#include <tuple>
//#define FETCH_DUMP_VERTEX 1
+using namespace llvm;
+using namespace SwrJit;
bool isComponentEnabled(ComponentEnable enableMask, uint8_t component);
CONVERT_NORMALIZED,
CONVERT_USCALED,
CONVERT_SSCALED,
+ CONVERT_SFIXED,
};
//////////////////////////////////////////////////////////////////////////
Value* GenerateCompCtrlVector(const ComponentControl ctrl);
- void JitLoadVertices(const FETCH_COMPILE_STATE &fetchState, Value* fetchInfo, Value* streams, Value* vIndices, Value* pVtxOut);
- void JitGatherVertices(const FETCH_COMPILE_STATE &fetchState, Value* fetchInfo, Value* streams, Value* vIndices, Value* pVtxOut);
+ void JitLoadVertices(const FETCH_COMPILE_STATE &fetchState, Value* streams, Value* vIndices, Value* pVtxOut);
+ void JitGatherVertices(const FETCH_COMPILE_STATE &fetchState, Value* streams, Value* vIndices, Value* pVtxOut);
bool IsOddFormat(SWR_FORMAT format);
bool IsUniformFormat(SWR_FORMAT format);
void UnpackComponents(SWR_FORMAT format, Value* vInput, Value* result[4]);
- void CreateGatherOddFormats(SWR_FORMAT format, Value* pBase, Value* offsets, Value* result[4]);
+ void CreateGatherOddFormats(SWR_FORMAT format, Value* pMask, Value* pBase, Value* offsets, Value* result[4]);
void ConvertFormat(SWR_FORMAT format, Value *texels[4]);
- void StoreSGVs(const FETCH_COMPILE_STATE& fetchState, Value* pFetchInfo, Value* pVtxOut);
+ Value* mpFetchInfo;
};
Function* FetchJit::Create(const FETCH_COMPILE_STATE& fetchState)
auto argitr = fetch->getArgumentList().begin();
// Fetch shader arguments
- Value* fetchInfo = &*argitr; ++argitr;
- fetchInfo->setName("fetchInfo");
+ mpFetchInfo = &*argitr; ++argitr;
+ mpFetchInfo->setName("fetchInfo");
Value* pVtxOut = &*argitr;
pVtxOut->setName("vtxOutput");
// this is just shorthand to tell LLVM to get a pointer to the base address of simdvertex
pVtxOut = BITCAST(pVtxOut, PointerType::get(VectorType::get(mFP32Ty, mVWidth), 0));
// SWR_FETCH_CONTEXT::pStreams
- Value* streams = LOAD(fetchInfo,{0, SWR_FETCH_CONTEXT_pStreams});
+ Value* streams = LOAD(mpFetchInfo,{0, SWR_FETCH_CONTEXT_pStreams});
streams->setName("pStreams");
// SWR_FETCH_CONTEXT::pIndices
- Value* indices = LOAD(fetchInfo,{0, SWR_FETCH_CONTEXT_pIndices});
+ Value* indices = LOAD(mpFetchInfo,{0, SWR_FETCH_CONTEXT_pIndices});
indices->setName("pIndices");
// SWR_FETCH_CONTEXT::pLastIndex
- Value* pLastIndex = LOAD(fetchInfo,{0, SWR_FETCH_CONTEXT_pLastIndex});
+ Value* pLastIndex = LOAD(mpFetchInfo,{0, SWR_FETCH_CONTEXT_pLastIndex});
pLastIndex->setName("pLastIndex");
default: SWR_ASSERT(0, "Unsupported index type"); vIndices = nullptr; break;
}
+ Value* vVertexId = vIndices;
+ if (fetchState.bVertexIDOffsetEnable)
+ {
+ // Assuming one of baseVertex or startVertex is 0, so adding both should be functionally correct
+ Value* vBaseVertex = VBROADCAST(LOAD(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_BaseVertex }));
+ Value* vStartVertex = VBROADCAST(LOAD(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_StartVertex }));
+ vVertexId = ADD(vIndices, vBaseVertex);
+ vVertexId = ADD(vVertexId, vStartVertex);
+ }
+
// store out vertex IDs
- STORE(vIndices, GEP(fetchInfo, { 0, SWR_FETCH_CONTEXT_VertexID }));
+ STORE(vVertexId, GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_VertexID }));
// store out cut mask if enabled
if (fetchState.bEnableCutIndex)
{
Value* vCutIndex = VIMMED1(fetchState.cutIndex);
