src/gallium/drivers/swr/rasterizer/core/api.cpp

   1 /****************************************************************************
   2 * Copyright (C) 2014-2016 Intel Corporation.   All Rights Reserved.
   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 (including the next
  12 * paragraph) shall be included in all copies or substantial portions of the
  13 * Software.
  14 *
  15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21 * IN THE SOFTWARE.
  22 *
  23 * @file api.cpp
  24 *
  25 * @brief API implementation
  26 *
  27 ******************************************************************************/
  28
  29 #include <cfloat>
  30 #include <cmath>
  31 #include <cstdio>
  32 #include <new>
  33
  34 #include "core/api.h"
  35 #include "core/backend.h"
  36 #include "core/context.h"
  37 #include "core/depthstencil.h"
  38 #include "core/frontend.h"
  39 #include "core/rasterizer.h"
  40 #include "core/rdtsc_core.h"
  41 #include "core/threads.h"
  42 #include "core/tilemgr.h"
  43 #include "core/clip.h"
  44 #include "core/utils.h"
  45
  46 #include "common/simdintrin.h"
  47 #include "common/os.h"
  48
  49 static const SWR_RECT g_MaxScissorRect = { 0, 0, KNOB_MAX_SCISSOR_X, KNOB_MAX_SCISSOR_Y };
  50
  51 void SetupDefaultState(SWR_CONTEXT *pContext);
  52
  53 static INLINE SWR_CONTEXT* GetContext(HANDLE hContext)
  54 {
  55     return (SWR_CONTEXT*)hContext;
  56 }
  57
  58 void WakeAllThreads(SWR_CONTEXT *pContext)
  59 {
  60     pContext->FifosNotEmpty.notify_all();
  61 }
  62
  63 //////////////////////////////////////////////////////////////////////////
  64 /// @brief Create SWR Context.
  65 /// @param pCreateInfo - pointer to creation info.
  66 HANDLE SwrCreateContext(
  67     SWR_CREATECONTEXT_INFO* pCreateInfo)
  68 {
  69     RDTSC_RESET();
  70     RDTSC_INIT(0);
  71
  72     void* pContextMem = AlignedMalloc(sizeof(SWR_CONTEXT), KNOB_SIMD_WIDTH * 4);
  73     memset(pContextMem, 0, sizeof(SWR_CONTEXT));
  74     SWR_CONTEXT *pContext = new (pContextMem) SWR_CONTEXT();
  75
  76     pContext->driverType = pCreateInfo->driver;
  77     pContext->privateStateSize = pCreateInfo->privateStateSize;
  78
  79     pContext->dcRing.Init(KNOB_MAX_DRAWS_IN_FLIGHT);
  80     pContext->dsRing.Init(KNOB_MAX_DRAWS_IN_FLIGHT);
  81
  82     pContext->pMacroTileManagerArray = (MacroTileMgr*)AlignedMalloc(sizeof(MacroTileMgr) * KNOB_MAX_DRAWS_IN_FLIGHT, 64);
  83     pContext->pDispatchQueueArray = (DispatchQueue*)AlignedMalloc(sizeof(DispatchQueue) * KNOB_MAX_DRAWS_IN_FLIGHT, 64);
  84
  85     for (uint32_t dc = 0; dc < KNOB_MAX_DRAWS_IN_FLIGHT; ++dc)
  86     {
  87         pContext->dcRing[dc].pArena = new CachingArena(pContext->cachingArenaAllocator);
  88         new (&pContext->pMacroTileManagerArray[dc]) MacroTileMgr(*pContext->dcRing[dc].pArena);
  89         new (&pContext->pDispatchQueueArray[dc]) DispatchQueue();
  90
  91         pContext->dsRing[dc].pArena = new CachingArena(pContext->cachingArenaAllocator);
  92     }
  93
  94     pContext->threadInfo.MAX_WORKER_THREADS        = KNOB_MAX_WORKER_THREADS;
  95     pContext->threadInfo.MAX_NUMA_NODES            = KNOB_MAX_NUMA_NODES;
  96     pContext->threadInfo.MAX_CORES_PER_NUMA_NODE   = KNOB_MAX_CORES_PER_NUMA_NODE;
  97     pContext->threadInfo.MAX_THREADS_PER_CORE      = KNOB_MAX_THREADS_PER_CORE;
  98     pContext->threadInfo.SINGLE_THREADED           = KNOB_SINGLE_THREADED;
  99
 100     if (pCreateInfo->pThreadInfo)
 101     {
 102         pContext->threadInfo = *pCreateInfo->pThreadInfo;
 103     }
 104
 105     memset(&pContext->WaitLock, 0, sizeof(pContext->WaitLock));
 106     memset(&pContext->FifosNotEmpty, 0, sizeof(pContext->FifosNotEmpty));
 107     new (&pContext->WaitLock) std::mutex();
 108     new (&pContext->FifosNotEmpty) std::condition_variable();
 109
 110     CreateThreadPool(pContext, &pContext->threadPool);
 111
 112     pContext->ppScratch = new uint8_t*[pContext->NumWorkerThreads];
 113     pContext->pStats = new SWR_STATS[pContext->NumWorkerThreads];
 114
 115     // Setup ArchRast thread contexts which includes +1 for API thread.
 116     pContext->pArContext = new HANDLE[pContext->NumWorkerThreads+1];
 117     pContext->pArContext[pContext->NumWorkerThreads] = ArchRast::CreateThreadContext();
 118
 119     // Allocate scratch space for workers.
 120     ///@note We could lazily allocate this but its rather small amount of memory.
 121     for (uint32_t i = 0; i < pContext->NumWorkerThreads; ++i)
 122     {
 123 #if defined(_WIN32)
 124         uint32_t numaNode = pContext->threadPool.pThreadData ?
 125             pContext->threadPool.pThreadData[i].numaId : 0;
 126         pContext->ppScratch[i] = (uint8_t*)VirtualAllocExNuma(
 127             GetCurrentProcess(), nullptr, 32 * sizeof(KILOBYTE),
 128             MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE,
 129             numaNode);
 130 #else
 131         pContext->ppScratch[i] = (uint8_t*)AlignedMalloc(32 * sizeof(KILOBYTE), KNOB_SIMD_WIDTH * 4);
 132 #endif
 133
 134         // Initialize worker thread context for ArchRast.
 135         pContext->pArContext[i] = ArchRast::CreateThreadContext();
 136     }
 137
 138     // State setup AFTER context is fully initialized
 139     SetupDefaultState(pContext);
 140
 141     // initialize hot tile manager
 142     pContext->pHotTileMgr = new HotTileMgr();
 143
 144     // initialize function pointer tables
 145     InitClearTilesTable();
 146
 147     // initialize callback functions
 148     pContext->pfnLoadTile = pCreateInfo->pfnLoadTile;
 149     pContext->pfnStoreTile = pCreateInfo->pfnStoreTile;
 150     pContext->pfnClearTile = pCreateInfo->pfnClearTile;
 151     pContext->pfnUpdateSoWriteOffset = pCreateInfo->pfnUpdateSoWriteOffset;
 152     pContext->pfnUpdateStats = pCreateInfo->pfnUpdateStats;
 153     pContext->pfnUpdateStatsFE = pCreateInfo->pfnUpdateStatsFE;
 154
 155     // pass pointer to bucket manager back to caller
 156 #ifdef KNOB_ENABLE_RDTSC
 157     pCreateInfo->pBucketMgr = &gBucketMgr;
 158 #endif
 159
 160     pCreateInfo->contextSaveSize = sizeof(API_STATE);
 161
 162     StartThreadPool(pContext, &pContext->threadPool);
 163
 164     return (HANDLE)pContext;
 165 }
 166
 167 void CopyState(DRAW_STATE& dst, const DRAW_STATE& src)
 168 {
 169     memcpy(&dst.state, &src.state, sizeof(API_STATE));
 170 }
 171
 172 template<bool IsDraw>
 173 void QueueWork(SWR_CONTEXT *pContext)
 174 {
 175     DRAW_CONTEXT* pDC = pContext->pCurDrawContext;
 176     uint32_t dcIndex = pDC->drawId % KNOB_MAX_DRAWS_IN_FLIGHT;
 177
 178     if (IsDraw)
 179     {
 180         pDC->pTileMgr = &pContext->pMacroTileManagerArray[dcIndex];
 181         pDC->pTileMgr->initialize();
 182     }
 183
 184     // Each worker thread looks at a DC for both FE and BE work at different times and so we
 185     // multiply threadDone by 2.  When the threadDone counter has reached 0 then all workers
 186     // have moved past this DC. (i.e. Each worker has checked this DC for both FE and BE work and
 187     // then moved on if all work is done.)
