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 //////////////////////////////////////////////////////////////////////////
  59 /// @brief Create SWR Context.
  60 /// @param pCreateInfo - pointer to creation info.
  61 HANDLE SwrCreateContext(
  62     SWR_CREATECONTEXT_INFO* pCreateInfo)
  63 {
  64     RDTSC_RESET();
  65     RDTSC_INIT(0);
  66
  67     void* pContextMem = AlignedMalloc(sizeof(SWR_CONTEXT), KNOB_SIMD_WIDTH * 4);
  68     memset(pContextMem, 0, sizeof(SWR_CONTEXT));
  69     SWR_CONTEXT *pContext = new (pContextMem) SWR_CONTEXT();
  70
  71     pContext->driverType = pCreateInfo->driver;
  72     pContext->privateStateSize = pCreateInfo->privateStateSize;
  73
  74     pContext->dcRing.Init(KNOB_MAX_DRAWS_IN_FLIGHT);
  75     pContext->dsRing.Init(KNOB_MAX_DRAWS_IN_FLIGHT);
  76
  77     pContext->pMacroTileManagerArray = (MacroTileMgr*)AlignedMalloc(sizeof(MacroTileMgr) * KNOB_MAX_DRAWS_IN_FLIGHT, 64);
  78     pContext->pDispatchQueueArray = (DispatchQueue*)AlignedMalloc(sizeof(DispatchQueue) * KNOB_MAX_DRAWS_IN_FLIGHT, 64);
  79
  80     pContext->threadInfo.MAX_WORKER_THREADS        = KNOB_MAX_WORKER_THREADS;
  81     pContext->threadInfo.MAX_NUMA_NODES            = KNOB_MAX_NUMA_NODES;
  82     pContext->threadInfo.MAX_CORES_PER_NUMA_NODE   = KNOB_MAX_CORES_PER_NUMA_NODE;
  83     pContext->threadInfo.MAX_THREADS_PER_CORE      = KNOB_MAX_THREADS_PER_CORE;
  84     pContext->threadInfo.SINGLE_THREADED           = KNOB_SINGLE_THREADED;
  85
  86     if (pCreateInfo->pThreadInfo)
  87     {
  88         pContext->threadInfo = *pCreateInfo->pThreadInfo;
  89     }
  90
  91     for (uint32_t dc = 0; dc < KNOB_MAX_DRAWS_IN_FLIGHT; ++dc)
  92     {
  93         pContext->dcRing[dc].pArena = new CachingArena(pContext->cachingArenaAllocator);
  94         new (&pContext->pMacroTileManagerArray[dc]) MacroTileMgr(*pContext->dcRing[dc].pArena);
  95         new (&pContext->pDispatchQueueArray[dc]) DispatchQueue();
  96
  97         pContext->dsRing[dc].pArena = new CachingArena(pContext->cachingArenaAllocator);
  98     }
  99
 100     if (!pContext->threadInfo.SINGLE_THREADED)
 101     {
 102         memset(&pContext->WaitLock, 0, sizeof(pContext->WaitLock));
 103         memset(&pContext->FifosNotEmpty, 0, sizeof(pContext->FifosNotEmpty));
 104         new (&pContext->WaitLock) std::mutex();
 105         new (&pContext->FifosNotEmpty) std::condition_variable();
 106
 107         CreateThreadPool(pContext, &pContext->threadPool);
 108     }
 109
 110     // Calling createThreadPool() above can set SINGLE_THREADED
 111     if (pContext->threadInfo.SINGLE_THREADED)
 112     {
 113         pContext->NumWorkerThreads = 1;
 114         pContext->NumFEThreads = 1;
 115         pContext->NumBEThreads = 1;
 116     }
 117
 118     // Allocate scratch space for workers.
 119     ///@note We could lazily allocate this but its rather small amount of memory.
 120     for (uint32_t i = 0; i < pContext->NumWorkerThreads; ++i)
 121     {
 122 #if defined(_WIN32)
 123         uint32_t numaNode = pContext->threadPool.pThreadData ?
 124             pContext->threadPool.pThreadData[i].numaId : 0;
 125         pContext->pScratch[i] = (uint8_t*)VirtualAllocExNuma(
 126             GetCurrentProcess(), nullptr, 32 * sizeof(KILOBYTE),
 127             MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE,
 128             numaNode);
 129 #else
 130         pContext->pScratch[i] = (uint8_t*)AlignedMalloc(32 * sizeof(KILOBYTE), KNOB_SIMD_WIDTH * 4);
 131 #endif
 132     }
 133
 134     // State setup AFTER context is fully initialized
 135     SetupDefaultState(pContext);
 136
 137     // initialize hot tile manager
 138     pContext->pHotTileMgr = new HotTileMgr();
 139
 140     // initialize function pointer tables
 141     InitClearTilesTable();
 142
 143     // initialize callback functions
 144     pContext->pfnLoadTile = pCreateInfo->pfnLoadTile;
 145     pContext->pfnStoreTile = pCreateInfo->pfnStoreTile;
 146     pContext->pfnClearTile = pCreateInfo->pfnClearTile;
 147     pContext->pfnUpdateSoWriteOffset = pCreateInfo->pfnUpdateSoWriteOffset;
 148     pContext->pfnUpdateStats = pCreateInfo->pfnUpdateStats;
 149     pContext->pfnUpdateStatsFE = pCreateInfo->pfnUpdateStatsFE;
 150
 151     // pass pointer to bucket manager back to caller
 152 #ifdef KNOB_ENABLE_RDTSC
 153     pCreateInfo->pBucketMgr = &gBucketMgr;
 154 #endif
 155
 156     pCreateInfo->contextSaveSize = sizeof(API_STATE);
 157
 158     return (HANDLE)pContext;
 159 }
 160
 161 void SwrDestroyContext(HANDLE hContext)
 162 {
 163     SWR_CONTEXT *pContext = GetContext(hContext);
 164     DestroyThreadPool(pContext, &pContext->threadPool);
 165
 166     // free the fifos
 167     for (uint32_t i = 0; i < KNOB_MAX_DRAWS_IN_FLIGHT; ++i)
 168     {
 169         delete pContext->dcRing[i].pArena;
 170         delete pContext->dsRing[i].pArena;
 171         pContext->pMacroTileManagerArray[i].~MacroTileMgr();
 172         pContext->pDispatchQueueArray[i].~DispatchQueue();
 173     }
 174
 175     AlignedFree(pContext->pDispatchQueueArray);
 176     AlignedFree(pContext->pMacroTileManagerArray);
 177
 178     // Free scratch space.
