Cold-Air-Networks/mirror-llvm-project
2
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 7

[offload] use argument pointer array in olLaunchKernel (#194333)

Browse source 
This PR changes olLaunchKernel to accept an array of pointers to
arguments:
```
  void *ArgPtrs[] = {&A, &B, &C};
  size_t ArgSizes[] = {sizeof(A), sizeof(B), sizeof(C)};

  olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, std::size(ArgPtrs), ArgPtrs, ArgSizes);
```

The newly proposed interface is implementable by existing and
anticipated
backends, is familiar to CUDA programmers, eliminates the extraneous
construction of a contiguous arguments buffer, replacing it with
constructing
an array of pointers, sidesteps the alignment requirements, does not
require reading program image metadata where it's impractical, and
enables
a compliant SYCL implementation to be built on top of it.

The ArgSizes array is required to support OpenCL, which does not have
native support for launching a kernel with an argument pointer array, or
a reliable way of retrieving argument sizes for a kernel.

## Mapping the proposed API to backends

CUDA and Level-Zero both support accepting an array of pointers to
kernel arguments, through `cuLaunchKernel`

(https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__EXEC.html#group__CUDA__EXEC_1gb8f3dc3031b40da29d5f9a7139e52e15)
and `zeCommandListAppendLaunchKernelWithArguments`

(https://oneapi-src.github.io/level-zero-spec/level-zero/latest/core/api.html#zecommandlistappendlaunchkernelwitharguments)
respectively.

For OpenCL, which requires the kernel arguments to be set separately
from the kernel launch, a potential implementation can extract the
number of arguments from the OpenCL API, and then iterate over the
argument pointer and size arrays:
```
  cl_uint num_args = 0;
  cl_int err = clGetKernelInfo(kernel, CL_KERNEL_NUM_ARGS,
                               sizeof(num_args), &num_args, NULL);
  for (cl_uint i = 0; i < num_args; ++i) {
    clSetKernelArg(kernel, i, ArgSizes[i], ArgPtrs[i]);
  }
```
The AMDGPU plugin needs to construct a contiguous buffer using the array
of argument pointers. To do this conversion, we need to have offsets
at which to place the arguments. Here, luckily, as mentioned before,
the AMD plugin already reads kernel metadata from the program image.
The implementation simply retrieves size and offset for each argument
from the kernel image, and then uses it to populate a buffer.
This commit is contained in:
Piotr Balcer 2026-05-29 19:54:22 +02:00 committed by GitHub
parent bd5c724752
commit a10ca8ed64
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
26 changed files with 439 additions and 279 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

5
llvm/include/llvm/Frontend/Offloading/Utility.h
View file

@ -15,6 +15,7 @@
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/SmallString.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringMap.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/IR/Module.h"
@ -154,6 +155,10 @@ struct AMDGPUKernelMetaData {
uint32_t WavefrontSize = KInvalidValue;
/// Maximum flat work-group size supported by the kernel in work-items.
uint32_t MaxFlatWorkgroupSize = KInvalidValue;
/// Per-argument {offset, size} in bytes, read from the ".args" array in code
/// object metadata. Explicit user arguments are first, followed by
/// hidden arguments.
SmallVector<std::pair<uint32_t, uint32_t>, 8> ArgMDs;
};
/// Reads AMDGPU specific metadata from the ELF file and propagates the

21
llvm/lib/Frontend/Offloading/Utility.cpp
View file

@ -315,6 +315,27 @@ private:
KernelData.WavefrontSize = V.second.getUInt();
} else if (IsKey(V.first, ".max_flat_workgroup_size")) {
KernelData.MaxFlatWorkgroupSize = V.second.getUInt();
} else if (IsKey(V.first, ".args")) {
auto ArgsArray = V.second.getArray();
for (auto ArgIt = ArgsArray.begin(), ArgEnd = ArgsArray.end();
ArgIt != ArgEnd; ++ArgIt) {
auto ArgMap = ArgIt->getMap();
auto OffsetIt = ArgMap.find(".offset");
if (OffsetIt == ArgMap.end())
return createStringError(
inconvertibleErrorCode(),
"Missing required .offset key in kernel argument metadata map");
auto SizeIt = ArgMap.find(".size");
if (SizeIt == ArgMap.end())
return createStringError(
inconvertibleErrorCode(),
"Missing required .size key in kernel argument metadata map");
KernelData.ArgMDs.emplace_back(OffsetIt->second.getUInt(),
SizeIt->second.getUInt());
}
}
return Error::success();

46
llvm/tools/llvm-gpu-loader/llvm-gpu-loader.cpp
View file

@ -72,24 +72,6 @@ static cl::list<std::string> Args(cl::ConsumeAfter,
cl::desc("<program arguments>..."),
cl::cat(LoaderCategory));
// The arguments to the '_begin' kernel.
struct BeginArgs {
int Argc;
void *Argv;
void *Envp;
};
// The arguments to the '_start' kernel.
struct StartArgs {
int Argc;
void *Argv;
void *Envp;
void *Ret;
};
// The arguments to the '_end' kernel.
struct EndArgs {};
[[noreturn]] static void handleError(Error E) {
outs().flush();
logAllUnhandledErrors(std::move(E), WithColor::error(errs(), "loader"));
@ -183,16 +165,23 @@ ol_device_handle_t getHostDevice() {
return Device;
}
template <typename Args>
template <typename... Args>
void launchKernel(ol_queue_handle_t Queue, ol_device_handle_t Device,
ol_program_handle_t Program, const char *Name,
ol_kernel_launch_size_args_t LaunchArgs, Args &KernelArgs) {
ol_kernel_launch_size_args_t LaunchArgs,
Args &...KernelArgs) {
ol_symbol_handle_t Kernel;
OFFLOAD_ERR(olGetSymbol(Program, Name, OL_SYMBOL_KIND_KERNEL, &Kernel));
OFFLOAD_ERR(olLaunchKernel(Queue, Device, Kernel, &KernelArgs,
std::is_empty_v<Args> ? 0 : sizeof(Args),
&LaunchArgs, /*Properties=*/nullptr));
if constexpr (sizeof...(Args) == 0) {
OFFLOAD_ERR(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr, 0,
nullptr, nullptr));
} else {
void *ArgPtrs[] = {static_cast<void *>(&KernelArgs)...};
size_t ArgSizes[] = {sizeof(KernelArgs)...};
OFFLOAD_ERR(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr,
sizeof...(Args), ArgPtrs, ArgSizes));
}
}
int main(int argc, const char **argv, const char **envp) {
@ -273,8 +262,8 @@ int main(int argc, const char **argv, const char **envp) {
OFFLOAD_ERR(olMemcpy(Queue, DevRet, Device, &Zero, Host, sizeof(int)));
ol_kernel_launch_size_args_t BeginLaunch{1, {1, 1, 1}, {1, 1, 1}, 0};
BeginArgs BeginArgs = {DevArgc, DevArgv, DevEnvp};
launchKernel(Queue, Device, Program, "_begin", BeginLaunch, BeginArgs);
launchKernel(Queue, Device, Program, "_begin", BeginLaunch, DevArgc, DevArgv,
DevEnvp);
OFFLOAD_ERR(olSyncQueue(Queue));
uint32_t Dims = (BlocksZ > 1) ? 3 : (BlocksY > 1) ? 2 : 1;
@ -282,12 +271,11 @@ int main(int argc, const char **argv, const char **envp) {
{BlocksX, BlocksY, BlocksZ},
{ThreadsX, ThreadsY, ThreadsZ},
/*SharedMemBytes=*/0};
StartArgs StartArgs = {DevArgc, DevArgv, DevEnvp, DevRet};
launchKernel(Queue, Device, Program, "_start", StartLaunch, StartArgs);
launchKernel(Queue, Device, Program, "_start", StartLaunch, DevArgc, DevArgv,
DevEnvp, DevRet);
ol_kernel_launch_size_args_t EndLaunch{1, {1, 1, 1}, {1, 1, 1}, 0};
EndArgs EndArgs = {};
launchKernel(Queue, Device, Program, "_end", EndLaunch, EndArgs);
launchKernel(Queue, Device, Program, "_end", EndLaunch);
int Ret;
OFFLOAD_ERR(olMemcpy(Queue, &Ret, Host, DevRet, Device, sizeof(int)));

