| 1 |
/* |
| 2 |
* Copyright (c) 2018-2021, Intel Corporation |
| 3 |
* |
| 4 |
* Permission is hereby granted, free of charge, to any person obtaining a |
| 5 |
* copy of this software and associated documentation files (the "Software"), |
| 6 |
* to deal in the Software without restriction, including without limitation |
| 7 |
* the rights to use, copy, modify, merge, publish, distribute, sublicense, |
| 8 |
* and/or sell copies of the Software, and to permit persons to whom the |
| 9 |
* Software is furnished to do so, subject to the following conditions: |
| 10 |
* |
| 11 |
* The above copyright notice and this permission notice shall be included |
| 12 |
* in all copies or substantial portions of the Software. |
| 13 |
* |
| 14 |
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS |
| 15 |
* OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 16 |
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
| 17 |
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR |
| 18 |
* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, |
| 19 |
* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR |
| 20 |
* OTHER DEALINGS IN THE SOFTWARE. |
| 21 |
*/ |
| 22 |
//! |
| 23 |
//! \file mos_gpucontext_specific.cpp |
| 24 |
//! \brief Container class for the Linux specific gpu context |
| 25 |
//! |
| 26 |
|
| 27 |
#include <unistd.h> |
| 28 |
#include "mos_context_specific.h" |
| 29 |
#include "mos_gpucontext_specific.h" |
| 30 |
#include "mos_graphicsresource_specific.h" |
| 31 |
#include "mos_commandbuffer_specific.h" |
| 32 |
#include "mos_util_devult_specific.h" |
| 33 |
#include "mos_cmdbufmgr.h" |
| 34 |
#include "mos_os_virtualengine.h" |
| 35 |
#include "mos_os_cp_interface_specific.h" |
| 36 |
|
| 37 |
#define MI_BATCHBUFFER_END 0x05000000 |
| 38 |
static pthread_mutex_t command_dump_mutex = PTHREAD_MUTEX_INITIALIZER; |
| 39 |
|
| 40 |
GpuContextSpecific::GpuContextSpecific( |
| 41 |
const MOS_GPU_NODE gpuNode, |
| 42 |
MOS_GPU_CONTEXT mosGpuCtx, |
| 43 |
CmdBufMgr *cmdBufMgr, |
| 44 |
GpuContext *reusedContext) |
| 45 |
{ |
| 46 |
MOS_OS_FUNCTION_ENTER; |
| 47 |
|
| 48 |
m_nodeOrdinal = gpuNode; |
| 49 |
m_cmdBufMgr = cmdBufMgr; |
| 50 |
m_gpuContext = mosGpuCtx; |
| 51 |
m_statusBufferResource = nullptr; |
| 52 |
m_maxPatchLocationsize = PATCHLOCATIONLIST_SIZE; |
| 53 |
|
| 54 |
if (reusedContext) |
| 55 |
{ |
| 56 |
MOS_OS_NORMALMESSAGE("gpucontex reusing not enabled on Linux."); |
| 57 |
} |
| 58 |
|
| 59 |
#if (_DEBUG || _RELEASE_INTERNAL) |
| 60 |
// get user engine instance setting from environment variable |
| 61 |
char *engineInstances = getenv("INTEL_ENGINE_INSTANCE"); |
| 62 |
if (engineInstances != nullptr) |
| 63 |
{ |
| 64 |
errno = 0; |
| 65 |
long int instance = strtol(engineInstances, nullptr, 16); |
| 66 |
/* Check for various possible errors. */ |
| 67 |
if ((errno == ERANGE && instance == LONG_MAX) || (instance < 0)) |
| 68 |
{ |
| 69 |
MOS_OS_NORMALMESSAGE("Invalid INTEL_ENGINE_INSTANCE setting.(%s)\n", engineInstances); |
| 70 |
m_engineInstanceSelect = 0x0; |
| 71 |
} |
| 72 |
else |
| 73 |
{ |
| 74 |
m_engineInstanceSelect = (uint32_t)instance; |
| 75 |
} |
| 76 |
} |
| 77 |
#endif |
| 78 |
} |
| 79 |
|
| 80 |
GpuContextSpecific::~GpuContextSpecific() |
| 81 |
{ |
| 82 |
MOS_OS_FUNCTION_ENTER; |
| 83 |
|
| 84 |
Clear(); |
| 85 |
} |
| 86 |
|
| 87 |
MOS_STATUS GpuContextSpecific::Init(OsContext *osContext, |
| 88 |
PMOS_INTERFACE osInterface, |
| 89 |
MOS_GPU_NODE GpuNode, |
| 90 |
PMOS_GPUCTX_CREATOPTIONS createOption) |
| 91 |
{ |
| 92 |
MOS_OS_FUNCTION_ENTER; |
| 93 |
|
| 94 |
MOS_OS_CHK_NULL_RETURN(osContext); |
| 95 |
|
| 96 |
if (m_cmdBufPoolMutex == nullptr) |
| 97 |
{ |
| 98 |
m_cmdBufPoolMutex = MosUtilities::MosCreateMutex(); |
| 99 |
} |
| 100 |
|
| 101 |
MOS_OS_CHK_NULL_RETURN(m_cmdBufPoolMutex); |
| 102 |
|
| 103 |
MosUtilities::MosLockMutex(m_cmdBufPoolMutex); |
| 104 |
|
| 105 |
m_cmdBufPool.clear(); |
| 106 |
|
| 107 |
MosUtilities::MosUnlockMutex(m_cmdBufPoolMutex); |
| 108 |
|
| 109 |
m_commandBufferSize = COMMAND_BUFFER_SIZE; |
| 110 |
|
| 111 |
m_nextFetchIndex = 0; |
| 112 |
|
| 113 |
m_cmdBufFlushed = true; |
| 114 |
|
| 115 |
m_osContext = osContext; |
| 116 |
|
| 117 |
MOS_OS_CHK_STATUS_RETURN(AllocateGPUStatusBuf()); |
| 118 |
|
| 119 |
m_commandBuffer = (PMOS_COMMAND_BUFFER)MOS_AllocAndZeroMemory(sizeof(MOS_COMMAND_BUFFER)); |
| 120 |
|
| 121 |
MOS_OS_CHK_NULL_RETURN(m_commandBuffer); |
| 122 |
|
| 123 |
m_IndirectHeapSize = 0; |
| 124 |
|
| 125 |
// each thread has its own GPU context, so do not need any lock as guarder here |
| 126 |
m_allocationList = (ALLOCATION_LIST *)MOS_AllocAndZeroMemory(sizeof(ALLOCATION_LIST) * ALLOCATIONLIST_SIZE); |
| 127 |
MOS_OS_CHK_NULL_RETURN(m_allocationList); |
| 128 |
m_maxNumAllocations = ALLOCATIONLIST_SIZE; |
| 129 |
|
| 130 |
m_patchLocationList = (PATCHLOCATIONLIST *)MOS_AllocAndZeroMemory(sizeof(PATCHLOCATIONLIST) * PATCHLOCATIONLIST_SIZE); |
| 131 |
MOS_OS_CHK_NULL_RETURN(m_patchLocationList); |
| 132 |
m_maxPatchLocationsize = PATCHLOCATIONLIST_SIZE; |
| 133 |
|
| 134 |
m_attachedResources = (PMOS_RESOURCE)MOS_AllocAndZeroMemory(sizeof(MOS_RESOURCE) * ALLOCATIONLIST_SIZE); |
| 135 |
MOS_OS_CHK_NULL_RETURN(m_attachedResources); |
| 136 |
|
| 137 |
m_writeModeList = (bool *)MOS_AllocAndZeroMemory(sizeof(bool) * ALLOCATIONLIST_SIZE); |
| 138 |
MOS_OS_CHK_NULL_RETURN(m_writeModeList); |
| 139 |
|
| 140 |
m_GPUStatusTag = 1; |
| 141 |
|
| 142 |
m_createOptionEnhanced = (MOS_GPUCTX_CREATOPTIONS_ENHANCED*)MOS_AllocAndZeroMemory(sizeof(MOS_GPUCTX_CREATOPTIONS_ENHANCED)); |
| 143 |
MOS_OS_CHK_NULL_RETURN(m_createOptionEnhanced); |
| 144 |
m_createOptionEnhanced->SSEUValue = createOption->SSEUValue; |
| 145 |
|
| 146 |
if (typeid(*createOption) == typeid(MOS_GPUCTX_CREATOPTIONS_ENHANCED)) |
| 147 |
{ |
| 148 |
PMOS_GPUCTX_CREATOPTIONS_ENHANCED createOptionEnhanced = static_cast<PMOS_GPUCTX_CREATOPTIONS_ENHANCED>(createOption); |
| 149 |
m_createOptionEnhanced->UsingSFC = createOptionEnhanced->UsingSFC; |
| 150 |
} |
| 151 |
|
| 152 |
for (int i=0; i<MAX_ENGINE_INSTANCE_NUM+1; i++) |
| 153 |
{ |
| 154 |
m_i915Context[i] = nullptr; |
| 155 |
} |
| 156 |
|
| 157 |
if (osInterface->ctxBasedScheduling) |
| 158 |
{ |
| 159 |
unsigned int nengine = 0; |
| 160 |
struct i915_engine_class_instance *engine_map = nullptr; |
| 161 |
|
| 162 |
MOS_TraceEventExt(EVENT_GPU_CONTEXT_CREATE, EVENT_TYPE_START, |
| 163 |
&GpuNode, sizeof(GpuNode), nullptr, 0); |
| 164 |
|
| 165 |
m_i915Context[0] = mos_context_create_shared(osInterface->pOsContext->bufmgr, |
| 166 |
osInterface->pOsContext->intel_context, |
| 167 |
I915_CONTEXT_CREATE_FLAGS_SINGLE_TIMELINE, |
| 168 |
false, |
| 169 |
(void *)engine_map, |
| 170 |
1, |
| 171 |
nengine, |
| 172 |
0); |
| 173 |
if (m_i915Context[0] == nullptr) |
| 174 |
{ |
| 175 |
MOS_OS_ASSERTMESSAGE("Failed to create context.\n"); |
| 176 |
return MOS_STATUS_UNKNOWN; |
| 177 |
} |
| 178 |
m_i915Context[0]->pOsContext = osInterface->pOsContext; |
| 179 |
|
| 180 |
m_i915ExecFlag = I915_EXEC_DEFAULT; |
| 181 |
|
| 182 |
if (mos_query_engines_count(osInterface->pOsContext->bufmgr, &nengine) || (nengine == 0)) |
| 183 |
{ |
| 184 |
MOS_OS_ASSERTMESSAGE("Failed to query engines count.\n"); |
| 185 |
return MOS_STATUS_UNKNOWN; |
| 186 |
} |
| 187 |
engine_map = (struct i915_engine_class_instance *)MOS_AllocAndZeroMemory(nengine * sizeof(struct i915_engine_class_instance)); |
| 188 |
MOS_OS_CHK_NULL_RETURN(engine_map); |
| 189 |
|
| 190 |
if (GpuNode == MOS_GPU_NODE_3D) |
| 191 |
{ |
| 192 |
__u16 engine_class = I915_ENGINE_CLASS_RENDER; |
| 193 |
__u64 caps = 0; |
| 194 |
|
| 195 |
if (mos_query_engines(osInterface->pOsContext->bufmgr, engine_class, caps, &nengine, (void *)engine_map)) |