 188     pContext->pCurDrawContext->threadsDone = pContext->NumFEThreads + pContext->NumBEThreads;
 189
 190     if (IsDraw)
 191     {
 192         InterlockedIncrement((volatile LONG*)&pContext->drawsOutstandingFE);
 193     }
 194
 195     _ReadWriteBarrier();
 196     {
 197         std::unique_lock<std::mutex> lock(pContext->WaitLock);
 198         pContext->dcRing.Enqueue();
 199     }
 200
 201     if (pContext->threadInfo.SINGLE_THREADED)
 202     {
 203         // flush denormals to 0
 204         uint32_t mxcsr = _mm_getcsr();
 205         _mm_setcsr(mxcsr | _MM_FLUSH_ZERO_ON | _MM_DENORMALS_ZERO_ON);
 206
 207         if (IsDraw)
 208         {
 209             uint32_t curDraw[2] = { pContext->pCurDrawContext->drawId, pContext->pCurDrawContext->drawId };
 210             WorkOnFifoFE(pContext, 0, curDraw[0]);
 211             WorkOnFifoBE(pContext, 0, curDraw[1], pContext->singleThreadLockedTiles, 0, 0);
 212         }
 213         else
 214         {
 215             uint32_t curDispatch = pContext->pCurDrawContext->drawId;
 216             WorkOnCompute(pContext, 0, curDispatch);
 217         }
 218
 219         // Dequeue the work here, if not already done, since we're single threaded (i.e. no workers).
 220         while (CompleteDrawContext(pContext, pContext->pCurDrawContext) > 0) {}
 221
 222         // restore csr
 223         _mm_setcsr(mxcsr);
 224     }
 225     else
 226     {
 227         AR_API_BEGIN(APIDrawWakeAllThreads, pDC->drawId);
 228         WakeAllThreads(pContext);
 229         AR_API_END(APIDrawWakeAllThreads, 1);
 230     }
 231
 232     // Set current draw context to NULL so that next state call forces a new draw context to be created and populated.
 233     pContext->pPrevDrawContext = pContext->pCurDrawContext;
 234     pContext->pCurDrawContext = nullptr;
 235 }
 236
 237 INLINE void QueueDraw(SWR_CONTEXT* pContext)
 238 {
 239     QueueWork<true>(pContext);
 240 }
 241
 242 INLINE void QueueDispatch(SWR_CONTEXT* pContext)
 243 {
 244     QueueWork<false>(pContext);
 245 }
 246
 247 DRAW_CONTEXT* GetDrawContext(SWR_CONTEXT *pContext, bool isSplitDraw = false)
 248 {
 249     AR_API_BEGIN(APIGetDrawContext, 0);
 250     // If current draw context is null then need to obtain a new draw context to use from ring.
 251     if (pContext->pCurDrawContext == nullptr)
 252     {
 253         // Need to wait for a free entry.
 254         while (pContext->dcRing.IsFull())
 255         {
 256             _mm_pause();
 257         }
 258
 259         uint64_t curDraw = pContext->dcRing.GetHead();
 260         uint32_t dcIndex = curDraw % KNOB_MAX_DRAWS_IN_FLIGHT;
 261
 262         if ((pContext->frameCount - pContext->lastFrameChecked) > 2 ||
 263             (curDraw - pContext->lastDrawChecked) > 0x10000)
 264         {
 265             // Take this opportunity to clean-up old arena allocations
 266             pContext->cachingArenaAllocator.FreeOldBlocks();
 267
 268             pContext->lastFrameChecked = pContext->frameCount;
 269             pContext->lastDrawChecked = curDraw;
 270         }
 271
 272         DRAW_CONTEXT* pCurDrawContext = &pContext->dcRing[dcIndex];
 273         pContext->pCurDrawContext = pCurDrawContext;
 274
 275         // Assign next available entry in DS ring to this DC.
 276         uint32_t dsIndex = pContext->curStateId % KNOB_MAX_DRAWS_IN_FLIGHT;
 277         pCurDrawContext->pState = &pContext->dsRing[dsIndex];
 278
 279         // Copy previous state to current state.
 280         if (pContext->pPrevDrawContext)
 281         {
 282             DRAW_CONTEXT* pPrevDrawContext = pContext->pPrevDrawContext;
 283
 284             // If we're splitting our draw then we can just use the same state from the previous
 285             // draw. In this case, we won't increment the DS ring index so the next non-split
 286             // draw can receive the state.
 287             if (isSplitDraw == false)
 288             {
 289                 CopyState(*pCurDrawContext->pState, *pPrevDrawContext->pState);
 290
 291                 // Should have been cleaned up previously
 292                 SWR_ASSERT(pCurDrawContext->pState->pArena->IsEmpty() == true);
 293
 294                 pCurDrawContext->pState->pPrivateState = nullptr;
 295
 296                 pContext->curStateId++;  // Progress state ring index forward.
 297             }
 298             else
 299             {
 300                 // If its a split draw then just copy the state pointer over
 301                 // since its the same draw.
 302                 pCurDrawContext->pState = pPrevDrawContext->pState;
 303                 SWR_ASSERT(pPrevDrawContext->cleanupState == false);
 304             }
 305         }
 306         else
 307         {
 308             SWR_ASSERT(pCurDrawContext->pState->pArena->IsEmpty() == true);
 309             pContext->curStateId++;  // Progress state ring index forward.
 310         }
 311
 312         SWR_ASSERT(pCurDrawContext->pArena->IsEmpty() == true);
 313
 314         pCurDrawContext->dependent = false;
 315         pCurDrawContext->pContext = pContext;
 316         pCurDrawContext->isCompute = false; // Dispatch has to set this to true.
 317
 318         pCurDrawContext->doneFE = false;
 319         pCurDrawContext->FeLock = 0;
 320         pCurDrawContext->threadsDone = 0;
 321         pCurDrawContext->retireCallback.pfnCallbackFunc = nullptr;
 322
 323         pCurDrawContext->dynState.Reset(pContext->NumWorkerThreads);
 324
 325         // Assign unique drawId for this DC
 326         pCurDrawContext->drawId = pContext->dcRing.GetHead();
 327
 328         pCurDrawContext->cleanupState = true;
 329     }
 330     else
 331     {
 332         SWR_ASSERT(isSplitDraw == false, "Split draw should only be used when obtaining a new DC");
 333     }
 334
 335     AR_API_END(APIGetDrawContext, 0);
 336     return pContext->pCurDrawContext;
 337 }
 338
 339 API_STATE* GetDrawState(SWR_CONTEXT *pContext)
 340 {
 341     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
 342     SWR_ASSERT(pDC->pState != nullptr);
 343
 344     return &pDC->pState->state;
 345 }
 346
 347 void SwrDestroyContext(HANDLE hContext)
 348 {
 349     SWR_CONTEXT *pContext = GetContext(hContext);
 350     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
 351
 352     pDC->FeWork.type = SHUTDOWN;
 353     pDC->FeWork.pfnWork = ProcessShutdown;
 354
 355     //enqueue
 356     QueueDraw(pContext);
 357
 358     DestroyThreadPool(pContext, &pContext->threadPool);
 359
 360     // free the fifos
 361     for (uint32_t i = 0; i < KNOB_MAX_DRAWS_IN_FLIGHT; ++i)
 362     {
 363         delete[] pContext->dcRing[i].dynState.pStats;
 364         delete pContext->dcRing[i].pArena;
 365         delete pContext->dsRing[i].pArena;
 366         pContext->pMacroTileManagerArray[i].~MacroTileMgr();
 367         pContext->pDispatchQueueArray[i].~DispatchQueue();
 368     }
 369
 370     AlignedFree(pContext->pDispatchQueueArray);
 371     AlignedFree(pContext->pMacroTileManagerArray);
 372
 373     // Free scratch space.