5
llvm/tools/llvm-gpu-loader/llvm-gpu-loader.h
View file

@ -133,9 +133,10 @@ ol_result_t (*olGetSymbol)(ol_program_handle_t Program, const char *Name,
ol_result_t (*olLaunchKernel)(
ol_queue_handle_t Queue, ol_device_handle_t Device,
ol_symbol_handle_t Kernel, const void *ArgumentsData, size_t ArgumentsSize,
ol_symbol_handle_t Kernel,
const ol_kernel_launch_size_args_t *LaunchSizeArgs,
const ol_kernel_launch_prop_t *Properties);
const ol_kernel_launch_prop_t *Properties, size_t NumArgs, void **ArgPtrs,
const size_t *ArgSizes);
ol_result_t (*olCreateQueue)(ol_device_handle_t Device,
ol_queue_handle_t *Queue);

5
offload/include/Shared/APITypes.h
View file

@ -105,8 +105,9 @@ struct KernelArgsTy {
uint64_t IsCUDA : 1; // Was this kernel spawned via CUDA.
uint64_t DynCGroupMemFallback : 2; // The fallback for dynamic cgroup mem.
uint64_t Cooperative : 1; // Was this kernel spawned as cooperative.
uint64_t Unused : 59;
} Flags = {0, 0, 0, 0, 0};
uint64_t IsPtrArgs : 1; // Arguments are laid out as an array of pointers.
uint64_t Unused : 58;
} Flags = {0, 0, 0, 0, 0, 0};
// User-requested number of blocks (for x,y,z dimension).
uint32_t UserNumBlocks[3] = {0, 0, 0};
// User-requested number of threads (for x,y,z dimension).

20
offload/liboffload/API/Kernel.td
View file

@ -42,22 +42,30 @@ def ol_kernel_launch_prop_t : Struct {
}
def olLaunchKernel : Function {
let desc = "Enqueue a kernel launch with the specified size and parameters.";
let desc = "Enqueue a kernel launch with arguments specified as an array of pointers and sizes.";
let details = [
"If a queue is not specified, kernel execution happens synchronously",
"ArgumentsData may be set to NULL (to indicate no parameters)"
"Each element of ArgPtrs points to the value of the corresponding kernel argument",
"Each element of ArgSizes specifies the size in bytes of the corresponding argument",
"NumArgs must match the number of arguments expected by the kernel",
"ArgPtrs and ArgSizes must both be NULL (when NumArgs == 0) or both be non-NULL"
];
let params = [
Param<"ol_queue_handle_t", "Queue", "handle of the queue", PARAM_IN_OPTIONAL>,
Param<"ol_device_handle_t", "Device", "handle of the device to execute on", PARAM_IN>,
Param<"ol_symbol_handle_t", "Kernel", "handle of the kernel", PARAM_IN>,
Param<"const void*", "ArgumentsData", "pointer to the kernel argument struct", PARAM_IN_OPTIONAL>,
Param<"size_t", "ArgumentsSize", "size of the kernel argument struct", PARAM_IN>,
Param<"const ol_kernel_launch_size_args_t*", "LaunchSizeArgs", "pointer to the struct containing launch size parameters", PARAM_IN>,
Param<"const ol_kernel_launch_prop_t *", "Properties", "Array of optional properties, last element must be OL_KERNEL_LAUNCH_PROP_END", PARAM_IN_OPTIONAL>
Param<"const ol_kernel_launch_prop_t *", "Properties", "Array of optional properties, last element must be OL_KERNEL_LAUNCH_PROP_END", PARAM_IN_OPTIONAL>,
Param<"size_t", "NumArgs", "number of kernel arguments", PARAM_IN>,
Param<"void**", "ArgPtrs", "array of pointers, each pointing to a kernel argument value", PARAM_IN_OPTIONAL>,
Param<"const size_t*", "ArgSizes", "array of kernel argument sizes", PARAM_IN_OPTIONAL>,
];
let returns = [
Return<"OL_ERRC_INVALID_ARGUMENT", ["`ArgumentsSize > 0 && ArgumentsData == NULL`"]>,
Return<"OL_ERRC_INVALID_ARGUMENT", [
"`(ArgPtrs == NULL) != (ArgSizes == NULL)`",
"`NumArgs > 0 && (ArgPtrs == NULL || ArgSizes == NULL)`",
"NumArgs does not match the number of arguments expected by the kernel",
]>,
Return<"OL_ERRC_INVALID_DEVICE", ["If Queue is non-null but does not belong to Device"]>,
Return<"OL_ERRC_SYMBOL_KIND", ["The provided symbol is not a kernel"]>,
];

18
offload/liboffload/src/OffloadImpl.cpp
View file

@ -1098,10 +1098,11 @@ Error olGetKernelMaxCooperativeGroupCount_impl(
}
Error olLaunchKernel_impl(ol_queue_handle_t Queue, ol_device_handle_t Device,
ol_symbol_handle_t Kernel, const void *ArgumentsData,
size_t ArgumentsSize,
ol_symbol_handle_t Kernel,
const ol_kernel_launch_size_args_t *LaunchSizeArgs,
const ol_kernel_launch_prop_t *Properties) {
const ol_kernel_launch_prop_t *Properties,
size_t NumArgs, void **ArgPtrs,
const size_t *ArgSizes) {
auto *DeviceImpl = Device->Device;
if (Queue && Device != Queue->Device) {
return createOffloadError(
@ -1115,6 +1116,7 @@ Error olLaunchKernel_impl(ol_queue_handle_t Queue, ol_device_handle_t Device,
auto *QueueImpl = Queue ? Queue->AsyncInfo : nullptr;
KernelArgsTy LaunchArgs{};
LaunchArgs.NumArgs = static_cast<uint32_t>(NumArgs);
LaunchArgs.UserNumBlocks[0] = LaunchSizeArgs->NumGroups.x;
LaunchArgs.UserNumBlocks[1] = LaunchSizeArgs->NumGroups.y;
LaunchArgs.UserNumBlocks[2] = LaunchSizeArgs->NumGroups.z;
@ -1138,12 +1140,10 @@ Error olLaunchKernel_impl(ol_queue_handle_t Queue, ol_device_handle_t Device,
}
AsyncInfoWrapperTy AsyncInfoWrapper(*DeviceImpl, QueueImpl);
KernelLaunchParamsTy Params;
Params.Data = const_cast<void *>(ArgumentsData);
Params.Size = ArgumentsSize;
LaunchArgs.ArgPtrs = reinterpret_cast<void **>(&Params);
// Don't do anything with pointer indirection; use arg data as-is
LaunchArgs.Flags.IsCUDA = true;
LaunchArgs.ArgPtrs = ArgPtrs;
LaunchArgs.ArgSizes =
reinterpret_cast<int64_t *>(const_cast<size_t *>(ArgSizes));
LaunchArgs.Flags.IsPtrArgs = true;
auto *KernelImpl = std::get<GenericKernelTy *>(Kernel->PluginImpl);
auto Err = KernelImpl->launch(*DeviceImpl, LaunchArgs.ArgPtrs, nullptr,