| 196 |
{ |
| 197 |
MOS_OS_ASSERTMESSAGE("Failed to query engines.\n"); |
| 198 |
MOS_SafeFreeMemory(engine_map); |
| 199 |
return MOS_STATUS_UNKNOWN; |
| 200 |
} |
| 201 |
|
| 202 |
if (mos_set_context_param_load_balance(m_i915Context[0], engine_map, nengine)) |
| 203 |
{ |
| 204 |
MOS_OS_ASSERTMESSAGE("Failed to set balancer extension.\n"); |
| 205 |
MOS_SafeFreeMemory(engine_map); |
| 206 |
return MOS_STATUS_UNKNOWN; |
| 207 |
} |
| 208 |
|
| 209 |
if (createOption->SSEUValue != 0) |
| 210 |
{ |
| 211 |
struct drm_i915_gem_context_param_sseu sseu; |
| 212 |
MOS_ZeroMemory(&sseu, sizeof(sseu)); |
| 213 |
sseu.flags = I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX; |
| 214 |
sseu.engine.engine_instance = m_i915ExecFlag; |
| 215 |
|
| 216 |
if (mos_get_context_param_sseu(m_i915Context[0], &sseu)) |
| 217 |
{ |
| 218 |
MOS_OS_ASSERTMESSAGE("Failed to get sseu configuration."); |
| 219 |
MOS_SafeFreeMemory(engine_map); |
| 220 |
return MOS_STATUS_UNKNOWN; |
| 221 |
} |
| 222 |
|
| 223 |
if (mos_hweight8(m_i915Context[0], sseu.subslice_mask) > createOption->packed.SubSliceCount) |
| 224 |
{ |
| 225 |
sseu.subslice_mask = mos_switch_off_n_bits(m_i915Context[0], sseu.subslice_mask, |
| 226 |
mos_hweight8(m_i915Context[0], sseu.subslice_mask)-createOption->packed.SubSliceCount); |
| 227 |
} |
| 228 |
|
| 229 |
if (mos_set_context_param_sseu(m_i915Context[0], sseu)) |
| 230 |
{ |
| 231 |
MOS_OS_ASSERTMESSAGE("Failed to set sseu configuration."); |
| 232 |
MOS_SafeFreeMemory(engine_map); |
| 233 |
return MOS_STATUS_UNKNOWN; |
| 234 |
} |
| 235 |
} |
| 236 |
} |
| 237 |
else if (GpuNode == MOS_GPU_NODE_COMPUTE) |
| 238 |
{ |
| 239 |
__u16 engine_class = 4; //To change later when linux define the name |
| 240 |
__u64 caps = 0; |
| 241 |
|
| 242 |
if (mos_query_engines(osInterface->pOsContext->bufmgr, engine_class, caps, &nengine, (void *)engine_map)) |
| 243 |
{ |
| 244 |
MOS_OS_ASSERTMESSAGE("Failed to query engines.\n"); |
| 245 |
MOS_SafeFreeMemory(engine_map); |
| 246 |
return MOS_STATUS_UNKNOWN; |
| 247 |
} |
| 248 |
|
| 249 |
#if (_DEBUG || _RELEASE_INTERNAL) |
| 250 |
SelectEngineInstanceByUser(engine_map, &nengine, m_engineInstanceSelect, GpuNode); |
| 251 |
#endif |
| 252 |
if (mos_set_context_param_load_balance(m_i915Context[0], engine_map, nengine)) |
| 253 |
{ |
| 254 |
MOS_OS_ASSERTMESSAGE("Failed to set balancer extension.\n"); |
| 255 |
MOS_SafeFreeMemory(engine_map); |
| 256 |
return MOS_STATUS_UNKNOWN; |
| 257 |
} |
| 258 |
} |
| 259 |
else if (GpuNode == MOS_GPU_NODE_VIDEO || GpuNode == MOS_GPU_NODE_VIDEO2 |
| 260 |
|| GpuNode == MOS_GPU_NODE_VE) |
| 261 |
{ |
| 262 |
__u16 engine_class = (GpuNode == MOS_GPU_NODE_VE)? I915_ENGINE_CLASS_VIDEO_ENHANCE : I915_ENGINE_CLASS_VIDEO; |
| 263 |
__u64 caps = 0; |
| 264 |
|
| 265 |
SetEngineQueryFlags(createOption, caps); |
| 266 |
|
| 267 |
if (mos_query_engines(osInterface->pOsContext->bufmgr, engine_class, caps, &nengine, (void *)engine_map)) |
| 268 |
{ |
| 269 |
MOS_OS_ASSERTMESSAGE("Failed to query engines.\n"); |
| 270 |
MOS_SafeFreeMemory(engine_map); |
| 271 |
return MOS_STATUS_UNKNOWN; |
| 272 |
} |
| 273 |
|
| 274 |
#if (_DEBUG || _RELEASE_INTERNAL) |
| 275 |
SelectEngineInstanceByUser(engine_map, &nengine, m_engineInstanceSelect, GpuNode); |
| 276 |
#endif |
| 277 |
if (mos_set_context_param_load_balance(m_i915Context[0], engine_map, nengine)) |
| 278 |
{ |
| 279 |
MOS_OS_ASSERTMESSAGE("Failed to set balancer extension.\n"); |
| 280 |
MOS_SafeFreeMemory(engine_map); |
| 281 |
return MOS_STATUS_UNKNOWN; |
| 282 |
} |
| 283 |
|
| 284 |
if (nengine >= 2) |
| 285 |
{ |
| 286 |
int i; |
| 287 |
//master queue |
| 288 |
m_i915Context[1] = mos_context_create_shared(osInterface->pOsContext->bufmgr, |
| 289 |
osInterface->pOsContext->intel_context, |
| 290 |
I915_CONTEXT_CREATE_FLAGS_SINGLE_TIMELINE, |
| 291 |
false, |
| 292 |
(void *)engine_map, |
| 293 |
1, |
| 294 |
1, |
| 295 |
0); |
| 296 |
if (m_i915Context[1] == nullptr) |
| 297 |
{ |
| 298 |
MOS_OS_ASSERTMESSAGE("Failed to create master context.\n"); |
| 299 |
MOS_SafeFreeMemory(engine_map); |
| 300 |
return MOS_STATUS_UNKNOWN; |
| 301 |
} |
| 302 |
m_i915Context[1]->pOsContext = osInterface->pOsContext; |
| 303 |
|
| 304 |
if (mos_set_context_param_load_balance(m_i915Context[1], engine_map, 1)) |
| 305 |
{ |
| 306 |
MOS_OS_ASSERTMESSAGE("Failed to set master context bond extension.\n"); |
| 307 |
MOS_SafeFreeMemory(engine_map); |
| 308 |
return MOS_STATUS_UNKNOWN; |
| 309 |
} |
| 310 |
|
| 311 |
//slave queue |
| 312 |
for (i=1; i<nengine; i++) |
| 313 |
{ |
| 314 |
m_i915Context[i+1] = mos_context_create_shared(osInterface->pOsContext->bufmgr, |
| 315 |
osInterface->pOsContext->intel_context, |
| 316 |
I915_CONTEXT_CREATE_FLAGS_SINGLE_TIMELINE, |
| 317 |
false, |
| 318 |
(void *)engine_map, |
| 319 |
1, |
| 320 |
1, |
| 321 |
0); |
| 322 |
if (m_i915Context[i+1] == nullptr) |
| 323 |
{ |
| 324 |
MOS_OS_ASSERTMESSAGE("Failed to create slave context.\n"); |
| 325 |
MOS_SafeFreeMemory(engine_map); |
| 326 |
return MOS_STATUS_UNKNOWN; |
| 327 |
} |
| 328 |
m_i915Context[i+1]->pOsContext = osInterface->pOsContext; |
| 329 |
|
| 330 |
if (mos_set_context_param_bond(m_i915Context[i+1], engine_map[0], &engine_map[i], 1) != S_SUCCESS) |
| 331 |
{ |
| 332 |
int err = errno; |
| 333 |
if (err == ENODEV) |
| 334 |
{ |
| 335 |
mos_context_destroy(m_i915Context[1]); |
| 336 |
mos_context_destroy(m_i915Context[i+1]); |
| 337 |
m_i915Context[i+1] = nullptr; |
| 338 |
break; |
| 339 |
} |
| 340 |
else |
| 341 |
{ |
| 342 |
MOS_OS_ASSERTMESSAGE("Failed to set slave context bond extension. errno=%d\n",err); |
| 343 |
MOS_SafeFreeMemory(engine_map); |
| 344 |
return MOS_STATUS_UNKNOWN; |
| 345 |
} |
| 346 |
} |
| 347 |
} |
| 348 |
if (i == nengine) |
| 349 |
{ |
| 350 |
osInterface->bParallelSubmission = false; |
| 351 |
} |
| 352 |
else |
| 353 |
{ |
| 354 |
osInterface->bParallelSubmission = true; |
| 355 |
//create context with different width |
| 356 |
for(i = 1; i < nengine; i++) |
| 357 |
{ |
| 358 |
unsigned int ctxWidth = i + 1; |
| 359 |
m_i915Context[i] = mos_context_create_shared(osInterface->pOsContext->bufmgr, |
| 360 |
osInterface->pOsContext->intel_context, |
| 361 |
0, // I915_CONTEXT_CREATE_FLAGS_SINGLE_TIMELINE not allowed for parallel submission |
| 362 |
false, |
| 363 |
(void *)engine_map, |
| 364 |
ctxWidth, |
| 365 |
1, |
| 366 |
0); |
| 367 |
if (mos_set_context_param_parallel(m_i915Context[i], engine_map, ctxWidth) != S_SUCCESS) |
| 368 |
{ |
| 369 |
MOS_OS_ASSERTMESSAGE("Failed to set parallel extension since discontinuous logical engine.\n"); |
| 370 |
mos_context_destroy(m_i915Context[i]); |
| 371 |
m_i915Context[i] = nullptr; |
| 372 |
break; |
| 373 |
} |
| 374 |
} |
| 375 |
} |
| 376 |
} |
| 377 |
} |
| 378 |
else if (GpuNode == MOS_GPU_NODE_BLT) |
| 379 |
{ |
| 380 |
__u16 engine_class = I915_ENGINE_CLASS_COPY; |
| 381 |
__u64 caps = 0; |
| 382 |
|
| 383 |
if (mos_query_engines(osInterface->pOsContext->bufmgr, engine_class, caps, &nengine, (void *)engine_map)) |
| 384 |
{ |
| 385 |
MOS_OS_ASSERTMESSAGE("Failed to query engines.\n"); |
| 386 |
MOS_SafeFreeMemory(engine_map); |
| 387 |
return MOS_STATUS_UNKNOWN; |
| 388 |
} |
| 389 |
|
| 390 |
if (mos_set_context_param_load_balance(m_i915Context[0], engine_map, nengine)) |
| 391 |
{ |
| 392 |
MOS_OS_ASSERTMESSAGE("Failed to set balancer extension.\n"); |
| 393 |
MOS_SafeFreeMemory(engine_map); |
| 394 |
return MOS_STATUS_UNKNOWN; |
| 395 |
} |
| 396 |
} |
| 397 |
else |
| 398 |
{ |
| 399 |
MOS_OS_ASSERTMESSAGE("Unknown engine class.\n"); |
| 400 |
MOS_SafeFreeMemory(engine_map); |
| 401 |
return MOS_STATUS_UNKNOWN; |
| 402 |
} |
| 403 |
MOS_SafeFreeMemory(engine_map); |
| 404 |
