compute_unit.cc revision 11657:5fad5a37d6fc 
1/*
2 * Copyright (c) 2011-2015 Advanced Micro Devices, Inc.
3 * All rights reserved.
4 *
5 * For use for simulation and test purposes only
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions are met:
9 *
10 * 1. Redistributions of source code must retain the above copyright notice,
11 * this list of conditions and the following disclaimer.
12 *
13 * 2. Redistributions in binary form must reproduce the above copyright notice,
14 * this list of conditions and the following disclaimer in the documentation
15 * and/or other materials provided with the distribution.
16 *
17 * 3. Neither the name of the copyright holder nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
22 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
25 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
26 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
27 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
28 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
29 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
30 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
31 * POSSIBILITY OF SUCH DAMAGE.
32 *
33 * Author: John Kalamatianos, Anthony Gutierrez
34 */
35#include "gpu-compute/compute_unit.hh"
36
37#include <limits>
38
39#include "base/output.hh"
40#include "debug/GPUDisp.hh"
41#include "debug/GPUExec.hh"
42#include "debug/GPUFetch.hh"
43#include "debug/GPUMem.hh"
44#include "debug/GPUPort.hh"
45#include "debug/GPUPrefetch.hh"
46#include "debug/GPUSync.hh"
47#include "debug/GPUTLB.hh"
48#include "gpu-compute/dispatcher.hh"
49#include "gpu-compute/gpu_dyn_inst.hh"
50#include "gpu-compute/gpu_static_inst.hh"
51#include "gpu-compute/ndrange.hh"
52#include "gpu-compute/shader.hh"
53#include "gpu-compute/simple_pool_manager.hh"
54#include "gpu-compute/vector_register_file.hh"
55#include "gpu-compute/wavefront.hh"
56#include "mem/page_table.hh"
57#include "sim/process.hh"
58
59ComputeUnit::ComputeUnit(const Params *p) : MemObject(p), fetchStage(p),
60    scoreboardCheckStage(p), scheduleStage(p), execStage(p),
61    globalMemoryPipe(p), localMemoryPipe(p), rrNextMemID(0), rrNextALUWp(0),
62    cu_id(p->cu_id), vrf(p->vector_register_file), numSIMDs(p->num_SIMDs),
63    spBypassPipeLength(p->spbypass_pipe_length),
64    dpBypassPipeLength(p->dpbypass_pipe_length),
65    issuePeriod(p->issue_period),
66    numGlbMemUnits(p->num_global_mem_pipes),
67    numLocMemUnits(p->num_shared_mem_pipes),
68    perLaneTLB(p->perLaneTLB), prefetchDepth(p->prefetch_depth),
69    prefetchStride(p->prefetch_stride), prefetchType(p->prefetch_prev_type),
70    xact_cas_mode(p->xactCasMode), debugSegFault(p->debugSegFault),
71    functionalTLB(p->functionalTLB), localMemBarrier(p->localMemBarrier),
72    countPages(p->countPages), barrier_id(0),
73    vrfToCoalescerBusWidth(p->vrf_to_coalescer_bus_width),
74    coalescerToVrfBusWidth(p->coalescer_to_vrf_bus_width),
75    req_tick_latency(p->mem_req_latency * p->clk_domain->clockPeriod()),
76    resp_tick_latency(p->mem_resp_latency * p->clk_domain->clockPeriod()),
77    _masterId(p->system->getMasterId(name() + ".ComputeUnit")),
78    lds(*p->localDataStore), globalSeqNum(0),  wavefrontSize(p->wfSize)
79{
80    /**
81     * This check is necessary because std::bitset only provides conversion
82     * to unsigned long or unsigned long long via to_ulong() or to_ullong().
83     * there are * a few places in the code where to_ullong() is used, however
84     * if VSZ is larger than a value the host can support then bitset will
85     * throw a runtime exception. we should remove all use of to_long() or
86     * to_ullong() so we can have VSZ greater than 64b, however until that is
87     * done this assert is required.
88     */
89    fatal_if(p->wfSize > std::numeric_limits<unsigned long long>::digits ||
90             p->wfSize <= 0,
91             "WF size is larger than the host can support");
92    fatal_if(!isPowerOf2(wavefrontSize),
93             "Wavefront size should be a power of 2");
94    // calculate how many cycles a vector load or store will need to transfer
95    // its data over the corresponding buses
96    numCyclesPerStoreTransfer =
97        (uint32_t)ceil((double)(wfSize() * sizeof(uint32_t)) /
98                (double)vrfToCoalescerBusWidth);
99
100    numCyclesPerLoadTransfer = (wfSize() * sizeof(uint32_t))
101                               / coalescerToVrfBusWidth;
102
103    lastVaddrWF.resize(numSIMDs);
104    wfList.resize(numSIMDs);
105
106    for (int j = 0; j < numSIMDs; ++j) {
107        lastVaddrWF[j].resize(p->n_wf);
108
109        for (int i = 0; i < p->n_wf; ++i) {
110            lastVaddrWF[j][i].resize(wfSize());
111
112            wfList[j].push_back(p->wavefronts[j * p->n_wf + i]);
113            wfList[j][i]->setParent(this);
114
115            for (int k = 0; k < wfSize(); ++k) {
116                lastVaddrWF[j][i][k] = 0;
117            }
118        }
119    }
120
121    lastVaddrSimd.resize(numSIMDs);
122
123    for (int i = 0; i < numSIMDs; ++i) {
124        lastVaddrSimd[i].resize(wfSize(), 0);
125    }
126
127    lastVaddrCU.resize(wfSize());
128
129    lds.setParent(this);
130
131    if (p->execPolicy == "OLDEST-FIRST") {
132        exec_policy = EXEC_POLICY::OLDEST;
133    } else if (p->execPolicy == "ROUND-ROBIN") {
134        exec_policy = EXEC_POLICY::RR;
135    } else {
136        fatal("Invalid WF execution policy (CU)\n");
137    }
138
139    memPort.resize(wfSize());
140
141    // resize the tlbPort vectorArray
142    int tlbPort_width = perLaneTLB ? wfSize() : 1;
143    tlbPort.resize(tlbPort_width);
144
145    cuExitCallback = new CUExitCallback(this);
146    registerExitCallback(cuExitCallback);
147
148    xactCasLoadMap.clear();
149    lastExecCycle.resize(numSIMDs, 0);
150
151    for (int i = 0; i < vrf.size(); ++i) {
152        vrf[i]->setParent(this);
153    }
154
155    numVecRegsPerSimd = vrf[0]->numRegs();
156}
157
158ComputeUnit::~ComputeUnit()
159{
160    // Delete wavefront slots
161    for (int j = 0; j < numSIMDs; ++j) {
162        for (int i = 0; i < shader->n_wf; ++i) {
163            delete wfList[j][i];
164        }
165        lastVaddrSimd[j].clear();
166    }
167    lastVaddrCU.clear();
168    readyList.clear();
169    waveStatusList.clear();
170    dispatchList.clear();
171    vectorAluInstAvail.clear();
172    delete cuExitCallback;
173    delete ldsPort;
174}
175
176void
177ComputeUnit::fillKernelState(Wavefront *w, NDRange *ndr)
178{
179    w->resizeRegFiles(ndr->q.cRegCount, ndr->q.sRegCount, ndr->q.dRegCount);
180
181    w->workGroupSz[0] = ndr->q.wgSize[0];
182    w->workGroupSz[1] = ndr->q.wgSize[1];
183    w->workGroupSz[2] = ndr->q.wgSize[2];
184    w->wgSz = w->workGroupSz[0] * w->workGroupSz[1] * w->workGroupSz[2];
185    w->gridSz[0] = ndr->q.gdSize[0];
186    w->gridSz[1] = ndr->q.gdSize[1];
187    w->gridSz[2] = ndr->q.gdSize[2];
188    w->kernelArgs = ndr->q.args;
189    w->privSizePerItem = ndr->q.privMemPerItem;
190    w->spillSizePerItem = ndr->q.spillMemPerItem;
191    w->roBase = ndr->q.roMemStart;
192    w->roSize = ndr->q.roMemTotal;
193    w->computeActualWgSz(ndr);
194}
195
196void
197ComputeUnit::updateEvents() {
198
199    if (!timestampVec.empty()) {
200        uint32_t vecSize = timestampVec.size();
201        uint32_t i = 0;
202        while (i < vecSize) {
203            if (timestampVec[i] <= shader->tick_cnt) {
204                std::pair<uint32_t, uint32_t> regInfo = regIdxVec[i];
205                vrf[regInfo.first]->markReg(regInfo.second, sizeof(uint32_t),
206                                            statusVec[i]);
207                timestampVec.erase(timestampVec.begin() + i);
208                regIdxVec.erase(regIdxVec.begin() + i);
209                statusVec.erase(statusVec.begin() + i);
210                --vecSize;
211                --i;
212            }
213            ++i;
