mirror of
https://github.com/intel/compute-runtime.git
synced 2026-01-06 02:18:05 +08:00
Reduce number of jumps in RelaxedOrdering scheduler
Related-To: NEO-7458 Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>
This commit is contained in:
Dunajski, Bartosz
committed by
Compute-Runtime-Automation
Compute-Runtime-Automation
parent
7dc4fd8dda
commit
3ff7a63145
@@ -111,10 +111,11 @@ void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchStaticRelaxedOrderingSch
|
||||
// 1. Init section
|
||||
{
|
||||
EncodeMiPredicate<GfxFamily>::encode(schedulerCmdStream, MiPredicateType::Disable);
|
||||
EncodeBatchBufferStartOrEnd<GfxFamily>::programConditionalDataRegBatchBufferStart(
|
||||
schedulerCmdStream,
|
||||
schedulerStartAddress + RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<GfxFamily>::semaphoreSectionJumpStart,
|
||||
CS_GPR_R1, 0, CompareOperation::Equal, false);
|
||||
|
||||
EncodeSetMMIO<GfxFamily>::encodeREG(schedulerCmdStream, CS_GPR_R0, CS_GPR_R9);
|
||||
EncodeSetMMIO<GfxFamily>::encodeREG(schedulerCmdStream, CS_GPR_R0 + 4, CS_GPR_R9 + 4);
|
||||
|
||||
EncodeBatchBufferStartOrEnd<GfxFamily>::programConditionalDataRegBatchBufferStart(schedulerCmdStream, 0, CS_GPR_R1, 0, CompareOperation::Equal, true);
|
||||
|
||||
LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R2, 0, true);
|
||||
LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R2 + 4, 0, true);
|
||||
@@ -130,6 +131,8 @@ void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchStaticRelaxedOrderingSch
|
||||
|
||||
// 2. Dispatch task section (loop start)
|
||||
{
|
||||
UNRECOVERABLE_IF(schedulerCmdStream.getUsed() != RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<GfxFamily>::loopStartSectionStart);
|
||||
|
||||
EncodeMiPredicate<GfxFamily>::encode(schedulerCmdStream, MiPredicateType::Disable);
|
||||
|
||||
LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R6, 8, true);
|
||||
@@ -158,15 +161,17 @@ void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchStaticRelaxedOrderingSch
|
||||
|
||||
// 3. Remove task section
|
||||
{
|
||||
UNRECOVERABLE_IF(schedulerCmdStream.getUsed() != RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<GfxFamily>::removeTaskSectionStart);
|
||||
|
||||
EncodeMiPredicate<GfxFamily>::encode(schedulerCmdStream, MiPredicateType::Disable);
|
||||
|
||||
EncodeMathMMIO<GfxFamily>::encodeDecrement(schedulerCmdStream, AluRegisters::R_1);
|
||||
EncodeMathMMIO<GfxFamily>::encodeDecrement(schedulerCmdStream, AluRegisters::R_2);
|
||||
|
||||
EncodeBatchBufferStartOrEnd<GfxFamily>::programConditionalDataRegBatchBufferStart(
|
||||
schedulerCmdStream,
|
||||
schedulerStartAddress + RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<GfxFamily>::semaphoreSectionJumpStart,
|
||||
CS_GPR_R1, 0, CompareOperation::Equal, false);
|
||||
EncodeSetMMIO<GfxFamily>::encodeREG(schedulerCmdStream, CS_GPR_R0, CS_GPR_R9);
|
||||
EncodeSetMMIO<GfxFamily>::encodeREG(schedulerCmdStream, CS_GPR_R0 + 4, CS_GPR_R9 + 4);
|
||||
|
||||
EncodeBatchBufferStartOrEnd<GfxFamily>::programConditionalDataRegBatchBufferStart(schedulerCmdStream, 0, CS_GPR_R1, 0, CompareOperation::Equal, true);
|
||||
|
||||
LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R7, 8, true);
|
||||
LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R7 + 4, 0, true);
|
||||
@@ -196,6 +201,8 @@ void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchStaticRelaxedOrderingSch
|
||||
|
||||
// 4. List loop check section
|
||||
{
|
||||
UNRECOVERABLE_IF(schedulerCmdStream.getUsed() != RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<GfxFamily>::tasksListLoopCheckSectionStart);
|
||||
|
||||
EncodeMiPredicate<GfxFamily>::encode(schedulerCmdStream, MiPredicateType::Disable);
|
||||
|
||||
EncodeMathMMIO<GfxFamily>::encodeIncrement(schedulerCmdStream, AluRegisters::R_2);
|
||||
@@ -211,6 +218,8 @@ void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchStaticRelaxedOrderingSch
|
||||
|
||||
// 5. Drain request section
|
||||
{
|
||||
UNRECOVERABLE_IF(schedulerCmdStream.getUsed() != RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<GfxFamily>::drainRequestSectionStart);
|
||||
|
||||
using MI_LOAD_REGISTER_IMM = typename GfxFamily::MI_LOAD_REGISTER_IMM;
|
||||
|
||||
EncodeDummyBlitWaArgs waArgs{false, const_cast<RootDeviceEnvironment *>(&this->rootDeviceEnvironment)};
|
||||
@@ -233,28 +242,28 @@ void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchStaticRelaxedOrderingSch
|
||||
CS_GPR_R5, 1, CompareOperation::Equal, false);
|
||||
}
|
||||
|
||||
// Exit Static scheduler
|
||||
// 6. Scheduler loop check section
|
||||
{
|
||||
UNRECOVERABLE_IF(schedulerCmdStream.getUsed() != RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<GfxFamily>::schedulerLoopCheckSectionStart);
|
||||
|
||||
// 6. Jump to scheduler loop check section (dynamic scheduler)
|
||||
EncodeSetMMIO<GfxFamily>::encodeREG(schedulerCmdStream, CS_GPR_R0, CS_GPR_R9);
|
||||
EncodeSetMMIO<GfxFamily>::encodeREG(schedulerCmdStream, CS_GPR_R0 + 4, CS_GPR_R9 + 4);
|
||||
EncodeBatchBufferStartOrEnd<GfxFamily>::programBatchBufferStart(&schedulerCmdStream, 0, false, true, false);
|
||||
LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R10, static_cast<uint32_t>(RelaxedOrderingHelper::DynamicSchedulerSizeAndOffsetSection<GfxFamily>::semaphoreSectionSize), true);
|
||||
LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R10 + 4, 0, true);
|
||||
|
||||
// 7. Jump to Semaphore section (dynamic scheduler)
|
||||
EncodeMiPredicate<GfxFamily>::encode(schedulerCmdStream, MiPredicateType::Disable);
|
||||
LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R10, static_cast<uint32_t>(RelaxedOrderingHelper::DynamicSchedulerSizeAndOffsetSection<GfxFamily>::schedulerLoopCheckSectionSize), true);
|
||||
EncodeAluHelper<GfxFamily, 4> aluHelper;
|
||||
aluHelper.setMocs(miMathMocs);
|
||||
aluHelper.setNextAlu(AluRegisters::OPCODE_LOAD, AluRegisters::R_SRCA, AluRegisters::R_9);
|
||||
aluHelper.setNextAlu(AluRegisters::OPCODE_LOAD, AluRegisters::R_SRCB, AluRegisters::R_10);
|
||||
aluHelper.setNextAlu(AluRegisters::OPCODE_ADD);
|
||||
aluHelper.setNextAlu(AluRegisters::OPCODE_STORE, AluRegisters::R_0, AluRegisters::R_ACCU);
|
||||
aluHelper.copyToCmdStream(schedulerCmdStream);
|
||||
|
||||
LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R10 + 4, 0, true);
|
||||
