWriteCombined)) {
if (flags & ~(hipHostAllocPortable | hipHostAllocMapped |
hipHostAllocWriteCombined | hipHostAllocUncached)) {
HIP_RETURN(hipErrorInvalidValue);
}
@@ -4300,4 +4322,4 @@ hipError_t hipMemGetHandleForAddressRange(void* handle, hipDeviceptr_t dptr, siz
HIP_RETURN(hipSuccess);
}
} // namespace hip
} // namespace hip