 374     for (uint32_t i = 0; i < pContext->NumWorkerThreads; ++i)
 375     {
 376 #if defined(_WIN32)
 377         VirtualFree(pContext->ppScratch[i], 0, MEM_RELEASE);
 378 #else
 379         AlignedFree(pContext->ppScratch[i]);
 380 #endif
 381
 382         ArchRast::DestroyThreadContext(pContext->pArContext[i]);
 383     }
 384
 385     delete[] pContext->ppScratch;
 386     delete[] pContext->pArContext;
 387     delete[] pContext->pStats;
 388
 389     delete(pContext->pHotTileMgr);
 390
 391     pContext->~SWR_CONTEXT();
 392     AlignedFree(GetContext(hContext));
 393 }
 394
 395 void SWR_API SwrSaveState(
 396     HANDLE hContext,
 397     void* pOutputStateBlock,
 398     size_t memSize)
 399 {
 400     SWR_CONTEXT *pContext = GetContext(hContext);
 401     auto pSrc = GetDrawState(pContext);
 402     SWR_ASSERT(pOutputStateBlock && memSize >= sizeof(*pSrc));
 403
 404     memcpy(pOutputStateBlock, pSrc, sizeof(*pSrc));
 405 }
 406
 407 void SWR_API SwrRestoreState(
 408     HANDLE hContext,
 409     const void* pStateBlock,
 410     size_t memSize)
 411 {
 412     SWR_CONTEXT *pContext = GetContext(hContext);
 413     auto pDst = GetDrawState(pContext);
 414     SWR_ASSERT(pStateBlock && memSize >= sizeof(*pDst));
 415
 416     memcpy(pDst, pStateBlock, sizeof(*pDst));
 417 }
 418
 419 void SetupDefaultState(SWR_CONTEXT *pContext)
 420 {
 421     API_STATE* pState = GetDrawState(pContext);
 422
 423     pState->rastState.cullMode = SWR_CULLMODE_NONE;
 424     pState->rastState.frontWinding = SWR_FRONTWINDING_CCW;
 425 }
 426
 427 void SwrSync(HANDLE hContext, PFN_CALLBACK_FUNC pfnFunc, uint64_t userData, uint64_t userData2, uint64_t userData3)
 428 {
 429     SWR_ASSERT(pfnFunc != nullptr);
 430
 431     SWR_CONTEXT *pContext = GetContext(hContext);
 432     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
 433
 434     AR_API_BEGIN(APISync, 0);
 435
 436     pDC->FeWork.type = SYNC;
 437     pDC->FeWork.pfnWork = ProcessSync;
 438
 439     // Setup callback function
 440     pDC->retireCallback.pfnCallbackFunc = pfnFunc;
 441     pDC->retireCallback.userData = userData;
 442     pDC->retireCallback.userData2 = userData2;
 443     pDC->retireCallback.userData3 = userData3;
 444
 445     //enqueue
 446     QueueDraw(pContext);
 447
 448     AR_API_END(APISync, 1);
 449 }
 450
 451 void SwrWaitForIdle(HANDLE hContext)
 452 {
 453     SWR_CONTEXT *pContext = GetContext(hContext);
 454
 455     AR_API_BEGIN(APIWaitForIdle, 0);
 456
 457     while (!pContext->dcRing.IsEmpty())
 458     {
 459         _mm_pause();
 460     }
 461
 462     AR_API_END(APIWaitForIdle, 1);
 463 }
 464
 465 void SwrWaitForIdleFE(HANDLE hContext)
 466 {
 467     SWR_CONTEXT *pContext = GetContext(hContext);
 468
 469     AR_API_BEGIN(APIWaitForIdle, 0);
 470
 471     while (pContext->drawsOutstandingFE > 0)
 472     {
 473         _mm_pause();
 474     }
 475
 476     AR_API_END(APIWaitForIdle, 1);
 477 }
 478
 479 void SwrSetVertexBuffers(
 480     HANDLE hContext,
 481     uint32_t numBuffers,
 482     const SWR_VERTEX_BUFFER_STATE* pVertexBuffers)
 483 {
 484     API_STATE* pState = GetDrawState(GetContext(hContext));
 485
 486     for (uint32_t i = 0; i < numBuffers; ++i)
 487     {
 488         const SWR_VERTEX_BUFFER_STATE *pVB = &pVertexBuffers[i];
 489         pState->vertexBuffers[pVB->index] = *pVB;
 490     }
 491 }
 492
 493 void SwrSetIndexBuffer(
 494     HANDLE hContext,
 495     const SWR_INDEX_BUFFER_STATE* pIndexBuffer)
 496 {
 497     API_STATE* pState = GetDrawState(GetContext(hContext));
 498
 499     pState->indexBuffer = *pIndexBuffer;
 500 }
 501
 502 void SwrSetFetchFunc(
 503     HANDLE hContext,
 504     PFN_FETCH_FUNC    pfnFetchFunc)
 505 {
 506     API_STATE* pState = GetDrawState(GetContext(hContext));
 507
 508     pState->pfnFetchFunc = pfnFetchFunc;
 509 }
 510
 511 void SwrSetSoFunc(
 512     HANDLE hContext,
 513     PFN_SO_FUNC    pfnSoFunc,
 514     uint32_t streamIndex)
 515 {
 516     API_STATE* pState = GetDrawState(GetContext(hContext));
 517
 518     SWR_ASSERT(streamIndex < MAX_SO_STREAMS);
 519
 520     pState->pfnSoFunc[streamIndex] = pfnSoFunc;
 521 }
 522
 523 void SwrSetSoState(
 524     HANDLE hContext,
 525     SWR_STREAMOUT_STATE* pSoState)
 526 {
 527     API_STATE* pState = GetDrawState(GetContext(hContext));
 528
 529     pState->soState = *pSoState;
 530 }
 531
 532 void SwrSetSoBuffers(
 533     HANDLE hContext,
 534     SWR_STREAMOUT_BUFFER* pSoBuffer,
 535     uint32_t slot)
 536 {
 537     API_STATE* pState = GetDrawState(GetContext(hContext));
 538
 539     SWR_ASSERT((slot < 4), "There are only 4 SO buffer slots [0, 3]\nSlot requested: %d", slot);
 540
 541     pState->soBuffer[slot] = *pSoBuffer;
 542 }
 543
 544 void SwrSetVertexFunc(
 545     HANDLE hContext,
 546     PFN_VERTEX_FUNC pfnVertexFunc)
 547 {
 548     API_STATE* pState = GetDrawState(GetContext(hContext));
 549
 550     pState->pfnVertexFunc = pfnVertexFunc;
 551 }
 552
 553 void SwrSetFrontendState(
 554     HANDLE hContext,
 555     SWR_FRONTEND_STATE *pFEState)
 556 {
 557     API_STATE* pState = GetDrawState(GetContext(hContext));
 558     pState->frontendState = *pFEState;
 559 }
 560
 561 void SwrSetGsState(
 562     HANDLE hContext,
 563     SWR_GS_STATE *pGSState)
 564 {
 565     API_STATE* pState = GetDrawState(GetContext(hContext));
 566     pState->gsState = *pGSState;
 567 }
 568
 569 void SwrSetGsFunc(
 570     HANDLE hContext,
 571     PFN_GS_FUNC pfnGsFunc)
 572 {
 573     API_STATE* pState = GetDrawState(GetContext(hContext));
 574     pState->pfnGsFunc = pfnGsFunc;
 575 }
 576
 577 void SwrSetCsFunc(
 578     HANDLE hContext,
 579     PFN_CS_FUNC pfnCsFunc,
 580     uint32_t totalThreadsInGroup,
 581     uint32_t totalSpillFillSize)
 582 {
 583     API_STATE* pState = GetDrawState(GetContext(hContext));
 584     pState->pfnCsFunc = pfnCsFunc;
 585     pState->totalThreadsInGroup = totalThreadsInGroup;
 586     pState->totalSpillFillSize = totalSpillFillSize;
 587 }
 588
 589 void SwrSetTsState(
 590     HANDLE hContext,
 591     SWR_TS_STATE *pState)
 592 {
 593     API_STATE* pApiState = GetDrawState(GetContext(hContext));
 594     pApiState->tsState = *pState;
 595 }
 596
 597 void SwrSetHsFunc(
 598     HANDLE hContext,
 599     PFN_HS_FUNC pfnFunc)
 600 {
 601     API_STATE* pApiState = GetDrawState(GetContext(hContext));
 602     pApiState->pfnHsFunc = pfnFunc;
 603 }
 604
 605 void SwrSetDsFunc(
 606     HANDLE hContext,
 607     PFN_DS_FUNC pfnFunc)
 608 {
 609     API_STATE* pApiState = GetDrawState(GetContext(hContext));
 610     pApiState->pfnDsFunc = pfnFunc;
 611 }
 612
 613 void SwrSetDepthStencilState(
 614     HANDLE hContext,
 615     SWR_DEPTH_STENCIL_STATE *pDSState)
 616 {
 617     API_STATE* pState = GetDrawState(GetContext(hContext));
 618
 619     pState->depthStencilState = *pDSState;
 620 }
 621
 622 void SwrSetBackendState(
 623     HANDLE hContext,
 624     SWR_BACKEND_STATE *pBEState)
 625 {
 626     API_STATE* pState = GetDrawState(GetContext(hContext));
 627
 628     pState->backendState = *pBEState;
 629 }
 630
 631 void SwrSetPixelShaderState(
 632     HANDLE hContext,
 633     SWR_PS_STATE *pPSState)
 634 {
 635     API_STATE *pState = GetDrawState(GetContext(hContext));
 636     pState->psState = *pPSState;
 637 }
 638
 639 void SwrSetBlendState(
 640     HANDLE hContext,
 641     SWR_BLEND_STATE *pBlendState)
 642 {
 643     API_STATE *pState = GetDrawState(GetContext(hContext));
 644     memcpy(&pState->blendState, pBlendState, sizeof(SWR_BLEND_STATE));
 645 }
 646
 647 void SwrSetBlendFunc(
 648     HANDLE hContext,
 649     uint32_t renderTarget,
 650     PFN_BLEND_JIT_FUNC pfnBlendFunc)
 651 {
 652     SWR_ASSERT(renderTarget < SWR_NUM_RENDERTARGETS);
 653     API_STATE *pState = GetDrawState(GetContext(hContext));
 654     pState->pfnBlendFunc[renderTarget] = pfnBlendFunc;
 655 }
 656
 657 // update guardband multipliers for the viewport
 658 void updateGuardbands(API_STATE *pState)
 659 {
 660     uint32_t numGbs = pState->gsState.emitsRenderTargetArrayIndex ? KNOB_NUM_VIEWPORTS_SCISSORS : 1;
 661
 662     for(uint32_t i = 0; i < numGbs; ++i)
 663     {
 664         // guardband center is viewport center
 665         pState->gbState.left[i] = KNOB_GUARDBAND_WIDTH / pState->vp[i].width;
 666         pState->gbState.right[i] = KNOB_GUARDBAND_WIDTH / pState->vp[i].width;
 667         pState->gbState.top[i] = KNOB_GUARDBAND_HEIGHT / pState->vp[i].height;
 668         pState->gbState.bottom[i] = KNOB_GUARDBAND_HEIGHT / pState->vp[i].height;
 669     }
 670 }
 671
 672 void SwrSetRastState(
 673     HANDLE hContext,
 674     const SWR_RASTSTATE *pRastState)
 675 {
 676     SWR_CONTEXT *pContext = GetContext(hContext);
 677     API_STATE* pState = GetDrawState(pContext);
 678
 679     memcpy(&pState->rastState, pRastState, sizeof(SWR_RASTSTATE));
 680 }
 681
 682 void SwrSetViewports(
 683     HANDLE hContext,
 684     uint32_t numViewports,
 685     const SWR_VIEWPORT* pViewports,
 686     const SWR_VIEWPORT_MATRICES* pMatrices)
 687 {
 688     SWR_ASSERT(numViewports <= KNOB_NUM_VIEWPORTS_SCISSORS,
 689         "Invalid number of viewports.");
 690
 691     SWR_CONTEXT *pContext = GetContext(hContext);
 692     API_STATE* pState = GetDrawState(pContext);
 693
 694     memcpy(&pState->vp[0], pViewports, sizeof(SWR_VIEWPORT) * numViewports);
 695
 696     if (pMatrices != nullptr)
 697     {
 698         // @todo Faster to copy portions of the SOA or just copy all of it?