MOS_TraceEventExt(EVENT_GPU_CONTEXT_CREATE, EVENT_TYPE_END, |
| 405 |
m_i915Context, sizeof(void *), |
| 406 |
&nengine, sizeof(nengine)); |
| 407 |
} |
| 408 |
return MOS_STATUS_SUCCESS; |
| 409 |
} |
| 410 |
|
| 411 |
void GpuContextSpecific::Clear() |
| 412 |
{ |
| 413 |
MOS_OS_FUNCTION_ENTER; |
| 414 |
|
| 415 |
MOS_TraceEventExt(EVENT_GPU_CONTEXT_DESTROY, EVENT_TYPE_START, |
| 416 |
m_i915Context, sizeof(void *), nullptr, 0); |
| 417 |
// hanlde the status buf bundled w/ the specified gpucontext |
| 418 |
if (m_statusBufferResource) |
| 419 |
{ |
| 420 |
if (m_statusBufferResource->Unlock(m_osContext) != MOS_STATUS_SUCCESS) |
| 421 |
{ |
| 422 |
MOS_OS_ASSERTMESSAGE("failed to unlock the status buf bundled w/ the specified gpucontext"); |
| 423 |
} |
| 424 |
m_statusBufferResource->Free(m_osContext, 0); |
| 425 |
MOS_Delete(m_statusBufferResource); |
| 426 |
} |
| 427 |
MOS_FreeMemAndSetNull(m_statusBufferMosResource); |
| 428 |
|
| 429 |
MosUtilities::MosLockMutex(m_cmdBufPoolMutex); |
| 430 |
|
| 431 |
if (m_cmdBufMgr) |
| 432 |
{ |
| 433 |
for (auto& curCommandBuffer : m_cmdBufPool) |
| 434 |
{ |
| 435 |
auto curCommandBufferSpecific = static_cast<CommandBufferSpecific *>(curCommandBuffer); |
| 436 |
if (curCommandBufferSpecific == nullptr) |
| 437 |
continue; |
| 438 |
curCommandBufferSpecific->waitReady(); // wait ready and return to comamnd buffer manager. |
| 439 |
m_cmdBufMgr->ReleaseCmdBuf(curCommandBuffer); |
| 440 |
} |
| 441 |
} |
| 442 |
|
| 443 |
m_cmdBufPool.clear(); |
| 444 |
|
| 445 |
MosUtilities::MosUnlockMutex(m_cmdBufPoolMutex); |
| 446 |
MosUtilities::MosDestroyMutex(m_cmdBufPoolMutex); |
| 447 |
m_cmdBufPoolMutex = nullptr; |
| 448 |
MOS_SafeFreeMemory(m_commandBuffer); |
| 449 |
MOS_SafeFreeMemory(m_allocationList); |
| 450 |
MOS_SafeFreeMemory(m_patchLocationList); |
| 451 |
MOS_SafeFreeMemory(m_attachedResources); |
| 452 |
MOS_SafeFreeMemory(m_writeModeList); |
| 453 |
MOS_SafeFreeMemory(m_createOptionEnhanced); |
| 454 |
|
| 455 |
for (int i=0; i<MAX_ENGINE_INSTANCE_NUM; i++) |
| 456 |
{ |
| 457 |
if (m_i915Context[i]) |
| 458 |
{ |
| 459 |
mos_context_destroy(m_i915Context[i]); |
| 460 |
m_i915Context[i] = nullptr; |
| 461 |
} |
| 462 |
} |
| 463 |
MOS_TraceEventExt(EVENT_GPU_CONTEXT_DESTROY, EVENT_TYPE_END, |
| 464 |
nullptr, 0, nullptr, 0); |
| 465 |
} |
| 466 |
|
| 467 |
MOS_STATUS GpuContextSpecific::RegisterResource( |
| 468 |
PMOS_RESOURCE osResource, |
| 469 |
bool writeFlag) |
| 470 |
{ |
| 471 |
MOS_OS_FUNCTION_ENTER; |
| 472 |
|
| 473 |
MOS_OS_CHK_NULL_RETURN(osResource); |
| 474 |
|
| 475 |
MOS_OS_CHK_NULL_RETURN(m_attachedResources); |
| 476 |
|
| 477 |
PMOS_RESOURCE registeredResources = m_attachedResources; |
| 478 |
uint32_t allocationIndex = 0; |
| 479 |
|
| 480 |
for ( allocationIndex = 0; allocationIndex < m_resCount; allocationIndex++, registeredResources++) |
| 481 |
{ |
| 482 |
if (osResource->bo == registeredResources->bo) |
| 483 |
{ |
| 484 |
break; |
| 485 |
} |
| 486 |
} |
| 487 |
|
| 488 |
// Allocation list to be updated |
| 489 |
if (allocationIndex < m_maxNumAllocations) |
| 490 |
{ |
| 491 |
// New buffer |
| 492 |
if (allocationIndex == m_resCount) |
| 493 |
{ |
| 494 |
m_resCount++; |
| 495 |
} |
| 496 |
|
| 497 |
// Set allocation |
| 498 |
if (m_gpuContext >= MOS_GPU_CONTEXT_MAX) |
| 499 |
{ |
| 500 |
MOS_OS_ASSERTMESSAGE("Gpu context exceeds max."); |
| 501 |
return MOS_STATUS_UNKNOWN; |
| 502 |
} |
| 503 |
|
| 504 |
osResource->iAllocationIndex[m_gpuContext] = (allocationIndex); |
| 505 |
m_attachedResources[allocationIndex] = *osResource; |
| 506 |
m_writeModeList[allocationIndex] |= writeFlag; |
| 507 |
m_allocationList[allocationIndex].hAllocation = &m_attachedResources[allocationIndex]; |
| 508 |
m_allocationList[allocationIndex].WriteOperation |= writeFlag; |
| 509 |
m_numAllocations = m_resCount; |
| 510 |
} |
| 511 |
else |
| 512 |
{ |
| 513 |
MOS_OS_ASSERTMESSAGE("Reached max # registrations."); |
| 514 |
return MOS_STATUS_UNKNOWN; |
| 515 |
} |
| 516 |
|
| 517 |
return MOS_STATUS_SUCCESS; |
| 518 |
} |
| 519 |
|
| 520 |
MOS_STATUS GpuContextSpecific::SetPatchEntry( |
| 521 |
PMOS_INTERFACE osInterface, |
| 522 |
PMOS_PATCH_ENTRY_PARAMS params) |
| 523 |
{ |
| 524 |
MOS_OS_FUNCTION_ENTER; |
| 525 |
|
| 526 |
MOS_OS_CHK_NULL_RETURN(m_patchLocationList); |
| 527 |
MOS_OS_CHK_NULL_RETURN(osInterface); |
| 528 |
MOS_OS_CHK_NULL_RETURN(params); |
| 529 |
|
| 530 |
m_patchLocationList[m_currentNumPatchLocations].AllocationIndex = params->uiAllocationIndex; |
| 531 |
m_patchLocationList[m_currentNumPatchLocations].AllocationOffset = params->uiResourceOffset; |
| 532 |
m_patchLocationList[m_currentNumPatchLocations].PatchOffset = params->uiPatchOffset; |
| 533 |
m_patchLocationList[m_currentNumPatchLocations].uiWriteOperation = params->bWrite ? true: false; |
| 534 |
m_patchLocationList[m_currentNumPatchLocations].cmdBo = |
| 535 |
params->cmdBuffer != nullptr ? params->cmdBuffer->OsResource.bo : nullptr; |
| 536 |
|
| 537 |
if (osInterface->osCpInterface && |
| 538 |
osInterface->osCpInterface->IsHMEnabled()) |
| 539 |
{ |
| 540 |
if (MOS_STATUS_SUCCESS != osInterface->osCpInterface->RegisterPatchForHM( |
| 541 |
(uint32_t *)(params->cmdBufBase + params->uiPatchOffset), |
| 542 |
params->bWrite, |
| 543 |
params->HwCommandType, |
| 544 |
params->forceDwordOffset, |
| 545 |
params->presResource, |
| 546 |
&m_patchLocationList[m_currentNumPatchLocations])) |
| 547 |
{ |
| 548 |
MOS_OS_ASSERTMESSAGE("Failed to RegisterPatchForHM."); |
| 549 |
} |
| 550 |
} |
| 551 |
|
| 552 |
m_currentNumPatchLocations++; |
| 553 |
|
| 554 |
return MOS_STATUS_SUCCESS; |
| 555 |
} |
| 556 |
|
| 557 |
MOS_STATUS GpuContextSpecific::GetCommandBuffer( |
| 558 |
PMOS_COMMAND_BUFFER comamndBuffer, |
| 559 |
uint32_t flags) |
| 560 |
{ |
| 561 |
MOS_OS_FUNCTION_ENTER; |
| 562 |
|
| 563 |
MOS_OS_CHK_NULL_RETURN(comamndBuffer); |
| 564 |
MOS_OS_CHK_NULL_RETURN(m_cmdBufMgr); |
| 565 |
MOS_OS_CHK_NULL_RETURN(m_commandBuffer); |
| 566 |
|
| 567 |
MOS_STATUS eStatus = MOS_STATUS_SUCCESS; |
| 568 |
CommandBuffer* cmdBuf = nullptr; |
| 569 |
|
| 570 |
uint32_t secondaryIdx = flags; |
| 571 |
bool isPrimaryCmdBuffer = (secondaryIdx == 0); |
| 572 |
bool hasSecondaryCmdBuffer = (!isPrimaryCmdBuffer && |
| 573 |
(m_secondaryCmdBufs.count(secondaryIdx) != 0)); |
| 574 |
|
| 575 |
bool needToAlloc = ((isPrimaryCmdBuffer && m_cmdBufFlushed) || |
| 576 |
(!isPrimaryCmdBuffer && !hasSecondaryCmdBuffer)); |
| 577 |
|
| 578 |
if (needToAlloc) |
| 579 |
{ |
| 580 |
MosUtilities::MosLockMutex(m_cmdBufPoolMutex); |
| 581 |
if (m_cmdBufPool.size() < MAX_CMD_BUF_NUM) |
| 582 |
{ |
| 583 |
cmdBuf = m_cmdBufMgr->PickupOneCmdBuf(m_commandBufferSize); |
| 584 |
if (cmdBuf == nullptr) |
| 585 |
{ |
| 586 |
MOS_OS_ASSERTMESSAGE("Invalid (nullptr) Pointer."); |
| 587 |
MosUtilities::MosUnlockMutex(m_cmdBufPoolMutex); |
| 588 |
return MOS_STATUS_NULL_POINTER; |
| 589 |
} |
| 590 |
if ((eStatus = cmdBuf->BindToGpuContext(this)) != MOS_STATUS_SUCCESS) |
| 591 |
{ |
| 592 |
MOS_OS_ASSERTMESSAGE("Invalid status of BindToGpuContext."); |
| 593 |
MosUtilities::MosUnlockMutex(m_cmdBufPoolMutex); |
| 594 |
return eStatus; |
| 595 |
} |
| 596 |
m_cmdBufPool.push_back(cmdBuf); |
| 597 |
} |
| 598 |
else if (m_cmdBufPool.size() == MAX_CMD_BUF_NUM && m_nextFetchIndex < m_cmdBufPool.size()) |
| 599 |
{ |
| 600 |
auto cmdBufOld = m_cmdBufPool[m_nextFetchIndex]; |
| 601 |
auto cmdBufSpecificOld = static_cast<CommandBufferSpecific *>(cmdBufOld); |
| 602 |
if (cmdBufSpecificOld == nullptr) |
| 603 |
{ |
| 604 |
MOS_OS_ASSERTMESSAGE("Invalid (nullptr) Pointer."); |
| 605 |
MosUtilities::MosUnlockMutex(m_cmdBufPoolMutex); |
| 606 |
return MOS_STATUS_NULL_POINTER; |
| 607 |
} |
| 608 |