214        }
215    }
216
217    for (int i = 0; i< numSIMDs; ++i) {
218        vrf[i]->updateEvents();
219    }
220}
221
222
223void
224ComputeUnit::startWavefront(Wavefront *w, int waveId, LdsChunk *ldsChunk,
225                            NDRange *ndr)
226{
227    static int _n_wave = 0;
228
229    VectorMask init_mask;
230    init_mask.reset();
231
232    for (int k = 0; k < wfSize(); ++k) {
233        if (k + waveId * wfSize() < w->actualWgSzTotal)
234            init_mask[k] = 1;
235    }
236
237    w->kernId = ndr->dispatchId;
238    w->wfId = waveId;
239    w->initMask = init_mask.to_ullong();
240
241    for (int k = 0; k < wfSize(); ++k) {
242        w->workItemId[0][k] = (k + waveId * wfSize()) % w->actualWgSz[0];
243        w->workItemId[1][k] = ((k + waveId * wfSize()) / w->actualWgSz[0]) %
244                             w->actualWgSz[1];
245        w->workItemId[2][k] = (k + waveId * wfSize()) /
246                              (w->actualWgSz[0] * w->actualWgSz[1]);
247
248        w->workItemFlatId[k] = w->workItemId[2][k] * w->actualWgSz[0] *
249            w->actualWgSz[1] + w->workItemId[1][k] * w->actualWgSz[0] +
250            w->workItemId[0][k];
251    }
252
253    w->barrierSlots = divCeil(w->actualWgSzTotal, wfSize());
254
255    w->barCnt.resize(wfSize(), 0);
256
257    w->maxBarCnt = 0;
258    w->oldBarrierCnt = 0;
259    w->barrierCnt = 0;
260
261    w->privBase = ndr->q.privMemStart;
262    ndr->q.privMemStart += ndr->q.privMemPerItem * wfSize();
263
264    w->spillBase = ndr->q.spillMemStart;
265    ndr->q.spillMemStart += ndr->q.spillMemPerItem * wfSize();
266
267    w->pushToReconvergenceStack(0, UINT32_MAX, init_mask.to_ulong());
268
269    // WG state
270    w->wgId = ndr->globalWgId;
271    w->dispatchId = ndr->dispatchId;
272    w->workGroupId[0] = w->wgId % ndr->numWg[0];
273    w->workGroupId[1] = (w->wgId / ndr->numWg[0]) % ndr->numWg[1];
274    w->workGroupId[2] = w->wgId / (ndr->numWg[0] * ndr->numWg[1]);
275
276    w->barrierId = barrier_id;
277    w->stalledAtBarrier = false;
278
279    // set the wavefront context to have a pointer to this section of the LDS
280    w->ldsChunk = ldsChunk;
281
282    int32_t refCount M5_VAR_USED =
283                    lds.increaseRefCounter(w->dispatchId, w->wgId);
284    DPRINTF(GPUDisp, "CU%d: increase ref ctr wg[%d] to [%d]\n",
285                    cu_id, w->wgId, refCount);
286
287    w->instructionBuffer.clear();
288
289    if (w->pendingFetch)
290        w->dropFetch = true;
291
292    // is this the last wavefront in the workgroup
293    // if set the spillWidth to be the remaining work-items
294    // so that the vector access is correct
295    if ((waveId + 1) * wfSize() >= w->actualWgSzTotal) {
296        w->spillWidth = w->actualWgSzTotal - (waveId * wfSize());
297    } else {
298        w->spillWidth = wfSize();
299    }
300
301    DPRINTF(GPUDisp, "Scheduling wfDynId/barrier_id %d/%d on CU%d: "
302            "WF[%d][%d]\n", _n_wave, barrier_id, cu_id, w->simdId, w->wfSlotId);
303
304    w->start(++_n_wave, ndr->q.code_ptr);
305}
306
307void
308ComputeUnit::StartWorkgroup(NDRange *ndr)
309{
310    // reserve the LDS capacity allocated to the work group
311    // disambiguated by the dispatch ID and workgroup ID, which should be
312    // globally unique
313    LdsChunk *ldsChunk = lds.reserveSpace(ndr->dispatchId, ndr->globalWgId,
314                                          ndr->q.ldsSize);
315
316    // Send L1 cache acquire
317    // isKernel + isAcquire = Kernel Begin
318    if (shader->impl_kern_boundary_sync) {
319        GPUDynInstPtr gpuDynInst = std::make_shared<GPUDynInst>(this,
320                                                                nullptr,
321                                                                nullptr, 0);
322
323        gpuDynInst->useContinuation = false;
324        gpuDynInst->memoryOrder = Enums::MEMORY_ORDER_SC_ACQUIRE;
325        gpuDynInst->scope = Enums::MEMORY_SCOPE_SYSTEM;
326        injectGlobalMemFence(gpuDynInst, true);
327    }
328
329    // calculate the number of 32-bit vector registers required by wavefront
330    int vregDemand = ndr->q.sRegCount + (2 * ndr->q.dRegCount);
331    int wave_id = 0;
332
333    // Assign WFs by spreading them across SIMDs, 1 WF per SIMD at a time
334    for (int m = 0; m < shader->n_wf * numSIMDs; ++m) {
335        Wavefront *w = wfList[m % numSIMDs][m / numSIMDs];
336        // Check if this wavefront slot is available:
337        // It must be stopped and not waiting
338        // for a release to complete S_RETURNING
339        if (w->status == Wavefront::S_STOPPED) {
340            fillKernelState(w, ndr);
341            // if we have scheduled all work items then stop
342            // scheduling wavefronts
343            if (wave_id * wfSize() >= w->actualWgSzTotal)
344                break;
345
346            // reserve vector registers for the scheduled wavefront
347            assert(vectorRegsReserved[m % numSIMDs] <= numVecRegsPerSimd);
348            uint32_t normSize = 0;
349
350            w->startVgprIndex = vrf[m % numSIMDs]->manager->
351                                    allocateRegion(vregDemand, &normSize);
352
353            w->reservedVectorRegs = normSize;
354            vectorRegsReserved[m % numSIMDs] += w->reservedVectorRegs;
355
356            startWavefront(w, wave_id, ldsChunk, ndr);
357            ++wave_id;
358        }
359    }
360    ++barrier_id;
361}
362
363int
364ComputeUnit::ReadyWorkgroup(NDRange *ndr)
365{
366    // Get true size of workgroup (after clamping to grid size)
367    int trueWgSize[3];
368    int trueWgSizeTotal = 1;
369
370    for (int d = 0; d < 3; ++d) {
371        trueWgSize[d] = std::min(ndr->q.wgSize[d], ndr->q.gdSize[d] -
372                                 ndr->wgId[d] * ndr->q.wgSize[d]);
373
374        trueWgSizeTotal *= trueWgSize[d];
375        DPRINTF(GPUDisp, "trueWgSize[%d] =  %d\n", d, trueWgSize[d]);
376    }
377
378    DPRINTF(GPUDisp, "trueWgSizeTotal =  %d\n", trueWgSizeTotal);
379
380    // calculate the number of 32-bit vector registers required by each
381    // work item of the work group
382    int vregDemandPerWI = ndr->q.sRegCount + (2 * ndr->q.dRegCount);
383    bool vregAvail = true;
384    int numWfs = (trueWgSizeTotal + wfSize() - 1) / wfSize();
385    int freeWfSlots = 0;
386    // check if the total number of VGPRs required by all WFs of the WG
387    // fit in the VRFs of all SIMD units
388    assert((numWfs * vregDemandPerWI) <= (numSIMDs * numVecRegsPerSimd));
389    int numMappedWfs = 0;
390    std::vector<int> numWfsPerSimd;
391    numWfsPerSimd.resize(numSIMDs, 0);
392    // find how many free WF slots we have across all SIMDs
393    for (int j = 0; j < shader->n_wf; ++j) {
394        for (int i = 0; i < numSIMDs; ++i) {
395            if (wfList[i][j]->status == Wavefront::S_STOPPED) {
396                // count the number of free WF slots
397                ++freeWfSlots;
398                if (numMappedWfs < numWfs) {
399                    // count the WFs to be assigned per SIMD
400                    numWfsPerSimd[i]++;
401                }
402                numMappedWfs++;
403            }
404        }
405    }
406
407    // if there are enough free WF slots then find if there are enough
408    // free VGPRs per SIMD based on the WF->SIMD mapping
409    if (freeWfSlots >= numWfs) {
410        for (int j = 0; j < numSIMDs; ++j) {
411            // find if there are enough free VGPR regions in the SIMD's VRF
412            // to accommodate the WFs of the new WG that would be mapped to
413            // this SIMD unit
414            vregAvail = vrf[j]->manager->canAllocate(numWfsPerSimd[j],
415                                                     vregDemandPerWI);
416
417            // stop searching if there is at least one SIMD
418            // whose VRF does not have enough free VGPR pools.