 699         memcpy(&pState->vpMatrices, pMatrices, sizeof(SWR_VIEWPORT_MATRICES));
 700     }
 701     else
 702     {
 703         // Compute default viewport transform.
 704         for (uint32_t i = 0; i < numViewports; ++i)
 705         {
 706             if (pContext->driverType == DX)
 707             {
 708                 pState->vpMatrices.m00[i] = pState->vp[i].width / 2.0f;
 709                 pState->vpMatrices.m11[i] = -pState->vp[i].height / 2.0f;
 710                 pState->vpMatrices.m22[i] = pState->vp[i].maxZ - pState->vp[i].minZ;
 711                 pState->vpMatrices.m30[i] = pState->vp[i].x + pState->vpMatrices.m00[i];
 712                 pState->vpMatrices.m31[i] = pState->vp[i].y - pState->vpMatrices.m11[i];
 713                 pState->vpMatrices.m32[i] = pState->vp[i].minZ;
 714             }
 715             else
 716             {
 717                 // Standard, with the exception that Y is inverted.
 718                 pState->vpMatrices.m00[i] = (pState->vp[i].width - pState->vp[i].x) / 2.0f;
 719                 pState->vpMatrices.m11[i] = (pState->vp[i].y - pState->vp[i].height) / 2.0f;
 720                 pState->vpMatrices.m22[i] = (pState->vp[i].maxZ - pState->vp[i].minZ) / 2.0f;
 721                 pState->vpMatrices.m30[i] = pState->vp[i].x + pState->vpMatrices.m00[i];
 722                 pState->vpMatrices.m31[i] = pState->vp[i].height + pState->vpMatrices.m11[i];
 723                 pState->vpMatrices.m32[i] = pState->vp[i].minZ + pState->vpMatrices.m22[i];
 724
 725                 // Now that the matrix is calculated, clip the view coords to screen size.
 726                 // OpenGL allows for -ve x,y in the viewport.
 727                 pState->vp[i].x = std::max(pState->vp[i].x, 0.0f);
 728                 pState->vp[i].y = std::max(pState->vp[i].y, 0.0f);
 729             }
 730         }
 731     }
 732
 733     updateGuardbands(pState);
 734 }
 735
 736 void SwrSetScissorRects(
 737     HANDLE hContext,
 738     uint32_t numScissors,
 739     const SWR_RECT* pScissors)
 740 {
 741     SWR_ASSERT(numScissors <= KNOB_NUM_VIEWPORTS_SCISSORS,
 742         "Invalid number of scissor rects.");
 743
 744     API_STATE* pState = GetDrawState(GetContext(hContext));
 745     memcpy(&pState->scissorRects[0], pScissors, numScissors * sizeof(pScissors[0]));
 746 };
 747
 748 void SetupMacroTileScissors(DRAW_CONTEXT *pDC)
 749 {
 750     API_STATE *pState = &pDC->pState->state;
 751     uint32_t numScissors = pState->gsState.emitsViewportArrayIndex ? KNOB_NUM_VIEWPORTS_SCISSORS : 1;
 752     pState->scissorsTileAligned = true;
 753
 754     for (uint32_t index = 0; index < numScissors; ++index)
 755     {
 756         SWR_RECT &scissorInFixedPoint = pState->scissorsInFixedPoint[index];
 757
 758         // Set up scissor dimensions based on scissor or viewport
 759         if (pState->rastState.scissorEnable)
 760         {
 761             scissorInFixedPoint = pState->scissorRects[index];
 762         }
 763         else
 764         {
 765             // the vp width and height must be added to origin un-rounded then the result round to -inf.
 766             // The cast to int works for rounding assuming all [left, right, top, bottom] are positive.
 767             scissorInFixedPoint.xmin = (int32_t)pState->vp[index].x;
 768             scissorInFixedPoint.xmax = (int32_t)(pState->vp[index].x + pState->vp[index].width);
 769             scissorInFixedPoint.ymin = (int32_t)pState->vp[index].y;
 770             scissorInFixedPoint.ymax = (int32_t)(pState->vp[index].y + pState->vp[index].height);
 771         }
 772
 773         // Clamp to max rect
 774         scissorInFixedPoint &= g_MaxScissorRect;
 775
 776         // Test for tile alignment
 777         bool tileAligned;
 778         tileAligned  = (scissorInFixedPoint.xmin % KNOB_TILE_X_DIM) == 0;
 779         tileAligned &= (scissorInFixedPoint.ymin % KNOB_TILE_Y_DIM) == 0;
 780         tileAligned &= (scissorInFixedPoint.xmax % KNOB_TILE_X_DIM) == 0;
 781         tileAligned &= (scissorInFixedPoint.xmax % KNOB_TILE_Y_DIM) == 0;
 782
 783         pState->scissorsTileAligned &= tileAligned;
 784
 785         // Scale to fixed point
 786         scissorInFixedPoint.xmin *= FIXED_POINT_SCALE;
 787         scissorInFixedPoint.xmax *= FIXED_POINT_SCALE;
 788         scissorInFixedPoint.ymin *= FIXED_POINT_SCALE;
 789         scissorInFixedPoint.ymax *= FIXED_POINT_SCALE;
 790
 791         // Make scissor inclusive
 792         scissorInFixedPoint.xmax -= 1;
 793         scissorInFixedPoint.ymax -= 1;
 794     }
 795 }
 796
 797 // templated backend function tables
 798 extern PFN_BACKEND_FUNC gBackendNullPs[SWR_MULTISAMPLE_TYPE_COUNT];
 799 extern PFN_BACKEND_FUNC gBackendSingleSample[SWR_INPUT_COVERAGE_COUNT][2][2];
 800 extern PFN_BACKEND_FUNC gBackendPixelRateTable[SWR_MULTISAMPLE_TYPE_COUNT][SWR_MSAA_SAMPLE_PATTERN_COUNT][SWR_INPUT_COVERAGE_COUNT][2][2][2];
 801 extern PFN_BACKEND_FUNC gBackendSampleRateTable[SWR_MULTISAMPLE_TYPE_COUNT][SWR_INPUT_COVERAGE_COUNT][2][2];
 802 void SetupPipeline(DRAW_CONTEXT *pDC)
 803 {
 804     DRAW_STATE* pState = pDC->pState;
 805     const SWR_RASTSTATE &rastState = pState->state.rastState;
 806     const SWR_PS_STATE &psState = pState->state.psState;
 807     BACKEND_FUNCS& backendFuncs = pState->backendFuncs;
 808     const uint32_t forcedSampleCount = (rastState.forcedSampleCount) ? 1 : 0;
 809
 810     // setup backend
 811     if (psState.pfnPixelShader == nullptr)
 812     {
 813         backendFuncs.pfnBackend = gBackendNullPs[pState->state.rastState.sampleCount];
 814     }
 815     else
 816     {
 817         const bool bMultisampleEnable = ((rastState.sampleCount > SWR_MULTISAMPLE_1X) || rastState.forcedSampleCount) ? 1 : 0;
 818         const uint32_t centroid = ((psState.barycentricsMask & SWR_BARYCENTRIC_CENTROID_MASK) > 0) ? 1 : 0;
 819         const uint32_t canEarlyZ = (psState.forceEarlyZ || (!psState.writesODepth && !psState.usesSourceDepth && !psState.usesUAV)) ? 1 : 0;
 820
 821         SWR_BARYCENTRICS_MASK barycentricsMask = (SWR_BARYCENTRICS_MASK)psState.barycentricsMask;
 822
 823         // select backend function
 824         switch(psState.shadingRate)
 825         {
 826         case SWR_SHADING_RATE_PIXEL:
 827             if(bMultisampleEnable)
 828             {
 829                 // always need to generate I & J per sample for Z interpolation
 830                 barycentricsMask = (SWR_BARYCENTRICS_MASK)(barycentricsMask | SWR_BARYCENTRIC_PER_SAMPLE_MASK);
 831                 backendFuncs.pfnBackend = gBackendPixelRateTable[rastState.sampleCount][rastState.samplePattern][psState.inputCoverage][centroid][forcedSampleCount][canEarlyZ];
 832             }
 833             else
 834             {
 835                 // always need to generate I & J per pixel for Z interpolation
 836                 barycentricsMask = (SWR_BARYCENTRICS_MASK)(barycentricsMask | SWR_BARYCENTRIC_PER_PIXEL_MASK);
 837                 backendFuncs.pfnBackend = gBackendSingleSample[psState.inputCoverage][centroid][canEarlyZ];
