rocr/aie: Detect AIE architecture and marketing name (#1459)
* rocr/aie: Detect AIE architecture and marketing name * rocr/aie: Modernize code, update comments
Αυτή η υποβολή περιλαμβάνεται σε:
Yiannis Papadopoulos
υποβλήθηκε από
GitHub
GitHub
γονέας
82d68fc772
υποβολή
37bbc9062a
+108
-9
@@ -47,6 +47,10 @@
|
||||
#include <sys/mman.h>
|
||||
#include <unistd.h>
|
||||
|
||||
#include <array>
|
||||
#include <filesystem>
|
||||
#include <fstream>
|
||||
#include <map>
|
||||
#include <memory>
|
||||
#include <string>
|
||||
|
||||
@@ -63,6 +67,37 @@ static_assert((sizeof(core::ShareableHandle::handle) >= sizeof(uint32_t)) &&
|
||||
(alignof(core::ShareableHandle::handle) >= alignof(uint32_t)),
|
||||
"ShareableHandle cannot store a XDNA handle");
|
||||
|
||||
/// @brief XDNA device type.
|
||||
enum class XDNADeviceType {
|
||||
Phx,
|
||||
Stx, // Strix Halo / Krackan
|
||||
Unknown // Unknown device
|
||||
};
|
||||
|
||||
/// @brief XDNA device ID.
|
||||
struct XDNADeviceId {
|
||||
uint16_t device;
|
||||
|
||||
bool operator<(const XDNADeviceId& other) const { return device < other.device; }
|
||||
};
|
||||
|
||||
/// @brief Supported XDNA devices.
|
||||
static const std::map<XDNADeviceId, XDNADeviceType> supported_xdna_devices = {
|
||||
{{0x1502}, XDNADeviceType::Phx}, // Phoenix
|
||||
{{0x17f0}, XDNADeviceType::Stx}, // Strix Halo / Krackan
|
||||
};
|
||||
|
||||
namespace fs = std::filesystem;
|
||||
|
||||
/// @brief Devnode path for XDNA devices.
|
||||
static const fs::path devnodes_path = "/dev/accel";
|
||||
/// @brief Sysfs path for XDNA devices.
|
||||
static const fs::path sysfs_path = "/sys/class/accel";
|
||||
/// @brief Devnode prefix for XDNA devices.
|
||||
static const std::string devnode_prefix = "accel";
|
||||
/// @brief Maximum devnode minor number for XDNA devices.
|
||||
static const uint32_t devnode_max_minor_num = 64;
|
||||
|
||||
/// @brief Index of the first operand in a command.
|
||||
///
|
||||
/// Before the operands there are:
|
||||
@@ -103,14 +138,12 @@ XdnaDriver::XdnaDriver(std::string devnode_name)
|
||||
: core::Driver(core::DriverType::XDNA, std::move(devnode_name)) {}
|
||||
|
||||
hsa_status_t XdnaDriver::DiscoverDriver(std::unique_ptr<core::Driver>& driver) {
|
||||
const int max_minor_num(64);
|
||||
static const std::string devnode_prefix("/dev/accel/accel");
|
||||
|
||||
for (int i = 0; i < max_minor_num; ++i) {
|
||||
auto tmp_driver = std::unique_ptr<Driver>(new XdnaDriver(devnode_prefix + std::to_string(i)));
|
||||
for (uint32_t i = 0; i < devnode_max_minor_num; ++i) {
|
||||
auto tmp_driver = std::make_unique<XdnaDriver>(devnode_prefix + std::to_string(i));
|
||||
if (tmp_driver->Open() == HSA_STATUS_SUCCESS) {
|
||||
if (tmp_driver->QueryKernelModeDriver(core::DriverQuery::GET_DRIVER_VERSION) ==
|
||||
HSA_STATUS_SUCCESS) {
|
||||
// XDNADriver supports only one XDNA device. Once found, the driver is initialized.
