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