419            // This is because a WG is scheduled only if ALL
420            // of its WFs can be scheduled
421            if (!vregAvail)
422                break;
423        }
424    }
425
426    DPRINTF(GPUDisp, "Free WF slots =  %d, VGPR Availability = %d\n",
427            freeWfSlots, vregAvail);
428
429    if (!vregAvail) {
430        ++numTimesWgBlockedDueVgprAlloc;
431    }
432
433    // Return true if enough WF slots to submit workgroup and if there are
434    // enough VGPRs to schedule all WFs to their SIMD units
435    if (!lds.canReserve(ndr->q.ldsSize)) {
436        wgBlockedDueLdsAllocation++;
437    }
438
439    // Return true if (a) there are enough free WF slots to submit
440    // workgrounp and (b) if there are enough VGPRs to schedule all WFs to their
441    // SIMD units and (c) if there is enough space in LDS
442    return freeWfSlots >= numWfs && vregAvail && lds.canReserve(ndr->q.ldsSize);
443}
444
445int
446ComputeUnit::AllAtBarrier(uint32_t _barrier_id, uint32_t bcnt, uint32_t bslots)
447{
448    DPRINTF(GPUSync, "CU%d: Checking for All At Barrier\n", cu_id);
449    int ccnt = 0;
450
451    for (int i_simd = 0; i_simd < numSIMDs; ++i_simd) {
452        for (int i_wf = 0; i_wf < shader->n_wf; ++i_wf) {
453            Wavefront *w = wfList[i_simd][i_wf];
454
455            if (w->status == Wavefront::S_RUNNING) {
456                DPRINTF(GPUSync, "Checking WF[%d][%d]\n", i_simd, i_wf);
457
458                DPRINTF(GPUSync, "wf->barrier_id = %d, _barrier_id = %d\n",
459                        w->barrierId, _barrier_id);
460
461                DPRINTF(GPUSync, "wf->barrier_cnt %d, bcnt = %d\n",
462                        w->barrierCnt, bcnt);
463            }
464
465            if (w->status == Wavefront::S_RUNNING &&
466                w->barrierId == _barrier_id && w->barrierCnt == bcnt &&
467                !w->outstandingReqs) {
468                ++ccnt;
469
470                DPRINTF(GPUSync, "WF[%d][%d] at barrier, increment ccnt to "
471                        "%d\n", i_simd, i_wf, ccnt);
472            }
473        }
474    }
475
476    DPRINTF(GPUSync, "CU%d: returning allAtBarrier ccnt = %d, bslots = %d\n",
477            cu_id, ccnt, bslots);
478
479    return ccnt == bslots;
480}
481
482//  Check if the current wavefront is blocked on additional resources.
483bool
484ComputeUnit::cedeSIMD(int simdId, int wfSlotId)
485{
486    bool cede = false;
487
488    // If --xact-cas-mode option is enabled in run.py, then xact_cas_ld
489    // magic instructions will impact the scheduling of wavefronts
490    if (xact_cas_mode) {
491        /*
492         * When a wavefront calls xact_cas_ld, it adds itself to a per address
493         * queue. All per address queues are managed by the xactCasLoadMap.
494         *
495         * A wavefront is not blocked if: it is not in ANY per address queue or
496         * if it is at the head of a per address queue.
497         */
498        for (auto itMap : xactCasLoadMap) {
499            std::list<waveIdentifier> curWaveIDQueue = itMap.second.waveIDQueue;
500
501            if (!curWaveIDQueue.empty()) {
502                for (auto it : curWaveIDQueue) {
503                    waveIdentifier cur_wave = it;
504
505                    if (cur_wave.simdId == simdId &&
506                        cur_wave.wfSlotId == wfSlotId) {
507                        // 2 possibilities
508                        // 1: this WF has a green light
509                        // 2: another WF has a green light
510                        waveIdentifier owner_wave = curWaveIDQueue.front();
511
512                        if (owner_wave.simdId != cur_wave.simdId ||
513                            owner_wave.wfSlotId != cur_wave.wfSlotId) {
514                            // possibility 2
515                            cede = true;
516                            break;
517                        } else {
518                            // possibility 1
519                            break;
520                        }
521                    }
522                }
523            }
524        }
525    }
526
527    return cede;
528}
529
530// Execute one clock worth of work on the ComputeUnit.
531void
532ComputeUnit::exec()
533{
534    updateEvents();
535    // Execute pipeline stages in reverse order to simulate
536    // the pipeline latency
537    globalMemoryPipe.exec();
538    localMemoryPipe.exec();
539    execStage.exec();
540    scheduleStage.exec();
541    scoreboardCheckStage.exec();
542    fetchStage.exec();
543
544    totalCycles++;
545}
546
547void
548ComputeUnit::init()
549{
550    // Initialize CU Bus models
551    glbMemToVrfBus.init(&shader->tick_cnt, shader->ticks(1));
552    locMemToVrfBus.init(&shader->tick_cnt, shader->ticks(1));
553    nextGlbMemBus = 0;
554    nextLocMemBus = 0;
555    fatal_if(numGlbMemUnits > 1,
556             "No support for multiple Global Memory Pipelines exists!!!");
557    vrfToGlobalMemPipeBus.resize(numGlbMemUnits);
558    for (int j = 0; j < numGlbMemUnits; ++j) {
559        vrfToGlobalMemPipeBus[j] = WaitClass();
560        vrfToGlobalMemPipeBus[j].init(&shader->tick_cnt, shader->ticks(1));
561    }
562
563    fatal_if(numLocMemUnits > 1,
564             "No support for multiple Local Memory Pipelines exists!!!");
565    vrfToLocalMemPipeBus.resize(numLocMemUnits);
566    for (int j = 0; j < numLocMemUnits; ++j) {
567        vrfToLocalMemPipeBus[j] = WaitClass();
568        vrfToLocalMemPipeBus[j].init(&shader->tick_cnt, shader->ticks(1));
569    }
570    vectorRegsReserved.resize(numSIMDs, 0);
571    aluPipe.resize(numSIMDs);
572    wfWait.resize(numSIMDs + numLocMemUnits + numGlbMemUnits);
573
574    for (int i = 0; i < numSIMDs + numLocMemUnits + numGlbMemUnits; ++i) {
575        wfWait[i] = WaitClass();
576        wfWait[i].init(&shader->tick_cnt, shader->ticks(1));
577    }
578
579    for (int i = 0; i < numSIMDs; ++i) {
580        aluPipe[i] = WaitClass();
581        aluPipe[i].init(&shader->tick_cnt, shader->ticks(1));
582    }
583
584    // Setup space for call args
585    for (int j = 0; j < numSIMDs; ++j) {
586        for (int i = 0; i < shader->n_wf; ++i) {
587            wfList[j][i]->initCallArgMem(shader->funcargs_size, wavefrontSize);
588        }
589    }
590
591    // Initializing pipeline resources
592    readyList.resize(numSIMDs + numGlbMemUnits + numLocMemUnits);
593    waveStatusList.resize(numSIMDs);
594
595    for (int j = 0; j < numSIMDs; ++j) {
596        for (int i = 0; i < shader->n_wf; ++i) {
597            waveStatusList[j].push_back(
598                std::make_pair(wfList[j][i], BLOCKED));
599        }
600    }
601
602    for (int j = 0; j < (numSIMDs + numGlbMemUnits + numLocMemUnits); ++j) {
603        dispatchList.push_back(std::make_pair((Wavefront*)nullptr, EMPTY));
604    }
605
606    fetchStage.init(this);
607    scoreboardCheckStage.init(this);
608    scheduleStage.init(this);
609    execStage.init(this);
610    globalMemoryPipe.init(this);
611    localMemoryPipe.init(this);
612    // initialize state for statistics calculation
613    vectorAluInstAvail.resize(numSIMDs, false);
614    shrMemInstAvail = 0;
615    glbMemInstAvail = 0;
616}
617
618bool
619ComputeUnit::DataPort::recvTimingResp(PacketPtr pkt)
620{
621    // Ruby has completed the memory op. Schedule the mem_resp_event at the
622    // appropriate cycle to process the timing memory response
623    // This delay represents the pipeline delay
624    SenderState *sender_state = safe_cast<SenderState*>(pkt->senderState);
625    int index = sender_state->port_index;
626    GPUDynInstPtr gpuDynInst = sender_state->_gpuDynInst;
627
628    // Is the packet returned a Kernel End or Barrier
629    if (pkt->req->isKernel() && pkt->req->isRelease()) {
630        Wavefront *w =
631            computeUnit->wfList[gpuDynInst->simdId][gpuDynInst->wfSlotId];
632
633        // Check if we are waiting on Kernel End Release
634        if (w->status == Wavefront::S_RETURNING) {
635            DPRINTF(GPUDisp, "CU%d: WF[%d][%d][wv=%d]: WG id completed %d\n",
636                    computeUnit->cu_id, w->simdId, w->wfSlotId,
637                    w->wfDynId, w->kernId);
638
639            computeUnit->shader->dispatcher->notifyWgCompl(w);
640            w->status = Wavefront::S_STOPPED;
641        } else {
642            w->outstandingReqs--;
643        }
644
645        DPRINTF(GPUSync, "CU%d: WF[%d][%d]: barrier_cnt = %d\n",
646                computeUnit->cu_id, gpuDynInst->simdId,
647                gpuDynInst->wfSlotId, w->barrierCnt);
648