EncodeBatchBufferStartOrEnd<GfxFamily>::programConditionalRegMemBatchBufferStart(schedulerCmdStream, 0, semaphoreGpuVa, CS_GPR_R11, CompareOperation::GreaterOrEqual, true);
|
||||
|
||||
EncodeAluHelper<GfxFamily, 4> aluHelper;
|
||||
aluHelper.setMocs(miMathMocs);
|
||||
aluHelper.setNextAlu(AluRegisters::OPCODE_LOAD, AluRegisters::R_SRCA, AluRegisters::R_9);
|
||||
aluHelper.setNextAlu(AluRegisters::OPCODE_LOAD, AluRegisters::R_SRCB, AluRegisters::R_10);
|
||||
aluHelper.setNextAlu(AluRegisters::OPCODE_ADD);
|
||||
aluHelper.setNextAlu(AluRegisters::OPCODE_STORE, AluRegisters::R_0, AluRegisters::R_ACCU);
|
||||
aluHelper.copyToCmdStream(schedulerCmdStream);
|
||||
EncodeBatchBufferStartOrEnd<GfxFamily>::programBatchBufferStart(&schedulerCmdStream, schedulerStartAddress + RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<GfxFamily>::loopStartSectionStart,
|
||||
false, false, false);
|
||||
}
|
||||
|
||||
EncodeBatchBufferStartOrEnd<GfxFamily>::programBatchBufferStart(&schedulerCmdStream, 0, false, true, false);
|
||||
UNRECOVERABLE_IF(schedulerCmdStream.getUsed() != RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<GfxFamily>::totalSize);
|
||||
}
|
||||
|
||||
template <typename GfxFamily, typename Dispatcher>
|
||||
@@ -265,23 +274,15 @@ void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchRelaxedOrderingScheduler
|
||||
|
||||
uint64_t schedulerStartVa = ringCommandStream.getCurrentGpuAddressPosition();
|
||||
|
||||
uint64_t schedulerLoopCheckVa = schedulerStartVa + RelaxedOrderingHelper::DynamicSchedulerSizeAndOffsetSection<GfxFamily>::schedulerLoopCheckSectionStart;
|
||||
uint64_t semaphoreSectionVa = schedulerStartVa + RelaxedOrderingHelper::DynamicSchedulerSizeAndOffsetSection<GfxFamily>::semaphoreSectionStart;
|
||||
|
||||
LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R9, static_cast<uint32_t>(schedulerLoopCheckVa & 0xFFFF'FFFFULL), true);
|
||||
LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R9 + 4, static_cast<uint32_t>(schedulerLoopCheckVa >> 32), true);
|
||||
LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R11, value, true);
|
||||
LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R9, static_cast<uint32_t>(semaphoreSectionVa & 0xFFFF'FFFFULL), true);
|
||||
LriHelper<GfxFamily>::program(&schedulerCmdStream, CS_GPR_R9 + 4, static_cast<uint32_t>(semaphoreSectionVa >> 32), true);
|
||||
|
||||
schedulerCmdStream.getSpace(sizeof(typename GfxFamily::MI_BATCH_BUFFER_START)); // skip patching
|
||||
|
||||
// 2. Scheduler loop check section
|
||||
{
|
||||
EncodeBatchBufferStartOrEnd<GfxFamily>::programConditionalDataMemBatchBufferStart(
|
||||
schedulerCmdStream, schedulerStartVa + RelaxedOrderingHelper::DynamicSchedulerSizeAndOffsetSection<GfxFamily>::endSectionStart,
|
||||
semaphoreGpuVa, value, CompareOperation::GreaterOrEqual, false);
|
||||
|
||||
schedulerCmdStream.getSpace(sizeof(typename GfxFamily::MI_BATCH_BUFFER_START)); // skip patching
|
||||
}
|
||||
|
||||
// 3. Semaphore section
|
||||
// 2. Semaphore section
|
||||
{
|
||||
using COMPARE_OPERATION = typename GfxFamily::MI_SEMAPHORE_WAIT::COMPARE_OPERATION;
|
||||
|
||||
@@ -850,21 +851,12 @@ void DirectSubmissionHw<GfxFamily, Dispatcher>::preinitializeRelaxedOrderingSect