 179     for (uint32_t i = 0; i < pContext->NumWorkerThreads; ++i)
 180     {
 181 #if defined(_WIN32)
 182         VirtualFree(pContext->pScratch[i], 0, MEM_RELEASE);
 183 #else
 184         AlignedFree(pContext->pScratch[i]);
 185 #endif
 186     }
 187
 188     delete(pContext->pHotTileMgr);
 189
 190     pContext->~SWR_CONTEXT();
 191     AlignedFree(GetContext(hContext));
 192 }
 193
 194 void CopyState(DRAW_STATE& dst, const DRAW_STATE& src)
 195 {
 196     memcpy(&dst.state, &src.state, sizeof(API_STATE));
 197 }
 198
 199 void WakeAllThreads(SWR_CONTEXT *pContext)
 200 {
 201     pContext->FifosNotEmpty.notify_all();
 202 }
 203
 204 template<bool IsDraw>
 205 void QueueWork(SWR_CONTEXT *pContext)
 206 {
 207     DRAW_CONTEXT* pDC = pContext->pCurDrawContext;
 208     uint32_t dcIndex = pDC->drawId % KNOB_MAX_DRAWS_IN_FLIGHT;
 209
 210     if (IsDraw)
 211     {
 212         pDC->pTileMgr = &pContext->pMacroTileManagerArray[dcIndex];
 213         pDC->pTileMgr->initialize();
 214     }
 215
 216     // Each worker thread looks at a DC for both FE and BE work at different times and so we
 217     // multiply threadDone by 2.  When the threadDone counter has reached 0 then all workers
 218     // have moved past this DC. (i.e. Each worker has checked this DC for both FE and BE work and
 219     // then moved on if all work is done.)
 220     pContext->pCurDrawContext->threadsDone = pContext->NumFEThreads + pContext->NumBEThreads;
 221
 222     if (IsDraw)
 223     {
 224         InterlockedIncrement((volatile LONG*)&pContext->drawsOutstandingFE);
 225     }
 226
 227     _ReadWriteBarrier();
 228     {
 229         std::unique_lock<std::mutex> lock(pContext->WaitLock);
 230         pContext->dcRing.Enqueue();
 231     }
 232
 233     if (pContext->threadInfo.SINGLE_THREADED)
 234     {
 235         // flush denormals to 0
 236         uint32_t mxcsr = _mm_getcsr();
 237         _mm_setcsr(mxcsr | _MM_FLUSH_ZERO_ON | _MM_DENORMALS_ZERO_ON);
 238
 239         if (IsDraw)
 240         {
 241             uint32_t curDraw[2] = { pContext->pCurDrawContext->drawId, pContext->pCurDrawContext->drawId };
 242             WorkOnFifoFE(pContext, 0, curDraw[0]);
 243             WorkOnFifoBE(pContext, 0, curDraw[1], pContext->singleThreadLockedTiles, 0, 0);
 244         }
 245         else
 246         {
 247             uint32_t curDispatch = pContext->pCurDrawContext->drawId;
 248             WorkOnCompute(pContext, 0, curDispatch);
 249         }
 250
 251         // Dequeue the work here, if not already done, since we're single threaded (i.e. no workers).
 252         while (CompleteDrawContext(pContext, pContext->pCurDrawContext) > 0) {}
 253
 254         // restore csr
 255         _mm_setcsr(mxcsr);
 256     }
 257     else
 258     {
 259         RDTSC_START(APIDrawWakeAllThreads);
 260         WakeAllThreads(pContext);
 261         RDTSC_STOP(APIDrawWakeAllThreads, 1, 0);
 262     }
 263
 264     // Set current draw context to NULL so that next state call forces a new draw context to be created and populated.
 265     pContext->pPrevDrawContext = pContext->pCurDrawContext;
 266     pContext->pCurDrawContext = nullptr;
 267 }
 268
 269 INLINE void QueueDraw(SWR_CONTEXT* pContext)
 270 {
 271     QueueWork<true>(pContext);
 272 }
 273
 274 INLINE void QueueDispatch(SWR_CONTEXT* pContext)
 275 {
 276     QueueWork<false>(pContext);
 277 }
 278
 279 DRAW_CONTEXT* GetDrawContext(SWR_CONTEXT *pContext, bool isSplitDraw = false)
 280 {
 281     RDTSC_START(APIGetDrawContext);
 282     // If current draw context is null then need to obtain a new draw context to use from ring.
 283     if (pContext->pCurDrawContext == nullptr)
 284     {
 285         // Need to wait for a free entry.
 286         while (pContext->dcRing.IsFull())
 287         {
 288             _mm_pause();
 289         }
 290
 291         uint64_t curDraw = pContext->dcRing.GetHead();
 292         uint32_t dcIndex = curDraw % KNOB_MAX_DRAWS_IN_FLIGHT;
 293
 294         if ((pContext->frameCount - pContext->lastFrameChecked) > 2 ||
 295             (curDraw - pContext->lastDrawChecked) > 0x10000)
 296         {
 297             // Take this opportunity to clean-up old arena allocations
 298             pContext->cachingArenaAllocator.FreeOldBlocks();
 299
 300             pContext->lastFrameChecked = pContext->frameCount;
 301             pContext->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_MATRICES* 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         // @todo Faster to copy portions of the SOA or just copy all of it?
 678         memcpy(&pState->vpMatrices, pMatrices, sizeof(SWR_VIEWPORT_MATRICES));
 679     }
 680     else
 681     {
 682         // Compute default viewport transform.
 683         for (uint32_t i = 0; i < numViewports; ++i)
 684         {
 685             if (pContext->driverType == DX)
 686             {
 687                 pState->vpMatrices.m00[i] = pState->vp[i].width / 2.0f;
 688                 pState->vpMatrices.m11[i] = -pState->vp[i].height / 2.0f;
 689                 pState->vpMatrices.m22[i] = pState->vp[i].maxZ - pState->vp[i].minZ;
 690                 pState->vpMatrices.m30[i] = pState->vp[i].x + pState->vpMatrices.m00[i];
 691                 pState->vpMatrices.m31[i] = pState->vp[i].y - pState->vpMatrices.m11[i];
 692                 pState->vpMatrices.m32[i] = pState->vp[i].minZ;
 693             }
 694             else
 695             {
 696                 // Standard, with the exception that Y is inverted.
 697                 pState->vpMatrices.m00[i] = (pState->vp[i].width - pState->vp[i].x) / 2.0f;
 698                 pState->vpMatrices.m11[i] = (pState->vp[i].y - pState->vp[i].height) / 2.0f;
 699                 pState->vpMatrices.m22[i] = (pState->vp[i].maxZ - pState->vp[i].minZ) / 2.0f;
 700                 pState->vpMatrices.m30[i] = pState->vp[i].x + pState->vpMatrices.m00[i];
 701                 pState->vpMatrices.m31[i] = pState->vp[i].height + pState->vpMatrices.m11[i];
 702                 pState->vpMatrices.m32[i] = pState->vp[i].minZ + pState->vpMatrices.m22[i];
 703
 704                 // Now that the matrix is calculated, clip the view coords to screen size.
