summaryrefslogtreecommitdiff
path: root/src/gpu-compute/gpu_dyn_inst.hh
blob: 9e63c445944d5016843d7e1c86af487cb2c375d2 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
/*
 * Copyright (c) 2015-2017 Advanced Micro Devices, Inc.
 * All rights reserved.
 *
 * For use for simulation and test purposes only
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 * 1. Redistributions of source code must retain the above copyright notice,
 * this list of conditions and the following disclaimer.
 *
 * 2. Redistributions in binary form must reproduce the above copyright notice,
 * this list of conditions and the following disclaimer in the documentation
 * and/or other materials provided with the distribution.
 *
 * 3. Neither the name of the copyright holder nor the names of its
 * contributors may be used to endorse or promote products derived from this
 * software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
 * POSSIBILITY OF SUCH DAMAGE.
 *
 * Authors: Anthony Gutierrez
 */

#ifndef __GPU_DYN_INST_HH__
#define __GPU_DYN_INST_HH__

#include <cstdint>
#include <string>

#include "base/logging.hh"
#include "enums/MemType.hh"
#include "enums/StorageClassType.hh"
#include "gpu-compute/compute_unit.hh"
#include "gpu-compute/gpu_exec_context.hh"

class GPUStaticInst;

template<typename T>
class AtomicOpAnd : public TypedAtomicOpFunctor<T>
{
  public:
    T a;

    AtomicOpAnd(T _a) : a(_a) { }
    void execute(T *b) { *b &= a; }
    AtomicOpFunctor* clone () { return new AtomicOpAnd(a); }
};

template<typename T>
class AtomicOpOr : public TypedAtomicOpFunctor<T>
{
  public:
    T a;
    AtomicOpOr(T _a) : a(_a) { }
    void execute(T *b) { *b |= a; }
    AtomicOpFunctor* clone () { return new AtomicOpOr(a); }
};

template<typename T>
class AtomicOpXor : public TypedAtomicOpFunctor<T>
{
  public:
    T a;
    AtomicOpXor(T _a) : a(_a) {}
    void execute(T *b) { *b ^= a; }
    AtomicOpFunctor* clone () { return new AtomicOpXor(a); }
};

template<typename T>
class AtomicOpCAS : public TypedAtomicOpFunctor<T>
{
  public:
    T c;
    T s;

    ComputeUnit *computeUnit;

    AtomicOpCAS(T _c, T _s, ComputeUnit *compute_unit)
      : c(_c), s(_s), computeUnit(compute_unit) { }

    void
    execute(T *b)
    {
        computeUnit->numCASOps++;

        if (*b == c) {
            *b = s;
        } else {
            computeUnit->numFailedCASOps++;
        }

        if (computeUnit->xact_cas_mode) {
            computeUnit->xactCasLoadMap.clear();
        }
    }
    AtomicOpFunctor* clone () { return new AtomicOpCAS(c, s, computeUnit); }
};

template<typename T>
class AtomicOpExch : public TypedAtomicOpFunctor<T>
{
  public:
    T a;
    AtomicOpExch(T _a) : a(_a) { }
    void execute(T *b) { *b = a; }
    AtomicOpFunctor* clone () { return new AtomicOpExch(a); }
};

template<typename T>
class AtomicOpAdd : public TypedAtomicOpFunctor<T>
{
  public:
    T a;
    AtomicOpAdd(T _a) : a(_a) { }
    void execute(T *b) { *b += a; }
    AtomicOpFunctor* clone () { return new AtomicOpAdd(a); }
};

template<typename T>
class AtomicOpSub : public TypedAtomicOpFunctor<T>
{
  public:
    T a;
    AtomicOpSub(T _a) : a(_a) { }
    void execute(T *b) { *b -= a; }
    AtomicOpFunctor* clone () { return new AtomicOpSub(a); }
};

template<typename T>
class AtomicOpInc : public TypedAtomicOpFunctor<T>
{
  public:
    AtomicOpInc() { }
    void execute(T *b) { *b += 1; }
    AtomicOpFunctor* clone () { return new AtomicOpInc(); }
};

template<typename T>
class AtomicOpDec : public TypedAtomicOpFunctor<T>
{
  public:
    AtomicOpDec() {}
    void execute(T *b) { *b -= 1; }
    AtomicOpFunctor* clone () { return new AtomicOpDec(); }
};

template<typename T>
class AtomicOpMax : public TypedAtomicOpFunctor<T>
{
  public:
    T a;
    AtomicOpMax(T _a) : a(_a) { }

    void
    execute(T *b)
    {
        if (a > *b)
            *b = a;
    }
    AtomicOpFunctor* clone () { return new AtomicOpMax(a); }
};

template<typename T>
class AtomicOpMin : public TypedAtomicOpFunctor<T>
{
  public:
    T a;
    AtomicOpMin(T _a) : a(_a) {}