649        if (gpuDynInst->useContinuation) {
650            assert(gpuDynInst->scope != Enums::MEMORY_SCOPE_NONE);
651            gpuDynInst->execContinuation(gpuDynInst->staticInstruction(),
652                                           gpuDynInst);
653        }
654
655        delete pkt->senderState;
656        delete pkt->req;
657        delete pkt;
658        return true;
659    } else if (pkt->req->isKernel() && pkt->req->isAcquire()) {
660        if (gpuDynInst->useContinuation) {
661            assert(gpuDynInst->scope != Enums::MEMORY_SCOPE_NONE);
662            gpuDynInst->execContinuation(gpuDynInst->staticInstruction(),
663                                           gpuDynInst);
664        }
665
666        delete pkt->senderState;
667        delete pkt->req;
668        delete pkt;
669        return true;
670    }
671
672    ComputeUnit::DataPort::MemRespEvent *mem_resp_event =
673        new ComputeUnit::DataPort::MemRespEvent(computeUnit->memPort[index],
674                                                pkt);
675
676    DPRINTF(GPUPort, "CU%d: WF[%d][%d]: index %d, addr %#x received!\n",
677            computeUnit->cu_id, gpuDynInst->simdId, gpuDynInst->wfSlotId,
678            index, pkt->req->getPaddr());
679
680    computeUnit->schedule(mem_resp_event,
681                          curTick() + computeUnit->resp_tick_latency);
682    return true;
683}
684
685void
686ComputeUnit::DataPort::recvReqRetry()
687{
688    int len = retries.size();
689
690    assert(len > 0);
691
692    for (int i = 0; i < len; ++i) {
693        PacketPtr pkt = retries.front().first;
694        GPUDynInstPtr gpuDynInst M5_VAR_USED = retries.front().second;
695        DPRINTF(GPUMem, "CU%d: WF[%d][%d]: retry mem inst addr %#x\n",
696                computeUnit->cu_id, gpuDynInst->simdId, gpuDynInst->wfSlotId,
697                pkt->req->getPaddr());
698
699        /** Currently Ruby can return false due to conflicts for the particular
700         *  cache block or address.  Thus other requests should be allowed to
701         *  pass and the data port should expect multiple retries. */
702        if (!sendTimingReq(pkt)) {
703            DPRINTF(GPUMem, "failed again!\n");
704            break;
705        } else {
706            DPRINTF(GPUMem, "successful!\n");
707            retries.pop_front();
708        }
709    }
710}
711
712bool
713ComputeUnit::SQCPort::recvTimingResp(PacketPtr pkt)
714{
715    computeUnit->fetchStage.processFetchReturn(pkt);
716
717    return true;
718}
719
720void
721ComputeUnit::SQCPort::recvReqRetry()
722{
723    int len = retries.size();
724
725    assert(len > 0);
726
727    for (int i = 0; i < len; ++i) {
728        PacketPtr pkt = retries.front().first;
729        Wavefront *wavefront M5_VAR_USED = retries.front().second;
730        DPRINTF(GPUFetch, "CU%d: WF[%d][%d]: retrying FETCH addr %#x\n",
731                computeUnit->cu_id, wavefront->simdId, wavefront->wfSlotId,
732                pkt->req->getPaddr());
733        if (!sendTimingReq(pkt)) {
734            DPRINTF(GPUFetch, "failed again!\n");
735            break;
736        } else {
737            DPRINTF(GPUFetch, "successful!\n");
738            retries.pop_front();
739        }
740    }
741}
742
743void
744ComputeUnit::sendRequest(GPUDynInstPtr gpuDynInst, int index, PacketPtr pkt)
745{
746    // There must be a way around this check to do the globalMemStart...
747    Addr tmp_vaddr = pkt->req->getVaddr();
748
749    updatePageDivergenceDist(tmp_vaddr);
750
751    pkt->req->setVirt(pkt->req->getAsid(), tmp_vaddr, pkt->req->getSize(),
752                      pkt->req->getFlags(), pkt->req->masterId(),
753                      pkt->req->getPC());
754
755    // figure out the type of the request to set read/write
756    BaseTLB::Mode TLB_mode;
757    assert(pkt->isRead() || pkt->isWrite());
758
759    // Check write before read for atomic operations
760    // since atomic operations should use BaseTLB::Write
761    if (pkt->isWrite()){
762        TLB_mode = BaseTLB::Write;
763    } else if (pkt->isRead()) {
764        TLB_mode = BaseTLB::Read;
765    } else {
766        fatal("pkt is not a read nor a write\n");
767    }
768
769    tlbCycles -= curTick();
770    ++tlbRequests;
771
772    int tlbPort_index = perLaneTLB ? index : 0;
773
774    if (shader->timingSim) {
775        if (debugSegFault) {
776            Process *p = shader->gpuTc->getProcessPtr();
777            Addr vaddr = pkt->req->getVaddr();
778            unsigned size = pkt->getSize();
779
780            if ((vaddr + size - 1) % 64 < vaddr % 64) {
781                panic("CU%d: WF[%d][%d]: Access to addr %#x is unaligned!\n",
782                      cu_id, gpuDynInst->simdId, gpuDynInst->wfSlotId, vaddr);
783            }
784
785            Addr paddr;
786
787            if (!p->pTable->translate(vaddr, paddr)) {
788                if (!p->fixupStackFault(vaddr)) {
789                    panic("CU%d: WF[%d][%d]: Fault on addr %#x!\n",
790                          cu_id, gpuDynInst->simdId, gpuDynInst->wfSlotId,
791                          vaddr);
792                }
793            }
794        }
795
796        // This is the SenderState needed upon return
797        pkt->senderState = new DTLBPort::SenderState(gpuDynInst, index);
798
799        // This is the senderState needed by the TLB hierarchy to function
800        TheISA::GpuTLB::TranslationState *translation_state =
801          new TheISA::GpuTLB::TranslationState(TLB_mode, shader->gpuTc, false,
802                                               pkt->senderState);
803
804        pkt->senderState = translation_state;
805
806        if (functionalTLB) {
807            tlbPort[tlbPort_index]->sendFunctional(pkt);
808
809            // update the hitLevel distribution
810            int hit_level = translation_state->hitLevel;
811            assert(hit_level != -1);
812            hitsPerTLBLevel[hit_level]++;
813
814            // New SenderState for the memory access
815            X86ISA::GpuTLB::TranslationState *sender_state =
816                safe_cast<X86ISA::GpuTLB::TranslationState*>(pkt->senderState);
817
818            delete sender_state->tlbEntry;
819            delete sender_state->saved;
820            delete sender_state;
821
822            assert(pkt->req->hasPaddr());
823            assert(pkt->req->hasSize());
824
825            uint8_t *tmpData = pkt->getPtr<uint8_t>();
826
827            // this is necessary because the GPU TLB receives packets instead
828            // of requests. when the translation is complete, all relevent
829            // fields in the request will be populated, but not in the packet.
830            // here we create the new packet so we can set the size, addr,
831            // and proper flags.
832            PacketPtr oldPkt = pkt;
833            pkt = new Packet(oldPkt->req, oldPkt->cmd);
834            delete oldPkt;
835            pkt->dataStatic(tmpData);
836
837
838            // New SenderState for the memory access
839            pkt->senderState = new ComputeUnit::DataPort::SenderState(gpuDynInst,
840                                                             index, nullptr);
841
842            gpuDynInst->memStatusVector[pkt->getAddr()].push_back(index);
843            gpuDynInst->tlbHitLevel[index] = hit_level;
844
845
846            // translation is done. Schedule the mem_req_event at the
847            // appropriate cycle to send the timing memory request to ruby
848            ComputeUnit::DataPort::MemReqEvent *mem_req_event =
849                new ComputeUnit::DataPort::MemReqEvent(memPort[index], pkt);
850
851            DPRINTF(GPUPort, "CU%d: WF[%d][%d]: index %d, addr %#x data "
852                    "scheduled\n", cu_id, gpuDynInst->simdId,
853                    gpuDynInst->wfSlotId, index, pkt->req->getPaddr());
854
855            schedule(mem_req_event, curTick() + req_tick_latency);
856        } else if (tlbPort[tlbPort_index]->isStalled()) {
857            assert(tlbPort[tlbPort_index]->retries.size() > 0);
858
859            DPRINTF(GPUTLB, "CU%d: WF[%d][%d]: Translation for addr %#x "
860                    "failed!\n", cu_id, gpuDynInst->simdId, gpuDynInst->wfSlotId,
861                    tmp_vaddr);
862
863            tlbPort[tlbPort_index]->retries.push_back(pkt);
864        } else if (!tlbPort[tlbPort_index]->sendTimingReq(pkt)) {
865            // Stall the data port;
866            // No more packet will be issued till
867            // ruby indicates resources are freed by
868            // a recvReqRetry() call back on this port.