 705                 // OpenGL allows for -ve x,y in the viewport.
 706                 pState->vp[i].x = std::max(pState->vp[i].x, 0.0f);
 707                 pState->vp[i].y = std::max(pState->vp[i].y, 0.0f);
 708             }
 709         }
 710     }
 711
 712     updateGuardband(pState);
 713 }
 714
 715 void SwrSetScissorRects(
 716     HANDLE hContext,
 717     uint32_t numScissors,
 718     const SWR_RECT* pScissors)
 719 {
 720     SWR_ASSERT(numScissors <= KNOB_NUM_VIEWPORTS_SCISSORS,
 721         "Invalid number of scissor rects.");
 722
 723     API_STATE* pState = GetDrawState(GetContext(hContext));
 724     memcpy(&pState->scissorRects[0], pScissors, numScissors * sizeof(pScissors[0]));
 725 };
 726
 727 void SetupMacroTileScissors(DRAW_CONTEXT *pDC)
 728 {
 729     API_STATE *pState = &pDC->pState->state;
 730
 731     // Set up scissor dimensions based on scissor or viewport
 732     if (pState->rastState.scissorEnable)
 733     {
 734         pState->scissorInFixedPoint = pState->scissorRects[0];
 735     }
 736     else
 737     {
 738         // the vp width and height must be added to origin un-rounded then the result round to -inf.
 739         // The cast to int works for rounding assuming all [left, right, top, bottom] are positive.
 740         pState->scissorInFixedPoint.xmin = (int32_t)pState->vp[0].x;
 741         pState->scissorInFixedPoint.xmax = (int32_t)(pState->vp[0].x + pState->vp[0].width);
 742         pState->scissorInFixedPoint.ymin = (int32_t)pState->vp[0].y;
 743         pState->scissorInFixedPoint.ymax = (int32_t)(pState->vp[0].y + pState->vp[0].height);
 744     }
 745
 746     // Clamp to max rect
 747     pState->scissorInFixedPoint &= g_MaxScissorRect;
 748
 749     // Scale to fixed point
 750     pState->scissorInFixedPoint.xmin *= FIXED_POINT_SCALE;
 751     pState->scissorInFixedPoint.xmax *= FIXED_POINT_SCALE;
 752     pState->scissorInFixedPoint.ymin *= FIXED_POINT_SCALE;
 753     pState->scissorInFixedPoint.ymax *= FIXED_POINT_SCALE;
 754
 755     // Make scissor inclusive
 756     pState->scissorInFixedPoint.xmax -= 1;
 757     pState->scissorInFixedPoint.ymax -= 1;
 758 }
 759
 760 // templated backend function tables
 761 extern PFN_BACKEND_FUNC gBackendNullPs[SWR_MULTISAMPLE_TYPE_COUNT];
 762 extern PFN_BACKEND_FUNC gBackendSingleSample[SWR_INPUT_COVERAGE_COUNT][2][2];
 763 extern PFN_BACKEND_FUNC gBackendPixelRateTable[SWR_MULTISAMPLE_TYPE_COUNT][SWR_MSAA_SAMPLE_PATTERN_COUNT][SWR_INPUT_COVERAGE_COUNT][2][2][2];
 764 extern PFN_BACKEND_FUNC gBackendSampleRateTable[SWR_MULTISAMPLE_TYPE_COUNT][SWR_INPUT_COVERAGE_COUNT][2][2];
 765 void SetupPipeline(DRAW_CONTEXT *pDC)
 766 {
 767     DRAW_STATE* pState = pDC->pState;
 768     const SWR_RASTSTATE &rastState = pState->state.rastState;
 769     const SWR_PS_STATE &psState = pState->state.psState;
 770     BACKEND_FUNCS& backendFuncs = pState->backendFuncs;
 771     const uint32_t forcedSampleCount = (rastState.forcedSampleCount) ? 1 : 0;
 772
 773     // setup backend
 774     if (psState.pfnPixelShader == nullptr)
 775     {
 776         backendFuncs.pfnBackend = gBackendNullPs[pState->state.rastState.sampleCount];
 777     }
 778     else
 779     {
 780         const bool bMultisampleEnable = ((rastState.sampleCount > SWR_MULTISAMPLE_1X) || rastState.forcedSampleCount) ? 1 : 0;
 781         const uint32_t centroid = ((psState.barycentricsMask & SWR_BARYCENTRIC_CENTROID_MASK) > 0) ? 1 : 0;
 782         const uint32_t canEarlyZ = (psState.forceEarlyZ || (!psState.writesODepth && !psState.usesSourceDepth && !psState.usesUAV)) ? 1 : 0;
 783
 784         SWR_BARYCENTRICS_MASK barycentricsMask = (SWR_BARYCENTRICS_MASK)psState.barycentricsMask;
 785
 786         // select backend function
 787         switch(psState.shadingRate)
 788         {
 789         case SWR_SHADING_RATE_PIXEL:
 790             if(bMultisampleEnable)
 791             {
 792                 // always need to generate I & J per sample for Z interpolation
 793                 barycentricsMask = (SWR_BARYCENTRICS_MASK)(barycentricsMask | SWR_BARYCENTRIC_PER_SAMPLE_MASK);
 794                 backendFuncs.pfnBackend = gBackendPixelRateTable[rastState.sampleCount][rastState.samplePattern][psState.inputCoverage][centroid][forcedSampleCount][canEarlyZ];
 795             }
 796             else
 797             {
 798                 // always need to generate I & J per pixel for Z interpolation
 799                 barycentricsMask = (SWR_BARYCENTRICS_MASK)(barycentricsMask | SWR_BARYCENTRIC_PER_PIXEL_MASK);
 800                 backendFuncs.pfnBackend = gBackendSingleSample[psState.inputCoverage][centroid][canEarlyZ];
 801             }
 802             break;
 803         case SWR_SHADING_RATE_SAMPLE:
 804             SWR_ASSERT(rastState.samplePattern == SWR_MSAA_STANDARD_PATTERN);
 805             // always need to generate I & J per sample for Z interpolation
 806             barycentricsMask = (SWR_BARYCENTRICS_MASK)(barycentricsMask | SWR_BARYCENTRIC_PER_SAMPLE_MASK);
 807             backendFuncs.pfnBackend = gBackendSampleRateTable[rastState.sampleCount][psState.inputCoverage][centroid][canEarlyZ];
 808             break;
 809         default:
 810             SWR_ASSERT(0 && "Invalid shading rate");
 811             break;
 812         }
 813     }
 814
 815     PFN_PROCESS_PRIMS pfnBinner;
 816     switch (pState->state.topology)
 817     {
 818     case TOP_POINT_LIST:
 819         pState->pfnProcessPrims = ClipPoints;
 820         pfnBinner = BinPoints;
 821         break;
 822     case TOP_LINE_LIST:
 823     case TOP_LINE_STRIP:
 824     case TOP_LINE_LOOP:
 825     case TOP_LINE_LIST_ADJ:
 826     case TOP_LISTSTRIP_ADJ:
 827         pState->pfnProcessPrims = ClipLines;
 828         pfnBinner = BinLines;
 829         break;
 830     default:
