#include "common/simdintrin.h"
#include "common/os.h"
+#include "archrast/archrast.h"
+
+static const SWR_RECT g_MaxScissorRect = { 0, 0, KNOB_MAX_SCISSOR_X, KNOB_MAX_SCISSOR_Y };
+
void SetupDefaultState(SWR_CONTEXT *pContext);
static INLINE SWR_CONTEXT* GetContext(HANDLE hContext)
pContext->pMacroTileManagerArray = (MacroTileMgr*)AlignedMalloc(sizeof(MacroTileMgr) * KNOB_MAX_DRAWS_IN_FLIGHT, 64);
pContext->pDispatchQueueArray = (DispatchQueue*)AlignedMalloc(sizeof(DispatchQueue) * KNOB_MAX_DRAWS_IN_FLIGHT, 64);
+ for (uint32_t dc = 0; dc < KNOB_MAX_DRAWS_IN_FLIGHT; ++dc)
+ {
+ pContext->dcRing[dc].pArena = new CachingArena(pContext->cachingArenaAllocator);
+ new (&pContext->pMacroTileManagerArray[dc]) MacroTileMgr(*pContext->dcRing[dc].pArena);
+ new (&pContext->pDispatchQueueArray[dc]) DispatchQueue();
+
+ pContext->dsRing[dc].pArena = new CachingArena(pContext->cachingArenaAllocator);
+ }
+
pContext->threadInfo.MAX_WORKER_THREADS = KNOB_MAX_WORKER_THREADS;
pContext->threadInfo.MAX_NUMA_NODES = KNOB_MAX_NUMA_NODES;
pContext->threadInfo.MAX_CORES_PER_NUMA_NODE = KNOB_MAX_CORES_PER_NUMA_NODE;
pContext->threadInfo = *pCreateInfo->pThreadInfo;
}
- for (uint32_t dc = 0; dc < KNOB_MAX_DRAWS_IN_FLIGHT; ++dc)
- {
- pContext->dcRing[dc].pArena = new CachingArena(pContext->cachingArenaAllocator);
- new (&pContext->pMacroTileManagerArray[dc]) MacroTileMgr(*pContext->dcRing[dc].pArena);
- new (&pContext->pDispatchQueueArray[dc]) DispatchQueue();
-
- pContext->dsRing[dc].pArena = new CachingArena(pContext->cachingArenaAllocator);
- }
+ memset(&pContext->WaitLock, 0, sizeof(pContext->WaitLock));
+ memset(&pContext->FifosNotEmpty, 0, sizeof(pContext->FifosNotEmpty));
+ new (&pContext->WaitLock) std::mutex();
+ new (&pContext->FifosNotEmpty) std::condition_variable();
- if (!pContext->threadInfo.SINGLE_THREADED)
- {
- memset(&pContext->WaitLock, 0, sizeof(pContext->WaitLock));
- memset(&pContext->FifosNotEmpty, 0, sizeof(pContext->FifosNotEmpty));
- new (&pContext->WaitLock) std::mutex();
- new (&pContext->FifosNotEmpty) std::condition_variable();
+ CreateThreadPool(pContext, &pContext->threadPool);
- CreateThreadPool(pContext, &pContext->threadPool);
- }
+ pContext->ppScratch = new uint8_t*[pContext->NumWorkerThreads];
+ pContext->pStats = new SWR_STATS[pContext->NumWorkerThreads];
- // Calling createThreadPool() above can set SINGLE_THREADED
- if (pContext->threadInfo.SINGLE_THREADED)
- {
- pContext->NumWorkerThreads = 1;
- pContext->NumFEThreads = 1;
- pContext->NumBEThreads = 1;
- }
+ // Setup ArchRast thread contexts which includes +1 for API thread.
+ pContext->pArContext = new HANDLE[pContext->NumWorkerThreads+1];
+ pContext->pArContext[pContext->NumWorkerThreads] = ArchRast::CreateThreadContext();
// Allocate scratch space for workers.
///@note We could lazily allocate this but its rather small amount of memory.
#if defined(_WIN32)
uint32_t numaNode = pContext->threadPool.pThreadData ?
pContext->threadPool.pThreadData[i].numaId : 0;
- pContext->pScratch[i] = (uint8_t*)VirtualAllocExNuma(
+ pContext->ppScratch[i] = (uint8_t*)VirtualAllocExNuma(
GetCurrentProcess(), nullptr, 32 * sizeof(KILOBYTE),
MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE,
numaNode);
#else
- pContext->pScratch[i] = (uint8_t*)AlignedMalloc(32 * sizeof(KILOBYTE), KNOB_SIMD_WIDTH * 4);
+ pContext->ppScratch[i] = (uint8_t*)AlignedMalloc(32 * sizeof(KILOBYTE), KNOB_SIMD_WIDTH * 4);
#endif
+
+ // Initialize worker thread context for ArchRast.