74
offload/plugins-nextgen/amdgpu/src/rtl.cpp
View file

@ -511,12 +511,14 @@ struct AMDGPUDeviceImageTy : public DeviceImageTy {
findDeviceSymbol(GenericDeviceTy &Device, StringRef SymbolName) const;
/// Get additional info for kernel, e.g., register spill counts
std::optional<offloading::amdgpu::AMDGPUKernelMetaData>
Expected<offloading::amdgpu::AMDGPUKernelMetaData>
getKernelInfo(StringRef Identifier) const {
auto It = KernelInfoMap.find(Identifier);
if (It == KernelInfoMap.end())
return {};
return Plugin::error(ErrorCode::INVALID_BINARY,
"could not find metadata for kernel %s",
Identifier.str().c_str());
return It->second;
}
@ -593,10 +595,10 @@ struct AMDGPUKernelTy : public GenericKernelTy {
ODBG(OLDT_Module) << "ELFABIVersion: " << AMDImage.getELFABIVersion();
// Get additional kernel info read from image
KernelInfo = AMDImage.getKernelInfo(getName());
if (!KernelInfo.has_value())
INFO(OMP_INFOTYPE_PLUGIN_KERNEL, Device.getDeviceId(),
"Could not read extra information for kernel %s.", getName());
auto KernelInfoOrErr = AMDImage.getKernelInfo(getName());
if (!KernelInfoOrErr)
return KernelInfoOrErr.takeError();
KernelInfo = std::move(*KernelInfoOrErr);
return Plugin::success();
}
@ -649,7 +651,7 @@ private:
uint32_t ImplicitArgsSize;
/// Additional Info for the AMD GPU Kernel
std::optional<offloading::amdgpu::AMDGPUKernelMetaData> KernelInfo;
offloading::amdgpu::AMDGPUKernelMetaData KernelInfo;
};
/// Class representing an HSA signal. Signals are used to define dependencies
@ -4191,11 +4193,39 @@ Error AMDGPUKernelTy::launchImpl(GenericDeviceTy &GenericDevice,
if (auto Err = GenericDevice.getDeviceStackSize(StackSize))
return Err;
// Copy the explicit arguments.
// TODO: We should expose the args memory manager alloc to the common part as
// alternative to copying them twice.
if (LaunchParams.Size)
std::memcpy(AllArgs, LaunchParams.Data, LaunchParams.Size);
// Copy explicit arguments.
size_t ExplicitEnd = 0;
if (KernelArgs.Flags.IsPtrArgs) {
if (KernelArgs.ArgPtrs) {
const auto &ArgMDs = KernelInfo.ArgMDs;
// ArgMDs might also contain hidden implicit arguments, so we can't check
// if user-provided NumArgs matches exactly.
if (KernelArgs.NumArgs > ArgMDs.size())
return Plugin::error(
ErrorCode::INVALID_ARGUMENT,
"number of arguments (%u) exceeds the number of arguments "
"expected by the kernel (%zu)",
KernelArgs.NumArgs, ArgMDs.size());
for (size_t I = 0; I < KernelArgs.NumArgs; I++) {
auto [Offset, Size] = ArgMDs[I];
std::memcpy(utils::advancePtr(AllArgs, Offset), KernelArgs.ArgPtrs[I],
Size);
}
if (KernelArgs.NumArgs) {
auto [Offset, Size] = ArgMDs[KernelArgs.NumArgs - 1];
ExplicitEnd = Offset + Size;
}
}
} else {
// TODO: We should expose the args memory manager alloc to the common part
// as alternative to copying them twice.
if (LaunchParams.Size)
std::memcpy(AllArgs, LaunchParams.Data, LaunchParams.Size);
ExplicitEnd = LaunchParams.Size;
}
AMDGPUDeviceTy &AMDGPUDevice = static_cast<AMDGPUDeviceTy &>(GenericDevice);
@ -4203,8 +4233,8 @@ Error AMDGPUKernelTy::launchImpl(GenericDeviceTy &GenericDevice,
if (auto Err = AMDGPUDevice.getStream(AsyncInfoWrapper, Stream))
return Err;
uint64_t ImplArgsOffset = llvm::alignTo(
LaunchParams.Size, alignof(hsa_utils::AMDGPUImplicitArgsTy));
uint64_t ImplArgsOffset =
llvm::alignTo(ExplicitEnd, alignof(hsa_utils::AMDGPUImplicitArgsTy));
if (ArgsSize > ImplArgsOffset) {
hsa_utils::AMDGPUImplicitArgsTy *ImplArgs =
reinterpret_cast<hsa_utils::AMDGPUImplicitArgsTy *>(
@ -4255,10 +4285,6 @@ Error AMDGPUKernelTy::printLaunchInfoDetails(GenericDeviceTy &GenericDevice,
if (!(getInfoLevel() & OMP_INFOTYPE_PLUGIN_KERNEL))
return Plugin::success();
// We don't have data to print additional info, but no hard error
if (!KernelInfo.has_value())
return Plugin::success();
// General Info
auto *NumGroups = NumBlocks;
auto *ThreadsPerGroup = NumThreads;
@ -4269,12 +4295,12 @@ Error AMDGPUKernelTy::printLaunchInfoDetails(GenericDeviceTy &GenericDevice,
// Details for AMDGPU kernels (read from image)
// https://www.llvm.org/docs/AMDGPUUsage.html#code-object-v4-metadata
auto GroupSegmentSize = (*KernelInfo).GroupSegmentList;
auto SGPRCount = (*KernelInfo).SGPRCount;
auto VGPRCount = (*KernelInfo).VGPRCount;
auto SGPRSpillCount = (*KernelInfo).SGPRSpillCount;
auto VGPRSpillCount = (*KernelInfo).VGPRSpillCount;
auto MaxFlatWorkgroupSize = (*KernelInfo).MaxFlatWorkgroupSize;
auto GroupSegmentSize = KernelInfo.GroupSegmentList;
auto SGPRCount = KernelInfo.SGPRCount;
auto VGPRCount = KernelInfo.VGPRCount;
auto SGPRSpillCount = KernelInfo.SGPRSpillCount;
auto VGPRSpillCount = KernelInfo.VGPRSpillCount;
auto MaxFlatWorkgroupSize = KernelInfo.MaxFlatWorkgroupSize;
// Prints additional launch info that contains the following.
// Num Args: The number of kernel arguments

6
offload/plugins-nextgen/common/src/PluginInterface.cpp
View file

@ -279,11 +279,13 @@ Error GenericKernelTy::launch(GenericDeviceTy &GenericDevice, void **ArgPtrs,
KernelLaunchParamsTy LaunchParams;
// Kernel languages don't use indirection.
// Kernel languages (.IsCUDA) don't use indirection, whereas dispatching with
// an array of kernel argument pointers (.IsPtrArgs) uses KernelArgs.ArgPtrs
// and KernelArgs.ArgSizes directly.
if (KernelArgs.Flags.IsCUDA) {
LaunchParams =
*reinterpret_cast<KernelLaunchParamsTy *>(KernelArgs.ArgPtrs);
} else {
} else if (!KernelArgs.Flags.IsPtrArgs) {
LaunchParams =
prepareArgs(GenericDevice, ArgPtrs, ArgOffsets, KernelArgs.NumArgs,
Args, Ptrs, *KernelLaunchEnvOrErr, KernelArgs.Version);

19
offload/plugins-nextgen/cuda/src/rtl.cpp
View file

@ -1472,17 +1472,23 @@ Error CUDAKernelTy::launchImpl(GenericDeviceTy &GenericDevice,
AsyncInfoWrapperTy &AsyncInfoWrapper) const {
CUDADeviceTy &CUDADevice = static_cast<CUDADeviceTy &>(GenericDevice);
// The args size passed in LaunchParams may have tail padding, which is not
// accepted by the CUDA driver.
if (ArgsSize > LaunchParams.Size)
return Plugin::error(ErrorCode::INVALID_ARGUMENT,
"mismatch in kernel arguments");
void **KernelParams = nullptr;
if (KernelArgs.Flags.IsPtrArgs) {
KernelParams = KernelArgs.ArgPtrs;
} else {
// The args size passed in LaunchParams may have tail padding,
// which is not accepted by the CUDA driver.
if (ArgsSize > LaunchParams.Size)
return Plugin::error(ErrorCode::INVALID_ARGUMENT,
"mismatch in kernel arguments");
}
CUstream Stream;
if (auto Err = CUDADevice.getStream(AsyncInfoWrapper, Stream))
return Err;
size_t ConfigArgsSize = ArgsSize;
// Valid only for a contiguous buffer passed through LaunchParams.
void *Config[] = {CU_LAUNCH_PARAM_BUFFER_POINTER, LaunchParams.Data,
CU_LAUNCH_PARAM_BUFFER_SIZE,
reinterpret_cast<void *>(&ConfigArgsSize),
@ -1514,7 +1520,8 @@ Error CUDAKernelTy::launchImpl(GenericDeviceTy &GenericDevice,
DynBlockMemSize, Stream,
&CoopAttr, 1};
CUresult Res = cuLaunchKernelEx(&LaunchConfig, Func, nullptr, Config);
CUresult Res = cuLaunchKernelEx(&LaunchConfig, Func, KernelParams,
KernelParams ? nullptr : Config);
// Register a callback to indicate when the kernel is complete.
if (GenericDevice.getRPCServer())

1
offload/plugins-nextgen/level_zero/dynamic_l0/L0DynWrapper.cpp
View file

@ -29,6 +29,7 @@ DLWRAP(zeModuleDestroy, 1)
DLWRAP(zeCommandListAppendBarrier, 4)
DLWRAP(zeCommandListAppendLaunchKernel, 6)
DLWRAP(zeCommandListAppendLaunchCooperativeKernel, 6)
DLWRAP(zeCommandListAppendLaunchKernelWithArguments, 9)
DLWRAP(zeCommandListAppendMemoryCopy, 7)
DLWRAP(zeCommandListAppendMemoryCopyRegion, 12)
DLWRAP(zeCommandListAppendMemoryFill, 8)

