compute-runtime/runtime/command_queue/gpgpu_walker.h

372 lines
16 KiB
C++

/*
* Copyright (c) 2018, Intel Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included
* in all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
* OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
* OTHER DEALINGS IN THE SOFTWARE.
*/
#pragma once
#include "runtime/built_ins/built_ins.h"
#include "runtime/context/context.h"
#include "runtime/command_queue/command_queue.h"
#include "runtime/command_stream/linear_stream.h"
#include "runtime/command_stream/preemption.h"
#include "runtime/device_queue/device_queue_hw.h"
#include "runtime/event/hw_timestamps.h"
#include "runtime/event/perf_counter.h"
#include "runtime/helpers/dispatch_info.h"
#include "runtime/helpers/kernel_commands.h"
#include "runtime/helpers/task_information.h"
#include "runtime/indirect_heap/indirect_heap.h"
#include "runtime/kernel/kernel.h"
#include "runtime/program/kernel_info.h"
#include "runtime/utilities/vec.h"
namespace OCLRT {
constexpr int32_t NUM_ALU_INST_FOR_READ_MODIFY_WRITE = 4;
constexpr int32_t L3SQC_BIT_LQSC_RO_PERF_DIS = 0x08000000;
constexpr int32_t L3SQC_REG4 = 0xB118;
constexpr int32_t GPGPU_WALKER_COOKIE_VALUE_BEFORE_WALKER = 0xFFFFFFFF;
constexpr int32_t GPGPU_WALKER_COOKIE_VALUE_AFTER_WALKER = 0x00000000;
constexpr int32_t CS_GPR_R0 = 0x2600;
constexpr int32_t CS_GPR_R1 = 0x2608;
constexpr int32_t ALU_OPCODE_LOAD = 0x080;
constexpr int32_t ALU_OPCODE_STORE = 0x180;
constexpr int32_t ALU_OPCODE_OR = 0x103;
constexpr int32_t ALU_OPCODE_AND = 0x102;
constexpr int32_t ALU_REGISTER_R_0 = 0x0;
constexpr int32_t ALU_REGISTER_R_1 = 0x1;
constexpr int32_t ALU_REGISTER_R_SRCA = 0x20;
constexpr int32_t ALU_REGISTER_R_SRCB = 0x21;
constexpr int32_t ALU_REGISTER_R_ACCU = 0x31;
constexpr uint32_t GP_THREAD_TIME_REG_ADDRESS_OFFSET_LOW = 0x23A8;
void computeWorkgroupSize1D(
uint32_t maxWorkGroupSize,
size_t workGroupSize[3],
const size_t workItems[3],
size_t simdSize);
void computeWorkgroupSizeND(
WorkSizeInfo wsInfo,
size_t workGroupSize[3],
const size_t workItems[3],
const uint32_t workDim);
void computeWorkgroupSize2D(
uint32_t maxWorkGroupSize,
size_t workGroupSize[3],
const size_t workItems[3],
size_t simdSize);
void computeWorkgroupSizeSquared(
uint32_t maxWorkGroupSize,
size_t workGroupSize[3],
const size_t workItems[3],
size_t simdSize,
const uint32_t workDim);
Vec3<size_t> computeWorkgroupSize(
const DispatchInfo &dispatchInfo);
Vec3<size_t> generateWorkgroupSize(
const DispatchInfo &dispatchInfo);
Vec3<size_t> computeWorkgroupsNumber(
const Vec3<size_t> gws,
const Vec3<size_t> lws);
Vec3<size_t> generateWorkgroupsNumber(
const Vec3<size_t> gws,
const Vec3<size_t> lws);
Vec3<size_t> generateWorkgroupsNumber(
const DispatchInfo &dispatchInfo);
inline uint32_t calculateDispatchDim(Vec3<size_t> dispatchSize, Vec3<size_t> dispatchOffset) {
return std::max(1U, std::max(dispatchSize.getSimplifiedDim(), dispatchOffset.getSimplifiedDim()));
}
Vec3<size_t> canonizeWorkgroup(
Vec3<size_t> workgroup);
void provideLocalWorkGroupSizeHints(Context *context, uint32_t maxWorkGroupSize, DispatchInfo dispatchInfo);
inline cl_uint computeDimensions(const size_t workItems[3]) {
return (workItems[2] > 1) ? 3 : (workItems[1] > 1) ? 2 : 1;
}
template <typename SizeAndAllocCalcT, typename... CalcArgsT>
IndirectHeap *allocateIndirectHeap(SizeAndAllocCalcT &&calc, CalcArgsT &&... args) {