 831         pState->pfnProcessPrims = ClipTriangles;
 832         pfnBinner = GetBinTrianglesFunc((rastState.conservativeRast > 0));
 833         break;
 834     };
 835
 836     // disable clipper if viewport transform is disabled
 837     if (pState->state.frontendState.vpTransformDisable)
 838     {
 839         pState->pfnProcessPrims = pfnBinner;
 840     }
 841
 842     if ((pState->state.psState.pfnPixelShader == nullptr) &&
 843         (pState->state.depthStencilState.depthTestEnable == FALSE) &&
 844         (pState->state.depthStencilState.depthWriteEnable == FALSE) &&
 845         (pState->state.depthStencilState.stencilTestEnable == FALSE) &&
 846         (pState->state.depthStencilState.stencilWriteEnable == FALSE) &&
 847         (pState->state.backendState.numAttributes == 0))
 848     {
 849         pState->pfnProcessPrims = nullptr;
 850     }
 851
 852     if (pState->state.soState.rasterizerDisable == true)
 853     {
 854         pState->pfnProcessPrims = nullptr;
 855     }
 856
 857     // set up the frontend attribute count
 858     pState->state.feNumAttributes = 0;
 859     const SWR_BACKEND_STATE& backendState = pState->state.backendState;
 860     if (backendState.swizzleEnable)
 861     {
 862         // attribute swizzling is enabled, iterate over the map and record the max attribute used
 863         for (uint32_t i = 0; i < backendState.numAttributes; ++i)
 864         {
 865             pState->state.feNumAttributes = std::max(pState->state.feNumAttributes, (uint32_t)backendState.swizzleMap[i].sourceAttrib + 1);
 866         }
 867     }
 868     else
 869     {
 870         pState->state.feNumAttributes = pState->state.backendState.numAttributes;
 871     }
 872
 873     if (pState->state.soState.soEnable)
 874     {
 875         uint32_t streamMasks = 0;
 876         for (uint32_t i = 0; i < 4; ++i)
 877         {
 878             streamMasks |= pState->state.soState.streamMasks[i];
 879         }
 880
 881         DWORD maxAttrib;
 882         if (_BitScanReverse(&maxAttrib, streamMasks))
 883         {
 884             pState->state.feNumAttributes = std::max(pState->state.feNumAttributes, (uint32_t)(maxAttrib + 1));
 885         }
 886     }
 887
 888     // complicated logic to test for cases where we don't need backing hottile memory for a draw
 889     // have to check for the special case where depth/stencil test is enabled but depthwrite is disabled.
 890     pState->state.depthHottileEnable = ((!(pState->state.depthStencilState.depthTestEnable &&
 891                                            !pState->state.depthStencilState.depthWriteEnable &&
 892                                            pState->state.depthStencilState.depthTestFunc == ZFUNC_ALWAYS)) &&
 893                                         (pState->state.depthStencilState.depthTestEnable ||
 894                                          pState->state.depthStencilState.depthWriteEnable)) ? true : false;
 895
 896     pState->state.stencilHottileEnable = (((!(pState->state.depthStencilState.stencilTestEnable &&
 897                                              !pState->state.depthStencilState.stencilWriteEnable &&
 898                                               pState->state.depthStencilState.stencilTestFunc == ZFUNC_ALWAYS)) ||
 899                                           // for stencil we have to check the double sided state as well
 900                                           (!(pState->state.depthStencilState.doubleSidedStencilTestEnable &&
 901                                              !pState->state.depthStencilState.stencilWriteEnable &&
 902                                               pState->state.depthStencilState.backfaceStencilTestFunc == ZFUNC_ALWAYS))) &&
 903                                           (pState->state.depthStencilState.stencilTestEnable  ||
 904                                            pState->state.depthStencilState.stencilWriteEnable)) ? true : false;
 905
 906     uint32_t numRTs = pState->state.psState.numRenderTargets;
 907     pState->state.colorHottileEnable = 0;
 908     if (psState.pfnPixelShader != nullptr)
 909     {
 910         for (uint32_t rt = 0; rt < numRTs; ++rt)
 911         {
 912             pState->state.colorHottileEnable |=
 913                 (!pState->state.blendState.renderTarget[rt].writeDisableAlpha ||
 914                  !pState->state.blendState.renderTarget[rt].writeDisableRed ||
 915                  !pState->state.blendState.renderTarget[rt].writeDisableGreen ||
 916                  !pState->state.blendState.renderTarget[rt].writeDisableBlue) ? (1 << rt) : 0;
 917         }
 918     }
 919
 920     // Setup depth quantization function
 921     if (pState->state.depthHottileEnable)
 922     {
 923         switch (pState->state.rastState.depthFormat)
 924         {
 925         case R32_FLOAT_X8X24_TYPELESS: pState->state.pfnQuantizeDepth = QuantizeDepth < R32_FLOAT_X8X24_TYPELESS > ; break;
 926         case R32_FLOAT: pState->state.pfnQuantizeDepth = QuantizeDepth < R32_FLOAT > ; break;
 927         case R24_UNORM_X8_TYPELESS: pState->state.pfnQuantizeDepth = QuantizeDepth < R24_UNORM_X8_TYPELESS > ; break;
 928         case R16_UNORM: pState->state.pfnQuantizeDepth = QuantizeDepth < R16_UNORM > ; break;
 929         default: SWR_ASSERT(false, "Unsupported depth format for depth quantiztion.");
 930             pState->state.pfnQuantizeDepth = QuantizeDepth < R32_FLOAT > ;
 931         }
 932     }
 933     else
 934     {
 935         // set up pass-through quantize if depth isn't enabled
 936         pState->state.pfnQuantizeDepth = QuantizeDepth < R32_FLOAT > ;
 937     }
 938 }
 939
 940 //////////////////////////////////////////////////////////////////////////
 941 /// @brief InitDraw
 942 /// @param pDC - Draw context to initialize for this draw.