19
offload/plugins-nextgen/level_zero/dynamic_l0/level_zero/ze_api.h
View file

@ -167,6 +167,7 @@ typedef enum _ze_structure_type_t {
ZE_STRUCTURE_TYPE_KERNEL_PROPERTIES = 0x1e,
ZE_STRUCTURE_TYPE_KERNEL_PREFERRED_GROUP_SIZE_PROPERTIES = 0x21,
ZE_STRUCTURE_TYPE_DEVICE_IP_VERSION_EXT = 0x1000f,
ZE_STRUCTURE_TYPE_COMMAND_LIST_APPEND_PARAM_COOPERATIVE_DESC = 0x00020036,
ZE_STRUCTURE_TYPE_RELAXED_ALLOCATION_LIMITS_EXP_DESC = 0x00020001,
ZE_STRUCTURE_TYPE_FORCE_UINT32 = 0x7fffffff
} ze_structure_type_t;
@ -489,6 +490,12 @@ typedef struct _ze_group_count_t {
uint32_t groupCountZ;
} ze_group_count_t;
typedef struct _ze_group_size_t {
uint32_t groupSizeX;
uint32_t groupSizeY;
uint32_t groupSizeZ;
} ze_group_size_t;
/* Memory allocation properties */
typedef struct _ze_memory_allocation_properties_t {
ze_structure_type_t stype;
@ -731,6 +738,18 @@ ZE_APIEXPORT ze_result_t ZE_APICALL zeCommandListAppendLaunchCooperativeKernel(
ze_command_list_handle_t hCommandList, ze_kernel_handle_t hKernel,
const ze_group_count_t *pLaunchFuncArgs, ze_event_handle_t hSignalEvent,
uint32_t numWaitEvents, ze_event_handle_t *phWaitEvents);
typedef struct _ze_command_list_append_launch_kernel_param_cooperative_desc_t {
ze_structure_type_t stype;
const void *pNext;
ze_bool_t isCooperative;
} ze_command_list_append_launch_kernel_param_cooperative_desc_t;
ZE_APIEXPORT ze_result_t ZE_APICALL
zeCommandListAppendLaunchKernelWithArguments(
ze_command_list_handle_t hCommandList, ze_kernel_handle_t hKernel,
const ze_group_count_t groupCounts, const ze_group_size_t groupSizes,
void **pArguments, const void *pNext, ze_event_handle_t hSignalEvent,
uint32_t numWaitEvents, ze_event_handle_t *phWaitEvents);
ZE_APIEXPORT ze_result_t ZE_APICALL zeCommandListAppendMemoryCopy(
ze_command_list_handle_t hCommandList, void *dstptr, const void *srcptr,
size_t size, ze_event_handle_t hSignalEvent, uint32_t numWaitEvents,

5
offload/plugins-nextgen/level_zero/include/L0Kernel.h
View file

@ -63,8 +63,9 @@ class L0KernelTy : public GenericKernelTy {
Error buildKernel(L0ProgramTy &Program);
Error readKernelProperties(L0ProgramTy &Program);
Error setKernelGroups(L0DeviceTy &l0Device, L0LaunchEnvTy &KEnv,
uint32_t NumThreads[3], uint32_t NumBlocks[3]) const;
ze_group_size_t createKernelGroups(L0DeviceTy &l0Device, L0LaunchEnvTy &KEnv,
uint32_t NumThreads[3],
uint32_t NumBlocks[3]) const;
Error setIndirectFlags(L0DeviceTy &l0Device, L0LaunchEnvTy &KEnv) const;
public:

