swr/rast: enable per-warp scratch space for CS

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>
This commit is contained in:
Tim Rowley 2017-04-19 17:03:32 -05:00
parent 0424e6249a
commit dabd0499a6
8 changed files with 33 additions and 8 deletions

View File

@ -592,12 +592,16 @@ void SwrSetCsFunc(
HANDLE hContext,
PFN_CS_FUNC pfnCsFunc,
uint32_t totalThreadsInGroup,
uint32_t totalSpillFillSize)
uint32_t totalSpillFillSize,
uint32_t scratchSpaceSizePerInstance,
uint32_t numInstances)
{
API_STATE* pState = GetDrawState(GetContext(hContext));
pState->pfnCsFunc = pfnCsFunc;
pState->totalThreadsInGroup = totalThreadsInGroup;
pState->totalSpillFillSize = totalSpillFillSize;
pState->scratchSpaceSize = scratchSpaceSizePerInstance;
pState->scratchSpaceNumInstances = numInstances;
}
void SwrSetTsState(

View File

@ -366,11 +366,16 @@ void SWR_API SwrSetGsFunc(
/// @param pfnCsFunc - Pointer to compute shader function
/// @param totalThreadsInGroup - product of thread group dimensions.
/// @param totalSpillFillSize - size in bytes needed for spill/fill.
/// @param scratchSpaceSizePerInstance - size of the scratch space needed per simd instance
/// @param numInstances - number of simd instances that are run per execution of the shader
void SWR_API SwrSetCsFunc(
HANDLE hContext,
PFN_CS_FUNC pfnCsFunc,
uint32_t totalThreadsInGroup,
uint32_t totalSpillFillSize);
uint32_t totalSpillFillSize,
uint32_t scratchSpaceSizePerInstance,
uint32_t numInstances
);
//////////////////////////////////////////////////////////////////////////
/// @brief Set tessellation state.

View File

@ -45,7 +45,7 @@ static PFN_CLEAR_TILES sClearTilesTable[NUM_SWR_FORMATS];
/// @param pDC - pointer to draw context (dispatch).
/// @param workerId - The unique worker ID that is assigned to this thread.
/// @param threadGroupId - the linear index for the thread group within the dispatch.
void ProcessComputeBE(DRAW_CONTEXT* pDC, uint32_t workerId, uint32_t threadGroupId, void*& pSpillFillBuffer)
void ProcessComputeBE(DRAW_CONTEXT* pDC, uint32_t workerId, uint32_t threadGroupId, void*& pSpillFillBuffer, void*& pScratchSpace)
{
SWR_CONTEXT *pContext = pDC->pContext;
@ -60,6 +60,12 @@ void ProcessComputeBE(DRAW_CONTEXT* pDC, uint32_t workerId, uint32_t threadGroup
{
pSpillFillBuffer = pDC->pArena->AllocAlignedSync(spillFillSize, KNOB_SIMD_BYTES);
}
size_t scratchSpaceSize = pDC->pState->state.scratchSpaceSize * pDC->pState->state.scratchSpaceNumInstances;
if (scratchSpaceSize && pScratchSpace == nullptr)
{
pScratchSpace = pDC->pArena->AllocAlignedSync(scratchSpaceSize, KNOB_SIMD_BYTES);
}
const API_STATE& state = GetApiState(pDC);
@ -70,6 +76,8 @@ void ProcessComputeBE(DRAW_CONTEXT* pDC, uint32_t workerId, uint32_t threadGroup
csContext.dispatchDims[2] = pTaskData->threadGroupCountZ;
csContext.pTGSM = pContext->ppScratch[workerId];
csContext.pSpillFillBuffer = (uint8_t*)pSpillFillBuffer;
csContext.pScratchSpace = (uint8_t*)pScratchSpace;
csContext.scratchSpacePerSimd = pDC->pState->state.scratchSpaceSize;
state.pfnCsFunc(GetPrivateState(pDC), &csContext);

View File

@ -34,7 +34,7 @@
#include "depthstencil.h"
#include "rdtsc_core.h"
void ProcessComputeBE(DRAW_CONTEXT* pDC, uint32_t workerId, uint32_t threadGroupId, void*& pSpillFillBuffer);
void ProcessComputeBE(DRAW_CONTEXT* pDC, uint32_t workerId, uint32_t threadGroupId, void*& pSpillFillBuffer, void*& pScratchSpace);
void ProcessSyncBE(DRAW_CONTEXT *pDC, uint32_t workerId, uint32_t macroTile, void *pUserData);
void ProcessClearBE(DRAW_CONTEXT *pDC, uint32_t workerId, uint32_t macroTile, void *pUserData);
void ProcessStoreTilesBE(DRAW_CONTEXT *pDC, uint32_t workerId, uint32_t macroTile, void *pData);

View File

@ -245,6 +245,8 @@ OSALIGNLINE(struct) API_STATE
PFN_CS_FUNC pfnCsFunc;
uint32_t totalThreadsInGroup;
uint32_t totalSpillFillSize;
uint32_t scratchSpaceSize;
uint32_t scratchSpaceNumInstances;
// FE - Frontend State
SWR_FRONTEND_STATE frontendState;

View File

@ -378,6 +378,11 @@ struct SWR_CS_CONTEXT
uint8_t* pTGSM; // Thread Group Shared Memory pointer.
uint8_t* pSpillFillBuffer; // Spill/fill buffer for barrier support
uint8_t* pScratchSpace; // Pointer to scratch space buffer used by the shader, shader is responsible
// for subdividing scratch space per instance/simd
uint32_t scratchSpacePerSimd; // Scratch space per work item x SIMD_WIDTH
};
// enums

View File

@ -726,10 +726,11 @@ void WorkOnCompute(
if (queue.getNumQueued() > 0)
{
void* pSpillFillBuffer = nullptr;
void* pScratchSpace = nullptr;
uint32_t threadGroupId = 0;
while (queue.getWork(threadGroupId))
{
queue.dispatch(pDC, workerId, threadGroupId, pSpillFillBuffer);
queue.dispatch(pDC, workerId, threadGroupId, pSpillFillBuffer, pScratchSpace);
queue.finishedWork();
}

View File

@ -151,7 +151,7 @@ private:
OSALIGNLINE(volatile LONG) mWorkItemsConsumed { 0 };
};
typedef void(*PFN_DISPATCH)(DRAW_CONTEXT* pDC, uint32_t workerId, uint32_t threadGroupId, void*& pSpillFillBuffer);
typedef void(*PFN_DISPATCH)(DRAW_CONTEXT* pDC, uint32_t workerId, uint32_t threadGroupId, void*& pSpillFillBuffer, void*& pScratchSpace);
//////////////////////////////////////////////////////////////////////////
/// DispatchQueue - work queue for dispatch
@ -231,10 +231,10 @@ public:
//////////////////////////////////////////////////////////////////////////
/// @brief Dispatches a unit of work
void dispatch(DRAW_CONTEXT* pDC, uint32_t workerId, uint32_t threadGroupId, void*& pSpillFillBuffer)
void dispatch(DRAW_CONTEXT* pDC, uint32_t workerId, uint32_t threadGroupId, void*& pSpillFillBuffer, void*& pScratchSpace)
{
SWR_ASSERT(mPfnDispatch != nullptr);
mPfnDispatch(pDC, workerId, threadGroupId, pSpillFillBuffer);
mPfnDispatch(pDC, workerId, threadGroupId, pSpillFillBuffer, pScratchSpace);
}
void* mpTaskData{ nullptr }; // The API thread will set this up and the callback task function will interpet this.