size_t alignment = MemoryConstants::pageSize;
size_t size = calc(std::forward<CalcArgsT>(args)...);
return new IndirectHeap(alignedMalloc(size, alignment), size);
}
template <typename GfxFamily>
class GpgpuWalkerHelper {
public:
static void addAluReadModifyWriteRegister(
LinearStream *pCommandStream,
uint32_t aluRegister,
uint32_t operation,
uint32_t mask);
static void applyWADisableLSQCROPERFforOCL(LinearStream *pCommandStream,
const Kernel &kernel,
bool disablePerfMode);
static size_t getSizeForWADisableLSQCROPERFforOCL(const Kernel *pKernel);
static size_t setGpgpuWalkerThreadData(
typename GfxFamily::GPGPU_WALKER *pCmd,
const size_t globalOffsets[3],
const size_t startWorkGroups[3],
const size_t numWorkGroups[3],
const size_t localWorkSizesIn[3],
uint32_t simd);
static void dispatchProfilingCommandsStart(
HwTimeStamps &hwTimeStamps,
OCLRT::LinearStream *commandStream);
static void dispatchProfilingCommandsEnd(
HwTimeStamps &hwTimeStamps,
OCLRT::LinearStream *commandStream);
static void dispatchPerfCountersNoopidRegisterCommands(
CommandQueue &commandQueue,
OCLRT::HwPerfCounter &hwPerfCounter,
OCLRT::LinearStream *commandStream,
bool start);
static void dispatchPerfCountersReadFreqRegisterCommands(
CommandQueue &commandQueue,
OCLRT::HwPerfCounter &hwPerfCounter,
OCLRT::LinearStream *commandStream,
bool start);
static void dispatchPerfCountersGeneralPurposeCounterCommands(
CommandQueue &commandQueue,
OCLRT::HwPerfCounter &hwPerfCounter,
OCLRT::LinearStream *commandStream,
bool start);
static void dispatchPerfCountersUserCounterCommands(
CommandQueue &commandQueue,
OCLRT::HwPerfCounter &hwPerfCounter,
OCLRT::LinearStream *commandStream,
bool start);
static void dispatchPerfCountersOABufferStateCommands(
CommandQueue &commandQueue,
OCLRT::HwPerfCounter &hwPerfCounter,
OCLRT::LinearStream *commandStream);
static void dispatchPerfCountersCommandsStart(
CommandQueue &commandQueue,
OCLRT::HwPerfCounter &hwPerfCounter,
OCLRT::LinearStream *commandStream);
static void dispatchPerfCountersCommandsEnd(
CommandQueue &commandQueue,
OCLRT::HwPerfCounter &hwPerfCounter,
OCLRT::LinearStream *commandStream);
static void dispatchWalker(
CommandQueue &commandQueue,
const MultiDispatchInfo &multiDispatchInfo,
cl_uint numEventsInWaitList,
const cl_event *eventWaitList,
KernelOperation **blockedCommandsData,
HwTimeStamps *hwTimeStamps,
OCLRT::HwPerfCounter *hwPerfCounter,
PreemptionMode preemptionMode,
bool blockQueue,
unsigned int commandType = 0);
static void dispatchWalker(
CommandQueue &commandQueue,
const Kernel &kernel,
cl_uint workDim,
const size_t globalOffsets[3],
const size_t workItems[3],
const size_t *localWorkSizesIn,
cl_uint numEventsInWaitList,
const cl_event *eventWaitList,
KernelOperation **blockedCommandsData,
HwTimeStamps *hwTimeStamps,
HwPerfCounter *hwPerfCounter,
PreemptionMode preemptionMode,
bool blockQueue);
static void dispatchScheduler(
CommandQueue &commandQueue,
DeviceQueueHw<GfxFamily> &devQueueHw,
PreemptionMode preemptionMode,
SchedulerKernel &scheduler);
};
template <typename GfxFamily, uint32_t eventType>
struct EnqueueOperation {
static_assert(eventType != CL_COMMAND_NDRANGE_KERNEL, "for eventType CL_COMMAND_NDRANGE_KERNEL use specialization class");
static_assert(eventType != CL_COMMAND_MARKER, "for eventType CL_COMMAND_MARKER use specialization class");
