shader.cc revision 11386:94c09b607a84
1/* 2 * Copyright (c) 2011-2015 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: Steve Reinhardt 34 */ 35 36#include "gpu-compute/shader.hh" 37 38#include <limits> 39 40#include "arch/x86/linux/linux.hh" 41#include "base/chunk_generator.hh" 42#include "debug/GPUDisp.hh" 43#include "debug/GPUMem.hh" 44#include "debug/HSAIL.hh" 45#include "gpu-compute/dispatcher.hh" 46#include "gpu-compute/gpu_static_inst.hh" 47#include "gpu-compute/qstruct.hh" 48#include "gpu-compute/wavefront.hh" 49#include "mem/packet.hh" 50#include "mem/ruby/system/RubySystem.hh" 51#include "sim/sim_exit.hh" 52 53Shader::Shader(const Params *p) : SimObject(p), 54 clock(p->clk_domain->clockPeriod()), cpuThread(nullptr), gpuTc(nullptr), 55 cpuPointer(p->cpu_pointer), tickEvent(this), timingSim(p->timing), 56 hsail_mode(SIMT), impl_kern_boundary_sync(p->impl_kern_boundary_sync), 57 separate_acquire_release(p->separate_acquire_release), coissue_return(1), 58 trace_vgpr_all(1), n_cu((p->CUs).size()), n_wf(p->n_wf), 59 globalMemSize(p->globalmem), nextSchedCu(0), sa_n(0), tick_cnt(0), 60 box_tick_cnt(0), start_tick_cnt(0) 61{ 62 63 cuList.resize(n_cu); 64 65 for (int i = 0; i < n_cu; ++i) { 66 cuList[i] = p->CUs[i]; 67 assert(i == cuList[i]->cu_id); 68 cuList[i]->shader = this; 69 } 70} 71 72Addr 73Shader::mmap(int length) 74{ 75 76 Addr start; 77 78 // round up length to the next page 79 length = roundUp(length, TheISA::PageBytes); 80 81 Process *proc = gpuTc->getProcessPtr(); 82 83 if (proc->mmapGrowsDown()) { 84 DPRINTF(HSAIL, "GROWS DOWN"); 85 start = proc->mmap_end - length; 86 proc->mmap_end = start; 87 } else { 88 DPRINTF(HSAIL, "GROWS UP"); 89 start = proc->mmap_end; 90 proc->mmap_end += length; 91 92 // assertion to make sure we don't overwrite the stack (it grows down) 93 assert(proc->mmap_end < proc->stack_base - proc->max_stack_size); 94 } 95 96 DPRINTF(HSAIL,"Shader::mmap start= %#x, %#x\n", start, length); 97 98 proc->allocateMem(start, length); 99 100 return start; 101} 102 103void 104Shader::init() 105{ 106 // grab the threadContext of the thread running on the CPU 107 assert(cpuPointer); 108 gpuTc = cpuPointer->getContext(0); 109 assert(gpuTc); 110} 111 112Shader::~Shader() 113{ 114 for (int j = 0; j < n_cu; ++j) 115 delete cuList[j]; 116} 117 118void 119Shader::updateThreadContext(int tid) { 120 // thread context of the thread which dispatched work 121 assert(cpuPointer); 122 gpuTc = cpuPointer->getContext(tid); 123 assert(gpuTc); 124} 125 126void 127Shader::hostWakeUp(BaseCPU *cpu) { 128 if (cpuPointer == cpu) { 129 if (gpuTc->status() == ThreadContext::Suspended) 130 cpu->activateContext(gpuTc->threadId()); 131 } else { 132 //Make sure both dispatcher and shader are trying to 133 //wakeup same host. Hack here to enable kernel launch 134 //from multiple CPUs 135 panic("Dispatcher wants to wakeup a different host"); 136 } 137} 138 139Shader* 140ShaderParams::create() 141{ 142 return new Shader(this); 143} 144 145void 146Shader::exec() 147{ 148 tick_cnt = curTick(); 149 box_tick_cnt = curTick() - start_tick_cnt; 150 151 // apply any scheduled adds 152 for (int i = 0; i < sa_n; ++i) { 153 if (sa_when[i] <= tick_cnt) { 154 *sa_val[i] += sa_x[i]; 155 sa_val.erase(sa_val.begin() + i); 156 sa_x.erase(sa_x.begin() + i); 157 sa_when.erase(sa_when.begin() + i); 158 --sa_n; 159 --i; 160 } 161 } 162 163 // clock all of the cu's 164 for (int i = 0; i < n_cu; ++i) 165 cuList[i]->exec(); 166} 167 168bool 169Shader::dispatch_workgroups(NDRange *ndr) 170{ 171 bool scheduledSomething = false; 172 int cuCount = 0; 173 int curCu = nextSchedCu; 174 175 while (cuCount < n_cu) { 176 //Every time we try a CU, update nextSchedCu 177 nextSchedCu = (nextSchedCu + 1) % n_cu; 178 179 // dispatch workgroup iff the following two conditions are met: 180 // (a) wg_rem is true - there are unassigned workgroups in the grid 181 // (b) there are enough free slots in cu cuList[i] for this wg 182 if (ndr->wg_disp_rem && cuList[curCu]->ReadyWorkgroup(ndr)) { 183 scheduledSomething = true; 184 DPRINTF(GPUDisp, "Dispatching a workgroup to CU %d\n", curCu); 185 186 // ticks() member function translates cycles to simulation ticks. 187 if (!tickEvent.scheduled()) { 188 schedule(tickEvent, curTick() + this->ticks(1)); 189 } 190 191 cuList[curCu]->StartWorkgroup(ndr); 192 ndr->wgId[0]++; 193 ndr->globalWgId++; 194 if (ndr->wgId[0] * ndr->q.wgSize[0] >= ndr->q.gdSize[0]) { 195 ndr->wgId[0] = 0; 196 ndr->wgId[1]++; 197 198 if (ndr->wgId[1] * ndr->q.wgSize[1] >= ndr->q.gdSize[1]) { 199 ndr->wgId[1] = 0; 200 ndr->wgId[2]++; 201 202 if (ndr->wgId[2] * ndr->q.wgSize[2] >= ndr->q.gdSize[2]) { 203 ndr->wg_disp_rem = false; 204 break; 205 } 206 } 207 } 208 } 209 210 ++cuCount; 211 curCu = nextSchedCu; 212 } 213 214 return scheduledSomething; 215} 216 217void 218Shader::handshake(GpuDispatcher *_dispatcher) 219{ 220 dispatcher = _dispatcher; 221} 222 223void 224Shader::doFunctionalAccess(RequestPtr req, MemCmd cmd, void *data, 225 bool suppress_func_errors, int cu_id) 226{ 227 unsigned block_size = RubySystem::getBlockSizeBytes(); 228 unsigned size = req->getSize(); 229 230 Addr tmp_addr; 231 BaseTLB::Mode trans_mode; 232 233 if (cmd == MemCmd::ReadReq) { 234 trans_mode = BaseTLB::Read; 235 } else if (cmd == MemCmd::WriteReq) { 236 trans_mode = BaseTLB::Write; 237 } else { 238 fatal("unexcepted MemCmd\n"); 239 } 240 241 tmp_addr = req->getVaddr(); 242 Addr split_addr = roundDown(tmp_addr + size - 1, block_size); 243 244 assert(split_addr <= tmp_addr || split_addr - tmp_addr < block_size); 245 246 // Misaligned access 247 if (split_addr > tmp_addr) { 248 RequestPtr req1, req2; 249 req->splitOnVaddr(split_addr, req1, req2); 250 251 252 PacketPtr pkt1 = new Packet(req2, cmd); 253 PacketPtr pkt2 = new Packet(req1, cmd); 254 255 functionalTLBAccess(pkt1, cu_id, trans_mode); 256 functionalTLBAccess(pkt2, cu_id, trans_mode); 257 258 PacketPtr new_pkt1 = new Packet(pkt1->req, cmd); 259 PacketPtr new_pkt2 = new Packet(pkt2->req, cmd); 260 261 new_pkt1->dataStatic(data); 262 new_pkt2->dataStatic((uint8_t*)data + req1->getSize()); 263 264 if (suppress_func_errors) { 265 new_pkt1->setSuppressFuncError(); 266 new_pkt2->setSuppressFuncError(); 267 } 268 269 // fixme: this should be cuList[cu_id] if cu_id != n_cu 270 // The latter requires a memPort in the dispatcher 271 cuList[0]->memPort[0]->sendFunctional(new_pkt1); 272 cuList[0]->memPort[0]->sendFunctional(new_pkt2); 273 274 delete new_pkt1; 275 delete new_pkt2; 276 delete pkt1; 277 delete pkt2; 278 } else { 279 PacketPtr pkt = new Packet(req, cmd); 280 functionalTLBAccess(pkt, cu_id, trans_mode); 281 PacketPtr new_pkt = new Packet(pkt->req, cmd); 282 