cmdBufSpecificOld->waitReady(); |
| 609 |
cmdBufSpecificOld->UnBindToGpuContext(); |
| 610 |
m_cmdBufMgr->ReleaseCmdBuf(cmdBufOld); // here just return old command buffer to available pool |
| 611 |
|
| 612 |
//pick up new comamnd buffer |
| 613 |
cmdBuf = m_cmdBufMgr->PickupOneCmdBuf(m_commandBufferSize); |
| 614 |
if (cmdBuf == nullptr) |
| 615 |
{ |
| 616 |
MOS_OS_ASSERTMESSAGE("Invalid (nullptr) Pointer."); |
| 617 |
MosUtilities::MosUnlockMutex(m_cmdBufPoolMutex); |
| 618 |
return MOS_STATUS_NULL_POINTER; |
| 619 |
} |
| 620 |
if ((eStatus = cmdBuf->BindToGpuContext(this)) != MOS_STATUS_SUCCESS) |
| 621 |
{ |
| 622 |
MOS_OS_ASSERTMESSAGE("Invalid status of BindToGpuContext."); |
| 623 |
MosUtilities::MosUnlockMutex(m_cmdBufPoolMutex); |
| 624 |
return eStatus; |
| 625 |
} |
| 626 |
m_cmdBufPool[m_nextFetchIndex] = cmdBuf; |
| 627 |
} |
| 628 |
else |
| 629 |
{ |
| 630 |
MOS_OS_ASSERTMESSAGE("Command buffer bool size exceed max."); |
| 631 |
MosUtilities::MosUnlockMutex(m_cmdBufPoolMutex); |
| 632 |
return MOS_STATUS_UNKNOWN; |
| 633 |
} |
| 634 |
MosUtilities::MosUnlockMutex(m_cmdBufPoolMutex); |
| 635 |
|
| 636 |
// util now, we got new command buffer from CmdBufMgr, next step to fill in the input command buffer |
| 637 |
MOS_OS_CHK_STATUS_RETURN(cmdBuf->GetResource()->ConvertToMosResource(&comamndBuffer->OsResource)); |
| 638 |
comamndBuffer->pCmdBase = (uint32_t *)cmdBuf->GetLockAddr(); |
| 639 |
comamndBuffer->pCmdPtr = (uint32_t *)cmdBuf->GetLockAddr(); |
| 640 |
comamndBuffer->iOffset = 0; |
| 641 |
comamndBuffer->iRemaining = cmdBuf->GetCmdBufSize(); |
| 642 |
comamndBuffer->iCmdIndex = m_nextFetchIndex; |
| 643 |
comamndBuffer->iVdboxNodeIndex = MOS_VDBOX_NODE_INVALID; |
| 644 |
comamndBuffer->iVeboxNodeIndex = MOS_VEBOX_NODE_INVALID; |
| 645 |
comamndBuffer->is1stLvlBB = true; |
| 646 |
comamndBuffer->Attributes.pAttriVe = nullptr; |
| 647 |
|
| 648 |
// zero comamnd buffer |
| 649 |
MOS_ZeroMemory(comamndBuffer->pCmdBase, comamndBuffer->iRemaining); |
| 650 |
comamndBuffer->iSubmissionType = SUBMISSION_TYPE_SINGLE_PIPE; |
| 651 |
MOS_ZeroMemory(&comamndBuffer->Attributes,sizeof(comamndBuffer->Attributes)); |
| 652 |
|
| 653 |
if (isPrimaryCmdBuffer) |
| 654 |
{ |
| 655 |
// update command buffer relared filed in GPU context |
| 656 |
m_cmdBufFlushed = false; |
| 657 |
|
| 658 |
// keep a copy in GPU context |
| 659 |
MOS_SecureMemcpy(m_commandBuffer, sizeof(MOS_COMMAND_BUFFER), comamndBuffer, sizeof(MOS_COMMAND_BUFFER)); |
| 660 |
} |
| 661 |
else |
| 662 |
{ |
| 663 |
PMOS_COMMAND_BUFFER tempCmdBuf = (PMOS_COMMAND_BUFFER)MOS_AllocAndZeroMemory(sizeof(MOS_COMMAND_BUFFER)); |
| 664 |
MOS_OS_CHK_NULL_RETURN(tempCmdBuf); |
| 665 |
m_secondaryCmdBufs[secondaryIdx] = tempCmdBuf; |
| 666 |
MOS_SecureMemcpy(tempCmdBuf, sizeof(MOS_COMMAND_BUFFER), comamndBuffer, sizeof(MOS_COMMAND_BUFFER)); |
| 667 |
} |
| 668 |
|
| 669 |
// Command buffers are treated as cyclical buffers, the CB after the just submitted one |
| 670 |
// has the minimal fence value that we should wait |
| 671 |
m_nextFetchIndex++; |
| 672 |
if (m_nextFetchIndex >= MAX_CMD_BUF_NUM) |
| 673 |
{ |
| 674 |
m_nextFetchIndex = 0; |
| 675 |
} |
| 676 |
} |
| 677 |
else |
| 678 |
{ |
| 679 |
// current command buffer still active, directly copy to comamndBuffer |
| 680 |
if (isPrimaryCmdBuffer) |
| 681 |
{ |
| 682 |
MOS_SecureMemcpy(comamndBuffer, sizeof(MOS_COMMAND_BUFFER), m_commandBuffer, sizeof(MOS_COMMAND_BUFFER)); |
| 683 |
} |
| 684 |
else |
| 685 |
{ |
| 686 |
MOS_SecureMemcpy(comamndBuffer, sizeof(MOS_COMMAND_BUFFER), m_secondaryCmdBufs[secondaryIdx], sizeof(MOS_COMMAND_BUFFER)); |
| 687 |
} |
| 688 |
} |
| 689 |
|
| 690 |
if (isPrimaryCmdBuffer) |
| 691 |
{ |
| 692 |
MOS_OS_CHK_STATUS_RETURN(RegisterResource(&m_commandBuffer->OsResource, false)); |
| 693 |
} |
| 694 |
else |
| 695 |
{ |
| 696 |
MOS_OS_CHK_STATUS_RETURN(RegisterResource(&m_secondaryCmdBufs[secondaryIdx]->OsResource, false)); |
| 697 |
} |
| 698 |
|
| 699 |
return MOS_STATUS_SUCCESS; |
| 700 |
} |
| 701 |
|
| 702 |
void GpuContextSpecific::ReturnCommandBuffer( |
| 703 |
PMOS_COMMAND_BUFFER cmdBuffer, |
| 704 |
uint32_t flags) |
| 705 |
{ |
| 706 |
MOS_OS_FUNCTION_ENTER; |
| 707 |
|
| 708 |
MOS_OS_ASSERT(cmdBuffer); |
| 709 |
MOS_OS_ASSERT(m_commandBuffer); |
| 710 |
|
| 711 |
bool isPrimaryCmdBuf = (flags == 0); |
| 712 |
|
| 713 |
if (isPrimaryCmdBuf) |
| 714 |
{ |
| 715 |
m_commandBuffer->iOffset = cmdBuffer->iOffset; |
| 716 |
m_commandBuffer->iRemaining = cmdBuffer->iRemaining; |
| 717 |
m_commandBuffer->pCmdPtr = cmdBuffer->pCmdPtr; |
| 718 |
m_commandBuffer->iVdboxNodeIndex = cmdBuffer->iVdboxNodeIndex; |
| 719 |
m_commandBuffer->iVeboxNodeIndex = cmdBuffer->iVeboxNodeIndex; |
| 720 |
} |
| 721 |
else |
| 722 |
{ |
| 723 |
uint32_t secondaryIdx = flags; |
| 724 |
MOS_OS_ASSERT(m_secondaryCmdBufs.count(secondaryIdx)); |
| 725 |
|
| 726 |
MOS_SecureMemcpy(m_secondaryCmdBufs[secondaryIdx], sizeof(MOS_COMMAND_BUFFER), cmdBuffer, sizeof(MOS_COMMAND_BUFFER)); |
| 727 |
} |
| 728 |
} |
| 729 |
|
| 730 |
MOS_STATUS GpuContextSpecific::ResetCommandBuffer() |
| 731 |
{ |
| 732 |
m_cmdBufFlushed = true; |
| 733 |
auto it = m_secondaryCmdBufs.begin(); |
| 734 |
while(it != m_secondaryCmdBufs.end()) |
| 735 |
{ |
| 736 |
MOS_FreeMemory(it->second); |
| 737 |
it++; |
| 738 |
} |
| 739 |
m_secondaryCmdBufs.clear(); |
| 740 |
return MOS_STATUS_SUCCESS; |
| 741 |
} |
| 742 |
|
| 743 |
MOS_STATUS GpuContextSpecific::SetIndirectStateSize(const uint32_t size) |
| 744 |
{ |
| 745 |
if(size < m_commandBufferSize) |
| 746 |
{ |
| 747 |
m_IndirectHeapSize = size; |
| 748 |
return MOS_STATUS_SUCCESS; |
| 749 |
} |
| 750 |
else |
| 751 |
{ |
| 752 |
MOS_OS_ASSERTMESSAGE("Indirect State Size if out of boundry!"); |
| 753 |
return MOS_STATUS_UNKNOWN; |
| 754 |
} |
| 755 |
} |
| 756 |
|
| 757 |
MOS_STATUS GpuContextSpecific::GetIndirectState( |
| 758 |
uint32_t *offset, |
| 759 |
uint32_t *size) |
| 760 |
{ |
| 761 |
MOS_OS_FUNCTION_ENTER; |
| 762 |
|
| 763 |
if (offset) |
| 764 |
{ |
| 765 |
*offset = m_commandBufferSize - m_IndirectHeapSize; |
| 766 |
} |
| 767 |
|
| 768 |
if (size) |
| 769 |
{ |
| 770 |
*size = m_IndirectHeapSize; |
| 771 |
} |
| 772 |
|
| 773 |
return MOS_STATUS_SUCCESS; |
| 774 |
} |
| 775 |
|
| 776 |
MOS_STATUS GpuContextSpecific::GetIndirectStatePointer( |
| 777 |
uint8_t **indirectState) |
| 778 |
{ |
| 779 |
MOS_OS_FUNCTION_ENTER; |
| 780 |
|
| 781 |
MOS_OS_CHK_NULL_RETURN(indirectState); |
| 782 |
|
| 783 |
*indirectState = (uint8_t *)m_commandBuffer->pCmdBase + m_commandBufferSize - m_IndirectHeapSize; |
| 784 |
|
| 785 |
return MOS_STATUS_SUCCESS; |
| 786 |
} |
| 787 |
|
| 788 |
MOS_STATUS GpuContextSpecific::ResizeCommandBufferAndPatchList( |
| 789 |
uint32_t requestedCommandBufferSize, |
| 790 |
uint32_t requestedPatchListSize, |
| 791 |
uint32_t flags) |
| 792 |
{ |
| 793 |
MOS_OS_FUNCTION_ENTER; |
| 794 |
|
| 795 |
// m_commandBufferSize is used for allocate command buffer and submit command buffer, in this moment, command buffer has not allocated yet. |
| 796 |
// Linux KMD requires command buffer size align to 8 bytes, or it will not execute the commands. |
| 797 |
m_commandBufferSize = MOS_ALIGN_CEIL(requestedCommandBufferSize, 8); |
| 798 |
|
| 799 |
if (requestedPatchListSize > m_maxPatchLocationsize) |
| 800 |
{ |
| 801 |
PPATCHLOCATIONLIST newPatchList = (PPATCHLOCATIONLIST)realloc(m_patchLocationList, sizeof(PATCHLOCATIONLIST) * requestedPatchListSize); |
| 802 |
MOS_OS_CHK_NULL_RETURN(newPatchList); |
| 803 |
|
| 804 |
m_patchLocationList = newPatchList; |
| 805 |
|
| 806 |
// now zero the extended portion |
| 807 |
MOS_ZeroMemory((m_patchLocationList + m_maxPatchLocationsize), sizeof(PATCHLOCATIONLIST) * (requestedPatchListSize - m_maxPatchLocationsize)); |