+ pContext->pArContext[i] = ArchRast::CreateThreadContext();
}
// State setup AFTER context is fully initialized
pContext->pfnClearTile = pCreateInfo->pfnClearTile;
pContext->pfnUpdateSoWriteOffset = pCreateInfo->pfnUpdateSoWriteOffset;
pContext->pfnUpdateStats = pCreateInfo->pfnUpdateStats;
+ pContext->pfnUpdateStatsFE = pCreateInfo->pfnUpdateStatsFE;
// pass pointer to bucket manager back to caller
#ifdef KNOB_ENABLE_RDTSC
// free the fifos
for (uint32_t i = 0; i < KNOB_MAX_DRAWS_IN_FLIGHT; ++i)
{
+ delete [] pContext->dcRing[i].dynState.pStats;
delete pContext->dcRing[i].pArena;
delete pContext->dsRing[i].pArena;
pContext->pMacroTileManagerArray[i].~MacroTileMgr();
for (uint32_t i = 0; i < pContext->NumWorkerThreads; ++i)
{
#if defined(_WIN32)
- VirtualFree(pContext->pScratch[i], 0, MEM_RELEASE);
+ VirtualFree(pContext->ppScratch[i], 0, MEM_RELEASE);
#else
- AlignedFree(pContext->pScratch[i]);
+ AlignedFree(pContext->ppScratch[i]);
#endif
+
+ ArchRast::DestroyThreadContext(pContext->pArContext[i]);
}
+ delete [] pContext->ppScratch;
+ delete [] pContext->pArContext;
+ delete [] pContext->pStats;
+
delete(pContext->pHotTileMgr);
pContext->~SWR_CONTEXT();
pContext->FifosNotEmpty.notify_all();
}
-static TileSet gSingleThreadLockedTiles;
-
template<bool IsDraw>
void QueueWork(SWR_CONTEXT *pContext)
{
{
uint32_t curDraw[2] = { pContext->pCurDrawContext->drawId, pContext->pCurDrawContext->drawId };
WorkOnFifoFE(pContext, 0, curDraw[0]);
- WorkOnFifoBE(pContext, 0, curDraw[1], gSingleThreadLockedTiles, 0, 0);
+ WorkOnFifoBE(pContext, 0, curDraw[1], pContext->singleThreadLockedTiles, 0, 0);
}
else
{
uint64_t curDraw = pContext->dcRing.GetHead();
uint32_t dcIndex = curDraw % KNOB_MAX_DRAWS_IN_FLIGHT;
- static uint64_t lastDrawChecked;
- static uint32_t lastFrameChecked;
- if ((pContext->frameCount - lastFrameChecked) > 2 ||
- (curDraw - lastDrawChecked) > 0x10000)
+ if ((pContext->frameCount - pContext->lastFrameChecked) > 2 ||
+ (curDraw - pContext->lastDrawChecked) > 0x10000)
{
// Take this opportunity to clean-up old arena allocations
pContext->cachingArenaAllocator.FreeOldBlocks();
- lastFrameChecked = pContext->frameCount;
- lastDrawChecked = curDraw;
+ pContext->lastFrameChecked = pContext->frameCount;
+ pContext->lastDrawChecked = curDraw;
}
DRAW_CONTEXT* pCurDrawContext = &pContext->dcRing[dcIndex];
pCurDrawContext->threadsDone = 0;
pCurDrawContext->retireCallback.pfnCallbackFunc = nullptr;
- memset(&pCurDrawContext->dynState, 0, sizeof(pCurDrawContext->dynState));
+ pCurDrawContext->dynState.Reset(pContext->NumWorkerThreads);
// Assign unique drawId for this DC
pCurDrawContext->drawId = pContext->dcRing.GetHead();
}
// update guardband multipliers for the viewport
-void updateGuardband(API_STATE *pState)
+void updateGuardbands(API_STATE *pState)
{
- // guardband center is viewport center
- pState->gbState.left = KNOB_GUARDBAND_WIDTH / pState->vp[0].width;
- pState->gbState.right = KNOB_GUARDBAND_WIDTH / pState->vp[0].width;
- pState->gbState.top = KNOB_GUARDBAND_HEIGHT / pState->vp[0].height;
- pState->gbState.bottom = KNOB_GUARDBAND_HEIGHT / pState->vp[0].height;
+ uint32_t numGbs = pState->gsState.emitsRenderTargetArrayIndex ? KNOB_NUM_VIEWPORTS_SCISSORS : 1;
+
+ for(uint32_t i = 0; i < numGbs; ++i)
+ {
+ // guardband center is viewport center
+ pState->gbState.left[i] = KNOB_GUARDBAND_WIDTH / pState->vp[i].width;
+ pState->gbState.right[i] = KNOB_GUARDBAND_WIDTH / pState->vp[i].width;
+ pState->gbState.top[i] = KNOB_GUARDBAND_HEIGHT / pState->vp[i].height;
+ pState->gbState.bottom[i] = KNOB_GUARDBAND_HEIGHT / pState->vp[i].height;
+ }
}
void SwrSetRastState(
if (pMatrices != nullptr)
{
- //memcpy(&pState->vpMatrix[0], pMatrices, sizeof(SWR_VIEWPORT_MATRIX) * numViewports);
// @todo Faster to copy portions of the SOA or just copy all of it?
