1/*
2 * Copyright (c) 2015-2017 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
18 * contributors may be used to endorse or promote products derived from this
19 * software 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 * Authors: John Kalamatianos,
34 *          Mark Wyse
35 */
36
37#include "gpu-compute/vector_register_file.hh"
38
39#include <string>
40
41#include "base/logging.hh"
42#include "gpu-compute/compute_unit.hh"
43#include "gpu-compute/gpu_dyn_inst.hh"
44#include "gpu-compute/shader.hh"
45#include "gpu-compute/simple_pool_manager.hh"
46#include "gpu-compute/wavefront.hh"
47#include "params/VectorRegisterFile.hh"
48
49VectorRegisterFile::VectorRegisterFile(const VectorRegisterFileParams *p)
50    : SimObject(p),
51      manager(new SimplePoolManager(p->min_alloc, p->num_regs_per_simd)),
52      simdId(p->simd_id), numRegsPerSimd(p->num_regs_per_simd),
53      vgprState(new VecRegisterState())
54{
55    fatal_if(numRegsPerSimd % 2, "VRF size is illegal\n");
56    fatal_if(simdId < 0, "Illegal SIMD id for VRF");
57
58    fatal_if(numRegsPerSimd % p->min_alloc, "Min VGPR region allocation is not "
59             "multiple of VRF size\n");
60
61    busy.clear();
62    busy.resize(numRegsPerSimd, 0);
63    nxtBusy.clear();
64    nxtBusy.resize(numRegsPerSimd, 0);
65
66    vgprState->init(numRegsPerSimd, p->wfSize);
67}
68
69void
70VectorRegisterFile::setParent(ComputeUnit *_computeUnit)
71{
72    computeUnit = _computeUnit;
73    vgprState->setParent(computeUnit);
74}
75
76uint8_t
77VectorRegisterFile::regNxtBusy(int idx, uint32_t operandSize) const
78{
79    uint8_t status = nxtBusy.at(idx);
80
81    if (operandSize > 4) {
82        status = status | (nxtBusy.at((idx + 1) % numRegs()));
83    }
84
85    return status;
86}
87
88uint8_t
89VectorRegisterFile::regBusy(int idx, uint32_t operandSize) const
90{
91    uint8_t status = busy.at(idx);
92
93    if (operandSize > 4) {
94        status = status | (busy.at((idx + 1) % numRegs()));
95    }
96
97    return status;
98}
99
100void
101VectorRegisterFile::preMarkReg(int regIdx, uint32_t operandSize, uint8_t value)
102{
103    nxtBusy.at(regIdx) = value;
104
105    if (operandSize > 4) {
106        nxtBusy.at((regIdx + 1) % numRegs()) = value;
107    }
108}
109
110void
111VectorRegisterFile::markReg(int regIdx, uint32_t operandSize, uint8_t value)
112{
113    busy.at(regIdx) = value;
114
115    if (operandSize > 4) {
116        busy.at((regIdx + 1) % numRegs()) = value;
117    }
118}
119
120bool
121VectorRegisterFile::operandsReady(Wavefront *w, GPUDynInstPtr ii) const
122{
123    for (int i = 0; i < ii->getNumOperands(); ++i) {
124        if (ii->isVectorRegister(i)) {
125            uint32_t vgprIdx = ii->getRegisterIndex(i, ii);
126            uint32_t pVgpr = w->remap(vgprIdx, ii->getOperandSize(i), 1);
127
128            if (regBusy(pVgpr, ii->getOperandSize(i)) == 1) {
129                if (ii->isDstOperand(i)) {
130                    w->numTimesBlockedDueWAXDependencies++;
131                } else if (ii->isSrcOperand(i)) {
132                    w->numTimesBlockedDueRAWDependencies++;
133                }
134
135                return false;
136            }
137
138            if (regNxtBusy(pVgpr, ii->getOperandSize(i)) == 1) {
139                if (ii->isDstOperand(i)) {
140                    w->numTimesBlockedDueWAXDependencies++;
141                } else if (ii->isSrcOperand(i)) {
142                    w->numTimesBlockedDueRAWDependencies++;
143                }
144
145                return false;
146            }
147        }
148    }
149
150    return true;
151}
152
153void
154VectorRegisterFile::exec(GPUDynInstPtr ii, Wavefront *w)
155{
156    bool loadInstr = ii->isLoad();