| 808 |
m_maxPatchLocationsize = requestedPatchListSize; |
| 809 |
} |
| 810 |
|
| 811 |
return MOS_STATUS_SUCCESS; |
| 812 |
} |
| 813 |
|
| 814 |
MOS_STATUS GpuContextSpecific::ResizeCommandBuffer(uint32_t requestedSize) |
| 815 |
{ |
| 816 |
MOS_OS_FUNCTION_ENTER; |
| 817 |
|
| 818 |
m_commandBufferSize = requestedSize; |
| 819 |
|
| 820 |
return MOS_STATUS_SUCCESS; |
| 821 |
} |
| 822 |
|
| 823 |
uint32_t GetVcsExecFlag(PMOS_INTERFACE osInterface, |
| 824 |
PMOS_COMMAND_BUFFER cmdBuffer, |
| 825 |
MOS_GPU_NODE gpuNode) |
| 826 |
{ |
| 827 |
if (osInterface == 0 || |
| 828 |
cmdBuffer == 0) |
| 829 |
{ |
| 830 |
MOS_OS_ASSERTMESSAGE("Input invalid(null) parameter."); |
| 831 |
return I915_EXEC_DEFAULT; |
| 832 |
} |
| 833 |
|
| 834 |
uint32_t vcsExecFlag = I915_EXEC_BSD | I915_EXEC_BSD_RING1; |
| 835 |
|
| 836 |
if (MOS_VDBOX_NODE_INVALID == cmdBuffer->iVdboxNodeIndex) |
| 837 |
{ |
| 838 |
// That's those case when BB did not have any VDBOX# specific commands. |
| 839 |
// Thus, we need to select VDBOX# here. Alternatively we can rely on KMD |
| 840 |
// to make balancing for us, i.e. rely on Virtual Engine support. |
| 841 |
cmdBuffer->iVdboxNodeIndex = osInterface->pfnGetVdboxNodeId(osInterface, cmdBuffer); |
| 842 |
if (MOS_VDBOX_NODE_INVALID == cmdBuffer->iVdboxNodeIndex) |
| 843 |
{ |
| 844 |
cmdBuffer->iVdboxNodeIndex = (gpuNode == MOS_GPU_NODE_VIDEO)? |
| 845 |
MOS_VDBOX_NODE_1: MOS_VDBOX_NODE_2; |
| 846 |
} |
| 847 |
} |
| 848 |
|
| 849 |
if (MOS_VDBOX_NODE_1 == cmdBuffer->iVdboxNodeIndex) |
| 850 |
{ |
| 851 |
vcsExecFlag = I915_EXEC_BSD | I915_EXEC_BSD_RING1; |
| 852 |
} |
| 853 |
else if (MOS_VDBOX_NODE_2 == cmdBuffer->iVdboxNodeIndex) |
| 854 |
{ |
| 855 |
vcsExecFlag = I915_EXEC_BSD | I915_EXEC_BSD_RING2; |
| 856 |
} |
| 857 |
|
| 858 |
return vcsExecFlag; |
| 859 |
} |
| 860 |
|
| 861 |
MOS_STATUS GpuContextSpecific::MapResourcesToAuxTable(mos_linux_bo *cmd_bo) |
| 862 |
{ |
| 863 |
MOS_OS_CHK_NULL_RETURN(cmd_bo); |
| 864 |
|
| 865 |
OsContextSpecific *osCtx = static_cast<OsContextSpecific*>(m_osContext); |
| 866 |
MOS_OS_CHK_NULL_RETURN(osCtx); |
| 867 |
|
| 868 |
AuxTableMgr *auxTableMgr = osCtx->GetAuxTableMgr(); |
| 869 |
if (auxTableMgr) |
| 870 |
{ |
| 871 |
// Map compress allocations to aux table if it is not mapped. |
| 872 |
for (uint32_t i = 0; i < m_numAllocations; i++) |
| 873 |
{ |
| 874 |
auto res = (PMOS_RESOURCE)m_allocationList[i].hAllocation; |
| 875 |
MOS_OS_CHK_NULL_RETURN(res); |
| 876 |
MOS_OS_CHK_STATUS_RETURN(auxTableMgr->MapResource(res->pGmmResInfo, res->bo)); |
| 877 |
} |
| 878 |
MOS_OS_CHK_STATUS_RETURN(auxTableMgr->EmitAuxTableBOList(cmd_bo)); |
| 879 |
} |
| 880 |
return MOS_STATUS_SUCCESS; |
| 881 |
} |
| 882 |
|
| 883 |
MOS_STATUS GpuContextSpecific::SubmitCommandBuffer( |
| 884 |
PMOS_INTERFACE osInterface, |
| 885 |
PMOS_COMMAND_BUFFER cmdBuffer, |
| 886 |
bool nullRendering) |
| 887 |
{ |
| 888 |
MOS_OS_FUNCTION_ENTER; |
| 889 |
|
| 890 |
MOS_TraceEventExt(EVENT_MOS_BATCH_SUBMIT, EVENT_TYPE_START, nullptr, 0, nullptr, 0); |
| 891 |
|
| 892 |
MOS_OS_CHK_NULL_RETURN(osInterface); |
| 893 |
PMOS_CONTEXT osContext = osInterface->pOsContext; |
| 894 |
MOS_OS_CHK_NULL_RETURN(osContext); |
| 895 |
MOS_OS_CHK_NULL_RETURN(cmdBuffer); |
| 896 |
MOS_OS_CHK_NULL_RETURN(m_patchLocationList); |
| 897 |
|
| 898 |
MOS_GPU_NODE gpuNode = OSKMGetGpuNode(m_gpuContext); |
| 899 |
uint32_t execFlag = gpuNode; |
| 900 |
MOS_STATUS eStatus = MOS_STATUS_SUCCESS; |
| 901 |
int32_t ret = 0; |
| 902 |
bool scalaEnabled = false; |
| 903 |
auto it = m_secondaryCmdBufs.begin(); |
| 904 |
|
| 905 |
// Command buffer object DRM pointer |
| 906 |
m_cmdBufFlushed = true; |
| 907 |
auto cmd_bo = cmdBuffer->OsResource.bo; |
| 908 |
|
| 909 |
// Map Resource to Aux if needed |
| 910 |
MapResourcesToAuxTable(cmd_bo); |
| 911 |
for(auto it : m_secondaryCmdBufs) |
| 912 |
{ |
| 913 |
MapResourcesToAuxTable(it.second->OsResource.bo); |
| 914 |
} |
| 915 |
|
| 916 |
if (m_secondaryCmdBufs.size() >= 2) |
| 917 |
{ |
| 918 |
scalaEnabled = true; |
| 919 |
cmdBuffer->iSubmissionType = SUBMISSION_TYPE_MULTI_PIPE_MASTER; |
| 920 |
} |
| 921 |
|
| 922 |
std::vector<PMOS_RESOURCE> mappedResList; |
| 923 |
std::vector<MOS_LINUX_BO *> skipSyncBoList; |
| 924 |
|
| 925 |
// Now, the patching will be done, based on the patch list. |
| 926 |
for (uint32_t patchIndex = 0; patchIndex < m_currentNumPatchLocations; patchIndex++) |
| 927 |
{ |
| 928 |
auto currentPatch = &m_patchLocationList[patchIndex]; |
| 929 |
MOS_OS_CHK_NULL_RETURN(currentPatch); |
| 930 |
|
| 931 |
auto tempCmdBo = currentPatch->cmdBo == nullptr ? cmd_bo : currentPatch->cmdBo; |
| 932 |
|
| 933 |
// Following are for Nested BB buffer, if it's nested BB, we need to ensure it's locked. |
| 934 |
if (tempCmdBo != cmd_bo) |
| 935 |
{ |
| 936 |
bool isSecondaryCmdBuf = false; |
| 937 |
it = m_secondaryCmdBufs.begin(); |
| 938 |
while(it != m_secondaryCmdBufs.end()) |
| 939 |
{ |
| 940 |
if (it->second->OsResource.bo == tempCmdBo) |
| 941 |
{ |
| 942 |
isSecondaryCmdBuf = true; |
| 943 |
break; |
| 944 |
} |
| 945 |
it++; |
| 946 |
} |
| 947 |
|
| 948 |
for(auto allocIdx = 0; allocIdx < m_numAllocations && (!isSecondaryCmdBuf); allocIdx++) |
| 949 |
{ |
| 950 |
auto tempRes = (PMOS_RESOURCE)m_allocationList[allocIdx].hAllocation; |
| 951 |
if (tempCmdBo == tempRes->bo) |
| 952 |
{ |
| 953 |
GraphicsResource::LockParams param; |
| 954 |
param.m_writeRequest = true; |
| 955 |
tempRes->pGfxResource->Lock(m_osContext, param); |
| 956 |
mappedResList.push_back(tempRes); |
| 957 |
break; |
| 958 |
} |
| 959 |
} |
| 960 |
} |
| 961 |
|
| 962 |
// This is the resource for which patching will be done |
| 963 |
auto resource = (PMOS_RESOURCE)m_allocationList[currentPatch->AllocationIndex].hAllocation; |
| 964 |
MOS_OS_CHK_NULL_RETURN(resource); |
| 965 |
|
| 966 |
// For now, we'll assume the system memory's DRM bo pointer |
| 967 |
// is NULL. If nullptr is detected, then the resource has been |
| 968 |
// placed inside the command buffer's indirect state area. |
| 969 |
// We'll simply set alloc_bo to the command buffer's bo pointer. |
| 970 |
MOS_OS_ASSERT(resource->bo); |
| 971 |
|
| 972 |
auto alloc_bo = (resource->bo) ? resource->bo : tempCmdBo; |
| 973 |
|
| 974 |
MOS_OS_CHK_STATUS_RETURN(osInterface->osCpInterface->PermeatePatchForHM( |
| 975 |
tempCmdBo->virt, |
| 976 |
currentPatch, |
| 977 |
resource)); |
| 978 |
|
| 979 |
uint64_t boOffset = alloc_bo->offset64; |
| 980 |
if (alloc_bo != tempCmdBo) |
| 981 |
{ |
| 982 |
auto item_ctx = osContext->contextOffsetList.begin(); |
| 983 |
for (; item_ctx != osContext->contextOffsetList.end(); item_ctx++) |
| 984 |
{ |
| 985 |
if (item_ctx->intel_context == osContext->intel_context && item_ctx->target_bo == alloc_bo) |
| 986 |
{ |
| 987 |
boOffset = item_ctx->offset64; |
| 988 |
break; |
| 989 |
} |
| 990 |
} |
| 991 |
} |
| 992 |
|
| 993 |
MOS_OS_CHK_NULL_RETURN(tempCmdBo->virt); |
| 994 |
if (osContext->bUse64BitRelocs) |
| 995 |
{ |
| 996 |
*((uint64_t *)((uint8_t *)tempCmdBo->virt + currentPatch->PatchOffset)) = |
| 997 |
boOffset + currentPatch->AllocationOffset; |
| 998 |
} |
| 999 |
else |
| 1000 |
{ |
| 1001 |
*((uint32_t *)((uint8_t *)tempCmdBo->virt + currentPatch->PatchOffset)) = |
| 1002 |
boOffset + currentPatch->AllocationOffset; |
| 1003 |
} |
| 1004 |
|
| 1005 |
if (scalaEnabled) |
| 1006 |
{ |
| 1007 |
it = m_secondaryCmdBufs.begin(); |
| 1008 |
while(it != m_secondaryCmdBufs.end()) |
| 1009 |
{ |
| 1010 |
if (it->second->OsResource.bo == tempCmdBo && |
| 1011 |
it->second->iSubmissionType & SUBMISSION_TYPE_MULTI_PIPE_SLAVE && |
| 1012 |
!mos_bo_is_exec_object_async(alloc_bo)) |
| 1013 |
{ |
| 1014 |