static_assert(eventType != CL_COMMAND_MIGRATE_MEM_OBJECTS, "for eventType CL_COMMAND_MIGRATE_MEM_OBJECTS use specialization class");
static size_t getTotalSizeRequiredCS(bool reserveProfilingCmdsSpace, bool reservePerfCounters, CommandQueue &commandQueue, const MultiDispatchInfo &multiDispatchInfo) {
size_t size = KernelCommandsHelper<GfxFamily>::getSizeRequiredCS() +
sizeof(typename GfxFamily::PIPE_CONTROL) * (KernelCommandsHelper<GfxFamily>::isPipeControlWArequired() ? 2 : 1);
if (reserveProfilingCmdsSpace) {
size += 2 * sizeof(typename GfxFamily::PIPE_CONTROL) + 4 * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
}
if (reservePerfCounters) {
//start cmds
//P_C: flush CS & TimeStamp BEGIN
size += 2 * sizeof(typename GfxFamily::PIPE_CONTROL);
//SRM NOOPID & Frequency
size += 2 * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
//gp registers
size += OCLRT::INSTR_GENERAL_PURPOSE_COUNTERS_COUNT * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
//report perf count
size += sizeof(typename GfxFamily::MI_REPORT_PERF_COUNT);
//user registers
size += commandQueue.getPerfCountersUserRegistersNumber() * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
//end cmds
//P_C: flush CS & TimeStamp END;
size += 2 * sizeof(typename GfxFamily::PIPE_CONTROL);
//OA buffer (status head, tail)
size += 3 * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
//report perf count
size += sizeof(typename GfxFamily::MI_REPORT_PERF_COUNT);
//gp registers
size += OCLRT::INSTR_GENERAL_PURPOSE_COUNTERS_COUNT * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
//SRM NOOPID & Frequency
size += 2 * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
//user registers
size += commandQueue.getPerfCountersUserRegistersNumber() * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
}
Device &device = commandQueue.getDevice();
for (auto &dispatchInfo : multiDispatchInfo) {
auto &kernel = *dispatchInfo.getKernel();
size += sizeof(typename GfxFamily::GPGPU_WALKER);
size += GpgpuWalkerHelper<GfxFamily>::getSizeForWADisableLSQCROPERFforOCL(&kernel);
size += PreemptionHelper::getPreemptionWaCsSize<GfxFamily>(device);
}
return size;
}
static size_t getSizeRequiredCS(bool reserveProfilingCmdsSpace, bool reservePerfCounters, CommandQueue &commandQueue, const Kernel *pKernel) {
size_t size = sizeof(typename GfxFamily::GPGPU_WALKER) + KernelCommandsHelper<GfxFamily>::getSizeRequiredCS() +
sizeof(typename GfxFamily::PIPE_CONTROL) * (KernelCommandsHelper<GfxFamily>::isPipeControlWArequired() ? 2 : 1);
size += PreemptionHelper::getPreemptionWaCsSize<GfxFamily>(commandQueue.getDevice());
if (reserveProfilingCmdsSpace) {
size += 2 * sizeof(typename GfxFamily::PIPE_CONTROL) + 2 * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
}
if (reservePerfCounters) {
//start cmds
//P_C: flush CS & TimeStamp BEGIN
size += 2 * sizeof(typename GfxFamily::PIPE_CONTROL);
//SRM NOOPID & Frequency
size += 2 * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
//gp registers
size += OCLRT::INSTR_GENERAL_PURPOSE_COUNTERS_COUNT * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
//report perf count
size += sizeof(typename GfxFamily::MI_REPORT_PERF_COUNT);
//user registers
size += commandQueue.getPerfCountersUserRegistersNumber() * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
//end cmds
//P_C: flush CS & TimeStamp END;
size += 2 * sizeof(typename GfxFamily::PIPE_CONTROL);
