intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-05 18:06:32 +08:00
Code Issues Packages Projects Releases Wiki Activity

feature: add missing experimental mcl api calls

Browse Source 
Related-To: NEO-10492

Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
This commit is contained in:
Zbigniew Zdanowicz
2025-07-11 14:35:04 +00:00
committed by Compute-Runtime-Automation
parent b39c5c151a
commit 1599e5bafe
10 changed files with 374 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files

146
level_zero/experimental/source/mutable_cmdlist/mutable_cmdlist_hw_experimental.inl
View File

@@ -51,6 +51,147 @@ constexpr uint32_t regToMMIO(MclAluReg reg) {
return 0;
}
template <GFXCORE_FAMILY gfxCoreFamily>
ze_result_t MutableCommandListCoreFamily<gfxCoreFamily>::appendMILoadRegImm(MclAluReg reg, uint32_t value) {
NEO::LriHelper<GfxFamily>::program(this->commandContainer.getCommandStream(),
regToMMIO(reg),
value,
false,
this->isCopyOnly(false));
return ZE_RESULT_SUCCESS;
}
template <GFXCORE_FAMILY gfxCoreFamily>
ze_result_t MutableCommandListCoreFamily<gfxCoreFamily>::appendMILoadRegReg(MclAluReg reg1, MclAluReg reg2) {
using MI_LOAD_REGISTER_REG = typename GfxFamily::MI_LOAD_REGISTER_REG;
MI_LOAD_REGISTER_REG cmd = GfxFamily::cmdInitLoadRegisterReg;
cmd.setSourceRegisterAddress(regToMMIO(reg2));
cmd.setDestinationRegisterAddress(regToMMIO(reg1));
auto buffer = this->commandContainer.getCommandStream()->getSpace(sizeof(cmd));
*reinterpret_cast<MI_LOAD_REGISTER_REG *>(buffer) = cmd;
return ZE_RESULT_SUCCESS;
}
template <GFXCORE_FAMILY gfxCoreFamily>
ze_result_t MutableCommandListCoreFamily<gfxCoreFamily>::appendMILoadRegMem(MclAluReg reg1, uint64_t address) {
using MI_LOAD_REGISTER_MEM = typename GfxFamily::MI_LOAD_REGISTER_MEM;
MI_LOAD_REGISTER_MEM cmd = GfxFamily::cmdInitLoadRegisterMem;
cmd.setRegisterAddress(regToMMIO(reg1));
cmd.setMemoryAddress(address);
auto buffer = this->commandContainer.getCommandStream()->getSpace(sizeof(cmd));
*reinterpret_cast<MI_LOAD_REGISTER_MEM *>(buffer) = cmd;
return ZE_RESULT_SUCCESS;
}
template <GFXCORE_FAMILY gfxCoreFamily>
ze_result_t MutableCommandListCoreFamily<gfxCoreFamily>::appendMIStoreRegMem(MclAluReg reg1, uint64_t address) {
using MI_STORE_REGISTER_MEM = typename GfxFamily::MI_STORE_REGISTER_MEM;
MI_STORE_REGISTER_MEM cmd0 = GfxFamily::cmdInitStoreRegisterMem;
cmd0.setRegisterAddress(regToMMIO(reg1));
cmd0.setMemoryAddress(address);
auto buffer = this->commandContainer.getCommandStream()->getSpace(sizeof(cmd0));
*reinterpret_cast<MI_STORE_REGISTER_MEM *>(buffer) = cmd0;
return ZE_RESULT_SUCCESS;
}
template <GFXCORE_FAMILY gfxCoreFamily>