skipSyncBoList.push_back(alloc_bo); |
| 1015 |
break; |
| 1016 |
} |
| 1017 |
it++; |
| 1018 |
} |
| 1019 |
} |
| 1020 |
else if (cmdBuffer->iSubmissionType & SUBMISSION_TYPE_MULTI_PIPE_SLAVE && |
| 1021 |
!mos_bo_is_exec_object_async(alloc_bo)) |
| 1022 |
{ |
| 1023 |
skipSyncBoList.push_back(alloc_bo); |
| 1024 |
} |
| 1025 |
|
| 1026 |
#if (_DEBUG || _RELEASE_INTERNAL) |
| 1027 |
{ |
| 1028 |
uint32_t evtData[] = {alloc_bo->handle, currentPatch->uiWriteOperation, currentPatch->AllocationOffset}; |
| 1029 |
MOS_TraceEventExt(EVENT_MOS_BATCH_SUBMIT, EVENT_TYPE_INFO, |
| 1030 |
evtData, sizeof(evtData), |
| 1031 |
&boOffset, sizeof(boOffset)); |
| 1032 |
} |
| 1033 |
#endif |
| 1034 |
if(mos_bo_is_softpin(alloc_bo)) |
| 1035 |
{ |
| 1036 |
if (alloc_bo != tempCmdBo) |
| 1037 |
{ |
| 1038 |
ret = mos_bo_add_softpin_target(tempCmdBo, alloc_bo, currentPatch->uiWriteOperation); |
| 1039 |
} |
| 1040 |
} |
| 1041 |
else |
| 1042 |
{ |
| 1043 |
// This call will patch the command buffer with the offsets of the indirect state region of the command buffer |
| 1044 |
ret = mos_bo_emit_reloc( |
| 1045 |
tempCmdBo, // Command buffer |
| 1046 |
currentPatch->PatchOffset, // Offset in the command buffer |
| 1047 |
alloc_bo, // Allocation object for which the patch will be made. |
| 1048 |
currentPatch->AllocationOffset, // Offset to the indirect state |
| 1049 |
I915_GEM_DOMAIN_RENDER, // Read domain |
| 1050 |
(currentPatch->uiWriteOperation) ? I915_GEM_DOMAIN_RENDER : 0x0, // Write domain |
| 1051 |
boOffset); |
| 1052 |
} |
| 1053 |
|
| 1054 |
if (ret != 0) |
| 1055 |
{ |
| 1056 |
MOS_OS_ASSERTMESSAGE("Error patching alloc_bo = 0x%x, cmd_bo = 0x%x.", |
| 1057 |
(uintptr_t)alloc_bo, |
| 1058 |
(uintptr_t)tempCmdBo); |
| 1059 |
return MOS_STATUS_UNKNOWN; |
| 1060 |
} |
| 1061 |
} |
| 1062 |
|
| 1063 |
for(auto res: mappedResList) |
| 1064 |
{ |
| 1065 |
res->pGfxResource->Unlock(m_osContext); |
| 1066 |
} |
| 1067 |
mappedResList.clear(); |
| 1068 |
|
| 1069 |
if (scalaEnabled) |
| 1070 |
{ |
| 1071 |
it = m_secondaryCmdBufs.begin(); |
| 1072 |
while(it != m_secondaryCmdBufs.end()) |
| 1073 |
{ |
| 1074 |
//Add Batch buffer End Command |
| 1075 |
uint32_t batchBufferEndCmd = MI_BATCHBUFFER_END; |
| 1076 |
if (MOS_FAILED(Mos_AddCommand( |
| 1077 |
it->second, |
| 1078 |
&batchBufferEndCmd, |
| 1079 |
sizeof(uint32_t)))) |
| 1080 |
{ |
| 1081 |
MOS_OS_ASSERTMESSAGE("Inserting BB_END failed!"); |
| 1082 |
return MOS_STATUS_UNKNOWN; |
| 1083 |
} |
| 1084 |
it++; |
| 1085 |
} |
| 1086 |
} |
| 1087 |
else |
| 1088 |
{ |
| 1089 |
//Add Batch buffer End Command |
| 1090 |
uint32_t batchBufferEndCmd = MI_BATCHBUFFER_END; |
| 1091 |
if (MOS_FAILED(Mos_AddCommand( |
| 1092 |
cmdBuffer, |
| 1093 |
&batchBufferEndCmd, |
| 1094 |
sizeof(uint32_t)))) |
| 1095 |
{ |
| 1096 |
MOS_OS_ASSERTMESSAGE("Inserting BB_END failed!"); |
| 1097 |
return MOS_STATUS_UNKNOWN; |
| 1098 |
} |
| 1099 |
} |
| 1100 |
// dump before cmd buffer unmap |
| 1101 |
MOS_TraceDumpExt("CmdBuffer", m_gpuContext, cmdBuffer->pCmdBase, cmdBuffer->iOffset); |
| 1102 |
|
| 1103 |
// Now, we can unmap the video command buffer, since we don't need CPU access anymore. |
| 1104 |
MOS_OS_CHK_NULL_RETURN(cmdBuffer->OsResource.pGfxResource); |
| 1105 |
cmdBuffer->OsResource.pGfxResource->Unlock(m_osContext); |
| 1106 |
|
| 1107 |
it = m_secondaryCmdBufs.begin(); |
| 1108 |
while(it != m_secondaryCmdBufs.end()) |
| 1109 |
{ |
| 1110 |
MOS_OS_CHK_NULL_RETURN(it->second->OsResource.pGfxResource); |
| 1111 |
it->second->OsResource.pGfxResource->Unlock(m_osContext); |
| 1112 |
|
| 1113 |
it++; |
| 1114 |
} |
| 1115 |
|
| 1116 |
int32_t perfData; |
| 1117 |
if (osContext->pPerfData != nullptr) |
| 1118 |
{ |
| 1119 |
perfData = *(int32_t *)(osContext->pPerfData); |
| 1120 |
} |
| 1121 |
else |
| 1122 |
{ |
| 1123 |
perfData = 0; |
| 1124 |
} |
| 1125 |
|
| 1126 |
drm_clip_rect_t *cliprects = nullptr; |
| 1127 |
int32_t num_cliprects = 0; |
| 1128 |
int32_t DR4 = osContext->uEnablePerfTag ? perfData : 0; |
| 1129 |
|
| 1130 |
//Since CB2 command is not supported, remove it and set cliprects to nullprt as default. |
| 1131 |
if ((gpuNode == MOS_GPU_NODE_VIDEO || gpuNode == MOS_GPU_NODE_VIDEO2) && |
| 1132 |
(cmdBuffer->iSubmissionType & SUBMISSION_TYPE_SINGLE_PIPE_MASK)) |
| 1133 |
{ |
| 1134 |
if (osContext->bKMDHasVCS2) |
| 1135 |
{ |
| 1136 |
if (osContext->bPerCmdBufferBalancing && osInterface->pfnGetVdboxNodeId) |
| 1137 |
{ |
| 1138 |
execFlag = GetVcsExecFlag(osInterface, cmdBuffer, gpuNode); |
| 1139 |
} |
| 1140 |
else if (gpuNode == MOS_GPU_NODE_VIDEO) |
| 1141 |
{ |
| 1142 |
execFlag = I915_EXEC_BSD | I915_EXEC_BSD_RING1; |
| 1143 |
} |
| 1144 |
else if (gpuNode == MOS_GPU_NODE_VIDEO2) |
| 1145 |
{ |
| 1146 |
execFlag = I915_EXEC_BSD | I915_EXEC_BSD_RING2; |
| 1147 |
} |
| 1148 |
else |
| 1149 |
{ |
| 1150 |
MOS_OS_ASSERTMESSAGE("Invalid gpuNode."); |
| 1151 |
} |
| 1152 |
} |
| 1153 |
else |
| 1154 |
{ |
| 1155 |
execFlag = I915_EXEC_BSD | I915_EXEC_BSD_RING1; |
| 1156 |
} |
| 1157 |
} |
| 1158 |
|
| 1159 |
#if (_DEBUG || _RELEASE_INTERNAL) |
| 1160 |
|
| 1161 |
MOS_LINUX_BO *bad_cmd_bo = nullptr; |
| 1162 |
MOS_LINUX_BO *nop_cmd_bo = nullptr; |
| 1163 |
uint32_t dwComponentTag = 0; |
| 1164 |
uint32_t dwCallType = 0; |
| 1165 |
|
| 1166 |
//dwComponentTag 3: decode,5: vpp,6: encode |
| 1167 |
//dwCallType 8: PAK(CODECHAL_ENCODE_PERFTAG_CALL_PAK_ENGINE) |
| 1168 |
// 34: PREENC |
| 1169 |
// 5: VPP |
| 1170 |
dwComponentTag = (perfData & 0xF000) >> 12; |
| 1171 |
dwCallType = (perfData & 0xFC) >> 2; |
| 1172 |
|
| 1173 |
if (osInterface->bTriggerCodecHang && |
| 1174 |
(dwComponentTag == 3 || (dwComponentTag == 6 && dwCallType == 8) || |
| 1175 |
(dwComponentTag == 6 && dwCallType == 34) || |
| 1176 |
(dwComponentTag == 5 && dwCallType == 5))) |
| 1177 |
{ |
| 1178 |
bad_cmd_bo = Mos_GetBadCommandBuffer_Linux(osInterface); |
| 1179 |
if (bad_cmd_bo) |
| 1180 |
{ |
| 1181 |
ret = mos_bo_mrb_exec(bad_cmd_bo, |
| 1182 |
4096, |
| 1183 |
nullptr, |
| 1184 |
0, |
| 1185 |
0, |
| 1186 |
execFlag); |
| 1187 |
} |
| 1188 |
else |
| 1189 |
{ |
| 1190 |
MOS_OS_ASSERTMESSAGE("Mos_GetBadCommandBuffer_Linux failed!"); |
| 1191 |
} |
| 1192 |
} |
| 1193 |
else if (osInterface->bTriggerVPHang == true) |
| 1194 |
{ |
| 1195 |
bad_cmd_bo = Mos_GetBadCommandBuffer_Linux(osInterface); |
| 1196 |
|
| 1197 |
if (bad_cmd_bo) |
| 1198 |
{ |
| 1199 |
ret = mos_bo_mrb_exec(bad_cmd_bo, |
| 1200 |
4096, |
| 1201 |
nullptr, |
| 1202 |
0, |
| 1203 |
0, |
| 1204 |
execFlag); |
| 1205 |
} |
| 1206 |
else |
| 1207 |
{ |
| 1208 |
MOS_OS_ASSERTMESSAGE("Mos_GetBadCommandBuffer_Linux failed!"); |
| 1209 |
} |
| 1210 |
|
| 1211 |
osInterface->bTriggerVPHang = false; |
| 1212 |
} |
| 1213 |
|
| 1214 |
nop_cmd_bo = nullptr; |
| 1215 |
if (nullRendering == true) |
| 1216 |
{ |
| 1217 |
nop_cmd_bo = Mos_GetNopCommandBuffer_Linux(osInterface); |
| 1218 |
|
| 1219 |
if (nop_cmd_bo) |
| 1220 |
{ |
| 1221 |
ret = mos_bo_mrb_exec(nop_cmd_bo, |
| 1222 |
4096, |
| 1223 |
nullptr, |
| 1224 |
0, |
| 1225 |
0, |
| 1226 |
execFlag); |
| 1227 |
} |
| 1228 |
else |
| 1229 |
{ |
| 1230 |
MOS_OS_ASSERTMESSAGE("Mos_GetNopCommandBuffer_Linux failed!"); |
| 1231 |
} |
| 1232 |
} |
| 1233 |
|
| 1234 |
#endif //(_DEBUG || _RELEASE_INTERNAL) |
| 1235 |
|
| 1236 |
if (gpuNode != I915_EXEC_RENDER && |
| 1237 |
osInterface->osCpInterface->IsTearDownHappen()) |
| 1238 |
{ |
| 1239 |
// skip PAK command when CP tear down happen to avoid of GPU hang |
| 1240 |
// conditonal batch buffer start PoC is in progress |
| 1241 |
} |
| 1242 |
else if (nullRendering == false) |
| 1243 |
{ |
| 1244 |
UnlockPendingOcaBuffers(cmdBuffer, osContext); |
| 1245 |
if (osInterface->ctxBasedScheduling && m_i915Context[0] != nullptr) |
