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

wsl/libhsakmt: fix default device in memory allocation

Browse Source 
Signed-off-by: Flora Cui <flora.cui@amd.com>
Reviewed-by: Tianci Yin <tianci.yin@amd.com>
Part-of: <http://10.67.69.192/wsl/rocr-runtime/-/merge_requests/82>
This commit is contained in:
Flora Cui
2025-06-30 13:51:12 +08:00
committed by Frank Min
parent 67332944d2
commit 75cbeda847
2 changed files with 23 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
libdrm.cpp
+4 -1
View File
@@ -102,7 +102,10 @@ HSAKMTAPI int amdgpu_bo_import(amdgpu_device_handle dev,
enum amdgpu_bo_handle_type type, enum amdgpu_bo_handle_type type,
uint32_t shared_handle, uint32_t shared_handle,
struct amdgpu_bo_import_result *output) { struct amdgpu_bo_import_result *output) {
HsaGraphicsResourceInfo GraphicsResourceInfo; wsl::thunk::WDDMDevice *pDevice = reinterpret_cast<wsl::thunk::WDDMDevice *>(dev);
HsaGraphicsResourceInfo GraphicsResourceInfo = {};
GraphicsResourceInfo.NodeId = pDevice->NodeId();
HSAKMT_STATUS ret = hsaKmtImportDMABufHandle(shared_handle, &GraphicsResourceInfo); HSAKMT_STATUS ret = hsaKmtImportDMABufHandle(shared_handle, &GraphicsResourceInfo);
if (ret == HSAKMT_STATUS_SUCCESS) { if (ret == HSAKMT_STATUS_SUCCESS) {
//use GpuMemory object's address as drm buf handle //use GpuMemory object's address as drm buf handle
memory.cpp
+19 -13
View File
@@ -164,7 +164,8 @@ HSAKMT_STATUS hsaKmtAllocMemoryAlignInternal(HSAuint32 PreferredNode,
} else } else
*MemoryAddress = nullptr; *MemoryAddress = nullptr;
wsl::thunk::WDDMDevice *dev = get_wddmdev(1); uint32_t node = (PreferredNode == 0) ? dxg_runtime->default_node : PreferredNode;
wsl::thunk::WDDMDevice *dev = get_wddmdev(node);
if (!dev) if (!dev)
return HSAKMT_STATUS_ERROR; return HSAKMT_STATUS_ERROR;
@@ -451,7 +452,7 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtRegisterGraphicsHandleToNodesExt(HSAuint64 Graphic
pr_debug("number of nodes %lu\n", NumberOfNodes); pr_debug("number of nodes %lu\n", NumberOfNodes);
GraphicsResourceInfo->NodeId = 1; GraphicsResourceInfo->NodeId = NodeArray[0];
return hsaKmtImportDMABufHandle(GraphicsResourceHandle, return hsaKmtImportDMABufHandle(GraphicsResourceHandle,
GraphicsResourceInfo, GraphicsResourceInfo,
RegisterFlags); RegisterFlags);
@@ -489,7 +490,7 @@ HSAKMT_STATUS hsaKmtImportDMABufHandle(int DMABufFd,
HSA_REGISTER_MEM_FLAGS RegisterFlags) { HSA_REGISTER_MEM_FLAGS RegisterFlags) {
CHECK_DXG_OPEN(); CHECK_DXG_OPEN();
wsl::thunk::WDDMDevice* dev = get_wddmdev(1); wsl::thunk::WDDMDevice* dev = get_wddmdev(GraphicsResourceInfo->NodeId);
wsl::thunk::GpuMemory *gpu_mem = nullptr; wsl::thunk::GpuMemory *gpu_mem = nullptr;
wsl::thunk::GpuMemoryCreateInfo create_info{}; wsl::thunk::GpuMemoryCreateInfo create_info{};
create_info.dmabuf_fd = DMABufFd; create_info.dmabuf_fd = DMABufFd;
@@ -529,7 +530,7 @@ HSAKMT_STATUS hsaKmtImportDMABufHandle(int DMABufFd,
(*allocation_map_)[MemoryAddress] = Allocation( (*allocation_map_)[MemoryAddress] = Allocation(