869            tlbPort[tlbPort_index]->stallPort();
870
871            DPRINTF(GPUTLB, "CU%d: WF[%d][%d]: Translation for addr %#x "
872                    "failed!\n", cu_id, gpuDynInst->simdId, gpuDynInst->wfSlotId,
873                    tmp_vaddr);
874
875            tlbPort[tlbPort_index]->retries.push_back(pkt);
876        } else {
877           DPRINTF(GPUTLB,
878                   "CU%d: WF[%d][%d]: Translation for addr %#x sent!\n",
879                   cu_id, gpuDynInst->simdId, gpuDynInst->wfSlotId, tmp_vaddr);
880        }
881    } else {
882        if (pkt->cmd == MemCmd::MemFenceReq) {
883            gpuDynInst->statusBitVector = VectorMask(0);
884        } else {
885            gpuDynInst->statusBitVector &= (~(1ll << index));
886        }
887
888        // New SenderState for the memory access
889        delete pkt->senderState;
890
891        // Because it's atomic operation, only need TLB translation state
892        pkt->senderState = new TheISA::GpuTLB::TranslationState(TLB_mode,
893                                                                shader->gpuTc);
894
895        tlbPort[tlbPort_index]->sendFunctional(pkt);
896
897        // the addr of the packet is not modified, so we need to create a new
898        // packet, or otherwise the memory access will have the old virtual
899        // address sent in the translation packet, instead of the physical
900        // address returned by the translation.
901        PacketPtr new_pkt = new Packet(pkt->req, pkt->cmd);
902        new_pkt->dataStatic(pkt->getPtr<uint8_t>());
903
904        // Translation is done. It is safe to send the packet to memory.
905        memPort[0]->sendFunctional(new_pkt);
906
907        DPRINTF(GPUMem, "CU%d: WF[%d][%d]: index %d: addr %#x\n", cu_id,
908                gpuDynInst->simdId, gpuDynInst->wfSlotId, index,
909                new_pkt->req->getPaddr());
910
911        // safe_cast the senderState
912        TheISA::GpuTLB::TranslationState *sender_state =
913             safe_cast<TheISA::GpuTLB::TranslationState*>(pkt->senderState);
914
915        delete sender_state->tlbEntry;
916        delete new_pkt;
917        delete pkt->senderState;
918        delete pkt->req;
919        delete pkt;
920    }
921}
922
923void
924ComputeUnit::sendSyncRequest(GPUDynInstPtr gpuDynInst, int index, PacketPtr pkt)
925{
926    ComputeUnit::DataPort::MemReqEvent *mem_req_event =
927        new ComputeUnit::DataPort::MemReqEvent(memPort[index], pkt);
928
929
930    // New SenderState for the memory access
931    pkt->senderState = new ComputeUnit::DataPort::SenderState(gpuDynInst, index,
932                                                              nullptr);
933
934    DPRINTF(GPUPort, "CU%d: WF[%d][%d]: index %d, addr %#x sync scheduled\n",
935            cu_id, gpuDynInst->simdId, gpuDynInst->wfSlotId, index,
936            pkt->req->getPaddr());
937
938    schedule(mem_req_event, curTick() + req_tick_latency);
939}
940
941void
942ComputeUnit::injectGlobalMemFence(GPUDynInstPtr gpuDynInst, bool kernelLaunch,
943                                  Request* req)
944{
945    if (!req) {
946        req = new Request(0, 0, 0, 0, masterId(), 0, gpuDynInst->wfDynId);
947    }
948    req->setPaddr(0);
949    if (kernelLaunch) {
950        req->setFlags(Request::KERNEL);
951    }
952
953    gpuDynInst->s_type = SEG_GLOBAL;
954
955    // for non-kernel MemFence operations, memorder flags are set depending
956    // on which type of request is currently being sent, so this
957    // should be set by the caller (e.g. if an inst has acq-rel
958    // semantics, it will send one acquire req an one release req)
959    gpuDynInst->setRequestFlags(req, kernelLaunch);
960
961    // a mem fence must correspond to an acquire/release request
962    assert(req->isAcquire() || req->isRelease());
963
964    // create packet
965    PacketPtr pkt = new Packet(req, MemCmd::MemFenceReq);
966
967    // set packet's sender state
968    pkt->senderState =
969        new ComputeUnit::DataPort::SenderState(gpuDynInst, 0, nullptr);
970
971    // send the packet
972    sendSyncRequest(gpuDynInst, 0, pkt);
973}
974
975const char*
976ComputeUnit::DataPort::MemRespEvent::description() const
977{
978    return "ComputeUnit memory response event";
979}
980
981void
982ComputeUnit::DataPort::MemRespEvent::process()
983{
984    DataPort::SenderState *sender_state =
985        safe_cast<DataPort::SenderState*>(pkt->senderState);
986
987    GPUDynInstPtr gpuDynInst = sender_state->_gpuDynInst;
988    ComputeUnit *compute_unit = dataPort->computeUnit;
989
990    assert(gpuDynInst);
991
992    DPRINTF(GPUPort, "CU%d: WF[%d][%d]: Response for addr %#x, index %d\n",
993            compute_unit->cu_id, gpuDynInst->simdId, gpuDynInst->wfSlotId,
994            pkt->req->getPaddr(), dataPort->index);
995
996    Addr paddr = pkt->req->getPaddr();
997
998    if (pkt->cmd != MemCmd::MemFenceResp) {
999        int index = gpuDynInst->memStatusVector[paddr].back();
1000
1001        DPRINTF(GPUMem, "Response for addr %#x, index %d\n",
1002                pkt->req->getPaddr(), index);
1003
1004        gpuDynInst->memStatusVector[paddr].pop_back();
1005        gpuDynInst->pAddr = pkt->req->getPaddr();
1006
1007        if (pkt->isRead() || pkt->isWrite()) {
1008
1009            if (gpuDynInst->n_reg <= MAX_REGS_FOR_NON_VEC_MEM_INST) {
1010                gpuDynInst->statusBitVector &= (~(1ULL << index));
1011            } else {
1012                assert(gpuDynInst->statusVector[index] > 0);
1013                gpuDynInst->statusVector[index]--;
1014
1015                if (!gpuDynInst->statusVector[index])
1016                    gpuDynInst->statusBitVector &= (~(1ULL << index));
1017            }
1018
1019            DPRINTF(GPUMem, "bitvector is now %#x\n",
1020                    gpuDynInst->statusBitVector);
1021
1022            if (gpuDynInst->statusBitVector == VectorMask(0)) {
1023                auto iter = gpuDynInst->memStatusVector.begin();
1024                auto end = gpuDynInst->memStatusVector.end();
1025
1026                while (iter != end) {
1027                    assert(iter->second.empty());
1028                    ++iter;
1029                }
1030
1031                gpuDynInst->memStatusVector.clear();
1032
1033                if (gpuDynInst->n_reg > MAX_REGS_FOR_NON_VEC_MEM_INST)
1034                    gpuDynInst->statusVector.clear();
1035
1036                if (gpuDynInst->m_op == Enums::MO_LD || MO_A(gpuDynInst->m_op)
1037                    || MO_ANR(gpuDynInst->m_op)) {
1038                    assert(compute_unit->globalMemoryPipe.isGMLdRespFIFOWrRdy());
1039
1040                    compute_unit->globalMemoryPipe.getGMLdRespFIFO()
1041                        .push(gpuDynInst);
1042                } else {
1043                    assert(compute_unit->globalMemoryPipe.isGMStRespFIFOWrRdy());
1044
1045                    compute_unit->globalMemoryPipe.getGMStRespFIFO()
1046                        .push(gpuDynInst);
1047                }
1048
1049                DPRINTF(GPUMem, "CU%d: WF[%d][%d]: packet totally complete\n",
1050                        compute_unit->cu_id, gpuDynInst->simdId,
1051                        gpuDynInst->wfSlotId);
1052
1053                // after clearing the status vectors,
1054                // see if there is a continuation to perform
1055                // the continuation may generate more work for
1056                // this memory request
1057                if (gpuDynInst->useContinuation) {
1058                    assert(gpuDynInst->scope != Enums::MEMORY_SCOPE_NONE);
1059                    gpuDynInst->execContinuation(gpuDynInst->staticInstruction(),
1060                                                 gpuDynInst);
1061                }
1062            }
1063        }
1064    } else {
1065        gpuDynInst->statusBitVector = VectorMask(0);
1066
1067        if (gpuDynInst->useContinuation) {
1068            assert(gpuDynInst->scope != Enums::MEMORY_SCOPE_NONE);
1069            gpuDynInst->execContinuation(gpuDynInst->staticInstruction(),
1070                                         gpuDynInst);
1071        }
1072    }
1073
1074    delete pkt->senderState;
1075    delete pkt->req;
1076    delete pkt;
1077}
1078
1079ComputeUnit*
1080ComputeUnitParams::create()
1081{
1082    return new ComputeUnit(this);
1083}
1084
1085bool
