2 * Copyright (c) 2014-2015 Advanced Micro Devices, Inc.
5 * For use for simulation and test purposes only
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions are met:
10 * 1. Redistributions of source code must retain the above copyright notice,
11 * this list of conditions and the following disclaimer.
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.
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.
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.
33 * Author: Sooraj Puthoor
36 #include "gpu-compute/local_memory_pipeline.hh"
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"
45 LocalMemPipeline::LocalMemPipeline(const ComputeUnitParams
* p
) :
46 computeUnit(nullptr), lmQueueSize(p
->local_mem_queue_size
)
51 LocalMemPipeline::init(ComputeUnit
*cu
)
54 _name
= computeUnit
->name() + ".LocalMemPipeline";
58 LocalMemPipeline::exec()
60 // apply any returned shared (LDS) memory operations
61 GPUDynInstPtr m
= !lmReturnedRequests
.empty() ?
62 lmReturnedRequests
.front() : nullptr;
64 bool accessVrf
= true;
65 if ((m
) && (m
->m_op
==Enums::MO_LD
|| MO_A(m
->m_op
))) {
66 Wavefront
*w
= computeUnit
->wfList
[m
->simdId
][m
->wfSlotId
];
69 w
->computeUnit
->vrf
[m
->simdId
]->
70 vrfOperandAccessReady(m
->seqNum(), w
, m
,
71 VrfAccessType::WRITE
);
74 if (!lmReturnedRequests
.empty() && m
->latency
.rdy() && accessVrf
&&
75 computeUnit
->locMemToVrfBus
.rdy() && (computeUnit
->shader
->coissue_return
76 || computeUnit
->wfWait
.at(m
->pipeId
).rdy())) {
77 if (m
->v_type
== VT_32
&& m
->m_type
== Enums::M_U8
)
78 doSmReturn
<uint32_t, uint8_t>(m
);
79 else if (m
->v_type
== VT_32
&& m
->m_type
== Enums::M_U16
)
80 doSmReturn
<uint32_t, uint16_t>(m
);
81 else if (m
->v_type
== VT_32
&& m
->m_type
== Enums::M_U32
)
82 doSmReturn
<uint32_t, uint32_t>(m
);
83 else if (m
->v_type
== VT_32
&& m
->m_type
== Enums::M_S8
)
84 doSmReturn
<int32_t, int8_t>(m
);
85 else if (m
->v_type
== VT_32
&& m
->m_type
== Enums::M_S16
)
86 doSmReturn
<int32_t, int16_t>(m
);
87 else if (m
->v_type
== VT_32
&& m
->m_type
== Enums::M_S32
)
88 doSmReturn
<int32_t, int32_t>(m
);
89 else if (m
->v_type
== VT_32
&& m
->m_type
== Enums::M_F16
)
90 doSmReturn
<float, Float16
>(m
);
91 else if (m
->v_type
== VT_32
&& m
->m_type
== Enums::M_F32
)
92 doSmReturn
<float, float>(m
);
93 else if (m
->v_type
== VT_64
&& m
->m_type
== Enums::M_U8
)
94 doSmReturn
<uint64_t, uint8_t>(m
);
95 else if (m
->v_type
== VT_64
&& m
->m_type
== Enums::M_U16
)
96 doSmReturn
<uint64_t, uint16_t>(m
);
97 else if (m
->v_type
== VT_64
&& m
->m_type
== Enums::M_U32
)
98 doSmReturn
<uint64_t, uint32_t>(m
);
99 else if (m
->v_type
== VT_64
&& m
->m_type
== Enums::M_U64
)
100 doSmReturn
<uint64_t, uint64_t>(m
);
101 else if (m
->v_type
== VT_64
&& m
->m_type
== Enums::M_S8
)
102 doSmReturn
<int64_t, int8_t>(m
);
103 else if (m
->v_type
== VT_64
&& m
->m_type
== Enums::M_S16
)
104 doSmReturn
<int64_t, int16_t>(m
);
105 else if (m
->v_type
== VT_64
&& m