 838             }
 839             break;
 840         case SWR_SHADING_RATE_SAMPLE:
 841             SWR_ASSERT(rastState.samplePattern == SWR_MSAA_STANDARD_PATTERN);
 842             // always need to generate I & J per sample for Z interpolation
 843             barycentricsMask = (SWR_BARYCENTRICS_MASK)(barycentricsMask | SWR_BARYCENTRIC_PER_SAMPLE_MASK);
 844             backendFuncs.pfnBackend = gBackendSampleRateTable[rastState.sampleCount][psState.inputCoverage][centroid][canEarlyZ];
 845             break;
 846         default:
 847             SWR_ASSERT(0 && "Invalid shading rate");
 848             break;
 849         }
 850     }
 851
 852     PFN_PROCESS_PRIMS pfnBinner;
 853     switch (pState->state.topology)
 854     {
 855     case TOP_POINT_LIST:
 856         pState->pfnProcessPrims = ClipPoints;
 857         pfnBinner = BinPoints;
 858         break;
 859     case TOP_LINE_LIST:
 860     case TOP_LINE_STRIP:
 861     case TOP_LINE_LOOP:
 862     case TOP_LINE_LIST_ADJ:
 863     case TOP_LISTSTRIP_ADJ:
 864         pState->pfnProcessPrims = ClipLines;
 865         pfnBinner = BinLines;
 866         break;
 867     default:
 868         pState->pfnProcessPrims = ClipTriangles;
 869         pfnBinner = GetBinTrianglesFunc((rastState.conservativeRast > 0));
 870         break;
 871     };
 872
 873     // disable clipper if viewport transform is disabled
 874     if (pState->state.frontendState.vpTransformDisable)
 875     {
 876         pState->pfnProcessPrims = pfnBinner;
 877     }
 878
 879     if ((pState->state.psState.pfnPixelShader == nullptr) &&
 880         (pState->state.depthStencilState.depthTestEnable == FALSE) &&
 881         (pState->state.depthStencilState.depthWriteEnable == FALSE) &&
 882         (pState->state.depthStencilState.stencilTestEnable == FALSE) &&
 883         (pState->state.depthStencilState.stencilWriteEnable == FALSE) &&
 884         (pState->state.backendState.numAttributes == 0))
 885     {
 886         pState->pfnProcessPrims = nullptr;
 887     }
 888
 889     if (pState->state.soState.rasterizerDisable == true)
 890     {
 891         pState->pfnProcessPrims = nullptr;
 892     }
 893
 894     // set up the frontend attribute count
 895     pState->state.feNumAttributes = 0;
 896     const SWR_BACKEND_STATE& backendState = pState->state.backendState;
 897     if (backendState.swizzleEnable)
 898     {
 899         // attribute swizzling is enabled, iterate over the map and record the max attribute used
 900         for (uint32_t i = 0; i < backendState.numAttributes; ++i)
 901         {
 902             pState->state.feNumAttributes = std::max(pState->state.feNumAttributes, (uint32_t)backendState.swizzleMap[i].sourceAttrib + 1);
 903         }
 904     }
 905     else
 906     {
 907         pState->state.feNumAttributes = pState->state.backendState.numAttributes;
 908     }
 909
 910     if (pState->state.soState.soEnable)
 911     {
 912         uint32_t streamMasks = 0;
 913         for (uint32_t i = 0; i < 4; ++i)
 914         {
 915             streamMasks |= pState->state.soState.streamMasks[i];
 916         }
 917
 918         DWORD maxAttrib;
 919         if (_BitScanReverse(&maxAttrib, streamMasks))
 920         {
 921             pState->state.feNumAttributes = std::max(pState->state.feNumAttributes, (uint32_t)(maxAttrib + 1));
 922         }
 923     }
 924
 925     // complicated logic to test for cases where we don't need backing hottile memory for a draw
 926     // have to check for the special case where depth/stencil test is enabled but depthwrite is disabled.
 927     pState->state.depthHottileEnable = ((!(pState->state.depthStencilState.depthTestEnable &&
 928                                            !pState->state.depthStencilState.depthWriteEnable &&
 929                                            pState->state.depthStencilState.depthTestFunc == ZFUNC_ALWAYS)) &&
 930                                         (pState->state.depthStencilState.depthTestEnable ||
 931                                          pState->state.depthStencilState.depthWriteEnable)) ? true : false;
 932
 933     pState->state.stencilHottileEnable = (((!(pState->state.depthStencilState.stencilTestEnable &&
 934                                              !pState->state.depthStencilState.stencilWriteEnable &&
 935                                               pState->state.depthStencilState.stencilTestFunc == ZFUNC_ALWAYS)) ||
 936                                           // for stencil we have to check the double sided state as well
 937                                           (!(pState->state.depthStencilState.doubleSidedStencilTestEnable &&
 938                                              !pState->state.depthStencilState.stencilWriteEnable &&
 939                                               pState->state.depthStencilState.backfaceStencilTestFunc == ZFUNC_ALWAYS))) &&
 940                                           (pState->state.depthStencilState.stencilTestEnable  ||
 941                                            pState->state.depthStencilState.stencilWriteEnable)) ? true : false;
 942
 943     uint32_t numRTs = pState->state.psState.numRenderTargets;
 944     pState->state.colorHottileEnable = 0;
 945     if (psState.pfnPixelShader != nullptr)
 946     {
 947         for (uint32_t rt = 0; rt < numRTs; ++rt)
 948         {
 949             pState->state.colorHottileEnable |=
 950                 (!pState->state.blendState.renderTarget[rt].writeDisableAlpha ||
 951                  !pState->state.blendState.renderTarget[rt].writeDisableRed ||
 952                  !pState->state.blendState.renderTarget[rt].writeDisableGreen ||
 953                  !pState->state.blendState.renderTarget[rt].writeDisableBlue) ? (1 << rt) : 0;
 954         }
 955     }
 956
 957     // Setup depth quantization function
 958     if (pState->state.depthHottileEnable)
 959     {
 960         switch (pState->state.rastState.depthFormat)
 961         {
 962         case R32_FLOAT_X8X24_TYPELESS: pState->state.pfnQuantizeDepth = QuantizeDepth < R32_FLOAT_X8X24_TYPELESS > ; break;
 963         case R32_FLOAT: pState->state.pfnQuantizeDepth = QuantizeDepth < R32_FLOAT > ; break;
 964         case R24_UNORM_X8_TYPELESS: pState->state.pfnQuantizeDepth = QuantizeDepth < R24_UNORM_X8_TYPELESS > ; break;
 965         case R16_UNORM: pState->state.pfnQuantizeDepth = QuantizeDepth < R16_UNORM > ; break;
 966         default: SWR_ASSERT(false, "Unsupported depth format for depth quantiztion.");
 967             pState->state.pfnQuantizeDepth = QuantizeDepth < R32_FLOAT > ;
 968         }
 969     }
 970     else
 971     {
 972         // set up pass-through quantize if depth isn't enabled
 973         pState->state.pfnQuantizeDepth = QuantizeDepth < R32_FLOAT > ;
 974     }
 975 }
 976
 977 //////////////////////////////////////////////////////////////////////////
 978 /// @brief InitDraw
 979 /// @param pDC - Draw context to initialize for this draw.