 943 void InitDraw(
 944     DRAW_CONTEXT *pDC,
 945     bool isSplitDraw)
 946 {
 947     // We don't need to re-setup the scissors/pipeline state again for split draw.
 948     if (isSplitDraw == false)
 949     {
 950         SetupMacroTileScissors(pDC);
 951         SetupPipeline(pDC);
 952     }
 953 }
 954
 955 //////////////////////////////////////////////////////////////////////////
 956 /// @brief We can split the draw for certain topologies for better performance.
 957 /// @param totalVerts - Total vertices for draw
 958 /// @param topology - Topology used for draw
 959 uint32_t MaxVertsPerDraw(
 960     DRAW_CONTEXT* pDC,
 961     uint32_t totalVerts,
 962     PRIMITIVE_TOPOLOGY topology)
 963 {
 964     API_STATE& state = pDC->pState->state;
 965
 966     uint32_t vertsPerDraw = totalVerts;
 967
 968     if (state.soState.soEnable)
 969     {
 970         return totalVerts;
 971     }
 972
 973     switch (topology)
 974     {
 975     case TOP_POINT_LIST:
 976     case TOP_TRIANGLE_LIST:
 977         vertsPerDraw = KNOB_MAX_PRIMS_PER_DRAW;
 978         break;
 979
 980     case TOP_PATCHLIST_1:
 981     case TOP_PATCHLIST_2:
 982     case TOP_PATCHLIST_3:
 983     case TOP_PATCHLIST_4:
 984     case TOP_PATCHLIST_5:
 985     case TOP_PATCHLIST_6:
 986     case TOP_PATCHLIST_7:
 987     case TOP_PATCHLIST_8:
 988     case TOP_PATCHLIST_9:
 989     case TOP_PATCHLIST_10:
 990     case TOP_PATCHLIST_11:
 991     case TOP_PATCHLIST_12:
 992     case TOP_PATCHLIST_13:
 993     case TOP_PATCHLIST_14:
 994     case TOP_PATCHLIST_15:
 995     case TOP_PATCHLIST_16:
 996     case TOP_PATCHLIST_17:
 997     case TOP_PATCHLIST_18:
 998     case TOP_PATCHLIST_19:
 999     case TOP_PATCHLIST_20:
1000     case TOP_PATCHLIST_21:
1001     case TOP_PATCHLIST_22:
1002     case TOP_PATCHLIST_23:
1003     case TOP_PATCHLIST_24:
1004     case TOP_PATCHLIST_25:
1005     case TOP_PATCHLIST_26:
1006     case TOP_PATCHLIST_27:
1007     case TOP_PATCHLIST_28:
1008     case TOP_PATCHLIST_29:
1009     case TOP_PATCHLIST_30:
1010     case TOP_PATCHLIST_31:
1011     case TOP_PATCHLIST_32:
1012         if (pDC->pState->state.tsState.tsEnable)
1013         {
1014             uint32_t vertsPerPrim = topology - TOP_PATCHLIST_BASE;
1015             vertsPerDraw = vertsPerPrim * KNOB_MAX_TESS_PRIMS_PER_DRAW;
1016         }
1017         break;
1018
1019     // The Primitive Assembly code can only handle 1 RECT at a time.
1020     case TOP_RECT_LIST:
1021         vertsPerDraw = 3;
1022         break;
1023
1024     default:
1025         // We are not splitting up draws for other topologies.
1026         break;
1027     }
1028
1029     return vertsPerDraw;
1030 }
1031
1032
1033 //////////////////////////////////////////////////////////////////////////
1034 /// @brief DrawInstanced
1035 /// @param hContext - Handle passed back from SwrCreateContext
1036 /// @param topology - Specifies topology for draw.
1037 /// @param numVerts - How many vertices to read sequentially from vertex data (per instance).
1038 /// @param startVertex - Specifies start vertex for draw. (vertex data)
1039 /// @param numInstances - How many instances to render.
1040 /// @param startInstance - Which instance to start sequentially fetching from in each buffer (instanced data)
1041 void DrawInstanced(
1042     HANDLE hContext,
1043     PRIMITIVE_TOPOLOGY topology,
1044     uint32_t numVertices,
1045     uint32_t startVertex,
1046     uint32_t numInstances = 1,
1047     uint32_t startInstance = 0)
1048 {
1049     if (KNOB_TOSS_DRAW)
1050     {
1051         return;
1052     }
1053
1054     RDTSC_START(APIDraw);
1055
1056     SWR_CONTEXT *pContext = GetContext(hContext);
1057     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1058
1059     uint32_t maxVertsPerDraw = MaxVertsPerDraw(pDC, numVertices, topology);
1060     uint32_t primsPerDraw = GetNumPrims(topology, maxVertsPerDraw);
1061     uint32_t remainingVerts = numVertices;
1062
1063     API_STATE    *pState = &pDC->pState->state;
1064     pState->topology = topology;
1065     pState->forceFront = false;
1066
1067     // disable culling for points/lines
1068     uint32_t oldCullMode = pState->rastState.cullMode;
1069     if (topology == TOP_POINT_LIST)
1070     {
1071         pState->rastState.cullMode = SWR_CULLMODE_NONE;
1072         pState->forceFront = true;
1073     }
1074
1075     int draw = 0;
1076     while (remainingVerts)
1077     {
1078         uint32_t numVertsForDraw = (remainingVerts < maxVertsPerDraw) ?
1079         remainingVerts : maxVertsPerDraw;
1080
1081         bool isSplitDraw = (draw > 0) ? true : false;
1082         DRAW_CONTEXT* pDC = GetDrawContext(pContext, isSplitDraw);
1083         InitDraw(pDC, isSplitDraw);
1084
1085         pDC->FeWork.type = DRAW;
1086         pDC->FeWork.pfnWork = GetProcessDrawFunc(
1087             false,  // IsIndexed
1088             false, // bEnableCutIndex
1089             pState->tsState.tsEnable,
1090             pState->gsState.gsEnable,
1091             pState->soState.soEnable,
1092             pDC->pState->pfnProcessPrims != nullptr);
1093         pDC->FeWork.desc.draw.numVerts = numVertsForDraw;
1094         pDC->FeWork.desc.draw.startVertex = startVertex;
1095         pDC->FeWork.desc.draw.numInstances = numInstances;
1096         pDC->FeWork.desc.draw.startInstance = startInstance;
1097         pDC->FeWork.desc.draw.startPrimID = draw * primsPerDraw;
1098         pDC->FeWork.desc.draw.startVertexID = draw * maxVertsPerDraw;
1099
1100         pDC->cleanupState = (remainingVerts == numVertsForDraw);
1101
1102         //enqueue DC
1103         QueueDraw(pContext);
1104
1105         remainingVerts -= numVertsForDraw;
1106         draw++;
1107     }
1108
1109     // restore culling state
1110     pDC = GetDrawContext(pContext);
1111     pDC->pState->state.rastState.cullMode = oldCullMode;
1112
1113     RDTSC_STOP(APIDraw, numVertices * numInstances, 0);
1114 }
1115
1116 //////////////////////////////////////////////////////////////////////////
1117 /// @brief SwrDraw
1118 /// @param hContext - Handle passed back from SwrCreateContext
1119 /// @param topology - Specifies topology for draw.