1086ComputeUnit::DTLBPort::recvTimingResp(PacketPtr pkt)
1087{
1088    Addr line = pkt->req->getPaddr();
1089
1090    DPRINTF(GPUTLB, "CU%d: DTLBPort received %#x->%#x\n", computeUnit->cu_id,
1091            pkt->req->getVaddr(), line);
1092
1093    assert(pkt->senderState);
1094    computeUnit->tlbCycles += curTick();
1095
1096    // pop off the TLB translation state
1097    TheISA::GpuTLB::TranslationState *translation_state =
1098               safe_cast<TheISA::GpuTLB::TranslationState*>(pkt->senderState);
1099
1100    // no PageFaults are permitted for data accesses
1101    if (!translation_state->tlbEntry->valid) {
1102        DTLBPort::SenderState *sender_state =
1103            safe_cast<DTLBPort::SenderState*>(translation_state->saved);
1104
1105        Wavefront *w M5_VAR_USED =
1106            computeUnit->wfList[sender_state->_gpuDynInst->simdId]
1107            [sender_state->_gpuDynInst->wfSlotId];
1108
1109        DPRINTFN("Wave %d couldn't tranlate vaddr %#x\n", w->wfDynId,
1110                 pkt->req->getVaddr());
1111    }
1112
1113    assert(translation_state->tlbEntry->valid);
1114
1115    // update the hitLevel distribution
1116    int hit_level = translation_state->hitLevel;
1117    computeUnit->hitsPerTLBLevel[hit_level]++;
1118
1119    delete translation_state->tlbEntry;
1120    assert(!translation_state->ports.size());
1121    pkt->senderState = translation_state->saved;
1122
1123    // for prefetch pkt
1124    BaseTLB::Mode TLB_mode = translation_state->tlbMode;
1125
1126    delete translation_state;
1127
1128    // use the original sender state to know how to close this transaction
1129    DTLBPort::SenderState *sender_state =
1130        safe_cast<DTLBPort::SenderState*>(pkt->senderState);
1131
1132    GPUDynInstPtr gpuDynInst = sender_state->_gpuDynInst;
1133    int mp_index = sender_state->portIndex;
1134    Addr vaddr = pkt->req->getVaddr();
1135    gpuDynInst->memStatusVector[line].push_back(mp_index);
1136    gpuDynInst->tlbHitLevel[mp_index] = hit_level;
1137
1138    MemCmd requestCmd;
1139
1140    if (pkt->cmd == MemCmd::ReadResp) {
1141        requestCmd = MemCmd::ReadReq;
1142    } else if (pkt->cmd == MemCmd::WriteResp) {
1143        requestCmd = MemCmd::WriteReq;
1144    } else if (pkt->cmd == MemCmd::SwapResp) {
1145        requestCmd = MemCmd::SwapReq;
1146    } else {
1147        panic("unsupported response to request conversion %s\n",
1148              pkt->cmd.toString());
1149    }
1150
1151    if (computeUnit->prefetchDepth) {
1152        int simdId = gpuDynInst->simdId;
1153        int wfSlotId = gpuDynInst->wfSlotId;
1154        Addr last = 0;
1155
1156        switch(computeUnit->prefetchType) {
1157        case Enums::PF_CU:
1158            last = computeUnit->lastVaddrCU[mp_index];
1159            break;
1160        case Enums::PF_PHASE:
1161            last = computeUnit->lastVaddrSimd[simdId][mp_index];
1162            break;
1163        case Enums::PF_WF:
1164            last = computeUnit->lastVaddrWF[simdId][wfSlotId][mp_index];
1165        default:
1166            break;
1167        }
1168
1169        DPRINTF(GPUPrefetch, "CU[%d][%d][%d][%d]: %#x was last\n",
1170                computeUnit->cu_id, simdId, wfSlotId, mp_index, last);
1171
1172        int stride = last ? (roundDown(vaddr, TheISA::PageBytes) -
1173                     roundDown(last, TheISA::PageBytes)) >> TheISA::PageShift
1174                     : 0;
1175
1176        DPRINTF(GPUPrefetch, "Stride is %d\n", stride);
1177
1178        computeUnit->lastVaddrCU[mp_index] = vaddr;
1179        computeUnit->lastVaddrSimd[simdId][mp_index] = vaddr;
1180        computeUnit->lastVaddrWF[simdId][wfSlotId][mp_index] = vaddr;
1181
1182        stride = (computeUnit->prefetchType == Enums::PF_STRIDE) ?
1183            computeUnit->prefetchStride: stride;
1184
1185        DPRINTF(GPUPrefetch, "%#x to: CU[%d][%d][%d][%d]\n", vaddr,
1186                computeUnit->cu_id, simdId, wfSlotId, mp_index);
1187
1188        DPRINTF(GPUPrefetch, "Prefetching from %#x:", vaddr);
1189
1190        // Prefetch Next few pages atomically
1191        for (int pf = 1; pf <= computeUnit->prefetchDepth; ++pf) {
1192            DPRINTF(GPUPrefetch, "%d * %d: %#x\n", pf, stride,
1193                    vaddr+stride*pf*TheISA::PageBytes);
1194
1195            if (!stride)
1196                break;
1197
1198            Request *prefetch_req = new Request(0, vaddr + stride * pf *
1199                                                TheISA::PageBytes,
1200                                                sizeof(uint8_t), 0,
1201                                                computeUnit->masterId(),
1202                                                0, 0, 0);
1203
1204            PacketPtr prefetch_pkt = new Packet(prefetch_req, requestCmd);
1205            uint8_t foo = 0;
1206            prefetch_pkt->dataStatic(&foo);
1207
1208            // Because it's atomic operation, only need TLB translation state
1209            prefetch_pkt->senderState =
1210                new TheISA::GpuTLB::TranslationState(TLB_mode,
1211                                                     computeUnit->shader->gpuTc,
1212                                                     true);
1213
1214            // Currently prefetches are zero-latency, hence the sendFunctional
1215            sendFunctional(prefetch_pkt);
1216
1217            /* safe_cast the senderState */
1218            TheISA::GpuTLB::TranslationState *tlb_state =
1219                 safe_cast<TheISA::GpuTLB::TranslationState*>(
1220                         prefetch_pkt->senderState);
1221
1222
1223            delete tlb_state->tlbEntry;
1224            delete tlb_state;
1225            delete prefetch_pkt->req;
1226            delete prefetch_pkt;
1227        }
1228    }
1229
1230    // First we must convert the response cmd back to a request cmd so that
1231    // the request can be sent through the cu's master port
1232    PacketPtr new_pkt = new Packet(pkt->req, requestCmd);
1233    new_pkt->dataStatic(pkt->getPtr<uint8_t>());
1234    delete pkt->senderState;
1235    delete pkt;
1236
1237    // New SenderState for the memory access
1238    new_pkt->senderState =
1239            new ComputeUnit::DataPort::SenderState(gpuDynInst, mp_index,
1240                                                   nullptr);
1241
1242    // translation is done. Schedule the mem_req_event at the appropriate
1243    // cycle to send the timing memory request to ruby
1244    ComputeUnit::DataPort::MemReqEvent *mem_req_event =
1245        new ComputeUnit::DataPort::MemReqEvent(computeUnit->memPort[mp_index],
1246                                               new_pkt);
1247
1248    DPRINTF(GPUPort, "CU%d: WF[%d][%d]: index %d, addr %#x data scheduled\n",
1249            computeUnit->cu_id, gpuDynInst->simdId,
1250            gpuDynInst->wfSlotId, mp_index, new_pkt->req->getPaddr());
1251
1252    computeUnit->schedule(mem_req_event, curTick() +
1253                          computeUnit->req_tick_latency);
1254
1255    return true;
1256}
1257
1258const char*
1259ComputeUnit::DataPort::MemReqEvent::description() const
1260{
1261    return "ComputeUnit memory request event";
1262}
1263
1264void
1265ComputeUnit::DataPort::MemReqEvent::process()
1266{
1267    SenderState *sender_state = safe_cast<SenderState*>(pkt->senderState);
1268    GPUDynInstPtr gpuDynInst = sender_state->_gpuDynInst;
1269    ComputeUnit *compute_unit M5_VAR_USED = dataPort->computeUnit;
1270
1271    if (!(dataPort->sendTimingReq(pkt))) {
1272        dataPort->retries.push_back(std::make_pair(pkt, gpuDynInst));
1273
1274        DPRINTF(GPUPort,
1275                "CU%d: WF[%d][%d]: index %d, addr %#x data req failed!\n",
1276                compute_unit->cu_id, gpuDynInst->simdId,
1277                gpuDynInst->wfSlotId, dataPort->index,
1278                pkt->req->getPaddr());
1279    } else {
1280        DPRINTF(GPUPort,
1281                "CU%d: WF[%d][%d]: index %d, addr %#x data req sent!\n",
1282                compute_unit->cu_id, gpuDynInst->simdId,
1283                gpuDynInst->wfSlotId, dataPort->index,
1284                pkt->req->getPaddr());
1285    }
1286}
1287
1288/*
1289 * The initial translation request could have been rejected,
1290 * if <retries> queue is not Retry sending the translation
1291 * request. sendRetry() is called from the peer port whenever
1292 * a translation completes.