 980 void InitDraw(
 981     DRAW_CONTEXT *pDC,
 982     bool isSplitDraw)
 983 {
 984     // We don't need to re-setup the scissors/pipeline state again for split draw.
 985     if (isSplitDraw == false)
 986     {
 987         SetupMacroTileScissors(pDC);
 988         SetupPipeline(pDC);
 989     }
 990 }
 991
 992 //////////////////////////////////////////////////////////////////////////
 993 /// @brief We can split the draw for certain topologies for better performance.
 994 /// @param totalVerts - Total vertices for draw
 995 /// @param topology - Topology used for draw
 996 uint32_t MaxVertsPerDraw(
 997     DRAW_CONTEXT* pDC,
 998     uint32_t totalVerts,
 999     PRIMITIVE_TOPOLOGY topology)
1000 {
1001     API_STATE& state = pDC->pState->state;
1002
1003     uint32_t vertsPerDraw = totalVerts;
1004
1005     if (state.soState.soEnable)
1006     {
1007         return totalVerts;
1008     }
1009
1010     switch (topology)
1011     {
1012     case TOP_POINT_LIST:
1013     case TOP_TRIANGLE_LIST:
1014         vertsPerDraw = KNOB_MAX_PRIMS_PER_DRAW;
1015         break;
1016
1017     case TOP_PATCHLIST_1:
1018     case TOP_PATCHLIST_2:
1019     case TOP_PATCHLIST_3:
1020     case TOP_PATCHLIST_4:
1021     case TOP_PATCHLIST_5:
1022     case TOP_PATCHLIST_6:
1023     case TOP_PATCHLIST_7:
1024     case TOP_PATCHLIST_8:
1025     case TOP_PATCHLIST_9:
1026     case TOP_PATCHLIST_10:
1027     case TOP_PATCHLIST_11:
1028     case TOP_PATCHLIST_12:
1029     case TOP_PATCHLIST_13:
1030     case TOP_PATCHLIST_14:
1031     case TOP_PATCHLIST_15:
1032     case TOP_PATCHLIST_16:
1033     case TOP_PATCHLIST_17:
1034     case TOP_PATCHLIST_18:
1035     case TOP_PATCHLIST_19:
1036     case TOP_PATCHLIST_20:
1037     case TOP_PATCHLIST_21:
1038     case TOP_PATCHLIST_22:
1039     case TOP_PATCHLIST_23:
1040     case TOP_PATCHLIST_24:
1041     case TOP_PATCHLIST_25:
1042     case TOP_PATCHLIST_26:
1043     case TOP_PATCHLIST_27:
1044     case TOP_PATCHLIST_28:
1045     case TOP_PATCHLIST_29:
1046     case TOP_PATCHLIST_30:
1047     case TOP_PATCHLIST_31:
1048     case TOP_PATCHLIST_32:
1049         if (pDC->pState->state.tsState.tsEnable)
1050         {
1051             uint32_t vertsPerPrim = topology - TOP_PATCHLIST_BASE;
1052             vertsPerDraw = vertsPerPrim * KNOB_MAX_TESS_PRIMS_PER_DRAW;
1053         }
1054         break;
1055
1056     // The Primitive Assembly code can only handle 1 RECT at a time.
1057     case TOP_RECT_LIST:
1058         vertsPerDraw = 3;
1059         break;
1060
1061     default:
1062         // We are not splitting up draws for other topologies.
1063         break;
1064     }
1065
1066     return vertsPerDraw;
1067 }
1068
1069
1070 //////////////////////////////////////////////////////////////////////////
1071 /// @brief DrawInstanced
1072 /// @param hContext - Handle passed back from SwrCreateContext
1073 /// @param topology - Specifies topology for draw.
1074 /// @param numVerts - How many vertices to read sequentially from vertex data (per instance).
1075 /// @param startVertex - Specifies start vertex for draw. (vertex data)
1076 /// @param numInstances - How many instances to render.
1077 /// @param startInstance - Which instance to start sequentially fetching from in each buffer (instanced data)
1078 void DrawInstanced(
1079     HANDLE hContext,
1080     PRIMITIVE_TOPOLOGY topology,
1081     uint32_t numVertices,
1082     uint32_t startVertex,
1083     uint32_t numInstances = 1,
1084     uint32_t startInstance = 0)
1085 {
1086     if (KNOB_TOSS_DRAW)
1087     {
1088         return;
1089     }
1090
1091     SWR_CONTEXT *pContext = GetContext(hContext);
1092     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1093
1094     AR_API_BEGIN(APIDraw, pDC->drawId);
1095
1096     uint32_t maxVertsPerDraw = MaxVertsPerDraw(pDC, numVertices, topology);
1097     uint32_t primsPerDraw = GetNumPrims(topology, maxVertsPerDraw);
1098     uint32_t remainingVerts = numVertices;
1099
1100     API_STATE    *pState = &pDC->pState->state;
1101     pState->topology = topology;
1102     pState->forceFront = false;
1103
1104     // disable culling for points/lines
1105     uint32_t oldCullMode = pState->rastState.cullMode;
1106     if (topology == TOP_POINT_LIST)
1107     {
1108         pState->rastState.cullMode = SWR_CULLMODE_NONE;
1109         pState->forceFront = true;
1110     }
1111
1112     int draw = 0;
1113     while (remainingVerts)
1114     {
1115         uint32_t numVertsForDraw = (remainingVerts < maxVertsPerDraw) ?
1116         remainingVerts : maxVertsPerDraw;
1117
1118         bool isSplitDraw = (draw > 0) ? true : false;
1119         DRAW_CONTEXT* pDC = GetDrawContext(pContext, isSplitDraw);
1120         InitDraw(pDC, isSplitDraw);
1121
1122         pDC->FeWork.type = DRAW;
1123         pDC->FeWork.pfnWork = GetProcessDrawFunc(
1124             false,  // IsIndexed
1125             false, // bEnableCutIndex
1126             pState->tsState.tsEnable,
1127             pState->gsState.gsEnable,
1128             pState->soState.soEnable,
1129             pDC->pState->pfnProcessPrims != nullptr);
1130         pDC->FeWork.desc.draw.numVerts = numVertsForDraw;
1131         pDC->FeWork.desc.draw.startVertex = startVertex;
1132         pDC->FeWork.desc.draw.numInstances = numInstances;
1133         pDC->FeWork.desc.draw.startInstance = startInstance;
1134         pDC->FeWork.desc.draw.startPrimID = draw * primsPerDraw;
1135         pDC->FeWork.desc.draw.startVertexID = draw * maxVertsPerDraw;
1136
1137         pDC->cleanupState = (remainingVerts == numVertsForDraw);
1138
1139         //enqueue DC
1140         QueueDraw(pContext);
1141
1142         remainingVerts -= numVertsForDraw;
1143         draw++;
1144     }
1145
1146     // restore culling state
1147     pDC = GetDrawContext(pContext);
1148     pDC->pState->state.rastState.cullMode = oldCullMode;
1149
1150     AR_API_END(APIDraw, numVertices * numInstances);
1151 }
1152
1153 //////////////////////////////////////////////////////////////////////////
1154 /// @brief SwrDraw
1155 /// @param hContext - Handle passed back from SwrCreateContext
1156 /// @param topology - Specifies topology for draw.
1157 /// @param startVertex - Specifies start vertex in vertex buffer for draw.
1158 /// @param primCount - Number of vertices.
1159 void SwrDraw(
1160     HANDLE hContext,
1161     PRIMITIVE_TOPOLOGY topology,
1162     uint32_t startVertex,
1163     uint32_t numVertices)
1164 {
1165     DrawInstanced(hContext, topology, numVertices, startVertex);
1166 }
1167
1168 //////////////////////////////////////////////////////////////////////////
1169 /// @brief SwrDrawInstanced
1170 /// @param hContext - Handle passed back from SwrCreateContext
1171 /// @param topology - Specifies topology for draw.
1172 /// @param numVertsPerInstance - How many vertices to read sequentially from vertex data.
1173 /// @param numInstances - How many instances to render.
1174 /// @param startVertex - Specifies start vertex for draw. (vertex data)
1175 /// @param startInstance - Which instance to start sequentially fetching from in each buffer (instanced data)
1176 void SwrDrawInstanced(
1177     HANDLE hContext,
1178     PRIMITIVE_TOPOLOGY topology,
1179     uint32_t numVertsPerInstance,
1180     uint32_t numInstances,
1181     uint32_t startVertex,
1182     uint32_t startInstance
1183     )
1184 {
1185     DrawInstanced(hContext, topology, numVertsPerInstance, startVertex, numInstances, startInstance);
1186 }
1187
1188 //////////////////////////////////////////////////////////////////////////
1189 /// @brief DrawIndexedInstanced
1190 /// @param hContext - Handle passed back from SwrCreateContext
1191 /// @param topology - Specifies topology for draw.