gpu_mem->GetGpuMemoryHandle(), MemoryAddress, (uint64_t)MemoryAddress, gpu_mem->GetGpuMemoryHandle(), MemoryAddress, (uint64_t)MemoryAddress,
gpu_mem->Size(), false, nullptr, gpu_mem->ClientSize(), gpu_mem->Size(), false, nullptr, gpu_mem->ClientSize(),
1, gpu_mem->Flags()); GraphicsResourceInfo->NodeId, gpu_mem->Flags());
GraphicsResourceInfo->MemoryAddress = MemoryAddress; GraphicsResourceInfo->MemoryAddress = MemoryAddress;
GraphicsResourceInfo->SizeInBytes = gpu_mem->ClientSize(); GraphicsResourceInfo->SizeInBytes = gpu_mem->ClientSize();
@@ -620,10 +621,21 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtDeregisterMemory(void *MemoryAddress) {
return HSAKMT_STATUS_SUCCESS; return HSAKMT_STATUS_SUCCESS;
} }
HSAKMT_STATUS HSAKMTAPI hsaKmtMapMemoryToGPU(void *MemoryAddress, HSAKMT_STATUS HSAKMTAPI hsaKmtMapMemoryToGPU(void *MemoryAddress,
HSAuint64 MemorySizeInBytes, HSAuint64 MemorySizeInBytes,
HSAuint64 *AlternateVAGPU) { HSAuint64 *AlternateVAGPU) {
HSAuint64 NumberOfNodes = 1;
HSAuint32 NodeArray[] = {dxg_runtime->default_node};
HsaMemMapFlags MemMapFlags;
MemMapFlags.Value = 0;
return hsaKmtMapMemoryToGPUNodes(MemoryAddress, MemorySizeInBytes, AlternateVAGPU,
MemMapFlags, NumberOfNodes, NodeArray);
}
HSAKMT_STATUS HSAKMTAPI hsaKmtMapMemoryToGPUNodes(
void *MemoryAddress, HSAuint64 MemorySizeInBytes, HSAuint64 *AlternateVAGPU,
HsaMemMapFlags MemMapFlags, HSAuint64 NumberOfNodes, HSAuint32 *NodeArray) {
CHECK_DXG_OPEN(); CHECK_DXG_OPEN();
if (!MemoryAddress || !AlternateVAGPU) { if (!MemoryAddress || !AlternateVAGPU) {
@@ -662,7 +674,7 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtMapMemoryToGPU(void *MemoryAddress,
if (code != ErrorCode::Success) if (code != ErrorCode::Success)
return HSAKMT_STATUS_ERROR; return HSAKMT_STATUS_ERROR;
wsl::thunk::WDDMDevice *dev = get_wddmdev(1); wsl::thunk::WDDMDevice *dev = gpu_mem->GetDevice();
if (!dev->WaitOnPagingFenceFromCpu()) if (!dev->WaitOnPagingFenceFromCpu())
return HSAKMT_STATUS_ERROR; return HSAKMT_STATUS_ERROR;
@@ -695,7 +707,7 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtMapMemoryToGPU(void *MemoryAddress,
} }
} }
wsl::thunk::WDDMDevice *dev = get_wddmdev(1); wsl::thunk::WDDMDevice *dev = get_wddmdev(NodeArray[0]);
if (!dev) if (!dev)
return HSAKMT_STATUS_ERROR; return HSAKMT_STATUS_ERROR;
@@ -730,12 +742,6 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtMapMemoryToGPU(void *MemoryAddress,
return HSAKMT_STATUS_SUCCESS; return HSAKMT_STATUS_SUCCESS;
} }
HSAKMT_STATUS HSAKMTAPI hsaKmtMapMemoryToGPUNodes(
void *MemoryAddress, HSAuint64 MemorySizeInBytes, HSAuint64 *AlternateVAGPU,
HsaMemMapFlags MemMapFlags, HSAuint64 NumberOfNodes, HSAuint32 *NodeArray) {
return hsaKmtMapMemoryToGPU(MemoryAddress, MemorySizeInBytes, AlternateVAGPU);
}
HSAKMT_STATUS HSAKMTAPI hsaKmtUnmapMemoryToGPU(void *MemoryAddress) { HSAKMT_STATUS HSAKMTAPI hsaKmtUnmapMemoryToGPU(void *MemoryAddress) {
CHECK_DXG_OPEN(); CHECK_DXG_OPEN();