1293 */
1294void
1295ComputeUnit::DTLBPort::recvReqRetry()
1296{
1297    int len = retries.size();
1298
1299    DPRINTF(GPUTLB, "CU%d: DTLB recvReqRetry - %d pending requests\n",
1300            computeUnit->cu_id, len);
1301
1302    assert(len > 0);
1303    assert(isStalled());
1304    // recvReqRetry is an indication that the resource on which this
1305    // port was stalling on is freed. So, remove the stall first
1306    unstallPort();
1307
1308    for (int i = 0; i < len; ++i) {
1309        PacketPtr pkt = retries.front();
1310        Addr vaddr M5_VAR_USED = pkt->req->getVaddr();
1311        DPRINTF(GPUTLB, "CU%d: retrying D-translaton for address%#x", vaddr);
1312
1313        if (!sendTimingReq(pkt)) {
1314            // Stall port
1315            stallPort();
1316            DPRINTF(GPUTLB, ": failed again\n");
1317            break;
1318        } else {
1319            DPRINTF(GPUTLB, ": successful\n");
1320            retries.pop_front();
1321        }
1322    }
1323}
1324
1325bool
1326ComputeUnit::ITLBPort::recvTimingResp(PacketPtr pkt)
1327{
1328    Addr line M5_VAR_USED = pkt->req->getPaddr();
1329    DPRINTF(GPUTLB, "CU%d: ITLBPort received %#x->%#x\n",
1330            computeUnit->cu_id, pkt->req->getVaddr(), line);
1331
1332    assert(pkt->senderState);
1333
1334    // pop off the TLB translation state
1335    TheISA::GpuTLB::TranslationState *translation_state =
1336                 safe_cast<TheISA::GpuTLB::TranslationState*>(pkt->senderState);
1337
1338    bool success = translation_state->tlbEntry->valid;
1339    delete translation_state->tlbEntry;
1340    assert(!translation_state->ports.size());
1341    pkt->senderState = translation_state->saved;
1342    delete translation_state;
1343
1344    // use the original sender state to know how to close this transaction
1345    ITLBPort::SenderState *sender_state =
1346        safe_cast<ITLBPort::SenderState*>(pkt->senderState);
1347
1348    // get the wavefront associated with this translation request
1349    Wavefront *wavefront = sender_state->wavefront;
1350    delete pkt->senderState;
1351
1352    if (success) {
1353        // pkt is reused in fetch(), don't delete it here.  However, we must
1354        // reset the command to be a request so that it can be sent through
1355        // the cu's master port
1356        assert(pkt->cmd == MemCmd::ReadResp);
1357        pkt->cmd = MemCmd::ReadReq;
1358
1359        computeUnit->fetchStage.fetch(pkt, wavefront);
1360    } else {
1361        if (wavefront->dropFetch) {
1362            assert(wavefront->instructionBuffer.empty());
1363            wavefront->dropFetch = false;
1364        }
1365
1366        wavefront->pendingFetch = 0;
1367    }
1368
1369    return true;
1370}
1371
1372/*
1373 * The initial translation request could have been rejected, if
1374 * <retries> queue is not empty. Retry sending the translation
1375 * request. sendRetry() is called from the peer port whenever
1376 * a translation completes.
1377 */
1378void
1379ComputeUnit::ITLBPort::recvReqRetry()
1380{
1381
1382    int len = retries.size();
1383    DPRINTF(GPUTLB, "CU%d: ITLB recvReqRetry - %d pending requests\n", len);
1384
1385    assert(len > 0);
1386    assert(isStalled());
1387
1388    // recvReqRetry is an indication that the resource on which this
1389    // port was stalling on is freed. So, remove the stall first
1390    unstallPort();
1391
1392    for (int i = 0; i < len; ++i) {
1393        PacketPtr pkt = retries.front();
1394        Addr vaddr M5_VAR_USED = pkt->req->getVaddr();
1395        DPRINTF(GPUTLB, "CU%d: retrying I-translaton for address%#x", vaddr);
1396
1397        if (!sendTimingReq(pkt)) {
1398            stallPort(); // Stall port
1399            DPRINTF(GPUTLB, ": failed again\n");
1400            break;
1401        } else {
1402            DPRINTF(GPUTLB, ": successful\n");
1403            retries.pop_front();
1404        }
1405    }
1406}
1407
1408void
1409ComputeUnit::regStats()
1410{
1411    MemObject::regStats();
1412
1413    tlbCycles
1414        .name(name() + ".tlb_cycles")
1415        .desc("total number of cycles for all uncoalesced requests")
1416        ;
1417
1418    tlbRequests
1419        .name(name() + ".tlb_requests")
1420        .desc("number of uncoalesced requests")
1421        ;
1422
1423    tlbLatency
1424        .name(name() + ".avg_translation_latency")
1425        .desc("Avg. translation latency for data translations")
1426        ;
1427
1428    tlbLatency = tlbCycles / tlbRequests;
1429
1430    hitsPerTLBLevel
1431       .init(4)
1432       .name(name() + ".TLB_hits_distribution")
1433       .desc("TLB hits distribution (0 for page table, x for Lx-TLB")
1434       ;
1435
1436    // fixed number of TLB levels
1437    for (int i = 0; i < 4; ++i) {
1438        if (!i)
1439            hitsPerTLBLevel.subname(i,"page_table");
1440        else
1441            hitsPerTLBLevel.subname(i, csprintf("L%d_TLB",i));
1442    }
1443
1444    execRateDist
1445        .init(0, 10, 2)
1446        .name(name() + ".inst_exec_rate")
1447        .desc("Instruction Execution Rate: Number of executed vector "
1448              "instructions per cycle")
1449        ;
1450
1451    ldsBankConflictDist
1452       .init(0, wfSize(), 2)
1453       .name(name() + ".lds_bank_conflicts")
1454       .desc("Number of bank conflicts per LDS memory packet")
1455       ;
1456
1457    ldsBankAccesses
1458        .name(name() + ".lds_bank_access_cnt")
1459        .desc("Total number of LDS bank accesses")
1460        ;
1461
1462    pageDivergenceDist
1463        // A wavefront can touch up to N pages per memory instruction where
1464        // N is equal to the wavefront size
1465        // The number of pages per bin can be configured (here it's 4).
1466       .init(1, wfSize(), 4)
1467       .name(name() + ".page_divergence_dist")
1468       .desc("pages touched per wf (over all mem. instr.)")