1192 /// @param numIndices - Number of indices to read sequentially from index buffer.
1193 /// @param indexOffset - Starting index into index buffer.
1194 /// @param baseVertex - Vertex in vertex buffer to consider as index "0". Note value is signed.
1195 /// @param numInstances - Number of instances to render.
1196 /// @param startInstance - Which instance to start sequentially fetching from in each buffer (instanced data)
1197 void DrawIndexedInstance(
1198     HANDLE hContext,
1199     PRIMITIVE_TOPOLOGY topology,
1200     uint32_t numIndices,
1201     uint32_t indexOffset,
1202     int32_t baseVertex,
1203     uint32_t numInstances = 1,
1204     uint32_t startInstance = 0)
1205 {
1206     if (KNOB_TOSS_DRAW)
1207     {
1208         return;
1209     }
1210
1211     SWR_CONTEXT *pContext = GetContext(hContext);
1212     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1213     API_STATE* pState = &pDC->pState->state;
1214
1215     AR_API_BEGIN(APIDrawIndexed, pDC->drawId);
1216     AR_API_EVENT(DrawIndexedInstance(topology, numIndices, indexOffset, baseVertex, numInstances, startInstance));
1217
1218     uint32_t maxIndicesPerDraw = MaxVertsPerDraw(pDC, numIndices, topology);
1219     uint32_t primsPerDraw = GetNumPrims(topology, maxIndicesPerDraw);
1220     uint32_t remainingIndices = numIndices;
1221
1222     uint32_t indexSize = 0;
1223     switch (pState->indexBuffer.format)
1224     {
1225     case R32_UINT: indexSize = sizeof(uint32_t); break;
1226     case R16_UINT: indexSize = sizeof(uint16_t); break;
1227     case R8_UINT: indexSize = sizeof(uint8_t); break;
1228     default:
1229         SWR_ASSERT(0);
1230     }
1231
1232     int draw = 0;
1233     uint8_t *pIB = (uint8_t*)pState->indexBuffer.pIndices;
1234     pIB += (uint64_t)indexOffset * (uint64_t)indexSize;
1235
1236     pState->topology = topology;
1237     pState->forceFront = false;
1238
1239     // disable culling for points/lines
1240     uint32_t oldCullMode = pState->rastState.cullMode;
1241     if (topology == TOP_POINT_LIST)
1242     {
1243         pState->rastState.cullMode = SWR_CULLMODE_NONE;
1244         pState->forceFront = true;
1245     }
1246
1247     while (remainingIndices)
1248     {
1249         uint32_t numIndicesForDraw = (remainingIndices < maxIndicesPerDraw) ?
1250         remainingIndices : maxIndicesPerDraw;
1251
1252         // When breaking up draw, we need to obtain new draw context for each iteration.
1253         bool isSplitDraw = (draw > 0) ? true : false;
1254         pDC = GetDrawContext(pContext, isSplitDraw);
1255         InitDraw(pDC, isSplitDraw);
1256
1257         pDC->FeWork.type = DRAW;
1258         pDC->FeWork.pfnWork = GetProcessDrawFunc(
1259             true,   // IsIndexed
1260             pState->frontendState.bEnableCutIndex,
1261             pState->tsState.tsEnable,
1262             pState->gsState.gsEnable,
1263             pState->soState.soEnable,
1264             pDC->pState->pfnProcessPrims != nullptr);
1265         pDC->FeWork.desc.draw.pDC = pDC;
1266         pDC->FeWork.desc.draw.numIndices = numIndicesForDraw;
1267         pDC->FeWork.desc.draw.pIB = (int*)pIB;
1268         pDC->FeWork.desc.draw.type = pDC->pState->state.indexBuffer.format;
1269
1270         pDC->FeWork.desc.draw.numInstances = numInstances;
1271         pDC->FeWork.desc.draw.startInstance = startInstance;
1272         pDC->FeWork.desc.draw.baseVertex = baseVertex;
1273         pDC->FeWork.desc.draw.startPrimID = draw * primsPerDraw;
1274
1275         pDC->cleanupState = (remainingIndices == numIndicesForDraw);
1276
1277         //enqueue DC
1278         QueueDraw(pContext);
1279
1280         pIB += maxIndicesPerDraw * indexSize;
1281         remainingIndices -= numIndicesForDraw;
1282         draw++;
1283     }
1284
1285     // restore culling state
1286     pDC = GetDrawContext(pContext);
1287     pDC->pState->state.rastState.cullMode = oldCullMode;
1288
1289     AR_API_END(APIDrawIndexed, numIndices * numInstances);
1290 }
1291
1292
1293 //////////////////////////////////////////////////////////////////////////
1294 /// @brief DrawIndexed
1295 /// @param hContext - Handle passed back from SwrCreateContext
1296 /// @param topology - Specifies topology for draw.
1297 /// @param numIndices - Number of indices to read sequentially from index buffer.
1298 /// @param indexOffset - Starting index into index buffer.
1299 /// @param baseVertex - Vertex in vertex buffer to consider as index "0". Note value is signed.
1300 void SwrDrawIndexed(
1301     HANDLE hContext,
1302     PRIMITIVE_TOPOLOGY topology,
1303     uint32_t numIndices,
1304     uint32_t indexOffset,
1305     int32_t baseVertex
1306     )
1307 {
1308     DrawIndexedInstance(hContext, topology, numIndices, indexOffset, baseVertex);
1309 }
1310
1311 //////////////////////////////////////////////////////////////////////////
1312 /// @brief SwrDrawIndexedInstanced
1313 /// @param hContext - Handle passed back from SwrCreateContext
1314 /// @param topology - Specifies topology for draw.
1315 /// @param numIndices - Number of indices to read sequentially from index buffer.
1316 /// @param numInstances - Number of instances to render.
1317 /// @param indexOffset - Starting index into index buffer.
1318 /// @param baseVertex - Vertex in vertex buffer to consider as index "0". Note value is signed.
1319 /// @param startInstance - Which instance to start sequentially fetching from in each buffer (instanced data)
1320 void SwrDrawIndexedInstanced(
1321     HANDLE hContext,
1322     PRIMITIVE_TOPOLOGY topology,
1323     uint32_t numIndices,
1324     uint32_t numInstances,
1325     uint32_t indexOffset,
1326     int32_t baseVertex,
1327     uint32_t startInstance)
1328 {
1329     DrawIndexedInstance(hContext, topology, numIndices, indexOffset, baseVertex, numInstances, startInstance);
1330 }
1331
1332 //////////////////////////////////////////////////////////////////////////
1333 /// @brief SwrInvalidateTiles
1334 /// @param hContext - Handle passed back from SwrCreateContext
1335 /// @param attachmentMask - The mask specifies which surfaces attached to the hottiles to invalidate.
1336 /// @param invalidateRect - The pixel-coordinate rectangle to invalidate.  This will be expanded to
1337 ///                         be hottile size-aligned.
1338 void SWR_API SwrInvalidateTiles(
1339     HANDLE hContext,
1340     uint32_t attachmentMask,
1341     const SWR_RECT& invalidateRect)
1342 {
1343     if (KNOB_TOSS_DRAW)
1344     {
1345         return;
1346     }
1347
1348     SWR_CONTEXT *pContext = GetContext(hContext);
1349     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1350
1351     pDC->FeWork.type = DISCARDINVALIDATETILES;
1352     pDC->FeWork.pfnWork = ProcessDiscardInvalidateTiles;
1353     pDC->FeWork.desc.discardInvalidateTiles.attachmentMask = attachmentMask;
1354     pDC->FeWork.desc.discardInvalidateTiles.rect = invalidateRect;
1355     pDC->FeWork.desc.discardInvalidateTiles.rect &= g_MaxScissorRect;
1356     pDC->FeWork.desc.discardInvalidateTiles.newTileState = SWR_TILE_INVALID;
1357     pDC->FeWork.desc.discardInvalidateTiles.createNewTiles = false;
1358     pDC->FeWork.desc.discardInvalidateTiles.fullTilesOnly = false;
1359
1360     //enqueue
1361     QueueDraw(pContext);
1362 }
1363
1364 //////////////////////////////////////////////////////////////////////////
1365 /// @brief SwrDiscardRect
1366 /// @param hContext - Handle passed back from SwrCreateContext
1367 /// @param attachmentMask - The mask specifies which surfaces attached to the hottiles to discard.
1368 /// @param rect - The pixel-coordinate rectangle to discard.  Only fully-covered hottiles will be
1369 ///               discarded.