1120 /// @param startVertex - Specifies start vertex in vertex buffer for draw.
1121 /// @param primCount - Number of vertices.
1122 void SwrDraw(
1123     HANDLE hContext,
1124     PRIMITIVE_TOPOLOGY topology,
1125     uint32_t startVertex,
1126     uint32_t numVertices)
1127 {
1128     DrawInstanced(hContext, topology, numVertices, startVertex);
1129 }
1130
1131 //////////////////////////////////////////////////////////////////////////
1132 /// @brief SwrDrawInstanced
1133 /// @param hContext - Handle passed back from SwrCreateContext
1134 /// @param topology - Specifies topology for draw.
1135 /// @param numVertsPerInstance - How many vertices to read sequentially from vertex data.
1136 /// @param numInstances - How many instances to render.
1137 /// @param startVertex - Specifies start vertex for draw. (vertex data)
1138 /// @param startInstance - Which instance to start sequentially fetching from in each buffer (instanced data)
1139 void SwrDrawInstanced(
1140     HANDLE hContext,
1141     PRIMITIVE_TOPOLOGY topology,
1142     uint32_t numVertsPerInstance,
1143     uint32_t numInstances,
1144     uint32_t startVertex,
1145     uint32_t startInstance
1146     )
1147 {
1148     DrawInstanced(hContext, topology, numVertsPerInstance, startVertex, numInstances, startInstance);
1149 }
1150
1151 //////////////////////////////////////////////////////////////////////////
1152 /// @brief DrawIndexedInstanced
1153 /// @param hContext - Handle passed back from SwrCreateContext
1154 /// @param topology - Specifies topology for draw.
1155 /// @param numIndices - Number of indices to read sequentially from index buffer.
1156 /// @param indexOffset - Starting index into index buffer.
1157 /// @param baseVertex - Vertex in vertex buffer to consider as index "0". Note value is signed.
1158 /// @param numInstances - Number of instances to render.
1159 /// @param startInstance - Which instance to start sequentially fetching from in each buffer (instanced data)
1160 void DrawIndexedInstance(
1161     HANDLE hContext,
1162     PRIMITIVE_TOPOLOGY topology,
1163     uint32_t numIndices,
1164     uint32_t indexOffset,
1165     int32_t baseVertex,
1166     uint32_t numInstances = 1,
1167     uint32_t startInstance = 0)
1168 {
1169     if (KNOB_TOSS_DRAW)
1170     {
1171         return;
1172     }
1173
1174     RDTSC_START(APIDrawIndexed);
1175
1176     SWR_CONTEXT *pContext = GetContext(hContext);
1177     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1178     API_STATE* pState = &pDC->pState->state;
1179
1180     uint32_t maxIndicesPerDraw = MaxVertsPerDraw(pDC, numIndices, topology);
1181     uint32_t primsPerDraw = GetNumPrims(topology, maxIndicesPerDraw);
1182     uint32_t remainingIndices = numIndices;
1183
1184     uint32_t indexSize = 0;
1185     switch (pState->indexBuffer.format)
1186     {
1187     case R32_UINT: indexSize = sizeof(uint32_t); break;
1188     case R16_UINT: indexSize = sizeof(uint16_t); break;
1189     case R8_UINT: indexSize = sizeof(uint8_t); break;
1190     default:
1191         SWR_ASSERT(0);
1192     }
1193
1194     int draw = 0;
1195     uint8_t *pIB = (uint8_t*)pState->indexBuffer.pIndices;
1196     pIB += (uint64_t)indexOffset * (uint64_t)indexSize;
1197
1198     pState->topology = topology;
1199     pState->forceFront = false;
1200
1201     // disable culling for points/lines
1202     uint32_t oldCullMode = pState->rastState.cullMode;
1203     if (topology == TOP_POINT_LIST)
1204     {
1205         pState->rastState.cullMode = SWR_CULLMODE_NONE;
1206         pState->forceFront = true;
1207     }
1208
1209     while (remainingIndices)
1210     {
1211         uint32_t numIndicesForDraw = (remainingIndices < maxIndicesPerDraw) ?
1212         remainingIndices : maxIndicesPerDraw;
1213
1214         // When breaking up draw, we need to obtain new draw context for each iteration.
1215         bool isSplitDraw = (draw > 0) ? true : false;
1216         pDC = GetDrawContext(pContext, isSplitDraw);
1217         InitDraw(pDC, isSplitDraw);
1218
1219         pDC->FeWork.type = DRAW;
1220         pDC->FeWork.pfnWork = GetProcessDrawFunc(
1221             true,   // IsIndexed
1222             pState->frontendState.bEnableCutIndex,
1223             pState->tsState.tsEnable,
1224             pState->gsState.gsEnable,
1225             pState->soState.soEnable,
1226             pDC->pState->pfnProcessPrims != nullptr);
1227         pDC->FeWork.desc.draw.pDC = pDC;
1228         pDC->FeWork.desc.draw.numIndices = numIndicesForDraw;
1229         pDC->FeWork.desc.draw.pIB = (int*)pIB;
1230         pDC->FeWork.desc.draw.type = pDC->pState->state.indexBuffer.format;
1231
1232         pDC->FeWork.desc.draw.numInstances = numInstances;
1233         pDC->FeWork.desc.draw.startInstance = startInstance;
1234         pDC->FeWork.desc.draw.baseVertex = baseVertex;
1235         pDC->FeWork.desc.draw.startPrimID = draw * primsPerDraw;
1236
1237         pDC->cleanupState = (remainingIndices == numIndicesForDraw);
1238
1239         //enqueue DC
1240         QueueDraw(pContext);
1241
1242         pIB += maxIndicesPerDraw * indexSize;
1243         remainingIndices -= numIndicesForDraw;
1244         draw++;
1245     }
1246
1247     // restore culling state
1248     pDC = GetDrawContext(pContext);
1249     pDC->pState->state.rastState.cullMode = oldCullMode;
1250
1251     RDTSC_STOP(APIDrawIndexed, numIndices * numInstances, 0);
1252 }
1253
1254
1255 //////////////////////////////////////////////////////////////////////////
1256 /// @brief DrawIndexed
1257 /// @param hContext - Handle passed back from SwrCreateContext
1258 /// @param topology - Specifies topology for draw.