1469       ;
1470
1471    controlFlowDivergenceDist
1472        .init(1, wfSize(), 4)
1473        .name(name() + ".warp_execution_dist")
1474        .desc("number of lanes active per instruction (oval all instructions)")
1475        ;
1476
1477    activeLanesPerGMemInstrDist
1478        .init(1, wfSize(), 4)
1479        .name(name() + ".gmem_lanes_execution_dist")
1480        .desc("number of active lanes per global memory instruction")
1481        ;
1482
1483    activeLanesPerLMemInstrDist
1484        .init(1, wfSize(), 4)
1485        .name(name() + ".lmem_lanes_execution_dist")
1486        .desc("number of active lanes per local memory instruction")
1487        ;
1488
1489    numInstrExecuted
1490        .name(name() + ".num_instr_executed")
1491        .desc("number of instructions executed")
1492        ;
1493
1494    numVecOpsExecuted
1495        .name(name() + ".num_vec_ops_executed")
1496        .desc("number of vec ops executed (e.g. WF size/inst)")
1497        ;
1498
1499    totalCycles
1500        .name(name() + ".num_total_cycles")
1501        .desc("number of cycles the CU ran for")
1502        ;
1503
1504    ipc
1505        .name(name() + ".ipc")
1506        .desc("Instructions per cycle (this CU only)")
1507        ;
1508
1509    vpc
1510        .name(name() + ".vpc")
1511        .desc("Vector Operations per cycle (this CU only)")
1512        ;
1513
1514    numALUInstsExecuted
1515        .name(name() + ".num_alu_insts_executed")
1516        .desc("Number of dynamic non-GM memory insts executed")
1517        ;
1518
1519    wgBlockedDueLdsAllocation
1520        .name(name() + ".wg_blocked_due_lds_alloc")
1521        .desc("Workgroup blocked due to LDS capacity")
1522        ;
1523
1524    ipc = numInstrExecuted / totalCycles;
1525    vpc = numVecOpsExecuted / totalCycles;
1526
1527    numTimesWgBlockedDueVgprAlloc
1528        .name(name() + ".times_wg_blocked_due_vgpr_alloc")
1529        .desc("Number of times WGs are blocked due to VGPR allocation per SIMD")
1530        ;
1531
1532    dynamicGMemInstrCnt
1533        .name(name() + ".global_mem_instr_cnt")
1534        .desc("dynamic global memory instructions count")
1535        ;
1536
1537    dynamicLMemInstrCnt
1538        .name(name() + ".local_mem_instr_cnt")
1539        .desc("dynamic local memory intruction count")
1540        ;
1541
1542    numALUInstsExecuted = numInstrExecuted - dynamicGMemInstrCnt -
1543        dynamicLMemInstrCnt;
1544
1545    completedWfs
1546        .name(name() + ".num_completed_wfs")
1547        .desc("number of completed wavefronts")
1548        ;
1549
1550    numCASOps
1551        .name(name() + ".num_CAS_ops")
1552        .desc("number of compare and swap operations")
1553        ;
1554
1555    numFailedCASOps
1556        .name(name() + ".num_failed_CAS_ops")
1557        .desc("number of compare and swap operations that failed")
1558        ;
1559
1560    // register stats of pipeline stages
1561    fetchStage.regStats();
1562    scoreboardCheckStage.regStats();
1563    scheduleStage.regStats();
1564    execStage.regStats();
1565
1566    // register stats of memory pipeline
1567    globalMemoryPipe.regStats();
1568    localMemoryPipe.regStats();
1569}
1570
1571void
1572ComputeUnit::updatePageDivergenceDist(Addr addr)
1573{
1574    Addr virt_page_addr = roundDown(addr, TheISA::PageBytes);
1575
1576    if (!pagesTouched.count(virt_page_addr))
1577        pagesTouched[virt_page_addr] = 1;
1578    else
1579        pagesTouched[virt_page_addr]++;
1580}
1581
1582void
1583ComputeUnit::CUExitCallback::process()
1584{
1585    if (computeUnit->countPages) {
1586        std::ostream *page_stat_file =
1587            simout.create(computeUnit->name().c_str())->stream();
1588
1589        *page_stat_file << "page, wavefront accesses, workitem accesses" <<
1590            std::endl;
1591
1592        for (auto iter : computeUnit->pageAccesses) {
1593            *page_stat_file << std::hex << iter.first << ",";
1594            *page_stat_file << std::dec << iter.second.first << ",";
1595            *page_stat_file << std::dec << iter.second.second << std::endl;
1596        }
1597    }
1598 }
1599
1600bool
1601ComputeUnit::isDone() const
1602{
1603    for (int i = 0; i < numSIMDs; ++i) {
1604        if (!isSimdDone(i)) {
1605            return false;
1606        }
1607    }
1608
1609    bool glbMemBusRdy = true;
1610    for (int j = 0; j < numGlbMemUnits; ++j) {
1611        glbMemBusRdy &= vrfToGlobalMemPipeBus[j].rdy();
1612    }
1613    bool locMemBusRdy = true;
1614    for (int j = 0; j < numLocMemUnits; ++j) {
1615        locMemBusRdy &= vrfToLocalMemPipeBus[j].rdy();
1616    }
1617
1618    if (!globalMemoryPipe.isGMLdRespFIFOWrRdy() ||
1619        !globalMemoryPipe.isGMStRespFIFOWrRdy() ||
1620        !globalMemoryPipe.isGMReqFIFOWrRdy() || !localMemoryPipe.isLMReqFIFOWrRdy()
1621        || !localMemoryPipe.isLMRespFIFOWrRdy() || !locMemToVrfBus.rdy() ||
1622        !glbMemToVrfBus.rdy() || !locMemBusRdy || !glbMemBusRdy) {
1623        return false;
1624    }
1625
1626    return true;
1627}
1628
1629int32_t
1630ComputeUnit::getRefCounter(const uint32_t dispatchId, const uint32_t wgId) const
1631{
1632    return lds.getRefCounter(dispatchId, wgId);
1633}
1634
1635bool
1636ComputeUnit::isSimdDone(uint32_t simdId) const
1637{
1638    assert(simdId < numSIMDs);
1639
1640    for (int i=0; i < numGlbMemUnits; ++i) {
1641        if (!vrfToGlobalMemPipeBus[i].rdy())
1642            return false;
1643    }
1644    for (int i=0; i < numLocMemUnits; ++i) {
1645        if (!vrfToLocalMemPipeBus[i].rdy())
1646            return false;
1647    }
1648    if (!aluPipe[simdId].rdy()) {
1649        return false;
1650    }
1651
1652    for (int i_wf = 0; i_wf < shader->n_wf; ++i_wf){
1653        if (wfList[simdId][i_wf]->status != Wavefront::S_STOPPED) {
1654            return false;
1655        }
1656    }
1657
1658    return true;
1659}
1660
1661/**
1662 * send a general request to the LDS
1663 * make sure to look at the return value here as your request might be
1664 * NACK'd and returning false means that you have to have some backup plan
1665 */
1666bool
1667ComputeUnit::sendToLds(GPUDynInstPtr gpuDynInst)
1668{
1669    // this is just a request to carry the GPUDynInstPtr
1670    // back and forth
1671    Request *newRequest = new Request();
1672    newRequest->setPaddr(0x0);
1673
1674    // ReadReq is not evaluted by the LDS but the Packet ctor requires this
1675    PacketPtr newPacket = new Packet(newRequest, MemCmd::ReadReq);
1676
1677    // This is the SenderState needed upon return
1678    newPacket->senderState = new LDSPort::SenderState(gpuDynInst);
1679
1680    return ldsPort->sendTimingReq(newPacket);
1681}
1682
1683/**
1684 * get the result of packets sent to the LDS when they return
1685 */
1686bool
1687ComputeUnit::LDSPort::recvTimingResp(PacketPtr packet)
1688{
1689    const ComputeUnit::LDSPort::SenderState *senderState =
1690        dynamic_cast<ComputeUnit::LDSPort::SenderState *>(packet->senderState);
1691
1692    fatal_if(!senderState, "did not get the right sort of sender state");
1693
1694    GPUDynInstPtr gpuDynInst = senderState->getMemInst();
1695
1696    delete packet->senderState;
1697    delete packet->req;
1698    delete packet;
1699
1700    computeUnit->localMemoryPipe.getLMRespFIFO().push(gpuDynInst);
1701    return true;
1702}
1703
1704/**
1705 * attempt to send this packet, either the port is already stalled, the request
1706 * is nack'd and must stall or the request goes through
1707 * when a request cannot be sent, add it to the retries queue
1708 */
1709bool
1710ComputeUnit::LDSPort::sendTimingReq(PacketPtr pkt)
1711{
1712    ComputeUnit::LDSPort::SenderState *sender_state =
1713            dynamic_cast<ComputeUnit::LDSPort::SenderState*>(pkt->senderState);
1714    fatal_if(!sender_state, "packet without a valid sender state");
1715
1716    GPUDynInstPtr gpuDynInst M5_VAR_USED = sender_state->getMemInst();
1717
1718    if (isStalled()) {
1719        fatal_if(retries.empty(), "must have retries waiting to be stalled");
1720
1721        retries.push(pkt);
1722
1723        DPRINTF(GPUPort, "CU%d: WF[%d][%d]: LDS send failed!\n",
1724                        computeUnit->cu_id, gpuDynInst->simdId,
1725                        gpuDynInst->wfSlotId);
1726        return false;
1727    } else if (!MasterPort::sendTimingReq(pkt)) {
1728        // need to stall the LDS port until a recvReqRetry() is received
1729        // this indicates that there is more space
1730        stallPort();
1731        retries.push(pkt);
1732
1733        DPRINTF(GPUPort, "CU%d: WF[%d][%d]: addr %#x lds req failed!\n",
1734                computeUnit->cu_id, gpuDynInst->simdId,
1735                gpuDynInst->wfSlotId, pkt->req->getPaddr());
1736        return false;
1737    } else {
1738        DPRINTF(GPUPort, "CU%d: WF[%d][%d]: addr %#x lds req sent!\n",
1739                computeUnit->cu_id, gpuDynInst->simdId,
1740                gpuDynInst->wfSlotId, pkt->req->getPaddr());
1741        return true;
1742    }
1743}
1744
1745/**
1746 * the bus is telling the port that there is now space so retrying stalled
1747 * requests should work now
1748 * this allows the port to have a request be nack'd and then have the receiver
1749 * say when there is space, rather than simply retrying the send every cycle
1750 */
1751void
1752ComputeUnit::LDSPort::recvReqRetry()
1753{
1754    auto queueSize = retries.size();
1755
1756    DPRINTF(GPUPort, "CU%d: LDSPort recvReqRetry - %d pending requests\n",
1757            computeUnit->cu_id, queueSize);
1758
1759    fatal_if(queueSize < 1,
1760             "why was there a recvReqRetry() with no pending reqs?");
1761    fatal_if(!isStalled(),
1762             "recvReqRetry() happened when the port was not stalled");
1763
1764    unstallPort();
1765
1766    while (!retries.empty()) {
1767        PacketPtr packet = retries.front();
1768
1769        DPRINTF(GPUPort, "CU%d: retrying LDS send\n", computeUnit->cu_id);
1770
1771        if (!MasterPort::sendTimingReq(packet)) {
1772            // Stall port
1773            stallPort();
1774            DPRINTF(GPUPort, ": LDS send failed again\n");
1775            break;
1776        } else {
1777            DPRINTF(GPUTLB, ": LDS send successful\n");
1778            retries.pop();
1779        }
1780    }
1781}
1782