memcpy(&pState->vpMatrices, pMatrices, sizeof(SWR_VIEWPORT_MATRICES));
}
}
}
- updateGuardband(pState);
+ updateGuardbands(pState);
}
void SwrSetScissorRects(
HANDLE hContext,
uint32_t numScissors,
- const BBOX* pScissors)
+ const SWR_RECT* pScissors)
{
SWR_ASSERT(numScissors <= KNOB_NUM_VIEWPORTS_SCISSORS,
"Invalid number of scissor rects.");
API_STATE* pState = GetDrawState(GetContext(hContext));
- memcpy(&pState->scissorRects[0], pScissors, numScissors * sizeof(BBOX));
+ memcpy(&pState->scissorRects[0], pScissors, numScissors * sizeof(pScissors[0]));
};
void SetupMacroTileScissors(DRAW_CONTEXT *pDC)
{
API_STATE *pState = &pDC->pState->state;
- uint32_t left, right, top, bottom;
+ uint32_t numScissors = pState->gsState.emitsViewportArrayIndex ? KNOB_NUM_VIEWPORTS_SCISSORS : 1;
+ pState->scissorsTileAligned = true;
- // Set up scissor dimensions based on scissor or viewport
- if (pState->rastState.scissorEnable)
+ for (uint32_t index = 0; index < numScissors; ++index)
{
- // scissor rect right/bottom edge are exclusive, core expects scissor dimensions to be inclusive, so subtract one pixel from right/bottom edges
- left = pState->scissorRects[0].left;
- right = pState->scissorRects[0].right;
- top = pState->scissorRects[0].top;
- bottom = pState->scissorRects[0].bottom;
- }
- else
- {
- // the vp width and height must be added to origin un-rounded then the result round to -inf.
- // The cast to int works for rounding assuming all [left, right, top, bottom] are positive.
- left = (int32_t)pState->vp[0].x;
- right = (int32_t)(pState->vp[0].x + pState->vp[0].width);
- top = (int32_t)pState->vp[0].y;
- bottom = (int32_t)(pState->vp[0].y + pState->vp[0].height);
- }
+ SWR_RECT &scissorInFixedPoint = pState->scissorsInFixedPoint[index];
- right = std::min<uint32_t>(right, KNOB_MAX_SCISSOR_X);
- bottom = std::min<uint32_t>(bottom, KNOB_MAX_SCISSOR_Y);
+ // Set up scissor dimensions based on scissor or viewport
+ if (pState->rastState.scissorEnable)
+ {
+ scissorInFixedPoint = pState->scissorRects[index];
+ }
+ else
+ {
+ // the vp width and height must be added to origin un-rounded then the result round to -inf.
+ // The cast to int works for rounding assuming all [left, right, top, bottom] are positive.