1370 void SWR_API SwrDiscardRect(
1371     HANDLE hContext,
1372     uint32_t attachmentMask,
1373     const SWR_RECT& rect)
1374 {
1375     if (KNOB_TOSS_DRAW)
1376     {
1377         return;
1378     }
1379
1380     SWR_CONTEXT *pContext = GetContext(hContext);
1381     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1382
1383     // Queue a load to the hottile
1384     pDC->FeWork.type = DISCARDINVALIDATETILES;
1385     pDC->FeWork.pfnWork = ProcessDiscardInvalidateTiles;
1386     pDC->FeWork.desc.discardInvalidateTiles.attachmentMask = attachmentMask;
1387     pDC->FeWork.desc.discardInvalidateTiles.rect = rect;
1388     pDC->FeWork.desc.discardInvalidateTiles.rect &= g_MaxScissorRect;
1389     pDC->FeWork.desc.discardInvalidateTiles.newTileState = SWR_TILE_RESOLVED;
1390     pDC->FeWork.desc.discardInvalidateTiles.createNewTiles = true;
1391     pDC->FeWork.desc.discardInvalidateTiles.fullTilesOnly = true;
1392
1393     //enqueue
1394     QueueDraw(pContext);
1395 }
1396
1397 //////////////////////////////////////////////////////////////////////////
1398 /// @brief SwrDispatch
1399 /// @param hContext - Handle passed back from SwrCreateContext
1400 /// @param threadGroupCountX - Number of thread groups dispatched in X direction
1401 /// @param threadGroupCountY - Number of thread groups dispatched in Y direction
1402 /// @param threadGroupCountZ - Number of thread groups dispatched in Z direction
1403 void SwrDispatch(
1404     HANDLE hContext,
1405     uint32_t threadGroupCountX,
1406     uint32_t threadGroupCountY,
1407     uint32_t threadGroupCountZ)
1408 {
1409     if (KNOB_TOSS_DRAW)
1410     {
1411         return;
1412     }
1413
1414     SWR_CONTEXT *pContext = GetContext(hContext);
1415     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1416
1417     AR_API_BEGIN(APIDispatch, pDC->drawId);
1418
1419     pDC->isCompute = true;      // This is a compute context.
1420
1421     COMPUTE_DESC* pTaskData = (COMPUTE_DESC*)pDC->pArena->AllocAligned(sizeof(COMPUTE_DESC), 64);
1422
1423     pTaskData->threadGroupCountX = threadGroupCountX;
1424     pTaskData->threadGroupCountY = threadGroupCountY;
1425     pTaskData->threadGroupCountZ = threadGroupCountZ;
1426
1427     uint32_t totalThreadGroups = threadGroupCountX * threadGroupCountY * threadGroupCountZ;
1428     uint32_t dcIndex = pDC->drawId % KNOB_MAX_DRAWS_IN_FLIGHT;
1429     pDC->pDispatch = &pContext->pDispatchQueueArray[dcIndex];
1430     pDC->pDispatch->initialize(totalThreadGroups, pTaskData, &ProcessComputeBE);
1431
1432     QueueDispatch(pContext);
1433     AR_API_END(APIDispatch, threadGroupCountX * threadGroupCountY * threadGroupCountZ);
1434 }
1435
1436 // Deswizzles, converts and stores current contents of the hot tiles to surface
1437 // described by pState
1438 void SWR_API SwrStoreTiles(
1439     HANDLE hContext,
1440     SWR_RENDERTARGET_ATTACHMENT attachment,
1441     SWR_TILE_STATE postStoreTileState,
1442     const SWR_RECT& storeRect)
1443 {
1444     if (KNOB_TOSS_DRAW)
1445     {
1446         return;
1447     }
1448
1449     SWR_CONTEXT *pContext = GetContext(hContext);
1450     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1451
1452     AR_API_BEGIN(APIStoreTiles, pDC->drawId);
1453
1454     pDC->FeWork.type = STORETILES;
1455     pDC->FeWork.pfnWork = ProcessStoreTiles;
1456     pDC->FeWork.desc.storeTiles.attachment = attachment;
1457     pDC->FeWork.desc.storeTiles.postStoreTileState = postStoreTileState;
1458     pDC->FeWork.desc.storeTiles.rect = storeRect;
1459     pDC->FeWork.desc.storeTiles.rect &= g_MaxScissorRect;
1460
1461     //enqueue
1462     QueueDraw(pContext);
1463
1464     AR_API_END(APIStoreTiles, 1);
1465 }
1466
1467 //////////////////////////////////////////////////////////////////////////
1468 /// @brief SwrClearRenderTarget - Clear attached render targets / depth / stencil
1469 /// @param hContext - Handle passed back from SwrCreateContext
1470 /// @param clearMask - combination of SWR_CLEAR_COLOR / SWR_CLEAR_DEPTH / SWR_CLEAR_STENCIL flags (or SWR_CLEAR_NONE)
1471 /// @param clearColor - color use for clearing render targets
1472 /// @param z - depth value use for clearing depth buffer
1473 /// @param stencil - stencil value used for clearing stencil buffer
1474 /// @param clearRect - The pixel-coordinate rectangle to clear in all cleared buffers
1475 void SWR_API SwrClearRenderTarget(
1476     HANDLE hContext,
1477     uint32_t clearMask,
1478     const float clearColor[4],
1479     float z,
1480     uint8_t stencil,
1481     const SWR_RECT& clearRect)
1482 {
1483     if (KNOB_TOSS_DRAW)
1484     {
1485         return;
1486     }
1487
1488     SWR_CONTEXT *pContext = GetContext(hContext);
1489     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1490
1491     AR_API_BEGIN(APIClearRenderTarget, pDC->drawId);
1492
1493     CLEAR_FLAGS flags;
1494     flags.bits = 0;
1495     flags.mask = clearMask;
1496
1497     pDC->FeWork.type = CLEAR;
1498     pDC->FeWork.pfnWork = ProcessClear;
1499     pDC->FeWork.desc.clear.rect = clearRect;
1500     pDC->FeWork.desc.clear.rect &= g_MaxScissorRect;
1501     pDC->FeWork.desc.clear.flags = flags;
1502     pDC->FeWork.desc.clear.clearDepth = z;
1503     pDC->FeWork.desc.clear.clearRTColor[0] = clearColor[0];
1504     pDC->FeWork.desc.clear.clearRTColor[1] = clearColor[1];
1505     pDC->FeWork.desc.clear.clearRTColor[2] = clearColor[2];
1506     pDC->FeWork.desc.clear.clearRTColor[3] = clearColor[3];
1507     pDC->FeWork.desc.clear.clearStencil = stencil;
1508
1509     // enqueue draw
1510     QueueDraw(pContext);
1511
1512     AR_API_END(APIClearRenderTarget, 1);
1513 }
1514
1515 //////////////////////////////////////////////////////////////////////////
1516 /// @brief Returns a pointer to the private context state for the current
1517 ///        draw operation. This is used for external componets such as the
1518 ///        sampler.
1519 ///        SWR is responsible for the allocation of the private context state.
1520 /// @param hContext - Handle passed back from SwrCreateContext
1521 VOID* SwrGetPrivateContextState(
1522     HANDLE hContext)
1523 {
1524     SWR_CONTEXT* pContext = GetContext(hContext);
1525     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1526     DRAW_STATE* pState = pDC->pState;
1527
1528     if (pState->pPrivateState == nullptr)
1529     {
1530         pState->pPrivateState = pState->pArena->AllocAligned(pContext->privateStateSize, KNOB_SIMD_WIDTH*sizeof(float));
1531     }
1532
1533     return pState->pPrivateState;
1534 }
1535
1536 //////////////////////////////////////////////////////////////////////////
1537 /// @brief Clients can use this to allocate memory for draw/dispatch
1538 ///        operations. The memory will automatically be freed once operation
1539 ///        has completed. Client can use this to allocate binding tables,
1540 ///        etc. needed for shader execution.
1541 /// @param hContext - Handle passed back from SwrCreateContext
1542 /// @param size - Size of allocation
1543 /// @param align - Alignment needed for allocation.
1544 VOID* SwrAllocDrawContextMemory(
1545     HANDLE hContext,
1546     uint32_t size,
1547     uint32_t align)
1548 {
1549     SWR_CONTEXT* pContext = GetContext(hContext);
1550     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1551
1552     return pDC->pState->pArena->AllocAligned(size, align);
1553 }
1554
1555 //////////////////////////////////////////////////////////////////////////
1556 /// @brief Enables stats counting
1557 /// @param hContext - Handle passed back from SwrCreateContext
1558 /// @param enable - If true then counts are incremented.
1559 void SwrEnableStats(
1560     HANDLE hContext,
1561     bool enable)
1562 {
1563     SWR_CONTEXT *pContext = GetContext(hContext);
1564     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1565
1566     pDC->pState->state.enableStats = enable;
1567 }
1568
1569 //////////////////////////////////////////////////////////////////////////
1570 /// @brief Mark end of frame - used for performance profiling
1571 /// @param hContext - Handle passed back from SwrCreateContext
1572 void SWR_API SwrEndFrame(
1573     HANDLE hContext)
1574 {
1575     RDTSC_ENDFRAME();
1576     SWR_CONTEXT *pContext = GetContext(hContext);
1577     pContext->frameCount++;
1578 }