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

rocr: support multiple driver types in agent initialization

Browse Source 
Modify agent initialization to support different driver types,
to enable KFD_VIRTIO dirver for CPU and GPU agent here.

1. Add driver_type parameter to CpuAgent and GpuAgent constructors
2. Update topology discovery to handle multiple driver types
3. Fix MakeMemoryResident return value check in VirtioDriver
4. Add helper function IsGPUDriver to check driver types
5. Update agent discovery to iterate through all available drivers

This change makes the runtime more flexible by removing hardcoded KFD
driver assumptions and properly handling different driver backends.

Signed-off-by: Honglei Huang <Honglei1.Huang@amd.com>


[ROCm/ROCR-Runtime commit: 20806577ce]
This commit is contained in:
Honglei Huang
2025-07-08 11:01:05 +08:00
committed by Huang, Honglei1
parent 8d7d06a867
commit 8f91cd2b03
7 changed files with 67 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_cpu_agent.cpp
+3 -4
View File
@@ -54,10 +54,9 @@
namespace rocr {
namespace AMD {
CpuAgent::CpuAgent(HSAuint32 node, const HsaNodeProperties &node_props)
: core::Agent(
core::Runtime::runtime_singleton_->AgentDriver(core::DriverType::KFD),
node, kAmdCpuDevice),
CpuAgent::CpuAgent(HSAuint32 node, const HsaNodeProperties& node_props,
core::DriverType driver_type)
: core::Agent(core::Runtime::runtime_singleton_->AgentDriver(driver_type), node, kAmdCpuDevice),
properties_(node_props) {
InitRegionList();
projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp
+2 -2
View File
@@ -93,8 +93,8 @@ namespace AMD {
const uint64_t CP_DMA_DATA_TRANSFER_CNT_MAX = (1 << 26);
GpuAgent::GpuAgent(HSAuint32 node, const HsaNodeProperties& node_props, bool xnack_mode,
uint32_t index)
: GpuAgentInt(node),
uint32_t index, core::DriverType driver_type)
: GpuAgentInt(node, driver_type),
properties_(node_props),
current_coherency_type_(HSA_AMD_COHERENCY_TYPE_COHERENT),
scratch_used_large_(0),
projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_topology.cpp
+33 -31
View File
@@ -78,17 +78,14 @@ namespace rocr {
namespace AMD {
// Anonymous namespace.
namespace {
#if _WIN32
constexpr size_t num_drivers = 0;
#elif __linux__
constexpr size_t num_drivers = 2
#ifdef HSAKMT_VIRTIO_ENABLED
+ 1
#endif
;
#endif
const std::array<std::function<hsa_status_t(std::unique_ptr<core::Driver>&)>, num_drivers>
const std::array<std::function<hsa_status_t(std::unique_ptr<core::Driver>&)>,
#if _WIN32
0
#elif __linux__
static_cast<size_t>(core::DriverType::NUM_DRIVER_TYPES)
#endif
>
discover_driver_funcs = {
#ifdef __linux__
KfdDriver::DiscoverDriver,
@@ -121,14 +118,14 @@ bool InitializeDriver(std::unique_ptr<core::Driver>& driver) {
return true;
}
void DiscoverCpu(HSAuint32 node_id, HsaNodeProperties& node_prop) {
CpuAgent* cpu = new CpuAgent(node_id, node_prop);
void DiscoverCpu(HSAuint32 node_id, HsaNodeProperties& node_prop, core::DriverType driver_type) {
CpuAgent* cpu = new CpuAgent(node_id, node_prop, driver_type);
cpu->Enable();
core::Runtime::runtime_singleton_->RegisterAgent(cpu, true);
}
GpuAgent* DiscoverGpu(HSAuint32 node_id, HsaNodeProperties& node_prop, bool xnack_mode,
bool enabled) {
bool enabled, core::DriverType driver_type) {
GpuAgent* gpu = nullptr;
if (node_prop.NumFComputeCores == 0) {
// Ignore non GPUs.