    void
    execute(T *b)
    {
        if (a < *b)
            *b = a;
    }
    AtomicOpFunctor* clone () { return new AtomicOpMin(a); }
};

typedef enum
{
    VT_32,
    VT_64,
} vgpr_type;

class GPUDynInst : public GPUExecContext
{
  public:
    GPUDynInst(ComputeUnit *_cu, Wavefront *_wf, GPUStaticInst *static_inst,
               uint64_t instSeqNum);
    ~GPUDynInst();
    void execute(GPUDynInstPtr gpuDynInst);
    int numSrcRegOperands();
    int numDstRegOperands();
    int getNumOperands();
    bool isVectorRegister(int operandIdx);
    bool isScalarRegister(int operandIdx);
    bool isCondRegister(int operandIdx);
    int getRegisterIndex(int operandIdx, GPUDynInstPtr gpuDynInst);
    int getOperandSize(int operandIdx);
    bool isDstOperand(int operandIdx);
    bool isSrcOperand(int operandIdx);

    const std::string &disassemble() const;

    uint64_t seqNum() const;

    Enums::StorageClassType executedAs();

    // The address of the memory operation
    std::vector<Addr> addr;
    Addr pAddr;

    // The data to get written
    uint8_t *d_data;
    // Additional data (for atomics)
    uint8_t *a_data;
    // Additional data (for atomics)
    uint8_t *x_data;
    // The execution mask
    VectorMask exec_mask;

    // The memory type (M_U32, M_S32, ...)
    Enums::MemType m_type;

    // The equivalency class
    int equiv;
    // The return VGPR type (VT_32 or VT_64)
    vgpr_type v_type;
    // Number of VGPR's accessed (1, 2, or 4)
    int n_reg;
    // The return VGPR index
    int dst_reg;
    // There can be max 4 dest regs>
    int dst_reg_vec[4];
    // SIMD where the WF of the memory instruction has been mapped to
    int simdId;
    // unique id of the WF where the memory instruction belongs to
    int wfDynId;
    // The kernel id of the requesting wf
    int kern_id;
    // The CU id of the requesting wf
    int cu_id;
    // HW slot id where the WF is mapped to inside a SIMD unit
    int wfSlotId;
    // execution pipeline id where the memory instruction has been scheduled
    int pipeId;
    // The execution time of this operation
    Tick time;
    // The latency of this operation
    WaitClass latency;
    // A list of bank conflicts for the 4 cycles.
    uint32_t bc[4];

    // A pointer to ROM
    uint8_t *rom;
    // The size of the READONLY segment
    int sz_rom;

    // Initiate the specified memory operation, by creating a
    // memory request and sending it off to the memory system.
    void initiateAcc(GPUDynInstPtr gpuDynInst);
    // Complete the specified memory operation, by writing
    // value back to the RF in the case of a load or atomic
    // return or, in the case of a store, we do nothing
    void completeAcc(GPUDynInstPtr gpuDynInst);

    void updateStats();

    GPUStaticInst* staticInstruction() { return _staticInst; }

    bool isALU() const;
    bool isBranch() const;
    bool isNop() const;
    bool isReturn() const;
    bool isUnconditionalJump() const;
    bool isSpecialOp() const;
    bool isWaitcnt() const;

    bool isBarrier() const;
    bool isMemFence() const;
    bool isMemRef() const;
    bool isFlat() const;
    bool isLoad() const;
    bool isStore() const;

    bool isAtomic() const;
    bool isAtomicNoRet() const;
    bool isAtomicRet() const;

    bool isScalar() const;
    bool readsSCC() const;
    bool writesSCC() const;
    bool readsVCC() const;
    bool writesVCC() const;

    bool isAtomicAnd() const;
    bool isAtomicOr() const;
    bool isAtomicXor() const;
    bool isAtomicCAS() const;
    bool isAtomicExch() const;
    bool isAtomicAdd() const;
    bool isAtomicSub() const;
    bool isAtomicInc() const;
    bool isAtomicDec() const;
    bool isAtomicMax() const;
    bool isAtomicMin() const;

    bool isArgLoad() const;
    bool isGlobalMem() const;
    bool isLocalMem() const;

    bool isArgSeg() const;
    bool isGlobalSeg() const;
    bool isGroupSeg() const;
    bool isKernArgSeg() const;
    bool isPrivateSeg() const;
    bool isReadOnlySeg() const;
    bool isSpillSeg() const;

    bool isWorkitemScope() const;
    bool isWavefrontScope() const;
    bool isWorkgroupScope() const;
    bool isDeviceScope() const;
    bool isSystemScope() const;
    bool isNoScope() const;

    bool isRelaxedOrder() const;
    bool isAcquire() const;
    bool isRelease() const;
    bool isAcquireRelease() const;
    bool isNoOrder() const;

    bool isGloballyCoherent() const;
    bool isSystemCoherent() const;