new_pkt->dataStatic(data); 283 284 if (suppress_func_errors) { 285 new_pkt->setSuppressFuncError(); 286 }; 287 288 // fixme: this should be cuList[cu_id] if cu_id != n_cu 289 // The latter requires a memPort in the dispatcher 290 cuList[0]->memPort[0]->sendFunctional(new_pkt); 291 292 delete new_pkt; 293 delete pkt; 294 } 295} 296 297bool 298Shader::busy() 299{ 300 for (int i_cu = 0; i_cu < n_cu; ++i_cu) { 301 if (!cuList[i_cu]->isDone()) { 302 return true; 303 } 304 } 305 306 return false; 307} 308 309void 310Shader::ScheduleAdd(uint32_t *val,Tick when,int x) 311{ 312 sa_val.push_back(val); 313 sa_when.push_back(tick_cnt + when); 314 sa_x.push_back(x); 315 ++sa_n; 316} 317 318Shader::TickEvent::TickEvent(Shader *_shader) 319 : Event(CPU_Tick_Pri), shader(_shader) 320{ 321} 322 323 324void 325Shader::TickEvent::process() 326{ 327 if (shader->busy()) { 328 shader->exec(); 329 shader->schedule(this, curTick() + shader->ticks(1)); 330 } 331} 332 333const char* 334Shader::TickEvent::description() const 335{ 336 return "Shader tick"; 337} 338 339void 340Shader::AccessMem(uint64_t address, void *ptr, uint32_t size, int cu_id, 341 MemCmd cmd, bool suppress_func_errors) 342{ 343 uint8_t *data_buf = (uint8_t*)ptr; 344 345 for (ChunkGenerator gen(address, size, RubySystem::getBlockSizeBytes()); 346 !gen.done(); gen.next()) { 347 Request *req = new Request(0, gen.addr(), gen.size(), 0, 348 cuList[0]->masterId(), 0, 0, 0); 349 350 doFunctionalAccess(req, cmd, data_buf, suppress_func_errors, cu_id); 351 data_buf += gen.size(); 352 delete req; 353 } 354} 355 356void 357Shader::ReadMem(uint64_t address, void *ptr, uint32_t size, int cu_id) 358{ 359 AccessMem(address, ptr, size, cu_id, MemCmd::ReadReq, false); 360} 361 362void 363Shader::ReadMem(uint64_t address, void *ptr, uint32_t size, int cu_id, 364 bool suppress_func_errors) 365{ 366 AccessMem(address, ptr, size, cu_id, MemCmd::ReadReq, suppress_func_errors); 367} 368 369void 370Shader::WriteMem(uint64_t address, void *ptr,uint32_t size, int cu_id) 371{ 372 AccessMem(address, ptr, size, cu_id, MemCmd::WriteReq, false); 373} 374 375void 376Shader::WriteMem(uint64_t address, void *ptr, uint32_t size, int cu_id, 377 bool suppress_func_errors) 378{ 379 AccessMem(address, ptr, size, cu_id, MemCmd::WriteReq, 380 suppress_func_errors); 381} 382 383/* 384 * Send a packet through the appropriate TLB functional port. 385 * If cu_id=n_cu, then this is the dispatcher's TLB. 386 * Otherwise it's the TLB of the cu_id compute unit. 387 */ 388void 389Shader::functionalTLBAccess(PacketPtr pkt, int cu_id, BaseTLB::Mode mode) 390{ 391 // update senderState. Need to know the gpuTc and the TLB mode 392 pkt->senderState = 393 new TheISA::GpuTLB::TranslationState(mode, gpuTc, false); 394 395 if (cu_id == n_cu) { 396 dispatcher->tlbPort->sendFunctional(pkt); 397 } else { 398 // even when the perLaneTLB flag is turned on 399 // it's ok tp send all accesses through lane 0 400 // since the lane # is not known here, 401 // This isn't important since these are functional accesses. 402 cuList[cu_id]->tlbPort[0]->sendFunctional(pkt); 403 } 404 405 /* safe_cast the senderState */ 406 TheISA::GpuTLB::TranslationState *sender_state = 407 safe_cast<TheISA::GpuTLB::TranslationState*>(pkt->senderState); 408 409 delete sender_state->tlbEntry; 410 delete pkt->senderState; 411} 412