//OA buffer (status head, tail)
size += 3 * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
//report perf count
size += sizeof(typename GfxFamily::MI_REPORT_PERF_COUNT);
//gp registers
size += OCLRT::INSTR_GENERAL_PURPOSE_COUNTERS_COUNT * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
//SRM NOOPID & Frequency
size += 2 * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
//user registers
size += commandQueue.getPerfCountersUserRegistersNumber() * sizeof(typename GfxFamily::MI_STORE_REGISTER_MEM);
}
size += GpgpuWalkerHelper<GfxFamily>::getSizeForWADisableLSQCROPERFforOCL(pKernel);
return size;
}
};
template <typename GfxFamily, uint32_t eventType>
LinearStream &getCommandStream(CommandQueue &commandQueue, bool reserveProfilingCmdsSpace, bool reservePerfCounterCmdsSpace, const Kernel *pKernel) {
auto expectedSizeCS = EnqueueOperation<GfxFamily, eventType>::getSizeRequiredCS(reserveProfilingCmdsSpace, reservePerfCounterCmdsSpace, commandQueue, pKernel);
return commandQueue.getCS(expectedSizeCS);
}
template <typename GfxFamily, uint32_t eventType>
LinearStream &getCommandStream(CommandQueue &commandQueue, bool reserveProfilingCmdsSpace, bool reservePerfCounterCmdsSpace, const MultiDispatchInfo &multiDispatchInfo) {
size_t expectedSizeCS = 0;
Kernel *parentKernel = multiDispatchInfo.size() > 0 ? multiDispatchInfo.begin()->getKernel() : nullptr;
for (auto &dispatchInfo : multiDispatchInfo) {
expectedSizeCS += EnqueueOperation<GfxFamily, eventType>::getSizeRequiredCS(reserveProfilingCmdsSpace, reservePerfCounterCmdsSpace, commandQueue, dispatchInfo.getKernel());
}
if (parentKernel && parentKernel->isParentKernel) {
SchedulerKernel &scheduler = BuiltIns::getInstance().getSchedulerKernel(parentKernel->getContext());
expectedSizeCS += EnqueueOperation<GfxFamily, eventType>::getSizeRequiredCS(reserveProfilingCmdsSpace, reservePerfCounterCmdsSpace, commandQueue, &scheduler);
}
return commandQueue.getCS(expectedSizeCS);
}
template <typename GfxFamily, IndirectHeap::Type heapType>
IndirectHeap &getIndirectHeap(CommandQueue &commandQueue, const MultiDispatchInfo &multiDispatchInfo) {
size_t expectedSize = 0;
IndirectHeap *ih = nullptr;
// clang-format off
switch (heapType) {
case IndirectHeap::DYNAMIC_STATE: expectedSize = KernelCommandsHelper<GfxFamily>::getTotalSizeRequiredDSH(multiDispatchInfo); break;
case IndirectHeap::INDIRECT_OBJECT: expectedSize = KernelCommandsHelper<GfxFamily>::getTotalSizeRequiredIOH(multiDispatchInfo); break;
case IndirectHeap::SURFACE_STATE: expectedSize = KernelCommandsHelper<GfxFamily>::getTotalSizeRequiredSSH(multiDispatchInfo); break;
}
// clang-format on
if (multiDispatchInfo.begin()->getKernel()->isParentKernel) {
if (heapType == IndirectHeap::SURFACE_STATE) {
expectedSize += KernelCommandsHelper<GfxFamily>::template getSizeRequiredForExecutionModel<heapType>(const_cast<const Kernel &>(*(multiDispatchInfo.begin()->getKernel())));
} else //if (heapType == IndirectHeap::DYNAMIC_STATE || heapType == IndirectHeap::INDIRECT_OBJECT)
{
DeviceQueueHw<GfxFamily> *pDevQueue = castToObject<DeviceQueueHw<GfxFamily>>(commandQueue.getContext().getDefaultDeviceQueue());
DEBUG_BREAK_IF(pDevQueue == nullptr);
ih = pDevQueue->getIndirectHeap(IndirectHeap::DYNAMIC_STATE);
}
}
if (ih == nullptr)
ih = &commandQueue.getIndirectHeap(heapType, expectedSize);
return *ih;
}
} // namespace OCLRT