src/gpu-compute/local_memory_pipeline.cc

   1 /*
   2  * Copyright (c) 2014-2015 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: Sooraj Puthoor
  34  */
  35
  36 #include "gpu-compute/local_memory_pipeline.hh"
  37
  38 #include "debug/GPUPort.hh"
  39 #include "gpu-compute/compute_unit.hh"
  40 #include "gpu-compute/gpu_dyn_inst.hh"
  41 #include "gpu-compute/shader.hh"
  42 #include "gpu-compute/vector_register_file.hh"
  43 #include "gpu-compute/wavefront.hh"
  44
  45 LocalMemPipeline::LocalMemPipeline(const ComputeUnitParams* p) :
  46     computeUnit(nullptr), lmQueueSize(p->local_mem_queue_size)
  47 {
  48 }
  49
  50 void
  51 LocalMemPipeline::init(ComputeUnit *cu)
  52 {
  53     computeUnit = cu;
  54     _name = computeUnit->name() + ".LocalMemPipeline";
  55 }
  56
  57 void
  58 LocalMemPipeline::exec()
  59 {
  60     // apply any returned shared (LDS) memory operations
  61     GPUDynInstPtr m = !lmReturnedRequests.empty() ?
  62         lmReturnedRequests.front() : nullptr;
  63
  64     bool accessVrf = true;
  65     Wavefront *w = nullptr;
  66
  67     if ((m) && (m->isLoad() || m->isAtomicRet())) {
  68         w = m->wavefront();
  69
  70         accessVrf =
  71             w->computeUnit->vrf[w->simdId]->
  72             vrfOperandAccessReady(m->seqNum(), w, m,
  73                                   VrfAccessType::WRITE);
  74     }
  75
  76     if (!lmReturnedRequests.empty() && m->latency.rdy() && accessVrf &&
  77         computeUnit->locMemToVrfBus.rdy() && (computeUnit->shader->coissue_return
  78                  || computeUnit->wfWait.at(m->pipeId).rdy())) {
  79
  80         lmReturnedRequests.pop();
  81         w = m->wavefront();
  82
  83         m->completeAcc(m);
  84
  85         // Decrement outstanding request count
  86         computeUnit->shader->ScheduleAdd(&w->outstandingReqs, m->time, -1);
  87
  88         if (m->isStore() || m->isAtomic()) {
  89             computeUnit->shader->ScheduleAdd(&w->outstandingReqsWrLm,
  90                                              m->time, -1);
  91         }
  92
  93         if (m->isLoad() || m->isAtomic()) {
  94             computeUnit->shader->ScheduleAdd(&w->outstandingReqsRdLm,
  95                                              m->time, -1);
  96         }
  97
  98         // Mark write bus busy for appropriate amount of time
  99         computeUnit->locMemToVrfBus.set(m->time);
 100         if (computeUnit->shader->coissue_return == 0)
 101             w->computeUnit->wfWait.at(m->pipeId).set(m->time);
 102     }
 103
 104     // If pipeline has executed a local memory instruction
 105     // execute local memory packet and issue the packets
 106     // to LDS
 107     if (!lmIssuedRequests.empty() && lmReturnedRequests.size() < lmQueueSize) {
 108
 109         GPUDynInstPtr m = lmIssuedRequests.front();
 110
 111         bool returnVal = computeUnit->sendToLds(m);
 112         if (!returnVal) {
 113             DPRINTF(GPUPort, "packet was nack'd and put in retry queue");
 114         }
 115         lmIssuedRequests.pop();
 116     }
 117 }
 118
 119 void
 120 LocalMemPipeline::regStats()
 121 {
 122     loadVrfBankConflictCycles
 123         .name(name() + ".load_vrf_bank_conflict_cycles")
 124         .desc("total number of cycles LDS data are delayed before updating "
 125               "the VRF")
 126         ;
 127 }