|
||||
uint64_t schedulerStartAddress = relaxedOrderingSchedulerAllocation->getGpuAddress();
|
||||
|
||||
// 1. Init section
|
||||
LriHelper<GfxFamily>::program(&schedulerStream, CS_GPR_R11, 0, true);
|
||||
LriHelper<GfxFamily>::program(&schedulerStream, CS_GPR_R9, 0, true);
|
||||
LriHelper<GfxFamily>::program(&schedulerStream, CS_GPR_R9 + 4, 0, true);
|
||||
EncodeBatchBufferStartOrEnd<GfxFamily>::programBatchBufferStart(&schedulerStream, schedulerStartAddress, false, false, false);
|
||||
|
||||
// 2. Scheduler loop check section
|
||||
{
|
||||
|
||||
EncodeBatchBufferStartOrEnd<GfxFamily>::programConditionalDataMemBatchBufferStart(schedulerStream, 0, 0, 0, CompareOperation::GreaterOrEqual, false);
|
||||
|
||||
EncodeBatchBufferStartOrEnd<GfxFamily>::programBatchBufferStart(&schedulerStream,
|
||||
schedulerStartAddress + RelaxedOrderingHelper::StaticSchedulerSizeAndOffsetSection<GfxFamily>::loopStartSectionStart,
|
||||
false, false, false);
|
||||
}
|
||||
|
||||
// 3. Semaphore section
|
||||
// 2. Semaphore section
|
||||
{
|
||||
using COMPARE_OPERATION = typename GfxFamily::MI_SEMAPHORE_WAIT::COMPARE_OPERATION;
|
||||
|
||||
@@ -873,7 +865,7 @@ void DirectSubmissionHw<GfxFamily, Dispatcher>::preinitializeRelaxedOrderingSect
|
||||
EncodeSemaphore<GfxFamily>::addMiSemaphoreWaitCommand(schedulerStream, 0, 0, COMPARE_OPERATION::COMPARE_OPERATION_SAD_GREATER_THAN_OR_EQUAL_SDD);
|
||||
}
|
||||
|
||||
// 4. End section
|
||||
// 3. End section
|
||||
{
|
||||
EncodeMiPredicate<GfxFamily>::encode(schedulerStream, MiPredicateType::Disable);
|
||||
|
||||
|
||||
@@ -62,7 +62,7 @@ struct StaticSchedulerSizeAndOffsetSection {
|
||||
using MI_BATCH_BUFFER_START = typename GfxFamily::MI_BATCH_BUFFER_START;
|
||||
|
||||
static constexpr uint64_t initSectionSize = EncodeBatchBufferStartOrEnd<GfxFamily>::getCmdSizeConditionalDataRegBatchBufferStart() + (6 * sizeof(MI_LOAD_REGISTER_IMM)) +
|
||||
EncodeMiPredicate<GfxFamily>::getCmdSize();
|
||||
EncodeMiPredicate<GfxFamily>::getCmdSize() + (2 * sizeof(MI_LOAD_REGISTER_REG));
|
||||
|
||||
static constexpr uint64_t loopStartSectionStart = initSectionSize;
|
||||
static constexpr uint64_t loopStartSectionSize = (4 * sizeof(MI_LOAD_REGISTER_IMM)) + EncodeAluHelper<GfxFamily, 10>::getCmdsSize() + sizeof(MI_BATCH_BUFFER_START) +
|
||||
@@ -70,7 +70,8 @@ struct StaticSchedulerSizeAndOffsetSection {
|
||||
|
||||
static constexpr uint64_t removeTaskSectionStart = loopStartSectionStart + loopStartSectionSize;
|
||||
static constexpr uint64_t removeStartSectionSize = (2 * EncodeMathMMIO<GfxFamily>::getCmdSizeForIncrementOrDecrement()) + EncodeBatchBufferStartOrEnd<GfxFamily>::getCmdSizeConditionalDataRegBatchBufferStart() +
|
||||
(4 * sizeof(MI_LOAD_REGISTER_IMM)) + EncodeAluHelper<GfxFamily, 14>::getCmdsSize() + EncodeMiPredicate<GfxFamily>::getCmdSize();
|
||||
(4 * sizeof(MI_LOAD_REGISTER_IMM)) + EncodeAluHelper<GfxFamily, 14>::getCmdsSize() + EncodeMiPredicate<GfxFamily>::getCmdSize() +
|
||||
(2 * sizeof(MI_LOAD_REGISTER_REG));
|
||||
|
||||
static constexpr uint64_t tasksListLoopCheckSectionStart = removeTaskSectionStart + removeStartSectionSize;