129
offload/plugins-nextgen/level_zero/src/L0Kernel.cpp
View file

@ -15,8 +15,6 @@
#include "L0Plugin.h"
#include "L0Program.h"
#include "llvm/ADT/ScopeExit.h"
namespace llvm::omp::target::plugin {
Error L0KernelTy::readKernelProperties(L0ProgramTy &Program) {
@ -85,17 +83,21 @@ Error L0KernelTy::initImpl(GenericDeviceTy &GenericDevice,
return Plugin::success();
}
using AppendLaunchFnTy = llvm::function_ref<Error(
ze_command_list_handle_t CmdList, ze_event_handle_t Event,
uint32_t NumWaitEvents, ze_event_handle_t *WaitEvents)>;
static Error launchKernelWithImmCmdList(L0DeviceTy &l0Device,
ze_kernel_handle_t zeKernel,
L0LaunchEnvTy &KEnv,
CommandModeTy CommandMode) {
CommandModeTy CommandMode,
AppendLaunchFnTy AppendLaunch) {
const auto DeviceId = l0Device.getDeviceId();
auto *IdStr = l0Device.getZeIdCStr();
auto CmdListOrErr = l0Device.getImmCmdList();
if (!CmdListOrErr)
return CmdListOrErr.takeError();
const ze_command_list_handle_t CmdList = *CmdListOrErr;
// Command queue is not used with immediate command list.
INFO(OMP_INFOTYPE_PLUGIN_KERNEL, DeviceId,
"Using immediate command list for kernel submission.\n");
@ -120,17 +122,8 @@ static Error launchKernelWithImmCmdList(L0DeviceTy &l0Device,
Error AllErrors = Error::success();
if (KEnv.IsCooperative) {
INFO(OMP_INFOTYPE_PLUGIN_KERNEL, DeviceId,
"Launching cooperative kernel " DPxMOD "\n", DPxPTR(zeKernel));
CALL_ZE_ACCUM_ERROR(AllErrors, zeCommandListAppendLaunchCooperativeKernel,
CmdList, zeKernel, &KEnv.GroupCounts, Event,
NumWaitEvents, WaitEvents);
} else {
CALL_ZE_ACCUM_ERROR(AllErrors, zeCommandListAppendLaunchKernel, CmdList,
zeKernel, &KEnv.GroupCounts, Event, NumWaitEvents,
WaitEvents);
}
if (auto Err = AppendLaunch(CmdList, Event, NumWaitEvents, WaitEvents))
AllErrors = joinErrors(std::move(AllErrors), std::move(Err));
KEnv.Lock.unlock();
if (AllErrors) {
@ -159,35 +152,33 @@ static Error launchKernelWithImmCmdList(L0DeviceTy &l0Device,
return Plugin::success();
}
Error L0KernelTy::setKernelGroups(L0DeviceTy &l0Device, L0LaunchEnvTy &KEnv,
uint32_t NumThreads[3],
uint32_t NumBlocks[3]) const {
ze_group_size_t L0KernelTy::createKernelGroups(L0DeviceTy &l0Device,
L0LaunchEnvTy &KEnv,
uint32_t NumThreads[3],
uint32_t NumBlocks[3]) const {
ze_group_size_t GroupSizes;
assert(NumThreads[0] > 0 && NumThreads[1] > 0 && NumThreads[2] > 0 &&
"Pre-computed ThreadLimit values must be non-zero");
assert(NumBlocks[0] > 0 && NumBlocks[1] > 0 && NumBlocks[2] > 0 &&
"Pre-computed NumTeams values must be non-zero");
uint32_t GroupSizes[3];
KEnv.GroupCounts = {NumBlocks[0], NumBlocks[1], NumBlocks[2]};
// Respect max group size attribute in the kernel.
uint32_t MaxGroupSize = KEnv.KernelPR.MaxThreadGroupSize;
GroupSizes[0] = std::min<uint32_t>(MaxGroupSize, NumThreads[0]);
GroupSizes[1] = std::min<uint32_t>(MaxGroupSize, NumThreads[1]);
GroupSizes[2] = std::min<uint32_t>(MaxGroupSize, NumThreads[2]);
GroupSizes.groupSizeX = std::min<uint32_t>(MaxGroupSize, NumThreads[0]);
GroupSizes.groupSizeY = std::min<uint32_t>(MaxGroupSize, NumThreads[1]);
GroupSizes.groupSizeZ = std::min<uint32_t>(MaxGroupSize, NumThreads[2]);
auto DeviceId = l0Device.getDeviceId();
INFO(OMP_INFOTYPE_PLUGIN_KERNEL, DeviceId,
"Team sizes = {%" PRIu32 ", %" PRIu32 ", %" PRIu32 "}\n", GroupSizes[0],
GroupSizes[1], GroupSizes[2]);
"Team sizes = {%" PRIu32 ", %" PRIu32 ", %" PRIu32 "}\n",
GroupSizes.groupSizeX, GroupSizes.groupSizeY, GroupSizes.groupSizeZ);
INFO(OMP_INFOTYPE_PLUGIN_KERNEL, DeviceId,
"Number of teams = {%" PRIu32 ", %" PRIu32 ", %" PRIu32 "}\n",
KEnv.GroupCounts.groupCountX, KEnv.GroupCounts.groupCountY,
KEnv.GroupCounts.groupCountZ);
CALL_ZE_RET_ERROR(zeKernelSetGroupSize, getZeKernel(), GroupSizes[0],
GroupSizes[1], GroupSizes[2]);
return Plugin::success();
return GroupSizes;
}
Error L0KernelTy::setIndirectFlags(L0DeviceTy &l0Device,
@ -251,8 +242,8 @@ Error L0KernelTy::launchImpl(GenericDeviceTy &GenericDevice,
// Protect from kernel preparation to submission as kernels are shared.
KEnv.Lock.lock();
if (auto Err = setKernelGroups(l0Device, KEnv, NumThreads, NumBlocks))
return Err;
ze_group_size_t GroupSizes =
createKernelGroups(l0Device, KEnv, NumThreads, NumBlocks);
// Validate cooperative kernel launch constraints
if (IsCooperative) {
@ -278,34 +269,74 @@ Error L0KernelTy::launchImpl(GenericDeviceTy &GenericDevice,
TotalGroupCount, MaxCooperativeGroupCount);
}
// Set kernel arguments.
uint32_t NumKernelArgs = KernelPR.NumKernelArgs;
if (NumKernelArgs > 0) {
if (!KernelPR.ArgSizes)
return Plugin::error(ErrorCode::INVALID_ARGUMENT,
"level zero plugin requires kernel argument sizes.");
// Use sizes from kernel properties.
// TODO: This is temporary workaround it will not work if there is
// padding/alignment between arguments.
char *Arg = static_cast<char *>(LaunchParams.Data);
for (uint32_t I = 0; I < NumKernelArgs; I++) {
uint32_t ArgSize = KernelPR.ArgSizes[I];
CALL_ZE_RET_ERROR(zeKernelSetArgumentValue, zeKernel, I, ArgSize, Arg);
// With pointer-array arguments, zeCommandListAppendLaunchKernelWithArguments
// folds group-size, per-argument set, and launch into a single call.
const bool IsPtrArgs = KernelArgs.Flags.IsPtrArgs;
if (IsPtrArgs) {
if (KernelArgs.NumArgs != KernelPR.NumKernelArgs)
return Plugin::error(
ErrorCode::INVALID_ARGUMENT,
"Number of arguments (%u) does not match the number of arguments "
"expected by the kernel (%u)",
KernelArgs.NumArgs, KernelPR.NumKernelArgs);
} else {
CALL_ZE_RET_ERROR(zeKernelSetGroupSize, zeKernel, GroupSizes.groupSizeX,
GroupSizes.groupSizeY, GroupSizes.groupSizeZ);
INFO(OMP_INFOTYPE_PLUGIN_KERNEL, DeviceId,
"Kernel Pointer argument %" PRIu32 " (value: " DPxMOD
") was set successfully for device %s.\n",
I, DPxPTR(Arg), IdStr);
Arg += ArgSize;
// Set kernel arguments.
uint32_t NumKernelArgs = KernelPR.NumKernelArgs;
if (NumKernelArgs > 0) {
if (!KernelPR.ArgSizes)
return Plugin::error(
ErrorCode::INVALID_ARGUMENT,
"level zero plugin requires kernel argument sizes.");
// Use sizes from kernel properties.
// TODO: This is temporary workaround it will not work if there is
// padding/alignment between arguments.
char *Arg = static_cast<char *>(LaunchParams.Data);
for (uint32_t I = 0; I < NumKernelArgs; I++) {
uint32_t ArgSize = KernelPR.ArgSizes[I];
CALL_ZE_RET_ERROR(zeKernelSetArgumentValue, zeKernel, I, ArgSize, Arg);
INFO(OMP_INFOTYPE_PLUGIN_KERNEL, DeviceId,
"Kernel Pointer argument %" PRIu32 " (value: " DPxMOD
") was set successfully for device %s.\n",
I, DPxPTR(Arg), IdStr);
Arg += ArgSize;
}
}
}
if (auto Err = setIndirectFlags(l0Device, KEnv))
return Err;
auto AppendLaunch = [&](ze_command_list_handle_t CmdList,
ze_event_handle_t Event, uint32_t NumWaitEvents,
ze_event_handle_t *WaitEvents) -> Error {
if (IsPtrArgs) {
ze_command_list_append_launch_kernel_param_cooperative_desc_t CoopDesc = {
ZE_STRUCTURE_TYPE_COMMAND_LIST_APPEND_PARAM_COOPERATIVE_DESC, nullptr,
static_cast<ze_bool_t>(IsCooperative)};
CALL_ZE_RET_ERROR(zeCommandListAppendLaunchKernelWithArguments, CmdList,
zeKernel, KEnv.GroupCounts, GroupSizes,
KernelArgs.ArgPtrs, IsCooperative ? &CoopDesc : nullptr,
Event, NumWaitEvents, WaitEvents);
} else if (IsCooperative) {
INFO(OMP_INFOTYPE_PLUGIN_KERNEL, DeviceId,
"Launching cooperative kernel " DPxMOD "\n", DPxPTR(zeKernel));
CALL_ZE_RET_ERROR(zeCommandListAppendLaunchCooperativeKernel, CmdList,
zeKernel, &KEnv.GroupCounts, Event, NumWaitEvents,
WaitEvents);
} else {
CALL_ZE_RET_ERROR(zeCommandListAppendLaunchKernel, CmdList, zeKernel,
&KEnv.GroupCounts, Event, NumWaitEvents, WaitEvents);
}
return Plugin::success();
};
// The next call should unlock the KernelLock internally.
return launchKernelWithImmCmdList(l0Device, zeKernel, KEnv,
Options.CommandMode);
Options.CommandMode, AppendLaunch);
}
Expected<uint32_t>

20
offload/unittests/Conformance/include/mathtest/DeviceContext.hpp
View file

@ -103,16 +103,22 @@ public:
"device");
if constexpr (sizeof...(Args) == 0) {
launchKernelImpl(Kernel.Handle, NumGroups, GroupSize, nullptr, 0);
launchKernelImpl(Kernel.Handle, NumGroups, GroupSize, 0, nullptr,
nullptr);
} else {
auto KernelArgs = makeKernelArgsPack(std::forward<ArgTypes>(Args)...);
static_assert(
(std::is_trivially_copyable_v<std::decay_t<ArgTypes>> && ...),
"Argument types provided to launchKernel must be trivially copyable");
launchKernelImpl(Kernel.Handle, NumGroups, GroupSize, &KernelArgs,
sizeof(KernelArgs));
ProvidedTypes ArgsTuple(std::forward<ArgTypes>(Args)...);
std::apply(
[&](auto &...A) {
void *ArgPtrs[] = {static_cast<void *>(&A)...};
size_t ArgSizes[] = {sizeof(A)...};
launchKernelImpl(Kernel.Handle, NumGroups, GroupSize,
sizeof...(Args), ArgPtrs, ArgSizes);
},
ArgsTuple);
}
}
@ -126,8 +132,8 @@ private:
llvm::StringRef KernelName) const noexcept;
void launchKernelImpl(ol_symbol_handle_t KernelHandle, uint32_t NumGroups,
uint32_t GroupSize, const void *KernelArgs,
std::size_t KernelArgsSize) const noexcept;
uint32_t GroupSize, size_t NumArgs, void **ArgPtrs,
const size_t *ArgSizes) const noexcept;
std::size_t GlobalDeviceId;
ol_device_handle_t DeviceHandle;