Value* cutMask = VMASK(ICMP_EQ(vIndices, vCutIndex));
- STORE(cutMask, GEP(fetchInfo, { 0, SWR_FETCH_CONTEXT_CutMask }));
+ STORE(cutMask, GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_CutMask }));
}
// Fetch attributes from memory and output to a simdvertex struct
// since VGATHER has a perf penalty on HSW vs BDW, allow client to choose which fetch method to use
- (fetchState.bDisableVGATHER) ? JitLoadVertices(fetchState, fetchInfo, streams, vIndices, pVtxOut)
- : JitGatherVertices(fetchState, fetchInfo, streams, vIndices, pVtxOut);
-
- // Store out SGVs if enabled
- StoreSGVs(fetchState, fetchInfo, pVtxOut);
+ (fetchState.bDisableVGATHER) ? JitLoadVertices(fetchState, streams, vIndices, pVtxOut)
+ : JitGatherVertices(fetchState, streams, vIndices, pVtxOut);
RET_VOID();
JitManager::DumpToFile(fetch, "src");
+#if defined(_DEBUG)
verifyFunction(*fetch);
+#endif
::FunctionPassManager setupPasses(JM()->mpCurrentModule);
return fetch;
}
-// store vertex ID and instance ID if enabled
-void FetchJit::StoreSGVs(const FETCH_COMPILE_STATE& fetchState, Value* pFetchInfo, Value* pVtxOut)
-{
- if (fetchState.InstanceIdEnable)
- {
- Value* pId = BITCAST(VBROADCAST(LOAD(GEP(pFetchInfo, { 0, SWR_FETCH_CONTEXT_CurInstance }))), mSimdFP32Ty);
- Value* pDest = GEP(pVtxOut, C(fetchState.InstanceIdElementOffset * 4 + fetchState.InstanceIdComponentNumber), "instanceID");
- STORE(pId, pDest);
- }
-
- if (fetchState.VertexIdEnable)
- {
- Value* pId = BITCAST(LOAD(GEP(pFetchInfo, { 0, SWR_FETCH_CONTEXT_VertexID })), mSimdFP32Ty);
- Value* pDest = GEP(pVtxOut, C(fetchState.VertexIdElementOffset * 4 + fetchState.VertexIdComponentNumber), "vertexID");
- STORE(pId, pDest);
- }
-}
-
//////////////////////////////////////////////////////////////////////////
/// @brief Loads attributes from memory using LOADs, shuffling the
/// components into SOA form.
/// @param streams - value pointer to the current vertex stream
/// @param vIndices - vector value of indices to load
/// @param pVtxOut - value pointer to output simdvertex struct
-void FetchJit::JitLoadVertices(const FETCH_COMPILE_STATE &fetchState, Value* fetchInfo, Value* streams, Value* vIndices, Value* pVtxOut)
+void FetchJit::JitLoadVertices(const FETCH_COMPILE_STATE &fetchState, Value* streams, Value* vIndices, Value* pVtxOut)
{
// Zack shuffles; a variant of the Charleston.
Constant* promoteMask = ConstantVector::get(pMask);
Constant* uwvec = UndefValue::get(VectorType::get(mFP32Ty, 4));
- Value* startVertex = LOAD(fetchInfo, {0, SWR_FETCH_CONTEXT_StartVertex});
- Value* startInstance = LOAD(fetchInfo, {0, SWR_FETCH_CONTEXT_StartInstance});
- Value* curInstance = LOAD(fetchInfo, {0, SWR_FETCH_CONTEXT_CurInstance});
- Value* vBaseVertex = VBROADCAST(LOAD(fetchInfo, {0, SWR_FETCH_CONTEXT_BaseVertex}));
+ Value* startVertex = LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_StartVertex});
+ Value* startInstance = LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_StartInstance});
+ Value* curInstance = LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_CurInstance});
+ Value* vBaseVertex = VBROADCAST(LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_BaseVertex}));
curInstance->setName("curInstance");
for(uint32_t nelt = 0; nelt < fetchState.numAttribs; ++nelt)
uint32_t numComponents = info.numComps;
uint32_t bpc = info.bpp / info.numComps; ///@todo Code below assumes all components are same size. Need to fix.