|
||||
driver = std::move(tmp_driver);
|
||||
return HSA_STATUS_SUCCESS;
|
||||
} else {
|
||||
@@ -147,7 +180,8 @@ hsa_status_t XdnaDriver::QueryKernelModeDriver(core::DriverQuery query) {
|
||||
}
|
||||
|
||||
hsa_status_t XdnaDriver::Open() {
|
||||
fd_ = open(devnode_name_.c_str(), O_RDWR | O_CLOEXEC);
|
||||
const auto devnode_path = devnodes_path / devnode_name_;
|
||||
fd_ = open(devnode_path.c_str(), O_RDWR | O_CLOEXEC);
|
||||
if (fd_ < 0) {
|
||||
return HSA_STATUS_ERROR_OUT_OF_RESOURCES;
|
||||
}
|
||||
@@ -182,13 +216,78 @@ hsa_status_t XdnaDriver::GetNodeProperties(HsaNodeProperties& node_props, uint32
|
||||
return HSA_STATUS_ERROR;
|
||||
}
|
||||
|
||||
// Right now can only target N-1 columns as that is the number of shim DMAs
|
||||
// in NPU1 devices.
|
||||
node_props.NumNeuralCores = (aie_metadata.cols - 1) * aie_metadata.core.row_count;
|
||||
const auto sysfs_device_path = sysfs_path / devnode_name_ / "device";
|
||||
|
||||
// Find device type.
|
||||
XDNADeviceType device_type = XDNADeviceType::Unknown;
|
||||
{
|
||||
const auto device_id_file = sysfs_device_path / "device";
|
||||
if (!fs::exists(device_id_file)) {
|
||||
assert(false && "Device file not found in sysfs.");
|
||||
return HSA_STATUS_ERROR;
|
||||
}
|
||||
|
||||
XDNADeviceId device_id = {};
|
||||
std::ifstream is(device_id_file);
|
||||
// Device ID is in hex.
|
||||
if (!(is >> std::hex >> device_id.device)) {
|
||||
assert(false && "Failed to read device ID from sysfs.");
|
||||
return HSA_STATUS_ERROR;
|
||||
}
|
||||
|
||||
const auto device_type_it = supported_xdna_devices.find(device_id);
|
||||
if (device_type_it == supported_xdna_devices.end()) {
|
||||
assert(false && "Unsupported XDNA device.");
|
||||
return HSA_STATUS_ERROR;
|
||||
}
|
||||
device_type = device_type_it->second;
|
||||
}
|
||||
|
||||
// Fill in node properties that depend on device type.
|
||||
std::fill_n(node_props.AMDName, HSA_PUBLIC_NAME_SIZE, 0);
|
||||
switch (device_type) {
|
||||
case XDNADeviceType::Phx: {
|
||||
constexpr std::string_view name("aie2");
|
||||
assert(name.size() < HSA_PUBLIC_NAME_SIZE);
|
||||
std::copy(name.begin(), name.end(), node_props.AMDName);
|
||||
// Only target N-1 columns as that is the number of shim DMAs in NPU1 devices.
|
||||
node_props.NumNeuralCores = (aie_metadata.cols - 1) * aie_metadata.core.row_count;
|
||||
} break;
|
||||
|
||||
case XDNADeviceType::Stx: {
|
||||
constexpr std::string_view name("aie2p");
|
||||
assert(name.size() < HSA_PUBLIC_NAME_SIZE);
|
||||
std::copy(name.begin(), name.end(), node_props.AMDName);
|
||||
node_props.NumNeuralCores = aie_metadata.cols * aie_metadata.core.row_count;
|
||||
} break;
|
||||
|
||||
default:
|
||||
assert(false && "Unsupported XDNA device.");
|
||||
return HSA_STATUS_ERROR;
|
||||
}
|
||||
|
||||
// Read device name from sysfs.
|
||||
{
|
||||
const auto device_name_file = sysfs_device_path / "vbnv";
|
||||
if (!fs::exists(device_name_file)) {
|
||||
assert(false && "Device file name not found in sysfs.");
|
||||
return HSA_STATUS_ERROR;
|
||||
}
|
||||
std::array<char, HSA_PUBLIC_NAME_SIZE> device_name = {};
|
||||
std::ifstream is(device_name_file);
|
||||
if (!is.getline(device_name.data(), device_name.size() - 1)) {
|
||||
assert(false && "Failed to read device name from sysfs.");
|
||||
return HSA_STATUS_ERROR;
|
||||
}
|
||||
// Convert device name from ASCII to UTF-16 for MarketingName.
|
||||
std::copy(device_name.begin(), device_name.end(), node_props.MarketingName);
|
||||
}
|
||||
|
||||
/// @todo XDNA driver currently only supports single-node AIE
|
||||
/// devices over PCIe. Update this once we can get topology
|
||||
/// information dynamically from the sysfs.