ze_result_t MutableCommandListCoreFamily<gfxCoreFamily>::appendMIMath(void *aluArray, size_t aluCount) {
using MI_MATH = typename GfxFamily::MI_MATH;
using MI_MATH_ALU_INST_INLINE = typename GfxFamily::MI_MATH_ALU_INST_INLINE;
constexpr int32_t aluRegisterRSrca = 0x20;
constexpr int32_t aluRegisterRSrcb = 0x21;
// Add instruction MI_MATH with 4 MI_MATH_ALU_INST_INLINE operands
auto pCmd3 = reinterpret_cast<uint32_t *>(this->commandContainer.getCommandStream()->getSpace(sizeof(MI_MATH) + aluCount * 4 * sizeof(MI_MATH_ALU_INST_INLINE)));
MI_MATH miMath;
miMath.DW0.Value = 0x0;
miMath.DW0.BitField.InstructionType = MI_MATH::COMMAND_TYPE_MI_COMMAND;
miMath.DW0.BitField.InstructionOpcode = MI_MATH::MI_COMMAND_OPCODE_MI_MATH;
// 0x3 - 5 Dwords length cmd (-2): 1 for MI_MATH, 4 for MI_MATH_ALU_INST_INLINE
miMath.DW0.BitField.DwordLength = static_cast<uint32_t>(aluCount * 4 - 1);
*reinterpret_cast<MI_MATH *>(pCmd3) = miMath;
pCmd3++;
MI_MATH_ALU_INST_INLINE *pAluParam = reinterpret_cast<MI_MATH_ALU_INST_INLINE *>(pCmd3);
MI_MATH_ALU_INST_INLINE aluParam;
zex_mcl_alu_operation_t *exAluArray = reinterpret_cast<zex_mcl_alu_operation_t *>(aluArray);
for (size_t loop = 0; loop < aluCount; loop++, exAluArray++) {
aluParam.DW0.Value = 0x0;
// Setup first operand of MI_MATH - load regSource1 into register A
if (exAluArray->regSource1 != zex_mcl_alu_reg_t::ZE_MCL_ALU_REG_NONE) {
switch (exAluArray->regSource1) {
case zex_mcl_alu_reg_t::ZE_MCL_ALU_REG_CONST0:
aluParam.DW0.BitField.ALUOpcode = static_cast<uint32_t>(zex_mcl_alu_op_type_t::ZE_MCL_ALU_OP_LOAD0);
break;
case zex_mcl_alu_reg_t::ZE_MCL_ALU_REG_CONST1:
aluParam.DW0.BitField.ALUOpcode = static_cast<uint32_t>(zex_mcl_alu_op_type_t::ZE_MCL_ALU_OP_LOAD1);
break;
default:
aluParam.DW0.BitField.ALUOpcode = static_cast<uint32_t>(zex_mcl_alu_op_type_t::ZE_MCL_ALU_OP_LOAD);
break;
}
aluParam.DW0.BitField.Operand1 = aluRegisterRSrca;
aluParam.DW0.BitField.Operand2 = static_cast<uint32_t>(exAluArray->regSource1) / 2;
} else {
aluParam.DW0.BitField.ALUOpcode = static_cast<uint32_t>(zex_mcl_alu_op_type_t::ZE_MCL_ALU_OP_NOOP);
aluParam.DW0.BitField.Operand1 = 0;
aluParam.DW0.BitField.Operand2 = 0;
}
*pAluParam = aluParam;
pAluParam++;
aluParam.DW0.Value = 0x0;
// Setup second operand of MI_MATH - load regSource2 into register B
if (exAluArray->regSource2 != zex_mcl_alu_reg_t::ZE_MCL_ALU_REG_NONE) {
switch (exAluArray->regSource2) {
case zex_mcl_alu_reg_t::ZE_MCL_ALU_REG_CONST0:
aluParam.DW0.BitField.ALUOpcode = static_cast<uint32_t>(zex_mcl_alu_op_type_t::ZE_MCL_ALU_OP_LOAD0);
break;
case zex_mcl_alu_reg_t::ZE_MCL_ALU_REG_CONST1:
aluParam.DW0.BitField.ALUOpcode = static_cast<uint32_t>(zex_mcl_alu_op_type_t::ZE_MCL_ALU_OP_LOAD1);