1259 /// @param numIndices - Number of indices to read sequentially from index buffer.
1260 /// @param indexOffset - Starting index into index buffer.
1261 /// @param baseVertex - Vertex in vertex buffer to consider as index "0". Note value is signed.
1262 void SwrDrawIndexed(
1263     HANDLE hContext,
1264     PRIMITIVE_TOPOLOGY topology,
1265     uint32_t numIndices,
1266     uint32_t indexOffset,
1267     int32_t baseVertex
1268     )
1269 {
1270     DrawIndexedInstance(hContext, topology, numIndices, indexOffset, baseVertex);
1271 }
1272
1273 //////////////////////////////////////////////////////////////////////////
1274 /// @brief SwrDrawIndexedInstanced
1275 /// @param hContext - Handle passed back from SwrCreateContext
1276 /// @param topology - Specifies topology for draw.
1277 /// @param numIndices - Number of indices to read sequentially from index buffer.
1278 /// @param numInstances - Number of instances to render.
1279 /// @param indexOffset - Starting index into index buffer.
1280 /// @param baseVertex - Vertex in vertex buffer to consider as index "0". Note value is signed.
1281 /// @param startInstance - Which instance to start sequentially fetching from in each buffer (instanced data)
1282 void SwrDrawIndexedInstanced(
1283     HANDLE hContext,
1284     PRIMITIVE_TOPOLOGY topology,
1285     uint32_t numIndices,
1286     uint32_t numInstances,
1287     uint32_t indexOffset,
1288     int32_t baseVertex,
1289     uint32_t startInstance)
1290 {
1291     DrawIndexedInstance(hContext, topology, numIndices, indexOffset, baseVertex, numInstances, startInstance);
1292 }
1293
1294 //////////////////////////////////////////////////////////////////////////
1295 /// @brief SwrInvalidateTiles
1296 /// @param hContext - Handle passed back from SwrCreateContext
1297 /// @param attachmentMask - The mask specifies which surfaces attached to the hottiles to invalidate.
1298 /// @param invalidateRect - The pixel-coordinate rectangle to invalidate.  This will be expanded to
1299 ///                         be hottile size-aligned.
1300 void SWR_API SwrInvalidateTiles(
1301     HANDLE hContext,
1302     uint32_t attachmentMask,
1303     const SWR_RECT& invalidateRect)
1304 {
1305     if (KNOB_TOSS_DRAW)
1306     {
1307         return;
1308     }
1309
1310     SWR_CONTEXT *pContext = GetContext(hContext);
1311     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1312
1313     pDC->FeWork.type = DISCARDINVALIDATETILES;
1314     pDC->FeWork.pfnWork = ProcessDiscardInvalidateTiles;
1315     pDC->FeWork.desc.discardInvalidateTiles.attachmentMask = attachmentMask;
1316     pDC->FeWork.desc.discardInvalidateTiles.rect = invalidateRect;
1317     pDC->FeWork.desc.discardInvalidateTiles.rect &= g_MaxScissorRect;
1318     pDC->FeWork.desc.discardInvalidateTiles.newTileState = SWR_TILE_INVALID;
1319     pDC->FeWork.desc.discardInvalidateTiles.createNewTiles = false;
1320     pDC->FeWork.desc.discardInvalidateTiles.fullTilesOnly = false;
1321
1322     //enqueue
1323     QueueDraw(pContext);
1324 }
1325
1326 //////////////////////////////////////////////////////////////////////////
1327 /// @brief SwrDiscardRect
1328 /// @param hContext - Handle passed back from SwrCreateContext
1329 /// @param attachmentMask - The mask specifies which surfaces attached to the hottiles to discard.
1330 /// @param rect - The pixel-coordinate rectangle to discard.  Only fully-covered hottiles will be
1331 ///               discarded.
1332 void SWR_API SwrDiscardRect(
1333     HANDLE hContext,
1334     uint32_t attachmentMask,
1335     const SWR_RECT& rect)
1336 {
1337     if (KNOB_TOSS_DRAW)
1338     {
1339         return;
1340     }
1341
1342     SWR_CONTEXT *pContext = GetContext(hContext);
1343     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1344
1345     // Queue a load to the hottile
1346     pDC->FeWork.type = DISCARDINVALIDATETILES;
1347     pDC->FeWork.pfnWork = ProcessDiscardInvalidateTiles;
1348     pDC->FeWork.desc.discardInvalidateTiles.attachmentMask = attachmentMask;
1349     pDC->FeWork.desc.discardInvalidateTiles.rect = rect;
1350     pDC->FeWork.desc.discardInvalidateTiles.rect &= g_MaxScissorRect;
1351     pDC->FeWork.desc.discardInvalidateTiles.newTileState = SWR_TILE_RESOLVED;
1352     pDC->FeWork.desc.discardInvalidateTiles.createNewTiles = true;
1353     pDC->FeWork.desc.discardInvalidateTiles.fullTilesOnly = true;
1354
1355     //enqueue
1356     QueueDraw(pContext);
1357 }
1358
1359 //////////////////////////////////////////////////////////////////////////
1360 /// @brief SwrDispatch
1361 /// @param hContext - Handle passed back from SwrCreateContext
1362 /// @param threadGroupCountX - Number of thread groups dispatched in X direction
1363 /// @param threadGroupCountY - Number of thread groups dispatched in Y direction
1364 /// @param threadGroupCountZ - Number of thread groups dispatched in Z direction
1365 void SwrDispatch(
1366     HANDLE hContext,
1367     uint32_t threadGroupCountX,
1368     uint32_t threadGroupCountY,
1369     uint32_t threadGroupCountZ)
1370 {
1371     if (KNOB_TOSS_DRAW)
1372     {
1373         return;
1374     }
1375
1376     RDTSC_START(APIDispatch);
1377     SWR_CONTEXT *pContext = GetContext(hContext);
1378     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1379
1380     pDC->isCompute = true;      // This is a compute context.