|
||||
node_props.NumIOLinks = 0;
|
||||
|
||||
return HSA_STATUS_SUCCESS;
|
||||
}
|
||||
|
||||
|
||||
@@ -42,9 +42,11 @@
|
||||
|
||||
#include "core/inc/amd_aie_agent.h"
|
||||
|
||||
#include <algorithm>
|
||||
#include <cstring>
|
||||
#include <functional>
|
||||
#include <string>
|
||||
#include <iterator>
|
||||
#include <string_view>
|
||||
|
||||
#include "core/inc/amd_aie_aql_queue.h"
|
||||
#include "core/inc/amd_memory_region.h"
|
||||
@@ -111,153 +113,146 @@ hsa_status_t AieAgent::GetInfo(hsa_agent_info_t attribute, void *value) const {
|
||||
const size_t attribute_ = static_cast<size_t>(attribute);
|
||||
|
||||
switch (attribute_) {
|
||||
case HSA_AGENT_INFO_NAME: {
|
||||
const std::string name_info_("aie2");
|
||||
assert(name_info_.size() < HSA_PUBLIC_NAME_SIZE);
|
||||
std::memset(value, 0, HSA_PUBLIC_NAME_SIZE);
|
||||
std::strncat(reinterpret_cast<char *>(value), name_info_.c_str(),
|
||||
name_info_.size());
|
||||
break;
|
||||
}
|
||||
case HSA_AGENT_INFO_VENDOR_NAME: {
|
||||
const std::string vendor_name_info_("AMD");
|
||||
assert(vendor_name_info_.size() < HSA_PUBLIC_NAME_SIZE);
|
||||
std::memset(value, 0, HSA_PUBLIC_NAME_SIZE);
|
||||
std::strncat(reinterpret_cast<char *>(value), vendor_name_info_.c_str(),
|
||||
vendor_name_info_.size());
|
||||
break;
|
||||
}
|
||||
case HSA_AGENT_INFO_FEATURE:
|
||||
*((hsa_agent_feature_t *)value) = HSA_AGENT_FEATURE_AGENT_DISPATCH;
|
||||
break;
|
||||
case HSA_AGENT_INFO_MACHINE_MODEL:
|
||||
*reinterpret_cast<hsa_machine_model_t *>(value) = HSA_MACHINE_MODEL_LARGE;
|
||||
break;
|
||||
case HSA_AGENT_INFO_BASE_PROFILE_DEFAULT_FLOAT_ROUNDING_MODES:
|
||||
case HSA_AGENT_INFO_DEFAULT_FLOAT_ROUNDING_MODE:
|
||||
// TODO: validate if this is true.
|
||||
*reinterpret_cast<hsa_default_float_rounding_mode_t *>(value) =
|
||||
HSA_DEFAULT_FLOAT_ROUNDING_MODE_NEAR;
|
||||
break;
|
||||
case HSA_AGENT_INFO_PROFILE:
|
||||
*reinterpret_cast<hsa_profile_t *>(value) = profile_;
|
||||
break;
|
||||
case HSA_AGENT_INFO_WAVEFRONT_SIZE:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AGENT_INFO_WORKGROUP_MAX_DIM:
|
||||
std::memset(value, 0, sizeof(uint16_t) * 3);
|
||||
break;
|
||||
case HSA_AGENT_INFO_WORKGROUP_MAX_SIZE:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AGENT_INFO_GRID_MAX_DIM:
|
||||
std::memset(value, 0, sizeof(uint16_t) * 3);
|
||||
break;
|
||||
case HSA_AGENT_INFO_GRID_MAX_SIZE:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AGENT_INFO_FBARRIER_MAX_SIZE:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AGENT_INFO_QUEUES_MAX:
|
||||
*reinterpret_cast<uint32_t *>(value) = max_queues_;
|
||||
break;
|
||||
case HSA_AGENT_INFO_QUEUE_MIN_SIZE:
|
||||
*reinterpret_cast<uint32_t *>(value) = min_aql_size_;
|
||||
break;
|
||||
case HSA_AGENT_INFO_QUEUE_MAX_SIZE:
|
||||
*reinterpret_cast<uint32_t *>(value) = max_aql_size_;
|
||||
break;
|
||||
case HSA_AGENT_INFO_QUEUE_TYPE:
|
||||
*reinterpret_cast<hsa_queue_type32_t *>(value) = HSA_QUEUE_TYPE_SINGLE;
|
||||
break;
|
||||
case HSA_AGENT_INFO_NODE:
|
||||
*reinterpret_cast<uint32_t *>(value) = node_id();
|
||||
break;
|
||||
case HSA_AGENT_INFO_DEVICE:
|
||||
*reinterpret_cast<hsa_device_type_t *>(value) = HSA_DEVICE_TYPE_AIE;
|
||||
break;
|
||||
case HSA_AGENT_INFO_CACHE_SIZE:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AGENT_INFO_VERSION_MAJOR:
|
||||
*reinterpret_cast<uint32_t *>(value) = 1;
|
||||
break;
|
||||
