mirror of
https://github.com/intel/compute-runtime.git
synced 2025-06-28 17:58:30 +08:00
54b2763466
General purpose register cannot be used for MI_MATH calculations. ALU registers must be used. To prevent passing general purpose register into the EncodeMath interface, enforce a ALU register type at compile time. Change-Id: I98aa8605cde27e7003029d33b3ef3bcfb2306878 Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
62 lines
2.5 KiB
C++
62 lines
2.5 KiB
C++
/*
|
|
* Copyright (C) 2018-2020 Intel Corporation
|
|
*
|
|
* SPDX-License-Identifier: MIT
|
|
*
|
|
*/
|
|
|
|
#include "shared/source/gen9/hw_cmds_base.h"
|
|
|
|
#include "opencl/source/command_queue/gpgpu_walker_bdw_plus.inl"
|
|
#include "opencl/source/command_queue/hardware_interface_bdw_plus.inl"
|
|
|
|
namespace NEO {
|
|
|
|
template <>
|
|
void GpgpuWalkerHelper<SKLFamily>::applyWADisableLSQCROPERFforOCL(NEO::LinearStream *pCommandStream, const Kernel &kernel, bool disablePerfMode) {
|
|
if (disablePerfMode) {
|
|
if (kernel.getKernelInfo().patchInfo.executionEnvironment->UsesFencesForReadWriteImages) {
|
|
// Set bit L3SQC_BIT_LQSC_RO_PERF_DIS in L3SQC_REG4
|
|
GpgpuWalkerHelper<SKLFamily>::addAluReadModifyWriteRegister(pCommandStream, L3SQC_REG4, AluRegisters::OPCODE_OR, L3SQC_BIT_LQSC_RO_PERF_DIS);
|
|
}
|
|
} else {
|
|
if (kernel.getKernelInfo().patchInfo.executionEnvironment->UsesFencesForReadWriteImages) {
|
|
// Add PIPE_CONTROL with CS_Stall to wait till GPU finishes its work
|
|
typedef typename SKLFamily::PIPE_CONTROL PIPE_CONTROL;
|
|
auto pCmd = reinterpret_cast<PIPE_CONTROL *>(pCommandStream->getSpace(sizeof(PIPE_CONTROL)));
|
|
*pCmd = SKLFamily::cmdInitPipeControl;
|
|
pCmd->setCommandStreamerStallEnable(true);
|
|
// Clear bit L3SQC_BIT_LQSC_RO_PERF_DIS in L3SQC_REG4
|
|
GpgpuWalkerHelper<SKLFamily>::addAluReadModifyWriteRegister(pCommandStream, L3SQC_REG4, AluRegisters::OPCODE_AND, ~L3SQC_BIT_LQSC_RO_PERF_DIS);
|
|
}
|
|
}
|
|
}
|
|
|
|
template <>
|
|
size_t GpgpuWalkerHelper<SKLFamily>::getSizeForWADisableLSQCROPERFforOCL(const Kernel *pKernel) {
|
|
typedef typename SKLFamily::MI_LOAD_REGISTER_REG MI_LOAD_REGISTER_REG;
|
|
typedef typename SKLFamily::MI_LOAD_REGISTER_IMM MI_LOAD_REGISTER_IMM;
|
|
typedef typename SKLFamily::PIPE_CONTROL PIPE_CONTROL;
|
|
typedef typename SKLFamily::MI_MATH MI_MATH;
|
|
typedef typename SKLFamily::MI_MATH_ALU_INST_INLINE MI_MATH_ALU_INST_INLINE;
|
|
size_t n = 0;
|
|
if (pKernel->getKernelInfo().patchInfo.executionEnvironment->UsesFencesForReadWriteImages) {
|
|
n += sizeof(PIPE_CONTROL) +
|
|
(2 * sizeof(MI_LOAD_REGISTER_REG) +
|
|
sizeof(MI_LOAD_REGISTER_IMM) +
|
|
sizeof(PIPE_CONTROL) +
|
|
sizeof(MI_MATH) +
|
|
NUM_ALU_INST_FOR_READ_MODIFY_WRITE * sizeof(MI_MATH_ALU_INST_INLINE)) *
|
|
2; // For 2 WADisableLSQCROPERFforOCL WAs
|
|
}
|
|
return n;
|
|
}
|
|
|
|
template class HardwareInterface<SKLFamily>;
|
|
|
|
template class GpgpuWalkerHelper<SKLFamily>;
|
|
|
|
template struct EnqueueOperation<SKLFamily>;
|
|
|
|
} // namespace NEO
|