1381
1382     COMPUTE_DESC* pTaskData = (COMPUTE_DESC*)pDC->pArena->AllocAligned(sizeof(COMPUTE_DESC), 64);
1383
1384     pTaskData->threadGroupCountX = threadGroupCountX;
1385     pTaskData->threadGroupCountY = threadGroupCountY;
1386     pTaskData->threadGroupCountZ = threadGroupCountZ;
1387
1388     uint32_t totalThreadGroups = threadGroupCountX * threadGroupCountY * threadGroupCountZ;
1389     uint32_t dcIndex = pDC->drawId % KNOB_MAX_DRAWS_IN_FLIGHT;
1390     pDC->pDispatch = &pContext->pDispatchQueueArray[dcIndex];
1391     pDC->pDispatch->initialize(totalThreadGroups, pTaskData);
1392
1393     QueueDispatch(pContext);
1394     RDTSC_STOP(APIDispatch, threadGroupCountX * threadGroupCountY * threadGroupCountZ, 0);
1395 }
1396
1397 // Deswizzles, converts and stores current contents of the hot tiles to surface
1398 // described by pState
1399 void SWR_API SwrStoreTiles(
1400     HANDLE hContext,
1401     SWR_RENDERTARGET_ATTACHMENT attachment,
1402     SWR_TILE_STATE postStoreTileState,
1403     const SWR_RECT& storeRect)
1404 {
1405     if (KNOB_TOSS_DRAW)
1406     {
1407         return;
1408     }
1409
1410     RDTSC_START(APIStoreTiles);
1411
1412     SWR_CONTEXT *pContext = GetContext(hContext);
1413     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1414
1415     pDC->FeWork.type = STORETILES;
1416     pDC->FeWork.pfnWork = ProcessStoreTiles;
1417     pDC->FeWork.desc.storeTiles.attachment = attachment;
1418     pDC->FeWork.desc.storeTiles.postStoreTileState = postStoreTileState;
1419     pDC->FeWork.desc.storeTiles.rect = storeRect;
1420     pDC->FeWork.desc.storeTiles.rect &= g_MaxScissorRect;
1421
1422     //enqueue
1423     QueueDraw(pContext);
1424
1425     RDTSC_STOP(APIStoreTiles, 0, 0);
1426 }
1427
1428 //////////////////////////////////////////////////////////////////////////
1429 /// @brief SwrClearRenderTarget - Clear attached render targets / depth / stencil
1430 /// @param hContext - Handle passed back from SwrCreateContext
1431 /// @param clearMask - combination of SWR_CLEAR_COLOR / SWR_CLEAR_DEPTH / SWR_CLEAR_STENCIL flags (or SWR_CLEAR_NONE)
1432 /// @param clearColor - color use for clearing render targets
1433 /// @param z - depth value use for clearing depth buffer
1434 /// @param stencil - stencil value used for clearing stencil buffer
1435 /// @param clearRect - The pixel-coordinate rectangle to clear in all cleared buffers
1436 void SWR_API SwrClearRenderTarget(
1437     HANDLE hContext,
1438     uint32_t clearMask,
1439     const float clearColor[4],
1440     float z,
1441     uint8_t stencil,
1442     const SWR_RECT& clearRect)
1443 {
1444     if (KNOB_TOSS_DRAW)
1445     {
1446         return;
1447     }
1448
1449     RDTSC_START(APIClearRenderTarget);
1450
1451     SWR_CONTEXT *pContext = GetContext(hContext);
1452     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1453
1454     CLEAR_FLAGS flags;
1455     flags.bits = 0;
1456     flags.mask = clearMask;
1457
1458     pDC->FeWork.type = CLEAR;
1459     pDC->FeWork.pfnWork = ProcessClear;
1460     pDC->FeWork.desc.clear.rect = clearRect;
1461     pDC->FeWork.desc.clear.rect &= g_MaxScissorRect;
1462     pDC->FeWork.desc.clear.flags = flags;
1463     pDC->FeWork.desc.clear.clearDepth = z;
1464     pDC->FeWork.desc.clear.clearRTColor[0] = clearColor[0];
1465     pDC->FeWork.desc.clear.clearRTColor[1] = clearColor[1];
1466     pDC->FeWork.desc.clear.clearRTColor[2] = clearColor[2];
1467     pDC->FeWork.desc.clear.clearRTColor[3] = clearColor[3];
1468     pDC->FeWork.desc.clear.clearStencil = stencil;
1469
1470     // enqueue draw
1471     QueueDraw(pContext);
1472
1473     RDTSC_STOP(APIClearRenderTarget, 0, pDC->drawId);
1474 }
1475
1476 //////////////////////////////////////////////////////////////////////////
1477 /// @brief Returns a pointer to the private context state for the current
1478 ///        draw operation. This is used for external componets such as the
1479 ///        sampler.
1480 ///        SWR is responsible for the allocation of the private context state.
1481 /// @param hContext - Handle passed back from SwrCreateContext
1482 VOID* SwrGetPrivateContextState(
1483     HANDLE hContext)
1484 {
1485     SWR_CONTEXT* pContext = GetContext(hContext);
1486     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1487     DRAW_STATE* pState = pDC->pState;
1488
1489     if (pState->pPrivateState == nullptr)
1490     {
1491         pState->pPrivateState = pState->pArena->AllocAligned(pContext->privateStateSize, KNOB_SIMD_WIDTH*sizeof(float));
1492     }
1493
1494     return pState->pPrivateState;
1495 }
1496
1497 //////////////////////////////////////////////////////////////////////////
1498 /// @brief Clients can use this to allocate memory for draw/dispatch
1499 ///        operations. The memory will automatically be freed once operation
1500 ///        has completed. Client can use this to allocate binding tables,
1501 ///        etc. needed for shader execution.
1502 /// @param hContext - Handle passed back from SwrCreateContext
1503 /// @param size - Size of allocation
1504 /// @param align - Alignment needed for allocation.
1505 VOID* SwrAllocDrawContextMemory(
1506     HANDLE hContext,
1507     uint32_t size,
1508     uint32_t align)
1509 {
1510     SWR_CONTEXT* pContext = GetContext(hContext);
1511     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1512
1513     return pDC->pState->pArena->AllocAligned(size, align);
1514 }
1515
1516 //////////////////////////////////////////////////////////////////////////
1517 /// @brief Enables stats counting
1518 /// @param hContext - Handle passed back from SwrCreateContext
1519 /// @param enable - If true then counts are incremented.
1520 void SwrEnableStats(
1521     HANDLE hContext,
1522     bool enable)
1523 {
1524     SWR_CONTEXT *pContext = GetContext(hContext);
1525     DRAW_CONTEXT* pDC = GetDrawContext(pContext);
1526
1527     pDC->pState->state.enableStats = enable;
1528 }
1529
1530 //////////////////////////////////////////////////////////////////////////
1531 /// @brief Mark end of frame - used for performance profiling
1532 /// @param hContext - Handle passed back from SwrCreateContext
1533 void SWR_API SwrEndFrame(
1534     HANDLE hContext)
1535 {
1536     RDTSC_ENDFRAME();
1537     SWR_CONTEXT *pContext = GetContext(hContext);
1538     pContext->frameCount++;
1539 }