case HSA_AGENT_INFO_VERSION_MINOR:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_CHIP_ID:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_CACHELINE_SIZE:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_COMPUTE_UNIT_COUNT:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_MAX_CLOCK_FREQUENCY:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_DRIVER_NODE_ID:
|
||||
*reinterpret_cast<uint32_t *>(value) = node_id();
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_MAX_ADDRESS_WATCH_POINTS:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_BDFID:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_NUM_SIMDS_PER_CU:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_NUM_SHADER_ENGINES:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_NUM_SHADER_ARRAYS_PER_SE:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_1D_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_1DA_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_1DB_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_2D_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_2DA_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_2DDEPTH_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_2DADEPTH_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_3D_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_ARRAY_MAX_LAYERS:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_PRODUCT_NAME: {
|
||||
const std::string product_name_info_("AIE-ML");
|
||||
assert(product_name_info_.size() < HSA_PUBLIC_NAME_SIZE);
|
||||
std::memset(value, 0, HSA_PUBLIC_NAME_SIZE);
|
||||
std::strncat(reinterpret_cast<char *>(value), product_name_info_.c_str(),
|
||||
product_name_info_.size());
|
||||
break;
|
||||
}
|
||||
case HSA_AMD_AGENT_INFO_UUID: {
|
||||
// At this point AIE devices do not support UUID's.
|
||||
char uuid_tmp[] = "AIE-XX";
|
||||
snprintf((char *)value, sizeof(uuid_tmp), "%s", uuid_tmp);
|
||||
break;
|
||||
}
|
||||
case HSA_AMD_AGENT_INFO_ASIC_REVISION:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_SVM_DIRECT_HOST_ACCESS:
|
||||
assert(regions_.size() != 0 && "No device local memory found!");
|
||||
*reinterpret_cast<bool *>(value) = true;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_MEMORY_PROPERTIES:
|
||||
std::memset(value, 0, sizeof(uint8_t) * 8);
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_CLOCK_COUNTERS:
|
||||
std::memset(value, 0, sizeof(hsa_amd_clock_counters_t));
|
||||
break;
|
||||
default:
|
||||
*reinterpret_cast<uint32_t *>(value) = 0;
|
||||
return HSA_STATUS_ERROR_INVALID_ARGUMENT;
|
||||
case HSA_AGENT_INFO_NAME:
|
||||
std::copy_n(node_props_.AMDName, HSA_PUBLIC_NAME_SIZE, static_cast<char*>(value));
|
||||
break;
|
||||
case HSA_AGENT_INFO_VENDOR_NAME: {
|
||||
constexpr std::string_view vendor_name_info("AMD");
|
||||
assert(vendor_name_info.size() < HSA_PUBLIC_NAME_SIZE);
|
||||
auto ptr = static_cast<char*>(value);
|
||||
std::copy(vendor_name_info.begin(), vendor_name_info.end(), ptr);
|
||||
std::fill(std::next(ptr, vendor_name_info.size()), std::next(ptr, HSA_PUBLIC_NAME_SIZE), 0);
|
||||
break;
|
||||
}
|
||||
case HSA_AGENT_INFO_FEATURE:
|
||||
*static_cast<hsa_agent_feature_t*>(value) = HSA_AGENT_FEATURE_AGENT_DISPATCH;
|
||||
break;
|
||||
case HSA_AGENT_INFO_MACHINE_MODEL:
|
||||
*static_cast<hsa_machine_model_t*>(value) = HSA_MACHINE_MODEL_LARGE;
|
||||
break;
|
||||
case HSA_AGENT_INFO_BASE_PROFILE_DEFAULT_FLOAT_ROUNDING_MODES:
|
||||
case HSA_AGENT_INFO_DEFAULT_FLOAT_ROUNDING_MODE:
|
||||
// TODO: validate if this is true.