@@ -136,7 +133,7 @@ GpuAgent* DiscoverGpu(HSAuint32 node_id, HsaNodeProperties& node_prop, bool xnac
}
try {
gpu = new GpuAgent(node_id, node_prop, xnack_mode,
core::Runtime::runtime_singleton_->gpu_agents().size());
core::Runtime::runtime_singleton_->gpu_agents().size(), driver_type);
const HsaVersionInfo& kfd_version = core::Runtime::runtime_singleton_->KfdVersion().version;
@@ -163,7 +160,7 @@ GpuAgent* DiscoverGpu(HSAuint32 node_id, HsaNodeProperties& node_prop, bool xnac
node_prop.Capability.ui32.SRAM_EDCSupport = 1;
delete gpu;
gpu = new GpuAgent(node_id, node_prop, xnack_mode,
core::Runtime::runtime_singleton_->gpu_agents().size());
core::Runtime::runtime_singleton_->gpu_agents().size(), driver_type);
}
}
} catch (const hsa_exception& e) {
@@ -268,24 +265,29 @@ void SurfaceGpuList(std::vector<int32_t>& gpu_list, bool xnack_mode, bool enable
const int32_t invalidIdx = -1;
int32_t list_sz = gpu_list.size();
HsaNodeProperties node_prop = {0};
const auto& gpu_driver = core::Runtime::runtime_singleton_->AgentDriver(core::DriverType::KFD);
for (int32_t idx = 0; idx < list_sz; idx++) {
if (gpu_list[idx] == invalidIdx) {
break;
for (const auto& gpu_driver : core::Runtime::runtime_singleton_->AgentDrivers()) {
if (!core::Runtime::IsGPUDriver(gpu_driver->kernel_driver_type_)) {
continue;
}
// Obtain properties of the node
hsa_status_t ret = gpu_driver.GetNodeProperties(node_prop, gpu_list[idx]);
assert(ret == HSA_STATUS_SUCCESS && "Error in getting Node Properties");
for (int32_t idx = 0; idx < list_sz; idx++) {
if (gpu_list[idx] == invalidIdx) {
break;
}
// disable interrupt signal for DTIF platform
if (core::Runtime::runtime_singleton_->flag().enable_dtif())
core::g_use_interrupt_wait = false;
// Obtain properties of the node
hsa_status_t ret = gpu_driver->GetNodeProperties(node_prop, gpu_list[idx]);
assert(ret == HSA_STATUS_SUCCESS && "Error in getting Node Properties");
// Instantiate a Gpu device. The IO links
// of this node have already been registered
assert((node_prop.NumFComputeCores != 0) && "Improper node used for GPU device discovery.");
DiscoverGpu(gpu_list[idx], node_prop, xnack_mode, enabled);
// disable interrupt signal for DTIF platform
if (core::Runtime::runtime_singleton_->flag().enable_dtif())
core::g_use_interrupt_wait = false;
// Instantiate a Gpu device. The IO links
// of this node have already been registered
assert((node_prop.NumFComputeCores != 0) && "Improper node used for GPU device discovery.");
DiscoverGpu(gpu_list[idx], node_prop, xnack_mode, enabled, gpu_driver->kernel_driver_type_);
}
}
}
@@ -346,7 +348,7 @@ bool BuildTopology() {
/// @todo: Add support for AIEs.
// Query if env ROCR_VISIBLE_DEVICES is defined. If defined
// determine number and order of GPU devices to be surfaced.
if (filter && driver->kernel_driver_type_ == core::DriverType::KFD) {
if (filter && (core::Runtime::IsGPUDriver(driver->kernel_driver_type_))) {
rvdFilter.BuildRvdTokenList();
rvdFilter.BuildDeviceUuidList(node_props_vec);
visibleCnt = rvdFilter.BuildUsrDeviceList();
@@ -361,7 +363,7 @@ bool BuildTopology() {
for (auto& node_props : node_props_vec) {
if (node_props.NumCPUCores) {
// Node has CPU cores so instantiate a CPU agent.
DiscoverCpu(node_id, node_props);
DiscoverCpu(node_id, node_props, driver->kernel_driver_type_);
}
if (node_props.NumNeuralCores) {