| 1246 |
{ |
| 1247 |
if (cmdBuffer->iSubmissionType & SUBMISSION_TYPE_MULTI_PIPE_MASK) |
| 1248 |
{ |
| 1249 |
if (scalaEnabled && !osInterface->bParallelSubmission) |
| 1250 |
{ |
| 1251 |
uint32_t secondaryIndex = 0; |
| 1252 |
it = m_secondaryCmdBufs.begin(); |
| 1253 |
while(it != m_secondaryCmdBufs.end()) |
| 1254 |
{ |
| 1255 |
if (it->second->iSubmissionType & SUBMISSION_TYPE_MULTI_PIPE_SLAVE) |
| 1256 |
{ |
| 1257 |
if(execFlag == MOS_GPU_NODE_VE) |
| 1258 |
{ |
| 1259 |
// decode excluded since init in other place |
| 1260 |
it->second->iSubmissionType |= (secondaryIndex << SUBMISSION_TYPE_MULTI_PIPE_SLAVE_INDEX_SHIFT); |
| 1261 |
secondaryIndex++; |
| 1262 |
} |
| 1263 |
} |
| 1264 |
|
| 1265 |
ret = SubmitPipeCommands(it->second, |
| 1266 |
it->second->OsResource.bo, |
| 1267 |
osContext, |
| 1268 |
skipSyncBoList, |
| 1269 |
execFlag, |
| 1270 |
DR4); |
| 1271 |
it++; |
| 1272 |
} |
| 1273 |
} |
| 1274 |
else if(scalaEnabled && osInterface->bParallelSubmission) |
| 1275 |
{ |
| 1276 |
ret = ParallelSubmitCommands(m_secondaryCmdBufs, |
| 1277 |
osContext, |
| 1278 |
execFlag, |
| 1279 |
DR4); |
| 1280 |
} |
| 1281 |
else |
| 1282 |
{ |
| 1283 |
ret = SubmitPipeCommands(cmdBuffer, |
| 1284 |
cmd_bo, |
| 1285 |
osContext, |
| 1286 |
skipSyncBoList, |
| 1287 |
execFlag, |
| 1288 |
DR4); |
| 1289 |
} |
| 1290 |
} |
| 1291 |
else |
| 1292 |
{ |
| 1293 |
ret = mos_bo_context_exec2(cmd_bo, |
| 1294 |
m_commandBufferSize, |
| 1295 |
m_i915Context[0], |
| 1296 |
cliprects, |
| 1297 |
num_cliprects, |
| 1298 |
DR4, |
| 1299 |
m_i915ExecFlag, |
| 1300 |
nullptr); |
| 1301 |
} |
| 1302 |
} |
| 1303 |
else |
| 1304 |
{ |
| 1305 |
ret = mos_bo_context_exec2(cmd_bo, |
| 1306 |
m_commandBufferSize, |
| 1307 |
osContext->intel_context, |
| 1308 |
cliprects, |
| 1309 |
num_cliprects, |
| 1310 |
DR4, |
| 1311 |
execFlag, |
| 1312 |
nullptr); |
| 1313 |
} |
| 1314 |
if (ret != 0) |
| 1315 |
{ |
| 1316 |
eStatus = MOS_STATUS_UNKNOWN; |
| 1317 |
} |
| 1318 |
} |
| 1319 |
|
| 1320 |
if (eStatus != MOS_STATUS_SUCCESS) |
| 1321 |
{ |
| 1322 |
MOS_OS_ASSERTMESSAGE("Command buffer submission failed!"); |
| 1323 |
} |
| 1324 |
|
| 1325 |
MOS_DEVULT_FuncCall(pfnUltGetCmdBuf, cmdBuffer); |
| 1326 |
|
| 1327 |
#if MOS_COMMAND_BUFFER_DUMP_SUPPORTED |
| 1328 |
pthread_mutex_lock(&command_dump_mutex); |
| 1329 |
if (osInterface->bDumpCommandBuffer) |
| 1330 |
{ |
| 1331 |
if (scalaEnabled) |
| 1332 |
{ |
| 1333 |
it = m_secondaryCmdBufs.begin(); |
| 1334 |
while(it != m_secondaryCmdBufs.end()) |
| 1335 |
{ |
| 1336 |
mos_bo_map(it->second->OsResource.bo, 0); |
| 1337 |
osInterface->pfnDumpCommandBuffer(osInterface, it->second); |
| 1338 |
mos_bo_unmap(it->second->OsResource.bo); |
| 1339 |
it++; |
| 1340 |
} |
| 1341 |
} |
| 1342 |
else |
| 1343 |
{ |
| 1344 |
mos_bo_map(cmd_bo, 0); |
| 1345 |
osInterface->pfnDumpCommandBuffer(osInterface, cmdBuffer); |
| 1346 |
mos_bo_unmap(cmd_bo); |
| 1347 |
} |
| 1348 |
} |
| 1349 |
pthread_mutex_unlock(&command_dump_mutex); |
| 1350 |
#endif // MOS_COMMAND_BUFFER_DUMP_SUPPORTED |
| 1351 |
|
| 1352 |
#if (_DEBUG || _RELEASE_INTERNAL) |
| 1353 |
if (bad_cmd_bo) |
| 1354 |
{ |
| 1355 |
mos_bo_wait_rendering(bad_cmd_bo); |
| 1356 |
mos_bo_unreference(bad_cmd_bo); |
| 1357 |
} |
| 1358 |
if (nop_cmd_bo) |
| 1359 |
{ |
| 1360 |
mos_bo_unreference(nop_cmd_bo); |
| 1361 |
} |
| 1362 |
#endif //(_DEBUG || _RELEASE_INTERNAL) |
| 1363 |
|
| 1364 |
//clear command buffer relocations to fix memory leak issue |
| 1365 |
mos_bo_clear_relocs(cmd_bo, 0); |
| 1366 |
it = m_secondaryCmdBufs.begin(); |
| 1367 |
while(it != m_secondaryCmdBufs.end()) |
| 1368 |
{ |
| 1369 |
mos_bo_clear_relocs(it->second->OsResource.bo, 0); |
| 1370 |
MOS_FreeMemory(it->second); |
| 1371 |
it++; |
| 1372 |
} |
| 1373 |
m_secondaryCmdBufs.clear(); |
| 1374 |
skipSyncBoList.clear(); |
| 1375 |
|
| 1376 |
// Reset resource allocation |
| 1377 |
m_numAllocations = 0; |
| 1378 |
MOS_ZeroMemory(m_allocationList, sizeof(ALLOCATION_LIST) * m_maxNumAllocations); |
| 1379 |
m_currentNumPatchLocations = 0; |
| 1380 |
MOS_ZeroMemory(m_patchLocationList, sizeof(PATCHLOCATIONLIST) * m_maxNumAllocations); |
| 1381 |
m_resCount = 0; |
| 1382 |
|
| 1383 |
MOS_ZeroMemory(m_writeModeList, sizeof(bool) * m_maxNumAllocations); |
| 1384 |
finish: |
| 1385 |
MOS_TraceEventExt(EVENT_MOS_BATCH_SUBMIT, EVENT_TYPE_END, &eStatus, sizeof(eStatus), nullptr, 0); |
| 1386 |
return eStatus; |
| 1387 |
} |
| 1388 |
|
| 1389 |
void GpuContextSpecific::UnlockPendingOcaBuffers(PMOS_COMMAND_BUFFER cmdBuffer, PMOS_CONTEXT mosContext) |
| 1390 |
{ |
| 1391 |
MOS_OS_CHK_NULL_NO_STATUS_RETURN(cmdBuffer); |
| 1392 |
MOS_OS_CHK_NULL_NO_STATUS_RETURN(mosContext); |
| 1393 |
MosOcaInterface *pOcaInterface = &MosOcaInterfaceSpecific::GetInstance(); |
| 1394 |
if (nullptr == pOcaInterface || !((MosOcaInterfaceSpecific*)pOcaInterface)->IsOcaEnabled()) |
| 1395 |
{ |
| 1396 |
// Will come here for UMD_OCA not being enabled case. |
| 1397 |
return; |
| 1398 |
} |
| 1399 |
|
| 1400 |
int count = 0; |
| 1401 |
struct MOS_OCA_EXEC_LIST_INFO *info = nullptr; |
| 1402 |
if ((cmdBuffer->iSubmissionType & SUBMISSION_TYPE_SINGLE_PIPE_MASK) && ((MosOcaInterfaceSpecific*)pOcaInterface)->IsOcaDumpExecListInfoEnabled()) |
| 1403 |
{ |
| 1404 |
info = mos_bo_get_softpin_targets_info(cmdBuffer->OsResource.bo, &count); |
| 1405 |
} |
| 1406 |
|
| 1407 |
pOcaInterface->UnlockPendingOcaBuffers(mosContext, info, count); |
| 1408 |
|
| 1409 |
if(info) |
| 1410 |
{ |
| 1411 |
free(info); |
| 1412 |
} |
| 1413 |
} |
| 1414 |
|
| 1415 |
int32_t GpuContextSpecific::SubmitPipeCommands( |
| 1416 |
MOS_COMMAND_BUFFER *cmdBuffer, |
| 1417 |
MOS_LINUX_BO *cmdBo, |
| 1418 |
PMOS_CONTEXT osContext, |
| 1419 |
const std::vector<MOS_LINUX_BO *> &skipSyncBoList, |
| 1420 |
uint32_t execFlag, |
| 1421 |
int32_t dr4) |
| 1422 |
{ |
| 1423 |
int32_t ret = 0; |
| 1424 |
int fence = -1; |
| 1425 |
unsigned int fenceFlag = 0; |
| 1426 |
|
| 1427 |
MOS_LINUX_CONTEXT *queue = m_i915Context[0]; |
| 1428 |
bool isVeboxSubmission = false; |
| 1429 |
|
| 1430 |
if (execFlag == MOS_GPU_NODE_VIDEO || execFlag == MOS_GPU_NODE_VIDEO2) |
| 1431 |
{ |
| 1432 |
execFlag = I915_EXEC_DEFAULT; |
| 1433 |
} |
| 1434 |
if (execFlag == MOS_GPU_NODE_VE) |
| 1435 |
{ |
| 1436 |
execFlag = I915_EXEC_DEFAULT; |
| 1437 |
isVeboxSubmission = true; |
| 1438 |
} |
| 1439 |
|
| 1440 |
if(cmdBuffer->iSubmissionType & SUBMISSION_TYPE_MULTI_PIPE_SLAVE) |
| 1441 |
{ |
| 1442 |
fence = osContext->submit_fence; |
| 1443 |
fenceFlag = I915_EXEC_FENCE_SUBMIT; |
| 1444 |
int slaveIndex = (cmdBuffer->iSubmissionType & SUBMISSION_TYPE_MULTI_PIPE_SLAVE_INDEX_MASK) >> SUBMISSION_TYPE_MULTI_PIPE_SLAVE_INDEX_SHIFT; |
| 1445 |
if(slaveIndex < 7) |
| 1446 |
{ |
| 1447 |
queue = m_i915Context[2 + slaveIndex]; //0 is for single pipe, 1 is for master, slave starts from 2 |
| 1448 |
} |
| 1449 |
else |
| 1450 |
{ |
| 1451 |
MOS_OS_ASSERTMESSAGE("slaveIndex value: %s is invalid!", slaveIndex); |
| 1452 |
return -1; |
| 1453 |
} |
| 1454 |
|
| 1455 |
if (isVeboxSubmission) |
| 1456 |
{ |
| 1457 |
queue = m_i915Context[cmdBuffer->iVeboxNodeIndex + 1]; |
| 1458 |
} |
| 1459 |
|
| 1460 |
for(auto bo: skipSyncBoList) |
| 1461 |
{ |
| 1462 |
mos_bo_set_exec_object_async(cmdBo, bo); |
| 1463 |
} |
| 1464 |
} |
| 1465 |
|
| 1466 |
if(cmdBuffer->iSubmissionType & SUBMISSION_TYPE_MULTI_PIPE_MASTER) |
| 1467 |
{ |
| 1468 |
//Only master pipe needs fence out flag |
| 1469 |
fenceFlag = I915_EXEC_FENCE_OUT; |
| 1470 |
queue = m_i915Context[1]; |
| 1471 |
} |
| 1472 |
|
| 1473 |
ret = mos_bo_context_exec2(cmdBo, |
| 1474 |
cmdBo->size, |
| 1475 |
queue, |
| 1476 |
nullptr, |
| 1477 |
0, |
| 1478 |
dr4, |
| 1479 |
execFlag | fenceFlag, |
