cl_driver.cc revision 11856 
1/*
2 * Copyright (c) 2012-2015 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: Anthony Gutierrez
34 */
35
36#include "gpu-compute/cl_driver.hh"
37
38#include <memory>
39
40#include "base/intmath.hh"
41#include "cpu/thread_context.hh"
42#include "gpu-compute/dispatcher.hh"
43#include "gpu-compute/hsa_code.hh"
44#include "gpu-compute/hsa_kernel_info.hh"
45#include "gpu-compute/hsa_object.hh"
46#include "params/ClDriver.hh"
47#include "sim/process.hh"
48#include "sim/syscall_emul_buf.hh"
49
50ClDriver::ClDriver(ClDriverParams *p)
51    : EmulatedDriver(p), hsaCode(0)
52{
53    for (const auto &codeFile : p->codefile)
54        codeFiles.push_back(&codeFile);
55
56    maxFuncArgsSize = 0;
57
58    for (int i = 0; i < codeFiles.size(); ++i) {
59        HsaObject *obj = HsaObject::createHsaObject(*codeFiles[i]);
60
61        for (int k = 0; k < obj->numKernels(); ++k) {
62            assert(obj->getKernel(k));
63            kernels.push_back(obj->getKernel(k));
64            kernels.back()->setReadonlyData((uint8_t*)obj->readonlyData);
65            int kern_funcargs_size = kernels.back()->funcarg_size;
66            maxFuncArgsSize = maxFuncArgsSize < kern_funcargs_size ?
67                kern_funcargs_size : maxFuncArgsSize;
68        }
69    }
70
71    int name_offs = 0;
72    int code_offs = 0;
73
74    for (int i = 0; i < kernels.size(); ++i) {
75        kernelInfo.push_back(HsaKernelInfo());
76        HsaCode *k = kernels[i];
77
78        k->generateHsaKernelInfo(&kernelInfo[i]);
79
80        kernelInfo[i].name_offs = name_offs;
81        kernelInfo[i].code_offs = code_offs;
82
83        name_offs += k->name().size() + 1;
84        code_offs += k->numInsts() * sizeof(TheGpuISA::RawMachInst);
85    }
86}
87
88void
89ClDriver::handshake(GpuDispatcher *_dispatcher)
90{
91    dispatcher = _dispatcher;
92    dispatcher->setFuncargsSize(maxFuncArgsSize);
93}
94
95int
96ClDriver::open(Process *p, ThreadContext *tc, int mode, int flags)
97{
98    std::shared_ptr<DeviceFDEntry> fdp;
99    fdp = std::make_shared<DeviceFDEntry>(this, filename);
100    int tgt_fd = p->fds->allocFD(fdp);
101    return tgt_fd;
102}
103
104int
105ClDriver::ioctl(Process *process, ThreadContext *tc, unsigned req)
106{
107    int index = 2;
108    Addr buf_addr = process->getSyscallArg(tc, index);
109
110    switch (req) {
111      case HSA_GET_SIZES:
112        {
113            TypedBufferArg<HsaDriverSizes> sizes(buf_addr);
114            sizes->num_kernels = kernels.size();
115            sizes->string_table_size = 0;
116            sizes->code_size = 0;
117            sizes->readonly_size = 0;
118
119            if (kernels.size() > 0) {
120                // all kernels will share the same read-only memory
121                sizes->readonly_size =
122                    kernels[0]->getSize(HsaCode::MemorySegment::READONLY);
123                // check our assumption
124                for (int i = 1; i<kernels.size(); ++i) {
125                    assert(sizes->readonly_size ==
126                    kernels[i]->getSize(HsaCode::MemorySegment::READONLY));
127                }
128            }
129
130            for (int i = 0; i < kernels.size(); ++i) {
131                HsaCode *k = kernels[i];
132                // add one for terminating '\0'
133                sizes->string_table_size += k->name().size() + 1;
134                sizes->code_size +=
135                    k->numInsts() * sizeof(TheGpuISA::RawMachInst);
136            }
137
138            sizes.copyOut(tc->getMemProxy());
139        }
140        break;
141
142      case HSA_GET_KINFO:
143        {
144            TypedBufferArg<HsaKernelInfo>
145                kinfo(buf_addr, sizeof(HsaKernelInfo) * kernels.size());
146
147            for (int i = 0; i < kernels.size(); ++i) {
148                HsaKernelInfo *ki = &kinfo[i];
149                ki->name_offs = kernelInfo[i].name_offs;
150                ki->code_offs = kernelInfo[i].code_offs;
151                ki->sRegCount = kernelInfo[i].sRegCount;