157    bool atomicInstr = ii->isAtomic() || ii->isMemFence();
158
159    bool loadNoArgInstr = loadInstr && !ii->isArgLoad();
160
161    // iterate over all register destination operands
162    for (int i = 0; i < ii->getNumOperands(); ++i) {
163        if (ii->isVectorRegister(i) && ii->isDstOperand(i)) {
164            uint32_t physReg = w->remap(ii->getRegisterIndex(i, ii),
165                                        ii->getOperandSize(i), 1);
166
167            // mark the destination vector register as busy
168            markReg(physReg, ii->getOperandSize(i), 1);
169            // clear the in-flight status of the destination vector register
170            preMarkReg(physReg, ii->getOperandSize(i), 0);
171
172            // FIXME: if we ever model correct timing behavior
173            // for load argument instructions then we should not
174            // set the destination register as busy now but when
175            // the data returns. Loads and Atomics should free
176            // their destination registers when the data returns,
177            // not now
178            if (!atomicInstr && !loadNoArgInstr) {
179                uint32_t pipeLen = ii->getOperandSize(i) <= 4 ?
180                    computeUnit->spBypassLength() :
181                    computeUnit->dpBypassLength();
182
183                // schedule an event for marking the register as ready
184                computeUnit->registerEvent(w->simdId, physReg,
185                                           ii->getOperandSize(i),
186                                           computeUnit->shader->tick_cnt +
187                                           computeUnit->shader->ticks(pipeLen),
188                                           0);
189            }
190        }
191    }
192}
193
194int
195VectorRegisterFile::exec(uint64_t dynamic_id, Wavefront *w,
196                         std::vector<uint32_t> &regVec, uint32_t operandSize,
197                         uint64_t timestamp)
198{
199    int delay = 0;
200
201    panic_if(regVec.size() <= 0, "Illegal VGPR vector size=%d\n",
202             regVec.size());
203
204    for (int i = 0; i < regVec.size(); ++i) {
205        // mark the destination VGPR as free when the timestamp expires
206        computeUnit->registerEvent(w->simdId, regVec[i], operandSize,
207                                   computeUnit->shader->tick_cnt + timestamp +
208                                   computeUnit->shader->ticks(delay), 0);
209    }
210
211    return delay;
212}
213
214void
215VectorRegisterFile::updateResources(Wavefront *w, GPUDynInstPtr ii)
216{
217    // iterate over all register destination operands
218    for (int i = 0; i < ii->getNumOperands(); ++i) {
219        if (ii->isVectorRegister(i) && ii->isDstOperand(i)) {
220            uint32_t physReg = w->remap(ii->getRegisterIndex(i, ii),
221                                        ii->getOperandSize(i), 1);
222            // set the in-flight status of the destination vector register
223            preMarkReg(physReg, ii->getOperandSize(i), 1);
224        }
225    }
226}
227
228bool
229VectorRegisterFile::vrfOperandAccessReady(uint64_t dynamic_id, Wavefront *w,
230                                          GPUDynInstPtr ii,
231                                          VrfAccessType accessType)
232{
233    bool ready = true;
234
235    return ready;
236}
237
238bool
239VectorRegisterFile::vrfOperandAccessReady(Wavefront *w, GPUDynInstPtr ii,
240                                          VrfAccessType accessType)
241{
242    bool ready = true;
243
244    return ready;
245}
246
247VectorRegisterFile*
248VectorRegisterFileParams::create()
249{
250    return new VectorRegisterFile(this);
251}
252