+ // load path doesn't support component packing
+ SWR_ASSERT(ied.ComponentPacking == ComponentEnable::XYZW, "Fetch load path doesn't support component packing.");
+
vectors.clear();
Value *vCurIndices;
case SWR_TYPE_SSCALED:
vec = SI_TO_FP(vec, VectorType::get(mFP32Ty, 4));
break;
+ case SWR_TYPE_SFIXED:
+ vec = FMUL(SI_TO_FP(vec, VectorType::get(mFP32Ty, 4)), VBROADCAST(C(1/65536.0f)));
+ break;
case SWR_TYPE_UNKNOWN:
case SWR_TYPE_UNUSED:
SWR_ASSERT(false, "Unsupported type %d!", info.type[0]);
bool FetchJit::IsOddFormat(SWR_FORMAT format)
{
const SWR_FORMAT_INFO& info = GetFormatInfo(format);
- if (info.bpc[0] != 8 && info.bpc[0] != 16 && info.bpc[0] != 32)
+ if (info.bpc[0] != 8 && info.bpc[0] != 16 && info.bpc[0] != 32 && info.bpc[0] != 64)
{
return true;
}
// gather for odd component size formats
// gather SIMD full pixels per lane then shift/mask to move each component to their
// own vector
-void FetchJit::CreateGatherOddFormats(SWR_FORMAT format, Value* pBase, Value* offsets, Value* result[4])
+void FetchJit::CreateGatherOddFormats(SWR_FORMAT format, Value* pMask, Value* pBase, Value* offsets, Value* result[4])
{
const SWR_FORMAT_INFO &info = GetFormatInfo(format);
result[comp] = VIMMED1((int)info.defaults[comp]);
}
+ // load the proper amount of data based on component size
+ PointerType* pLoadTy = nullptr;
+ switch (info.bpp)
+ {
+ case 8: pLoadTy = Type::getInt8PtrTy(JM()->mContext); break;
+ case 16: pLoadTy = Type::getInt16PtrTy(JM()->mContext); break;
+ case 24:
+ case 32: pLoadTy = Type::getInt32PtrTy(JM()->mContext); break;
+ default: SWR_ASSERT(0);
+ }
+
+ // allocate temporary memory for masked off lanes
+ Value* pTmp = ALLOCA(pLoadTy->getElementType());
+
// gather SIMD pixels
for (uint32_t e = 0; e < JM()->mVWidth; ++e)
{
- Value* elemOffset = VEXTRACT(offsets, C(e));
- Value* load = GEP(pBase, elemOffset);
+ Value* pElemOffset = VEXTRACT(offsets, C(e));
+ Value* pLoad = GEP(pBase, pElemOffset);
+ Value* pLaneMask = VEXTRACT(pMask, C(e));
- // load the proper amount of data based on component size
- switch (info.bpp)
- {
- case 8: load = POINTER_CAST(load, Type::getInt8PtrTy(JM()->mContext)); break;
- case 16: load = POINTER_CAST(load, Type::getInt16PtrTy(JM()->mContext)); break;
- case 32: load = POINTER_CAST(load, Type::getInt32PtrTy(JM()->mContext)); break;
- default: SWR_ASSERT(0);
- }
+ pLoad = POINTER_CAST(pLoad, pLoadTy);
+
+ // mask in tmp pointer for disabled lanes
+ pLoad = SELECT(pLaneMask, pLoad, pTmp);
// load pixel
- Value *val = LOAD(load);
+ Value *val = LOAD(pLoad);
// zero extend to 32bit integer
val = INT_CAST(val, mInt32Ty, false);
//////////////////////////////////////////////////////////////////////////
/// @brief Loads attributes from memory using AVX2 GATHER(s)
/// @param fetchState - info about attributes to be fetched from memory
-/// @param fetchInfo - first argument passed to fetch shader
/// @param streams - value pointer to the current vertex stream
/// @param vIndices - vector value of indices to gather
/// @param pVtxOut - value pointer to output simdvertex struct
-void FetchJit::JitGatherVertices(const FETCH_COMPILE_STATE &fetchState, Value* fetchInfo,