break;
default:
aluParam.DW0.BitField.ALUOpcode = static_cast<uint32_t>(zex_mcl_alu_op_type_t::ZE_MCL_ALU_OP_LOAD);
break;
}
aluParam.DW0.BitField.Operand1 = aluRegisterRSrcb;
aluParam.DW0.BitField.Operand2 = static_cast<uint32_t>(exAluArray->regSource2) / 2;
} else {
aluParam.DW0.BitField.ALUOpcode = static_cast<uint32_t>(zex_mcl_alu_op_type_t::ZE_MCL_ALU_OP_NOOP);
aluParam.DW0.BitField.Operand1 = 0;
aluParam.DW0.BitField.Operand2 = 0;
}
*pAluParam = aluParam;
pAluParam++;
// Setup third operand of MI_MATH - "Operation" on registers A and B
aluParam.DW0.Value = 0x0;
aluParam.DW0.BitField.ALUOpcode = static_cast<uint32_t>(exAluArray->opType);
aluParam.DW0.BitField.Operand1 = 0;
aluParam.DW0.BitField.Operand2 = 0;
*pAluParam = aluParam;
pAluParam++;
aluParam.DW0.Value = 0x0;
// Setup fourth operand of MI_MATH - store result into regDest
if (exAluArray->flag >= zex_mcl_alu_flag_t::ZE_MCL_ALU_FLAG_NACC) {
aluParam.DW0.BitField.ALUOpcode = static_cast<uint32_t>(zex_mcl_alu_op_type_t::ZE_MCL_ALU_OP_STOREINV);
aluParam.DW0.BitField.Operand2 = static_cast<uint32_t>(exAluArray->flag) - 0x100;
} else {
aluParam.DW0.BitField.ALUOpcode = static_cast<uint32_t>(zex_mcl_alu_op_type_t::ZE_MCL_ALU_OP_STORE);
aluParam.DW0.BitField.Operand2 = static_cast<uint32_t>(exAluArray->flag);
}
aluParam.DW0.BitField.Operand1 = static_cast<uint32_t>(exAluArray->regDest) / 2;
*pAluParam = aluParam;
pAluParam++;
}
return ZE_RESULT_SUCCESS;
}
template <GFXCORE_FAMILY gfxCoreFamily>
ze_result_t MutableCommandListCoreFamily<gfxCoreFamily>::appendJump(Label *label, const InterfaceOperandDescriptor *condition) {
auto cs = this->commandContainer.getCommandStream();
@@ -60,8 +201,9 @@ ze_result_t MutableCommandListCoreFamily<gfxCoreFamily>::appendJump(Label *label
if (condition->memory != nullptr) {
if (condition->flags & InterfaceOperandDescriptor::Flags::usesVariable) {
auto variable = reinterpret_cast<Variable *>(condition->memory);
if (false == variable->isType(VariableType::buffer))
if (false == variable->isType(VariableType::buffer)) {
return ZE_RESULT_ERROR_INVALID_ARGUMENT;
}
appendMILoadRegVariable(MclAluReg::mclAluRegPredicate2, variable);
} else {
@@ -69,7 +211,7 @@ ze_result_t MutableCommandListCoreFamily<gfxCoreFamily>::appendJump(Label *label
NEO::EncodeSetMMIO<GfxFamily>::encodeMEM(*cs, regToMMIO(MclAluReg::mclAluRegPredicate2), memAddr, getBase()->isCopyOnly(false));
}
} else {
auto regMMIO = static_cast<uint32_t>(condition->offset);
auto regMMIO = regToMMIO(static_cast<MclAluReg>(condition->offset));
NEO::EncodeSetMMIO<GfxFamily>::encodeREG(*cs, regToMMIO(MclAluReg::mclAluRegPredicate2), regMMIO, getBase()->isCopyOnly(false));
}