27
offload/unittests/Conformance/include/mathtest/Support.hpp
View file

@ -87,33 +87,6 @@ using KernelSignatureOf = detail::KernelSignatureOfImpl<
template <auto Func>
using KernelSignatureOf_t = typename KernelSignatureOf<Func>::type;
//===----------------------------------------------------------------------===//
// Kernel Argument Packing
//===----------------------------------------------------------------------===//
template <typename... ArgTypes> struct KernelArgsPack;
template <typename ArgType> struct KernelArgsPack<ArgType> {
std::decay_t<ArgType> Arg;
constexpr KernelArgsPack(ArgType &&Arg) : Arg(std::forward<ArgType>(Arg)) {}
};
template <typename ArgType0, typename ArgType1, typename... ArgTypes>
struct KernelArgsPack<ArgType0, ArgType1, ArgTypes...> {
std::decay_t<ArgType0> Arg0;
KernelArgsPack<ArgType1, ArgTypes...> Args;
constexpr KernelArgsPack(ArgType0 &&Arg0, ArgType1 &&Arg1, ArgTypes &&...Args)
: Arg0(std::forward<ArgType0>(Arg0)),
Args(std::forward<ArgType1>(Arg1), std::forward<ArgTypes>(Args)...) {}
};
template <typename... ArgTypes>
KernelArgsPack<ArgTypes...> makeKernelArgsPack(ArgTypes &&...Args) {
return KernelArgsPack<ArgTypes...>(std::forward<ArgTypes>(Args)...);
}
//===----------------------------------------------------------------------===//
// Configuration Helpers
//===----------------------------------------------------------------------===//

11
offload/unittests/Conformance/lib/DeviceContext.cpp
View file

@ -286,17 +286,18 @@ DeviceContext::getKernelHandle(ol_program_handle_t ProgramHandle,
return Handle;
}
void DeviceContext::launchKernelImpl(
ol_symbol_handle_t KernelHandle, uint32_t NumGroups, uint32_t GroupSize,
const void *KernelArgs, std::size_t KernelArgsSize) const noexcept {
void DeviceContext::launchKernelImpl(ol_symbol_handle_t KernelHandle,
uint32_t NumGroups, uint32_t GroupSize,
size_t NumArgs, void **ArgPtrs,
const size_t *ArgSizes) const noexcept {
ol_kernel_launch_size_args_t LaunchSizeArgs;
LaunchSizeArgs.Dimensions = 1;
LaunchSizeArgs.NumGroups = {NumGroups, 1, 1};
LaunchSizeArgs.GroupSize = {GroupSize, 1, 1};
LaunchSizeArgs.DynSharedMemory = 0;
OL_CHECK(olLaunchKernel(nullptr, DeviceHandle, KernelHandle, KernelArgs,
KernelArgsSize, &LaunchSizeArgs, nullptr));
OL_CHECK(olLaunchKernel(nullptr, DeviceHandle, KernelHandle, &LaunchSizeArgs,
nullptr, NumArgs, ArgPtrs, ArgSizes));
}
[[nodiscard]] llvm::StringRef DeviceContext::getName() const noexcept {

2
offload/unittests/OffloadAPI/device_code/CMakeLists.txt
View file

@ -3,6 +3,7 @@ add_offload_test_device_code(bar.cpp bar)
# Compile with optimizations to eliminate AMDGPU implicit arguments.
add_offload_test_device_code(noargs.cpp noargs -O3)
add_offload_test_device_code(multiargs.cpp multiargs -O3)
add_offload_test_device_code(composite.cpp composite -O3)
add_offload_test_device_code(byte.cpp byte)
add_offload_test_device_code(localmem.cpp localmem)
add_offload_test_device_code(localmem_reduction.cpp localmem_reduction)
@ -18,6 +19,7 @@ add_custom_target(offload_device_binaries DEPENDS
bar.bin
noargs.bin
multiargs.bin
composite.bin
byte.bin
localmem.bin
localmem_reduction.bin

10
offload/unittests/OffloadAPI/device_code/composite.cpp Normal file
View file

@ -0,0 +1,10 @@
#include <gpuintrin.h>
struct Foo {
uint32_t a;
uint32_t b;
};
extern "C" __gpu_kernel void composite(uint8_t N, Foo F, uint32_t *Out) {
Out[__gpu_thread_id(0)] = N + F.a + F.b + __gpu_thread_id(0);
}

4
offload/unittests/OffloadAPI/device_code/multiargs.cpp
View file

@ -1,3 +1,5 @@
#include <gpuintrin.h>
extern "C" __gpu_kernel void multiargs(char, int *, short) { (void)0; }
extern "C" __gpu_kernel void multiargs(char A, int *B, short C) {
B[__gpu_thread_id(0)] = A + C + __gpu_thread_id(0);
}

9
offload/unittests/OffloadAPI/event/olGetEventElapsedTime.cpp
View file

@ -41,12 +41,11 @@ struct olGetEventElapsedTimeTest : OffloadQueueTest {
}
void launchFoo() {
struct {
void *Mem;
} Args{Mem};
void *ArgPtrs[] = {&Mem};
size_t ArgSizes[] = {sizeof(Mem)};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &Args, sizeof(Args),
&LaunchArgs, nullptr));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr,
1, ArgPtrs, ArgSizes));
}
std::unique_ptr<llvm::MemoryBuffer> DeviceBin;