+ scissorInFixedPoint.xmin = (int32_t)pState->vp[index].x;
+ scissorInFixedPoint.xmax = (int32_t)(pState->vp[index].x + pState->vp[index].width);
+ scissorInFixedPoint.ymin = (int32_t)pState->vp[index].y;
+ scissorInFixedPoint.ymax = (int32_t)(pState->vp[index].y + pState->vp[index].height);
+ }
- if (left > KNOB_MAX_SCISSOR_X || top > KNOB_MAX_SCISSOR_Y)
- {
- pState->scissorInFixedPoint.left = 0;
- pState->scissorInFixedPoint.right = 0;
- pState->scissorInFixedPoint.top = 0;
- pState->scissorInFixedPoint.bottom = 0;
- }
- else
- {
- pState->scissorInFixedPoint.left = left * FIXED_POINT_SCALE;
- pState->scissorInFixedPoint.right = right * FIXED_POINT_SCALE - 1;
- pState->scissorInFixedPoint.top = top * FIXED_POINT_SCALE;
- pState->scissorInFixedPoint.bottom = bottom * FIXED_POINT_SCALE - 1;
+ // Clamp to max rect
+ scissorInFixedPoint &= g_MaxScissorRect;
+
+ // Test for tile alignment
+ bool tileAligned;
+ tileAligned = (scissorInFixedPoint.xmin % KNOB_TILE_X_DIM) == 0;
+ tileAligned &= (scissorInFixedPoint.ymin % KNOB_TILE_Y_DIM) == 0;
+ tileAligned &= (scissorInFixedPoint.xmax % KNOB_TILE_X_DIM) == 0;
+ tileAligned &= (scissorInFixedPoint.xmax % KNOB_TILE_Y_DIM) == 0;
+
+ pState->scissorsTileAligned &= tileAligned;
+
+ // Scale to fixed point
+ scissorInFixedPoint.xmin *= FIXED_POINT_SCALE;
+ scissorInFixedPoint.xmax *= FIXED_POINT_SCALE;
+ scissorInFixedPoint.ymin *= FIXED_POINT_SCALE;
+ scissorInFixedPoint.ymax *= FIXED_POINT_SCALE;
+
+ // Make scissor inclusive
+ scissorInFixedPoint.xmax -= 1;
+ scissorInFixedPoint.ymax -= 1;
}
}
const bool bMultisampleEnable = ((rastState.sampleCount > SWR_MULTISAMPLE_1X) || rastState.forcedSampleCount) ? 1 : 0;
const uint32_t centroid = ((psState.barycentricsMask & SWR_BARYCENTRIC_CENTROID_MASK) > 0) ? 1 : 0;
const uint32_t canEarlyZ = (psState.forceEarlyZ || (!psState.writesODepth && !psState.usesSourceDepth && !psState.usesUAV)) ? 1 : 0;
-
+
SWR_BARYCENTRICS_MASK barycentricsMask = (SWR_BARYCENTRICS_MASK)psState.barycentricsMask;
// select backend function
DRAW_CONTEXT* pDC = GetDrawContext(pContext);
API_STATE* pState = &pDC->pState->state;
+ AR_BEGIN(AR_API_CTX, APIDrawIndexed, pDC->drawId);
+ AR_EVENT(AR_API_CTX, DrawIndexedInstance(topology, numIndices, indexOffset, baseVertex, numInstances, startInstance));
+
uint32_t maxIndicesPerDraw = MaxVertsPerDraw(pDC, numIndices, topology);
uint32_t primsPerDraw = GetNumPrims(topology, maxIndicesPerDraw);
uint32_t remainingIndices = numIndices;
pDC = GetDrawContext(pContext);
pDC->pState->state.rastState.cullMode = oldCullMode;
+ AR_END(AR_API_CTX, APIDrawIndexed, numIndices * numInstances);
RDTSC_STOP(APIDrawIndexed, numIndices * numInstances, 0);
}
/// @brief SwrInvalidateTiles
/// @param hContext - Handle passed back from SwrCreateContext
/// @param attachmentMask - The mask specifies which surfaces attached to the hottiles to invalidate.
-void SwrInvalidateTiles(
+/// @param invalidateRect - The pixel-coordinate rectangle to invalidate. This will be expanded to
+/// be hottile size-aligned.
+void SWR_API SwrInvalidateTiles(
HANDLE hContext,
- uint32_t attachmentMask)
+ uint32_t attachmentMask,
+ const SWR_RECT& invalidateRect)
{
if (KNOB_TOSS_DRAW)
{
pDC->FeWork.type = DISCARDINVALIDATETILES;
pDC->FeWork.pfnWork = ProcessDiscardInvalidateTiles;
pDC->FeWork.desc.discardInvalidateTiles.attachmentMask = attachmentMask;
- memset(&pDC->FeWork.desc.discardInvalidateTiles.rect, 0, sizeof(SWR_RECT));
+ pDC->FeWork.desc.discardInvalidateTiles.rect = invalidateRect;
+ pDC->FeWork.desc.discardInvalidateTiles.rect &= g_MaxScissorRect;
pDC->FeWork.desc.discardInvalidateTiles.newTileState = SWR_TILE_INVALID;
pDC->FeWork.desc.discardInvalidateTiles.createNewTiles = false;
pDC->FeWork.desc.discardInvalidateTiles.fullTilesOnly = false;
/// @brief SwrDiscardRect
/// @param hContext - Handle passed back from SwrCreateContext
/// @param attachmentMask - The mask specifies which surfaces attached to the hottiles to discard.