| 1480 |
&fence); |
| 1481 |
|
| 1482 |
if(cmdBuffer->iSubmissionType & SUBMISSION_TYPE_MULTI_PIPE_MASTER) |
| 1483 |
{ |
| 1484 |
osContext->submit_fence = fence; |
| 1485 |
} |
| 1486 |
|
| 1487 |
if(cmdBuffer->iSubmissionType & SUBMISSION_TYPE_MULTI_PIPE_FLAGS_LAST_PIPE) |
| 1488 |
{ |
| 1489 |
close(fence); |
| 1490 |
} |
| 1491 |
|
| 1492 |
return ret; |
| 1493 |
} |
| 1494 |
|
| 1495 |
int32_t GpuContextSpecific::ParallelSubmitCommands( |
| 1496 |
std::map<uint32_t, PMOS_COMMAND_BUFFER> secondaryCmdBufs, |
| 1497 |
PMOS_CONTEXT osContext, |
| 1498 |
uint32_t execFlag, |
| 1499 |
int32_t dr4) |
| 1500 |
{ |
| 1501 |
int32_t ret = 0; |
| 1502 |
int fence = -1; |
| 1503 |
unsigned int fenceFlag = 0; |
| 1504 |
auto it = m_secondaryCmdBufs.begin(); |
| 1505 |
MOS_LINUX_BO *cmdBos[MAX_PARALLEN_CMD_BO_NUM]; |
| 1506 |
int numBos = 0; // exclude FE bo |
| 1507 |
|
| 1508 |
MOS_LINUX_CONTEXT *queue = m_i915Context[0]; |
| 1509 |
bool isVeboxSubmission = false; |
| 1510 |
|
| 1511 |
if (execFlag == MOS_GPU_NODE_VIDEO || execFlag == MOS_GPU_NODE_VIDEO2) |
| 1512 |
{ |
| 1513 |
execFlag = I915_EXEC_DEFAULT; |
| 1514 |
} |
| 1515 |
if (execFlag == MOS_GPU_NODE_VE) |
| 1516 |
{ |
| 1517 |
execFlag = I915_EXEC_DEFAULT; |
| 1518 |
isVeboxSubmission = true; |
| 1519 |
} |
| 1520 |
|
| 1521 |
while(it != m_secondaryCmdBufs.end()) |
| 1522 |
{ |
| 1523 |
if(it->second->iSubmissionType & SUBMISSION_TYPE_MULTI_PIPE_ALONE) |
| 1524 |
{ |
| 1525 |
fenceFlag = I915_EXEC_FENCE_OUT; |
| 1526 |
queue = m_i915Context[0]; |
| 1527 |
|
| 1528 |
ret = mos_bo_context_exec2(it->second->OsResource.bo, |
| 1529 |
it->second->OsResource.bo->size, |
| 1530 |
queue, |
| 1531 |
nullptr, |
| 1532 |
0, |
| 1533 |
dr4, |
| 1534 |
execFlag | fenceFlag, |
| 1535 |
&fence); |
| 1536 |
|
| 1537 |
osContext->submit_fence = fence; |
| 1538 |
} |
| 1539 |
|
| 1540 |
if((it->second->iSubmissionType & SUBMISSION_TYPE_MULTI_PIPE_MASTER) |
| 1541 |
|| (it->second->iSubmissionType & SUBMISSION_TYPE_MULTI_PIPE_SLAVE)) |
| 1542 |
{ |
| 1543 |
cmdBos[numBos++] = it->second->OsResource.bo; |
| 1544 |
|
| 1545 |
if(it->second->iSubmissionType & SUBMISSION_TYPE_MULTI_PIPE_FLAGS_LAST_PIPE) |
| 1546 |
{ |
| 1547 |
queue = m_i915Context[numBos - 1]; |
| 1548 |
MOS_OS_CHK_NULL_RETURN(queue); |
| 1549 |
if(-1 != fence) |
| 1550 |
{ |
| 1551 |
fenceFlag = I915_EXEC_FENCE_IN; |
| 1552 |
} |
| 1553 |
|
| 1554 |
ret = mos_bo_context_exec3(cmdBos, |
| 1555 |
numBos, |
| 1556 |
queue, |
| 1557 |
nullptr, |
| 1558 |
0, |
| 1559 |
dr4, |
| 1560 |
execFlag | fenceFlag, |
| 1561 |
&fence); |
| 1562 |
|
| 1563 |
for(int i = 0; i < numBos; i++) |
| 1564 |
{ |
| 1565 |
cmdBos[i] = nullptr; |
| 1566 |
} |
| 1567 |
numBos = 0; |
| 1568 |
|
| 1569 |
if(-1 != fence) |
| 1570 |
{ |
| 1571 |
close(fence); |
| 1572 |
} |
| 1573 |
} |
| 1574 |
} |
| 1575 |
|
| 1576 |
it++; |
| 1577 |
} |
| 1578 |
|
| 1579 |
return ret; |
| 1580 |
} |
| 1581 |
|
| 1582 |
void GpuContextSpecific::IncrementGpuStatusTag() |
| 1583 |
{ |
| 1584 |
m_GPUStatusTag = m_GPUStatusTag % UINT_MAX + 1; |
| 1585 |
if (m_GPUStatusTag == 0) |
| 1586 |
{ |
| 1587 |
m_GPUStatusTag = 1; |
| 1588 |
} |
| 1589 |
} |
| 1590 |
|
| 1591 |
void GpuContextSpecific::ResetGpuContextStatus() |
| 1592 |
{ |
| 1593 |
MOS_ZeroMemory(m_allocationList, sizeof(ALLOCATION_LIST) * ALLOCATIONLIST_SIZE); |
| 1594 |
m_numAllocations = 0; |
| 1595 |
MOS_ZeroMemory(m_patchLocationList, sizeof(PATCHLOCATIONLIST) * PATCHLOCATIONLIST_SIZE); |
| 1596 |
m_currentNumPatchLocations = 0; |
| 1597 |
|
| 1598 |
MOS_ZeroMemory(m_attachedResources, sizeof(MOS_RESOURCE) * ALLOCATIONLIST_SIZE); |
| 1599 |
m_resCount = 0; |
| 1600 |
|
| 1601 |
MOS_ZeroMemory(m_writeModeList, sizeof(bool) * ALLOCATIONLIST_SIZE); |
| 1602 |
|
| 1603 |
if ((m_cmdBufFlushed == true) && m_commandBuffer->OsResource.bo) |
| 1604 |
{ |
| 1605 |
m_commandBuffer->OsResource.bo = nullptr; |
| 1606 |
} |
| 1607 |
} |
| 1608 |
|
| 1609 |
MOS_STATUS GpuContextSpecific::AllocateGPUStatusBuf() |
| 1610 |
{ |
| 1611 |
MOS_OS_FUNCTION_ENTER; |
| 1612 |
|
| 1613 |
m_statusBufferMosResource = (MOS_RESOURCE_HANDLE)MOS_AllocAndZeroMemory(sizeof(MOS_RESOURCE)); |
| 1614 |
MOS_OS_CHK_NULL_RETURN(m_statusBufferMosResource); |
| 1615 |
|
| 1616 |
GraphicsResource::CreateParams params; |
| 1617 |
params.m_tileType = MOS_TILE_LINEAR; |
| 1618 |
params.m_type = MOS_GFXRES_BUFFER; |
| 1619 |
params.m_format = Format_Buffer; |
| 1620 |
params.m_width = sizeof(MOS_GPU_STATUS_DATA); |
| 1621 |
params.m_height = 1; |
| 1622 |
params.m_depth = 1; |
| 1623 |
params.m_arraySize = 1; |
| 1624 |
params.m_name = "GPU Status Buffer"; |
| 1625 |
|
| 1626 |
GraphicsResource *graphicsResource = GraphicsResource::CreateGraphicResource(GraphicsResource::osSpecificResource); |
| 1627 |
MOS_OS_CHK_NULL_RETURN(graphicsResource); |
| 1628 |
|
| 1629 |
MOS_OS_CHK_STATUS_RETURN(graphicsResource->Allocate(m_osContext, params)); |
| 1630 |
|
| 1631 |
GraphicsResource::LockParams lockParams; |
| 1632 |
lockParams.m_writeRequest = true; |
| 1633 |
auto gpuStatusData = (MOS_GPU_STATUS_DATA *)graphicsResource->Lock(m_osContext, lockParams); |
| 1634 |
if (gpuStatusData == nullptr) |
| 1635 |
{ |
| 1636 |
MOS_OS_ASSERTMESSAGE("Unable to lock gpu eStatus buffer for read."); |
| 1637 |
graphicsResource->Free(m_osContext); |
| 1638 |
MOS_Delete(graphicsResource); |
| 1639 |
return MOS_STATUS_UNKNOWN; |
| 1640 |
} |
| 1641 |
|
| 1642 |
m_statusBufferResource = graphicsResource; |
| 1643 |
return MOS_STATUS_SUCCESS; |
| 1644 |
} |
| 1645 |
|
| 1646 |
#if (_DEBUG || _RELEASE_INTERNAL) |
| 1647 |
bool GpuContextSpecific::SelectEngineInstanceByUser(struct i915_engine_class_instance *engineMap, |
| 1648 |
uint32_t *engineNum, uint32_t userEngineInstance, MOS_GPU_NODE gpuNode) |
| 1649 |
{ |
| 1650 |
uint32_t engineInstance = 0x0; |
| 1651 |
|
| 1652 |
if(gpuNode == MOS_GPU_NODE_COMPUTE) |
| 1653 |
{ |
| 1654 |
engineInstance = (userEngineInstance >> ENGINE_INSTANCE_SELECT_COMPUTE_INSTANCE_SHIFT) |
| 1655 |
& (ENGINE_INSTANCE_SELECT_ENABLE_MASK >> (MAX_ENGINE_INSTANCE_NUM - *engineNum)); |
| 1656 |
} |
| 1657 |
else if(gpuNode == MOS_GPU_NODE_VE) |
| 1658 |
{ |
| 1659 |
engineInstance = (userEngineInstance >> ENGINE_INSTANCE_SELECT_VEBOX_INSTANCE_SHIFT) |
| 1660 |
& (ENGINE_INSTANCE_SELECT_ENABLE_MASK >> (MAX_ENGINE_INSTANCE_NUM - *engineNum)); |
| 1661 |
} |
| 1662 |
else if(gpuNode == MOS_GPU_NODE_VIDEO || gpuNode == MOS_GPU_NODE_VIDEO2) |
| 1663 |
{ |
| 1664 |
engineInstance = (userEngineInstance >> ENGINE_INSTANCE_SELECT_VDBOX_INSTANCE_SHIFT) |
| 1665 |
& (ENGINE_INSTANCE_SELECT_ENABLE_MASK >> (MAX_ENGINE_INSTANCE_NUM - *engineNum)); |
| 1666 |
} |
| 1667 |
else |
| 1668 |
{ |
| 1669 |
MOS_OS_NORMALMESSAGE("Invalid gpu node in use."); |
| 1670 |
} |
| 1671 |
|
| 1672 |
if(engineInstance) |
| 1673 |
{ |
| 1674 |
auto unSelectIndex = 0; |
| 1675 |
for(auto bit = 0; bit < *engineNum; bit++) |
| 1676 |
{ |
| 1677 |
if(((engineInstance >> bit) & 0x1) && (bit > unSelectIndex)) |
| 1678 |
{ |
| 1679 |
engineMap[unSelectIndex].engine_class = engineMap[bit].engine_class; |
| 1680 |
engineMap[unSelectIndex].engine_instance = engineMap[bit].engine_instance; |
| 1681 |
engineMap[bit].engine_class = 0; |
| 1682 |
engineMap[bit].engine_instance = 0; |
| 1683 |
unSelectIndex++; |
| 1684 |
} |
| 1685 |
else if(((engineInstance >> bit) & 0x1) && (bit == unSelectIndex)) |
| 1686 |
{ |
| 1687 |
unSelectIndex++; |
| 1688 |
} |
| 1689 |
else if(!((engineInstance >> bit) & 0x1)) |
| 1690 |
{ |
| 1691 |
engineMap[bit].engine_class = 0; |
| 1692 |
engineMap[bit].engine_instance = 0; |
| 1693 |
} |
| 1694 |
} |
| 1695 |
*engineNum = unSelectIndex; |
| 1696 |
} |
| 1697 |
return engineInstance; |
| 1698 |
} |
| 1699 |
#endif |