|
||||
*static_cast<hsa_default_float_rounding_mode_t*>(value) =
|
||||
HSA_DEFAULT_FLOAT_ROUNDING_MODE_NEAR;
|
||||
break;
|
||||
case HSA_AGENT_INFO_PROFILE:
|
||||
*static_cast<hsa_profile_t*>(value) = profile_;
|
||||
break;
|
||||
case HSA_AGENT_INFO_WAVEFRONT_SIZE:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AGENT_INFO_WORKGROUP_MAX_DIM:
|
||||
std::memset(value, 0, sizeof(uint16_t) * 3);
|
||||
break;
|
||||
case HSA_AGENT_INFO_WORKGROUP_MAX_SIZE:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AGENT_INFO_GRID_MAX_DIM:
|
||||
std::memset(value, 0, sizeof(uint16_t) * 3);
|
||||
break;
|
||||
case HSA_AGENT_INFO_GRID_MAX_SIZE:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AGENT_INFO_FBARRIER_MAX_SIZE:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AGENT_INFO_QUEUES_MAX:
|
||||
*static_cast<uint32_t*>(value) = max_queues_;
|
||||
break;
|
||||
case HSA_AGENT_INFO_QUEUE_MIN_SIZE:
|
||||
*static_cast<uint32_t*>(value) = min_aql_size_;
|
||||
break;
|
||||
case HSA_AGENT_INFO_QUEUE_MAX_SIZE:
|
||||
*static_cast<uint32_t*>(value) = max_aql_size_;
|
||||
break;
|
||||
case HSA_AGENT_INFO_QUEUE_TYPE:
|
||||
*static_cast<hsa_queue_type32_t*>(value) = HSA_QUEUE_TYPE_SINGLE;
|
||||
break;
|
||||
case HSA_AGENT_INFO_NODE:
|
||||
*static_cast<uint32_t*>(value) = node_id();
|
||||
break;
|
||||
case HSA_AGENT_INFO_DEVICE:
|
||||
*static_cast<hsa_device_type_t*>(value) = HSA_DEVICE_TYPE_AIE;
|
||||
break;
|
||||
case HSA_AGENT_INFO_CACHE_SIZE:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AGENT_INFO_VERSION_MAJOR:
|
||||
*static_cast<uint32_t*>(value) = 1;
|
||||
break;
|
||||
case HSA_AGENT_INFO_VERSION_MINOR:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_CHIP_ID:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_CACHELINE_SIZE:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_COMPUTE_UNIT_COUNT:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_MAX_CLOCK_FREQUENCY:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_DRIVER_NODE_ID:
|
||||
*static_cast<uint32_t*>(value) = node_id();
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_MAX_ADDRESS_WATCH_POINTS:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_BDFID:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_NUM_SIMDS_PER_CU:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_NUM_SHADER_ENGINES:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_NUM_SHADER_ARRAYS_PER_SE:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_1D_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_1DA_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_1DB_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_2D_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_2DA_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_2DDEPTH_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_2DADEPTH_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_3D_MAX_ELEMENTS:
|
||||
case HSA_EXT_AGENT_INFO_IMAGE_ARRAY_MAX_LAYERS:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_PRODUCT_NAME:
|
||||
// Copy MarketingName which is 7-bit ASCII stored in UTF-16 array
|
||||
std::copy_n(node_props_.MarketingName, HSA_PUBLIC_NAME_SIZE, static_cast<char*>(value));
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_UUID: {
|
||||
// At this point AIE devices do not support UUID's.
|
||||
constexpr std::string_view uuid = "AIE-XX";
|
||||
auto ptr = static_cast<char*>(value);
|
||||
std::copy(uuid.begin(), uuid.end(), ptr);
|
||||
*std::next(ptr, uuid.size()) = '\0';
|
||||
break;
|
||||
}
|
||||
case HSA_AMD_AGENT_INFO_ASIC_REVISION:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_SVM_DIRECT_HOST_ACCESS:
|
||||
assert(regions_.size() != 0 && "No device local memory found!");
|
||||
*static_cast<bool*>(value) = true;
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_MEMORY_PROPERTIES:
|
||||
std::memset(value, 0, sizeof(uint8_t) * 8);
|
||||
break;
|
||||
case HSA_AMD_AGENT_INFO_CLOCK_COUNTERS:
|
||||
std::memset(value, 0, sizeof(hsa_amd_clock_counters_t));
|
||||
break;
|
||||
default:
|
||||
*static_cast<uint32_t*>(value) = 0;
|
||||
return HSA_STATUS_ERROR_INVALID_ARGUMENT;
|
||||
}
|
||||
|
||||
return HSA_STATUS_SUCCESS;
|
||||
|
||||
Αναφορά σε νέο ζήτημα
Block a user