-/// @param rect - if rect is all zeros, the entire attachment surface will be discarded
-void SwrDiscardRect(
+/// @param rect - The pixel-coordinate rectangle to discard. Only fully-covered hottiles will be
+/// discarded.
+void SWR_API SwrDiscardRect(
HANDLE hContext,
uint32_t attachmentMask,
- SWR_RECT rect)
+ const SWR_RECT& rect)
{
if (KNOB_TOSS_DRAW)
{
pDC->FeWork.pfnWork = ProcessDiscardInvalidateTiles;
pDC->FeWork.desc.discardInvalidateTiles.attachmentMask = attachmentMask;
pDC->FeWork.desc.discardInvalidateTiles.rect = rect;
+ pDC->FeWork.desc.discardInvalidateTiles.rect &= g_MaxScissorRect;
pDC->FeWork.desc.discardInvalidateTiles.newTileState = SWR_TILE_RESOLVED;
pDC->FeWork.desc.discardInvalidateTiles.createNewTiles = true;
pDC->FeWork.desc.discardInvalidateTiles.fullTilesOnly = true;
// Deswizzles, converts and stores current contents of the hot tiles to surface
// described by pState
-void SwrStoreTiles(
+void SWR_API SwrStoreTiles(
HANDLE hContext,
SWR_RENDERTARGET_ATTACHMENT attachment,
- SWR_TILE_STATE postStoreTileState)
+ SWR_TILE_STATE postStoreTileState,
+ const SWR_RECT& storeRect)
{
if (KNOB_TOSS_DRAW)
{
SWR_CONTEXT *pContext = GetContext(hContext);
DRAW_CONTEXT* pDC = GetDrawContext(pContext);
- SetupMacroTileScissors(pDC);
-
pDC->FeWork.type = STORETILES;
pDC->FeWork.pfnWork = ProcessStoreTiles;
pDC->FeWork.desc.storeTiles.attachment = attachment;
pDC->FeWork.desc.storeTiles.postStoreTileState = postStoreTileState;
+ pDC->FeWork.desc.storeTiles.rect = storeRect;
+ pDC->FeWork.desc.storeTiles.rect &= g_MaxScissorRect;
//enqueue
QueueDraw(pContext);
RDTSC_STOP(APIStoreTiles, 0, 0);
}
-void SwrClearRenderTarget(
+//////////////////////////////////////////////////////////////////////////
+/// @brief SwrClearRenderTarget - Clear attached render targets / depth / stencil
+/// @param hContext - Handle passed back from SwrCreateContext
+/// @param clearMask - combination of SWR_CLEAR_COLOR / SWR_CLEAR_DEPTH / SWR_CLEAR_STENCIL flags (or SWR_CLEAR_NONE)
+/// @param clearColor - color use for clearing render targets
+/// @param z - depth value use for clearing depth buffer
+/// @param stencil - stencil value used for clearing stencil buffer
+/// @param clearRect - The pixel-coordinate rectangle to clear in all cleared buffers
+void SWR_API SwrClearRenderTarget(
HANDLE hContext,
uint32_t clearMask,
const float clearColor[4],
float z,
- uint8_t stencil)
+ uint8_t stencil,
+ const SWR_RECT& clearRect)
{
if (KNOB_TOSS_DRAW)
{
RDTSC_START(APIClearRenderTarget);
SWR_CONTEXT *pContext = GetContext(hContext);
-
DRAW_CONTEXT* pDC = GetDrawContext(pContext);
- SetupMacroTileScissors(pDC);
-
CLEAR_FLAGS flags;
+ flags.bits = 0;
flags.mask = clearMask;
pDC->FeWork.type = CLEAR;
pDC->FeWork.pfnWork = ProcessClear;
+ pDC->FeWork.desc.clear.rect = clearRect;
+ pDC->FeWork.desc.clear.rect &= g_MaxScissorRect;
pDC->FeWork.desc.clear.flags = flags;
pDC->FeWork.desc.clear.clearDepth = z;
pDC->FeWork.desc.clear.clearRTColor[0] = clearColor[0];