1/*
2 * Copyright (C) 2021 Intel Corporation
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 */
7
8#include "opencl/source/command_queue/gpgpu_walker.h"
9
10namespace NEO {
11template <typename GfxFamily>
12void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
13    LinearStream &commandStream,
14    DeviceQueueHw<GfxFamily> &devQueueHw,
15    PreemptionMode preemptionMode,
16    SchedulerKernel &scheduler,
17    IndirectHeap *ssh,
18    IndirectHeap *dsh,
19    bool isCcsUsed) {
20
21    const auto &kernelInfo = scheduler.getKernelInfo();
22
23    using INTERFACE_DESCRIPTOR_DATA = typename GfxFamily::INTERFACE_DESCRIPTOR_DATA;
24    using GPGPU_WALKER = typename GfxFamily::GPGPU_WALKER;
25    using MI_BATCH_BUFFER_START = typename GfxFamily::MI_BATCH_BUFFER_START;
26
27    const auto &hwInfo = devQueueHw.getDevice().getHardwareInfo();
28    NEO::PipeControlArgs args;
29    MemorySynchronizationCommands<GfxFamily>::addPipeControl(commandStream, args);
30
31    uint32_t interfaceDescriptorIndex = devQueueHw.schedulerIDIndex;
32    const size_t offsetInterfaceDescriptorTable = devQueueHw.colorCalcStateSize;
33    const size_t offsetInterfaceDescriptor = offsetInterfaceDescriptorTable;
34    const size_t totalInterfaceDescriptorTableSize = devQueueHw.interfaceDescriptorEntries * sizeof(INTERFACE_DESCRIPTOR_DATA);
35
36    // Program media interface descriptor load
37    HardwareCommandsHelper<GfxFamily>::sendMediaInterfaceDescriptorLoad(
38        commandStream,
39        offsetInterfaceDescriptor,
40        totalInterfaceDescriptorTableSize);
41
42    DEBUG_BREAK_IF(offsetInterfaceDescriptorTable % 64 != 0);
43
44    // Determine SIMD size
45    uint32_t simd = kernelInfo.getMaxSimdSize();
46    DEBUG_BREAK_IF(simd != PARALLEL_SCHEDULER_COMPILATION_SIZE_20);
47
48    // Patch our kernel constants
49    scheduler.setGlobalWorkOffsetValues(0, 0, 0);
50    scheduler.setGlobalWorkSizeValues(static_cast<uint32_t>(scheduler.getGws()), 1, 1);
51    scheduler.setLocalWorkSizeValues(static_cast<uint32_t>(scheduler.getLws()), 1, 1);
52    scheduler.setLocalWorkSize2Values(static_cast<uint32_t>(scheduler.getLws()), 1, 1);
53    scheduler.setEnqueuedLocalWorkSizeValues(static_cast<uint32_t>(scheduler.getLws()), 1, 1);
54    scheduler.setNumWorkGroupsValues(static_cast<uint32_t>(scheduler.getGws() / scheduler.getLws()), 0, 0);
55    scheduler.setWorkDim(1);
56
57    // Send our indirect object data
58    size_t localWorkSizes[3] = {scheduler.getLws(), 1, 1};
59
60    // Create indirectHeap for IOH that is located at the end of device enqueue DSH
61    size_t curbeOffset = devQueueHw.setSchedulerCrossThreadData(scheduler);
62    IndirectHeap indirectObjectHeap(dsh->getCpuBase(), dsh->getMaxAvailableSpace());
63    indirectObjectHeap.getSpace(curbeOffset);
64    IndirectHeap *ioh = &indirectObjectHeap;
65
66    // Program the walker.  Invokes execution so all state should already be programmed
67    auto pGpGpuWalkerCmd = commandStream.getSpaceForCmd<GPGPU_WALKER>();
68    GPGPU_WALKER cmdWalker = GfxFamily::cmdInitGpgpuWalker;
69
70    bool inlineDataProgrammingRequired = HardwareCommandsHelper<GfxFamily>::inlineDataProgrammingRequired(scheduler);
71    auto kernelUsesLocalIds = HardwareCommandsHelper<GfxFamily>::kernelUsesLocalIds(scheduler);
72
73    HardwareCommandsHelper<GfxFamily>::sendIndirectState(
74        commandStream,
75        *dsh,
76        *ioh,
77        *ssh,
78        scheduler,
79        scheduler.getKernelStartOffset(true, kernelUsesLocalIds, isCcsUsed),
80        simd,
81        localWorkSizes,
82        offsetInterfaceDescriptorTable,
83        interfaceDescriptorIndex,
84        preemptionMode,
85        &cmdWalker,
86        nullptr,
87        true,
88        devQueueHw.getDevice());
89
90    // Implement enabling special WA DisableLSQCROPERFforOCL if needed
91    GpgpuWalkerHelper<GfxFamily>::applyWADisableLSQCROPERFforOCL(&commandStream, scheduler, true);
92
93    size_t globalOffsets[3] = {0, 0, 0};
94    size_t workGroups[3] = {(scheduler.getGws() / scheduler.getLws()), 1, 1};
95    GpgpuWalkerHelper<GfxFamily>::setGpgpuWalkerThreadData(&cmdWalker, kernelInfo.kernelDescriptor, globalOffsets, globalOffsets, workGroups, localWorkSizes,
96                                                           simd, 1, true, inlineDataProgrammingRequired, 0u);
97    *pGpGpuWalkerCmd = cmdWalker;
98
99    // Implement disabling special WA DisableLSQCROPERFforOCL if needed
100    GpgpuWalkerHelper<GfxFamily>::applyWADisableLSQCROPERFforOCL(&commandStream, scheduler, false);
101
102    // Do not put BB_START only when returning in first Scheduler run
103    if (devQueueHw.getSchedulerReturnInstance() != 1) {
104        args.dcFlushEnable = MemorySynchronizationCommands<GfxFamily>::isDcFlushAllowed(true, hwInfo);
105        MemorySynchronizationCommands<GfxFamily>::addPipeControl(commandStream, args);
106
107        // Add BB Start Cmd to the SLB in the Primary Batch Buffer
108        auto bbStart = commandStream.getSpaceForCmd<MI_BATCH_BUFFER_START>();
109        MI_BATCH_BUFFER_START cmdBbStart = GfxFamily::cmdInitBatchBufferStart;
110        cmdBbStart.setSecondLevelBatchBuffer(MI_BATCH_BUFFER_START::SECOND_LEVEL_BATCH_BUFFER_FIRST_LEVEL_BATCH);
111        uint64_t slbAddress = devQueueHw.getSlbBuffer()->getGpuAddress();
112        cmdBbStart.setBatchBufferStartAddress(slbAddress);
113        *bbStart = cmdBbStart;
114    }
115}
116} // namespace NEO