    /*
     * Loads/stores/atomics may have acquire/release semantics associated
     * withthem. Some protocols want to see the acquire/release as separate
     * requests from the load/store/atomic. We implement that separation
     * using continuations (i.e., a function pointer with an object associated
     * with it). When, for example, the front-end generates a store with
     * release semantics, we will first issue a normal store and set the
     * continuation in the GPUDynInst to a function that generate a
     * release request. That continuation will be called when the normal
     * store completes (in ComputeUnit::DataPort::recvTimingResponse). The
     * continuation will be called in the context of the same GPUDynInst
     * that generated the initial store.
     */
    std::function<void(GPUStaticInst*, GPUDynInstPtr)> execContinuation;

    // when true, call execContinuation when response arrives
    bool useContinuation;

    template<typename c0> AtomicOpFunctor*
    makeAtomicOpFunctor(c0 *reg0, c0 *reg1)
    {
        if (isAtomicAnd()) {
            return new AtomicOpAnd<c0>(*reg0);
        } else if (isAtomicOr()) {
            return new AtomicOpOr<c0>(*reg0);
        } else if (isAtomicXor()) {
            return new AtomicOpXor<c0>(*reg0);
        } else if (isAtomicCAS()) {
            return new AtomicOpCAS<c0>(*reg0, *reg1, cu);
        } else if (isAtomicExch()) {
            return new AtomicOpExch<c0>(*reg0);
        } else if (isAtomicAdd()) {
            return new AtomicOpAdd<c0>(*reg0);
        } else if (isAtomicSub()) {
            return new AtomicOpSub<c0>(*reg0);
        } else if (isAtomicInc()) {
            return new AtomicOpInc<c0>();
        } else if (isAtomicDec()) {
            return new AtomicOpDec<c0>();
        } else if (isAtomicMax()) {
            return new AtomicOpMax<c0>(*reg0);
        } else if (isAtomicMin()) {
            return new AtomicOpMin<c0>(*reg0);
        } else {
            fatal("Unrecognized atomic operation");
        }
    }

    void
    setRequestFlags(RequestPtr req, bool setMemOrder=true)
    {
        // currently these are the easy scopes to deduce
        if (isPrivateSeg()) {
            req->setMemSpaceConfigFlags(Request::PRIVATE_SEGMENT);
        } else if (isSpillSeg()) {
            req->setMemSpaceConfigFlags(Request::SPILL_SEGMENT);
        } else if (isGlobalSeg()) {
            req->setMemSpaceConfigFlags(Request::GLOBAL_SEGMENT);
        } else if (isReadOnlySeg()) {
            req->setMemSpaceConfigFlags(Request::READONLY_SEGMENT);
        } else if (isGroupSeg()) {
            req->setMemSpaceConfigFlags(Request::GROUP_SEGMENT);
        } else if (isFlat()) {
            panic("TODO: translate to correct scope");
        } else {
            fatal("%s has bad segment type\n", disassemble());
        }

        if (isWavefrontScope()) {
            req->setMemSpaceConfigFlags(Request::SCOPE_VALID |
                                        Request::WAVEFRONT_SCOPE);
        } else if (isWorkgroupScope()) {
            req->setMemSpaceConfigFlags(Request::SCOPE_VALID |
                                        Request::WORKGROUP_SCOPE);
        } else if (isDeviceScope()) {
            req->setMemSpaceConfigFlags(Request::SCOPE_VALID |
                                        Request::DEVICE_SCOPE);
        } else if (isSystemScope()) {
            req->setMemSpaceConfigFlags(Request::SCOPE_VALID |
                                        Request::SYSTEM_SCOPE);
        } else if (!isNoScope() && !isWorkitemScope()) {
            fatal("%s has bad scope type\n", disassemble());
        }

        if (setMemOrder) {
            // set acquire and release flags
            if (isAcquire()) {
                req->setFlags(Request::ACQUIRE);
            } else if (isRelease()) {
                req->setFlags(Request::RELEASE);
            } else if (isAcquireRelease()) {
                req->setFlags(Request::ACQUIRE | Request::RELEASE);
            } else if (!isNoOrder()) {
                fatal("%s has bad memory order\n", disassemble());
            }
        }

        // set atomic type
        // currently, the instruction genenerator only produces atomic return
        // but a magic instruction can produce atomic no return
        if (isAtomicRet()) {
            req->setFlags(Request::ATOMIC_RETURN_OP);
        } else if (isAtomicNoRet()) {
            req->setFlags(Request::ATOMIC_NO_RETURN_OP);
        }
    }

    // Map returned packets and the addresses they satisfy with which lane they
    // were requested from
    typedef std::unordered_map<Addr, std::vector<int>> StatusVector;
    StatusVector memStatusVector;

    // Track the status of memory requests per lane, a bit per lane
    VectorMask statusBitVector;
    // for ld_v# or st_v#
    std::vector<int> statusVector;
    std::vector<int> tlbHitLevel;

  private:
    GPUStaticInst *_staticInst;
    uint64_t _seqNum;
};

#endif // __GPU_DYN_INST_HH__