->m_type
== Enums::M_S32
)
106 doSmReturn
<int64_t, int32_t>(m
);
107 else if (m
->v_type
== VT_64
&& m
->m_type
== Enums::M_S64
)
108 doSmReturn
<int64_t, int64_t>(m
);
109 else if (m
->v_type
== VT_64
&& m
->m_type
== Enums::M_F16
)
110 doSmReturn
<double, Float16
>(m
);
111 else if (m
->v_type
== VT_64
&& m
->m_type
== Enums::M_F32
)
112 doSmReturn
<double, float>(m
);
113 else if (m
->v_type
== VT_64
&& m
->m_type
== Enums::M_F64
)
114 doSmReturn
<double, double>(m
);
117 // If pipeline has executed a local memory instruction
118 // execute local memory packet and issue the packets
120 if (!lmIssuedRequests
.empty() && lmReturnedRequests
.size() < lmQueueSize
) {
122 GPUDynInstPtr m
= lmIssuedRequests
.front();
124 bool returnVal
= computeUnit
->sendToLds(m
);
126 DPRINTF(GPUPort
, "packet was nack'd and put in retry queue");
128 lmIssuedRequests
.pop();
132 template<typename c0
, typename c1
>
134 LocalMemPipeline::doSmReturn(GPUDynInstPtr m
)
136 lmReturnedRequests
.pop();
137 Wavefront
*w
= computeUnit
->wfList
[m
->simdId
][m
->wfSlotId
];
139 // Return data to registers
140 if (m
->m_op
== Enums::MO_LD
|| MO_A(m
->m_op
)) {
141 std::vector
<uint32_t> regVec
;
142 for (int k
= 0; k
< m
->n_reg
; ++k
) {
143 int dst
= m
->dst_reg
+k
;
145 if (m
->n_reg
> MAX_REGS_FOR_NON_VEC_MEM_INST
)
146 dst
= m
->dst_reg_vec
[k
];
147 // virtual->physical VGPR mapping
148 int physVgpr
= w
->remap(dst
,sizeof(c0
),1);
149 // save the physical VGPR index
150 regVec
.push_back(physVgpr
);
151 c1
*p1
= &((c1
*)m
->d_data
)[k
* w
->computeUnit
->wfSize()];
153 for (int i
= 0; i
< w
->computeUnit
->wfSize(); ++i
) {
154 if (m
->exec_mask
[i
]) {
155 // write the value into the physical VGPR. This is a purely
156 // functional operation. No timing is modeled.
157 w
->computeUnit
->vrf
[w
->simdId
]->write
<c0
>(physVgpr
,
164 // Schedule the write operation of the load data on the VRF. This simply
165 // models the timing aspect of the VRF write operation. It does not
166 // modify the physical VGPR.
167 loadVrfBankConflictCycles
+=
168 w
->computeUnit
->vrf
[w
->simdId
]->exec(m
->seqNum(), w
,
169 regVec
, sizeof(c0
), m
->time
);
172 // Decrement outstanding request count
173 computeUnit
->shader
->ScheduleAdd(&w
->outstandingReqs
, m
->time
, -1);
175 if (m
->m_op
== Enums::MO_ST
|| MO_A(m
->m_op
) || MO_ANR(m
->m_op
)
177 computeUnit
->shader
->ScheduleAdd(&w
->outstandingReqsWrLm
,
181 if (m
->m_op
== Enums::MO_LD
|| MO_A(m
->m_op
) || MO_ANR(m
->m_op
)) {
182 computeUnit
->shader
->ScheduleAdd(&w
->outstandingReqsRdLm
,
186 // Mark write bus busy for appropriate amount of time
187 computeUnit
->locMemToVrfBus
.set(m
->time
);
188 if (computeUnit
->shader
->coissue_return
== 0)
189 w
->computeUnit
->wfWait
.at(m
->pipeId
).set(m
->time
);
193 LocalMemPipeline::regStats()
195 loadVrfBankConflictCycles
196 .name(name() + ".load_vrf_bank_conflict_cycles")
197 .desc("total number of cycles LDS data are delayed before updating "