152                ki->dRegCount = kernelInfo[i].dRegCount;
153                ki->cRegCount = kernelInfo[i].cRegCount;
154                ki->static_lds_size  = kernelInfo[i].static_lds_size;
155                ki->private_mem_size = kernelInfo[i].private_mem_size;
156                ki->spill_mem_size   = kernelInfo[i].spill_mem_size;
157            }
158
159            kinfo.copyOut(tc->getMemProxy());
160        }
161        break;
162
163      case HSA_GET_STRINGS:
164        {
165            int string_table_size = 0;
166            for (int i = 0; i < kernels.size(); ++i) {
167                HsaCode *k = kernels[i];
168                string_table_size += k->name().size() + 1;
169            }
170
171            BufferArg buf(buf_addr, string_table_size);
172            char *bufp = (char*)buf.bufferPtr();
173
174            for (int i = 0; i < kernels.size(); ++i) {
175                HsaCode *k = kernels[i];
176                const char *n = k->name().c_str();
177
178                // idiomatic string copy
179                while ((*bufp++ = *n++));
180            }
181
182            assert(bufp - (char *)buf.bufferPtr() == string_table_size);
183
184            buf.copyOut(tc->getMemProxy());
185        }
186        break;
187
188      case HSA_GET_READONLY_DATA:
189        {
190            // we can pick any kernel --- they share the same
191            // readonly segment (this assumption is checked in GET_SIZES)
192            uint64_t size =
193                kernels.back()->getSize(HsaCode::MemorySegment::READONLY);
194            BufferArg data(buf_addr, size);
195            char *datap = (char *)data.bufferPtr();
196            memcpy(datap,
197                   kernels.back()->readonly_data,
198                   size);
199            data.copyOut(tc->getMemProxy());
200        }
201        break;
202
203      case HSA_GET_CODE:
204        {
205            // set hsaCode pointer
206            hsaCode = buf_addr;
207            int code_size = 0;
208
209            for (int i = 0; i < kernels.size(); ++i) {
210                HsaCode *k = kernels[i];
211                code_size += k->numInsts() * sizeof(TheGpuISA::RawMachInst);
212            }
213
214            TypedBufferArg<TheGpuISA::RawMachInst> buf(buf_addr, code_size);
215            TheGpuISA::RawMachInst *bufp = buf;
216
217            int buf_idx = 0;
218
219            for (int i = 0; i < kernels.size(); ++i) {
220                HsaCode *k = kernels[i];
221
222                for (int j = 0; j < k->numInsts(); ++j) {
223                    bufp[buf_idx] = k->insts()->at(j);
224                    ++buf_idx;
225                }
226            }
227
228            buf.copyOut(tc->getMemProxy());
229        }
230        break;
231
232      case HSA_GET_CU_CNT:
233        {
234            BufferArg buf(buf_addr, sizeof(uint32_t));
235            *((uint32_t*)buf.bufferPtr()) = dispatcher->getNumCUs();
236            buf.copyOut(tc->getMemProxy());
237        }
238        break;
239
240      case HSA_GET_VSZ:
241        {
242            BufferArg buf(buf_addr, sizeof(uint32_t));
243            *((uint32_t*)buf.bufferPtr()) = dispatcher->wfSize();
244            buf.copyOut(tc->getMemProxy());
245        }
246        break;
247      case HSA_GET_HW_STATIC_CONTEXT_SIZE:
248        {
249            BufferArg buf(buf_addr, sizeof(uint32_t));
250            *((uint32_t*)buf.bufferPtr()) = dispatcher->getStaticContextSize();
251            buf.copyOut(tc->getMemProxy());
252        }
253        break;
254
255      default:
256        fatal("ClDriver: bad ioctl %d\n", req);
257    }
258
259    return 0;
260}
261
262const char*
263ClDriver::codeOffToKernelName(uint64_t code_ptr)
264{
265    assert(hsaCode);
266    uint32_t code_offs = code_ptr - hsaCode;
267
268    for (int i = 0; i < kernels.size(); ++i) {
269        if (code_offs == kernelInfo[i].code_offs) {
270            return kernels[i]->name().c_str();
271        }
272    }
273
274    return nullptr;
275}
276
277ClDriver*
278ClDriverParams::create()
279{
280    return new ClDriver(this);
281}
282