185
offload/unittests/OffloadAPI/kernel/olLaunchKernel.cpp
View file

@ -19,6 +19,7 @@
KERNEL_TEST(Foo, foo)
KERNEL_TEST(NoArgs, noargs)
KERNEL_TEST(MultiArgs, multiargs)
KERNEL_TEST(Composite, composite)
KERNEL_TEST(Byte, byte)
KERNEL_TEST(LocalMem, localmem)
KERNEL_TEST(LocalMemReduction, localmem_reduction)
@ -60,12 +61,12 @@ TEST_P(olLaunchKernelFooTest, Success) {
void *Mem;
ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
LaunchArgs.GroupSize.x * sizeof(uint32_t), &Mem));
struct {
void *Mem;
} Args{Mem};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &Args, sizeof(Args),
&LaunchArgs, nullptr));
void *ArgPtrs[] = {&Mem};
size_t ArgSizes[] = {sizeof(Mem)};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr,
std::size(ArgPtrs), ArgPtrs, ArgSizes));
ASSERT_SUCCESS(olSyncQueue(Queue));
@ -86,12 +87,11 @@ TEST_P(olLaunchKernelFooTest, SuccessThreaded) {
ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_DEVICE, Size, &DevAlloc));
ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_HOST, Size, &HstAlloc));
struct {
void *Mem;
} Args{DevAlloc};
void *ArgPtrs[] = {&DevAlloc};
size_t ArgSizes[] = {sizeof(DevAlloc)};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &Args, sizeof(Args),
&LaunchArgs, nullptr));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr,
std::size(ArgPtrs), ArgPtrs, ArgSizes));
ASSERT_SUCCESS(olMemcpy(Queue, HstAlloc, Host, DevAlloc, Device, Size));
@ -108,22 +108,63 @@ TEST_P(olLaunchKernelFooTest, SuccessThreaded) {
}
TEST_P(olLaunchKernelNoArgsTest, Success) {
ASSERT_SUCCESS(
olLaunchKernel(Queue, Device, Kernel, nullptr, 0, &LaunchArgs, nullptr));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr, 0,
nullptr, nullptr));
ASSERT_SUCCESS(olSyncQueue(Queue));
}
TEST_P(olLaunchKernelMultiArgsTest, Success) {
struct {
char A;
int *B;
short C;
} Args{0, nullptr, 0};
void *Mem;
ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
LaunchArgs.GroupSize.x * sizeof(int), &Mem));
char A = 3;
int *B = (int *)Mem;
short C = 5;
void *ArgPtrs[] = {&A, &B, &C};
size_t ArgSizes[] = {sizeof(A), sizeof(B), sizeof(C)};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr,
std::size(ArgPtrs), ArgPtrs, ArgSizes));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &Args, sizeof(Args),
&LaunchArgs, nullptr));
ASSERT_SUCCESS(olSyncQueue(Queue));
int *Data = (int *)Mem;
for (uint32_t i = 0; i < LaunchArgs.GroupSize.x; i++)
ASSERT_EQ(Data[i], A + C + static_cast<int>(i));
ASSERT_SUCCESS(olMemFree(Mem));
}
struct Foo {
uint32_t a;
uint32_t b;
};
TEST_P(olLaunchKernelCompositeTest, Success) {
void *Mem;
ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
LaunchArgs.GroupSize.x * sizeof(uint32_t), &Mem));
uint8_t N = 1;
Foo F{2, 3};
uint32_t *Out = (uint32_t *)Mem;
void *ArgPtrs[] = {&N, &F, &Out};
size_t ArgSizes[] = {sizeof(N), sizeof(F), sizeof(Out)};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr,
std::size(ArgPtrs), ArgPtrs, ArgSizes));
ASSERT_SUCCESS(olSyncQueue(Queue));
uint32_t *Data = (uint32_t *)Mem;
for (uint32_t i = 0; i < LaunchArgs.GroupSize.x; i++)
ASSERT_EQ(Data[i], N + F.a + F.b + i);
ASSERT_SUCCESS(olMemFree(Mem));
}
TEST_P(olLaunchKernelFooTest, SuccessSynchronous) {
@ -131,12 +172,11 @@ TEST_P(olLaunchKernelFooTest, SuccessSynchronous) {
ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
LaunchArgs.GroupSize.x * sizeof(uint32_t), &Mem));
struct {
void *Mem;
} Args{Mem};
void *ArgPtrs[] = {&Mem};
size_t ArgSizes[] = {sizeof(Mem)};
ASSERT_SUCCESS(olLaunchKernel(nullptr, Device, Kernel, &Args, sizeof(Args),
&LaunchArgs, nullptr));
ASSERT_SUCCESS(olLaunchKernel(nullptr, Device, Kernel, &LaunchArgs, nullptr,
std::size(ArgPtrs), ArgPtrs, ArgSizes));
uint32_t *Data = (uint32_t *)Mem;
for (uint32_t i = 0; i < 64; i++) {
@ -146,6 +186,18 @@ TEST_P(olLaunchKernelFooTest, SuccessSynchronous) {
ASSERT_SUCCESS(olMemFree(Mem));
}
TEST_P(olLaunchKernelByteTest, Success) {
unsigned char C = 42;
void *ArgPtrs[] = {&C};
size_t ArgSizes[] = {sizeof(C)};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr,
std::size(ArgPtrs), ArgPtrs, ArgSizes));
ASSERT_SUCCESS(olSyncQueue(Queue));
}
TEST_P(olLaunchKernelLocalMemTest, Success) {
SKIP_KNOWN_FAILURE(LevelZero{"unsupported DynSharedMemory"});
@ -157,12 +209,12 @@ TEST_P(olLaunchKernelLocalMemTest, Success) {
LaunchArgs.GroupSize.x * LaunchArgs.NumGroups.x *
sizeof(uint32_t),
&Mem));
struct {
void *Mem;
} Args{Mem};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &Args, sizeof(Args),
&LaunchArgs, nullptr));
void *ArgPtrs[] = {&Mem};
size_t ArgSizes[] = {sizeof(Mem)};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr,
std::size(ArgPtrs), ArgPtrs, ArgSizes));
ASSERT_SUCCESS(olSyncQueue(Queue));
@ -182,12 +234,12 @@ TEST_P(olLaunchKernelLocalMemReductionTest, Success) {
void *Mem;
ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
LaunchArgs.NumGroups.x * sizeof(uint32_t), &Mem));
struct {
void *Mem;
} Args{Mem};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &Args, sizeof(Args),
&LaunchArgs, nullptr));
void *ArgPtrs[] = {&Mem};
size_t ArgSizes[] = {sizeof(Mem)};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr,
std::size(ArgPtrs), ArgPtrs, ArgSizes));
ASSERT_SUCCESS(olSyncQueue(Queue));
@ -207,12 +259,12 @@ TEST_P(olLaunchKernelLocalMemStaticTest, Success) {
void *Mem;
ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
LaunchArgs.NumGroups.x * sizeof(uint32_t), &Mem));
struct {
void *Mem;
} Args{Mem};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &Args, sizeof(Args),
&LaunchArgs, nullptr));
void *ArgPtrs[] = {&Mem};
size_t ArgSizes[] = {sizeof(Mem)};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr,
std::size(ArgPtrs), ArgPtrs, ArgSizes));
ASSERT_SUCCESS(olSyncQueue(Queue));
@ -254,8 +306,12 @@ TEST_P(olLaunchKernelSingleCounterSyncEventTest, SuccessSyncEvent) {
ArgsSingleCounter Args{LoopRange, NumberToAdd, (uint32_t *)InitValuePassed,
(uint32_t *)ResNum};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &Args, sizeof(Args),
&LaunchArgs, nullptr));
void *ArgPtrs[] = {&Args.InitLoop, &Args.Addend, &Args.InitVal, &Args.Out};
size_t ArgSizes[] = {sizeof(Args.InitLoop), sizeof(Args.Addend),
sizeof(Args.InitVal), sizeof(Args.Out)};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr,
std::size(ArgPtrs), ArgPtrs, ArgSizes));
uint32_t FinalResVal = 0;
ASSERT_SUCCESS(olMemcpy(Queue, &FinalResVal, Host, ResNum, Device, Size));
@ -307,21 +363,26 @@ TEST_P(olLaunchKernelSingleCounterSyncEventTest, SuccessTwoQueues) {
ArgsSingleCounter Args{LoopRange, NumberToAdd, (uint32_t *)InitValuePassed,
(uint32_t *)ResNum1};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &Args, sizeof(Args),
&LaunchArgs, nullptr));
void *ArgPtrs[] = {&Args.InitLoop, &Args.Addend, &Args.InitVal, &Args.Out};
size_t ArgSizes[] = {sizeof(Args.InitLoop), sizeof(Args.Addend),
sizeof(Args.InitVal), sizeof(Args.Out)};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr,
std::size(ArgPtrs), ArgPtrs, ArgSizes));
ol_event_handle_t Event = nullptr;
ASSERT_SUCCESS(olCreateEvent(Queue, OL_EVENT_FLAGS_NONE, &Event));
ASSERT_SUCCESS(olWaitEvents(Queue2, &Event, 1));
ArgsSingleCounter Args2{LoopRange, NumberToAdd, (uint32_t *)ResNum1,
(uint32_t *)ResNum2};
void *ArgPtrs2[] = {&Args2.InitLoop, &Args2.Addend, &Args2.InitVal,
&Args2.Out};
// At the beginning of the kernel, ResNum from the first queue is saved
// locally as the initial value. If operations enqueued before Event have not
// been completed by the time the kernel is executed using the second queue,
// the FinalResVal would be incorrect.
ASSERT_SUCCESS(olLaunchKernel(Queue2, Device, Kernel, &Args2, sizeof(Args2),
&LaunchArgs, nullptr));
ASSERT_SUCCESS(olLaunchKernel(Queue2, Device, Kernel, &LaunchArgs, nullptr,
std::size(ArgPtrs2), ArgPtrs2, ArgSizes));
ASSERT_SUCCESS(olSyncQueue(Queue2));
uint32_t FinalResVal = 0;
@ -339,15 +400,15 @@ TEST_P(olLaunchKernelGlobalTest, Success) {
void *Mem;
ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
LaunchArgs.GroupSize.x * sizeof(uint32_t), &Mem));
struct {
void *Mem;
} Args{Mem};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernels[0], nullptr, 0,
&LaunchArgs, nullptr));
void *ArgPtrs[] = {&Mem};
size_t ArgSizes[] = {sizeof(Mem)};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernels[0], &LaunchArgs, nullptr,
0, nullptr, nullptr));
ASSERT_SUCCESS(olSyncQueue(Queue));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernels[1], &Args, sizeof(Args),
&LaunchArgs, nullptr));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernels[1], &LaunchArgs, nullptr,
std::size(ArgPtrs), ArgPtrs, ArgSizes));
ASSERT_SUCCESS(olSyncQueue(Queue));
uint32_t *Data = (uint32_t *)Mem;
@ -362,9 +423,9 @@ TEST_P(olLaunchKernelGlobalTest, InvalidNotAKernel) {
ol_symbol_handle_t Global = nullptr;
ASSERT_SUCCESS(
olGetSymbol(Program, "global", OL_SYMBOL_KIND_GLOBAL_VARIABLE, &Global));
ASSERT_ERROR(
OL_ERRC_SYMBOL_KIND,
olLaunchKernel(Queue, Device, Global, nullptr, 0, &LaunchArgs, nullptr));
ASSERT_ERROR(OL_ERRC_SYMBOL_KIND,
olLaunchKernel(Queue, Device, Global, &LaunchArgs, nullptr, 0,
nullptr, nullptr));
}
TEST_P(olLaunchKernelGlobalCtorTest, Success) {
@ -373,12 +434,12 @@ TEST_P(olLaunchKernelGlobalCtorTest, Success) {
void *Mem;
ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
LaunchArgs.GroupSize.x * sizeof(uint32_t), &Mem));
struct {
void *Mem;
} Args{Mem};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &Args, sizeof(Args),
&LaunchArgs, nullptr));
void *ArgPtrs[] = {&Mem};
size_t ArgSizes[] = {sizeof(Mem)};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr,
std::size(ArgPtrs), ArgPtrs, ArgSizes));
ASSERT_SUCCESS(olSyncQueue(Queue));
uint32_t *Data = (uint32_t *)Mem;
@ -393,7 +454,7 @@ TEST_P(olLaunchKernelGlobalDtorTest, Success) {
// TODO: We can't inspect the result of a destructor yet, once we
// find/implement a way, update this test. For now we just check that nothing
// crashes
ASSERT_SUCCESS(
olLaunchKernel(Queue, Device, Kernel, nullptr, 0, &LaunchArgs, nullptr));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr, 0,
nullptr, nullptr));
ASSERT_SUCCESS(olSyncQueue(Queue));
}