|
||||
static constexpr uint64_t tasksListLoopCheckSectionSize = EncodeMathMMIO<GfxFamily>::getCmdSizeForIncrementOrDecrement() + EncodeBatchBufferStartOrEnd<GfxFamily>::getCmdSizeConditionalRegRegBatchBufferStart() +
|
||||
@@ -78,14 +79,12 @@ struct StaticSchedulerSizeAndOffsetSection {
|
||||
|
||||
static constexpr uint64_t drainRequestSectionStart = tasksListLoopCheckSectionStart + tasksListLoopCheckSectionSize;
|
||||
static constexpr uint64_t drainRequestSectionSize = sizeof(typename GfxFamily::MI_ARB_CHECK) + (2 * EncodeBatchBufferStartOrEnd<GfxFamily>::getCmdSizeConditionalDataRegBatchBufferStart());
|
||||
static constexpr uint64_t schedulerLoopCheckSectionJumpStart = drainRequestSectionStart + drainRequestSectionSize;
|
||||
static constexpr uint64_t schedulerLoopCheckSectionJumpSize = 2 * sizeof(MI_LOAD_REGISTER_REG) + sizeof(MI_BATCH_BUFFER_START);
|
||||
|
||||
static constexpr uint64_t semaphoreSectionJumpStart = schedulerLoopCheckSectionJumpStart + schedulerLoopCheckSectionJumpSize;
|
||||
static constexpr uint64_t semaphoreSectionJumpSize = EncodeMiPredicate<GfxFamily>::getCmdSize() + (2 * sizeof(MI_LOAD_REGISTER_IMM)) + EncodeAluHelper<GfxFamily, 4>::getCmdsSize() +
|
||||
sizeof(MI_BATCH_BUFFER_START);
|
||||
static constexpr uint64_t schedulerLoopCheckSectionStart = drainRequestSectionStart + drainRequestSectionSize;
|
||||
static constexpr uint64_t schedulerLoopCheckSectionSize = (2 * sizeof(MI_LOAD_REGISTER_IMM)) + EncodeAluHelper<GfxFamily, 4>::getCmdsSize() +
|
||||
EncodeBatchBufferStartOrEnd<GfxFamily>::getCmdSizeConditionalRegMemBatchBufferStart() + sizeof(MI_BATCH_BUFFER_START);
|
||||
|
||||
static constexpr uint64_t totalSize = semaphoreSectionJumpStart + semaphoreSectionJumpSize;
|
||||
static constexpr uint64_t totalSize = schedulerLoopCheckSectionStart + schedulerLoopCheckSectionSize;
|
||||
};
|
||||
|
||||
template <typename GfxFamily>
|
||||
@@ -93,12 +92,9 @@ struct DynamicSchedulerSizeAndOffsetSection {
|
||||
using MI_LOAD_REGISTER_IMM = typename GfxFamily::MI_LOAD_REGISTER_IMM;
|
||||
using MI_BATCH_BUFFER_START = typename GfxFamily::MI_BATCH_BUFFER_START;
|
||||
|
||||
static constexpr uint64_t initSectionSize = (2 * sizeof(MI_LOAD_REGISTER_IMM)) + sizeof(MI_BATCH_BUFFER_START);
|
||||
static constexpr uint64_t initSectionSize = (3 * sizeof(MI_LOAD_REGISTER_IMM)) + sizeof(MI_BATCH_BUFFER_START);
|
||||
|
||||
static constexpr uint64_t schedulerLoopCheckSectionStart = initSectionSize;
|
||||
static constexpr uint64_t schedulerLoopCheckSectionSize = EncodeBatchBufferStartOrEnd<GfxFamily>::getCmdSizeConditionalDataMemBatchBufferStart() + sizeof(MI_BATCH_BUFFER_START);
|
||||
|
||||
static constexpr uint64_t semaphoreSectionStart = schedulerLoopCheckSectionStart + schedulerLoopCheckSectionSize;
|
||||
static constexpr uint64_t semaphoreSectionStart = initSectionSize;
|
||||
static constexpr uint64_t semaphoreSectionSize = EncodeSemaphore<GfxFamily>::getSizeMiSemaphoreWait() + EncodeMiPredicate<GfxFamily>::getCmdSize();
|
||||
|
||||
static constexpr uint64_t endSectionStart = semaphoreSectionStart + semaphoreSectionSize;
|
||||
|
||||
Reference in New Issue
Block a user