16
offload/unittests/OffloadAPI/kernel/olLaunchKernelCooperative.cpp
View file

@ -56,14 +56,14 @@ TEST_P(olLaunchKernelCooperativeTest, SuccessCooperative) {
{OL_KERNEL_LAUNCH_PROP_TYPE_IS_COOPERATIVE, &IsCooperative},
OL_KERNEL_LAUNCH_PROP_END};
ASSERT_SUCCESS(
olLaunchKernel(Queue, Device, Kernel, nullptr, 0, &LaunchArgs, Props));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, Props, 0,
nullptr, nullptr));
ASSERT_SUCCESS(olSyncQueue(Queue));
}
TEST_P(olLaunchKernelCooperativeTest, SuccessNonCooperative) {
ASSERT_SUCCESS(
olLaunchKernel(Queue, Device, Kernel, nullptr, 0, &LaunchArgs, nullptr));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr, 0,
nullptr, nullptr));
ASSERT_SUCCESS(olSyncQueue(Queue));
}
@ -79,8 +79,8 @@ TEST_P(olLaunchKernelCooperativeTest, TooManyGroups) {
{OL_KERNEL_LAUNCH_PROP_TYPE_IS_COOPERATIVE, &IsCooperative},
OL_KERNEL_LAUNCH_PROP_END};
ASSERT_ANY_ERROR(
olLaunchKernel(Queue, Device, Kernel, nullptr, 0, &LaunchArgs, Props));
ASSERT_ANY_ERROR(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, Props, 0,
nullptr, nullptr));
}
TEST_P(olLaunchKernelCooperativeTest, SynchronousLaunch) {
@ -95,8 +95,8 @@ TEST_P(olLaunchKernelCooperativeTest, SynchronousLaunch) {
{OL_KERNEL_LAUNCH_PROP_TYPE_IS_COOPERATIVE, &IsCooperative},
OL_KERNEL_LAUNCH_PROP_END};
ASSERT_SUCCESS(
olLaunchKernel(nullptr, Device, Kernel, nullptr, 0, &LaunchArgs, Props));
ASSERT_SUCCESS(olLaunchKernel(nullptr, Device, Kernel, &LaunchArgs, Props, 0,
nullptr, nullptr));
}
TEST_P(olLaunchKernelCooperativeTest, InvalidNullHandleKernel) {

13
offload/unittests/OffloadAPI/memory/olMemcpy.cpp
View file

@ -163,15 +163,14 @@ TEST_P(olMemcpyGlobalTest, SuccessWrite) {
ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
LaunchArgs.GroupSize.x * sizeof(uint32_t),
&DestMem));
struct {
void *Mem;
} Args{DestMem};
void *ArgPtrs[] = {&DestMem};
size_t ArgSizes[] = {sizeof(DestMem)};
ASSERT_SUCCESS(
olMemcpy(Queue, Addr, Device, SourceMem, Host, 64 * sizeof(uint32_t)));
ASSERT_SUCCESS(olSyncQueue(Queue));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, ReadKernel, &Args, sizeof(Args),
&LaunchArgs, nullptr));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, ReadKernel, &LaunchArgs, nullptr,
1, ArgPtrs, ArgSizes));
ASSERT_SUCCESS(olSyncQueue(Queue));
uint32_t *DestData = (uint32_t *)DestMem;
@ -188,8 +187,8 @@ TEST_P(olMemcpyGlobalTest, SuccessRead) {
LaunchArgs.GroupSize.x * sizeof(uint32_t),
&DestMem));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, WriteKernel, nullptr, 0,
&LaunchArgs, nullptr));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, WriteKernel, &LaunchArgs,
nullptr, 0, nullptr, nullptr));
ASSERT_SUCCESS(olSyncQueue(Queue));
ASSERT_SUCCESS(
olMemcpy(Queue, DestMem, Host, Addr, Device, 64 * sizeof(uint32_t)));

9
offload/unittests/OffloadAPI/queue/olLaunchHostFunction.cpp
View file

@ -84,11 +84,10 @@ TEST_P(olLaunchHostFunctionKernelTest, SuccessBlocking) {
},
const_cast<bool *>(&Block)));
struct {
void *Mem;
} Args{Mem};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &Args, sizeof(Args),
&LaunchArgs, nullptr));
void *ArgPtrs[] = {&Mem};
size_t ArgSizes[] = {sizeof(Mem)};
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr, 1,
ArgPtrs, ArgSizes));
std::this_thread::sleep_for(std::chrono::milliseconds(500));
for (uint32_t i = 0; i < 64; i++) {

39
offload/unittests/OffloadAPI/queue/olWaitEvents.cpp
View file

@ -38,21 +38,20 @@ TEST_P(olWaitEventsTest, Success) {
void *Mem;
ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
NUM_KERNELS * sizeof(uint32_t), &Mem));
struct {
uint32_t Idx;
void *Mem;
} Args{0, Mem};
uint32_t Idx = 0;
void *ArgPtrs[] = {&Idx, &Mem};
size_t ArgSizes[] = {sizeof(Idx), sizeof(Mem)};
for (size_t I = 0; I < NUM_KERNELS; I++) {
Args.Idx = I;
Idx = I;
ASSERT_SUCCESS(olCreateQueue(Device, &Queues[I]));
if (I > 0)
ASSERT_SUCCESS(olWaitEvents(Queues[I], &Events[I - 1], 1));
ASSERT_SUCCESS(olLaunchKernel(Queues[I], Device, Kernel, &Args,
sizeof(Args), &LaunchArgs, nullptr));
ASSERT_SUCCESS(olLaunchKernel(Queues[I], Device, Kernel, &LaunchArgs,
nullptr, 2, ArgPtrs, ArgSizes));
ASSERT_SUCCESS(olCreateEvent(Queues[I], OL_EVENT_FLAGS_NONE, &Events[I]));
}
@ -74,19 +73,18 @@ TEST_P(olWaitEventsTest, SuccessSingleQueue) {
void *Mem;
ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
NUM_KERNELS * sizeof(uint32_t), &Mem));
struct {
uint32_t Idx;
void *Mem;
} Args{0, Mem};
uint32_t Idx = 0;
void *ArgPtrs[] = {&Idx, &Mem};
size_t ArgSizes[] = {sizeof(Idx), sizeof(Mem)};
for (size_t I = 0; I < NUM_KERNELS; I++) {
Args.Idx = I;
Idx = I;
if (I > 0)
ASSERT_SUCCESS(olWaitEvents(Queue, &Events[I - 1], 1));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &Args, sizeof(Args),
&LaunchArgs, nullptr));
ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &LaunchArgs, nullptr,
2, ArgPtrs, ArgSizes));
ASSERT_SUCCESS(olCreateEvent(Queue, OL_EVENT_FLAGS_NONE, &Events[I]));
}
@ -106,21 +104,20 @@ TEST_P(olWaitEventsTest, SuccessMultipleEvents) {
void *Mem;
ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
NUM_KERNELS * sizeof(uint32_t), &Mem));
struct {
uint32_t Idx;
void *Mem;
} Args{0, Mem};
uint32_t Idx = 0;
void *ArgPtrs[] = {&Idx, &Mem};
size_t ArgSizes[] = {sizeof(Idx), sizeof(Mem)};
for (size_t I = 0; I < NUM_KERNELS; I++) {
Args.Idx = I;
Idx = I;
ASSERT_SUCCESS(olCreateQueue(Device, &Queues[I]));
if (I > 0)
ASSERT_SUCCESS(olWaitEvents(Queues[I], Events, I));
ASSERT_SUCCESS(olLaunchKernel(Queues[I], Device, Kernel, &Args,
sizeof(Args), &LaunchArgs, nullptr));
ASSERT_SUCCESS(olLaunchKernel(Queues[I], Device, Kernel, &LaunchArgs,
nullptr, 2, ArgPtrs, ArgSizes));
ASSERT_SUCCESS(olCreateEvent(Queues[I], OL_EVENT_FLAGS_NONE, &Events[I]));
}