+void FetchJit::JitGatherVertices(const FETCH_COMPILE_STATE &fetchState,
Value* streams, Value* vIndices, Value* pVtxOut)
{
uint32_t currentVertexElement = 0;
uint32_t outputElt = 0;
Value* vVertexElements[4];
- Value* startVertex = LOAD(fetchInfo, {0, SWR_FETCH_CONTEXT_StartVertex});
- Value* startInstance = LOAD(fetchInfo, {0, SWR_FETCH_CONTEXT_StartInstance});
- Value* curInstance = LOAD(fetchInfo, {0, SWR_FETCH_CONTEXT_CurInstance});
- Value* vBaseVertex = VBROADCAST(LOAD(fetchInfo, {0, SWR_FETCH_CONTEXT_BaseVertex}));
+ Value* startVertex = LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_StartVertex});
+ Value* startInstance = LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_StartInstance});
+ Value* curInstance = LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_CurInstance});
+ Value* vBaseVertex = VBROADCAST(LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_BaseVertex}));
curInstance->setName("curInstance");
for(uint32_t nInputElt = 0; nInputElt < fetchState.numAttribs; ++nInputElt)
{
const INPUT_ELEMENT_DESC& ied = fetchState.layout[nInputElt];
+
+ // skip element if all components are disabled
+ if (ied.ComponentPacking == ComponentEnable::NONE)
+ {
+ continue;
+ }
+
const SWR_FORMAT_INFO &info = GetFormatInfo((SWR_FORMAT)ied.Format);
SWR_ASSERT((info.bpp != 0), "Unsupported format in JitGatherVertices.");
uint32_t bpc = info.bpp / info.numComps; ///@todo Code below assumes all components are same size. Need to fix.
// is the element is <= the partially valid size
Value* vElementInBoundsMask = ICMP_SLE(vBpp, SUB(vPartialVertexSize, vAlignmentOffsets));
+ // override cur indices with 0 if pitch is 0
+ Value* pZeroPitchMask = ICMP_EQ(vStride, VIMMED1(0));
+ vCurIndices = SELECT(pZeroPitchMask, VIMMED1(0), vCurIndices);
+
// are vertices partially OOB?
Value* vMaxVertex = VBROADCAST(maxVertex);
Value* vPartialOOBMask = ICMP_EQ(vCurIndices, vMaxVertex);
// blend in any partially OOB indices that have valid elements
vGatherMask = SELECT(vPartialOOBMask, vElementInBoundsMask, vGatherMask);
+ Value* pMask = vGatherMask;
vGatherMask = VMASK(vGatherMask);
// calculate the actual offsets into the VB
// Special gather/conversion for formats without equal component sizes
if (IsOddFormat((SWR_FORMAT)ied.Format))
{
- // Only full 4 component fetch is supported for odd formats
- SWR_ASSERT(compMask == XYZW);
Value* pResults[4];
- CreateGatherOddFormats((SWR_FORMAT)ied.Format, pStreamBase, vOffsets, pResults);
+ CreateGatherOddFormats((SWR_FORMAT)ied.Format, pMask, pStreamBase, vOffsets, pResults);
ConvertFormat((SWR_FORMAT)ied.Format, pResults);
- StoreVertexElements(pVtxOut, outputElt++, 4, pResults);
- currentVertexElement = 0;
+ for (uint32_t c = 0; c < 4; ++c)
+ {
+ if (isComponentEnabled(compMask, c))
+ {
+ vVertexElements[currentVertexElement++] = pResults[c];
+ if (currentVertexElement > 3)
+ {
+ StoreVertexElements(pVtxOut, outputElt++, 4, vVertexElements);
+ // reset to the next vVertexElement to output
+ currentVertexElement = 0;
+ }
+ }
+ }
}
else if(info.type[0] == SWR_TYPE_FLOAT)
{
}
}
break;
+ case 64:
+ {
+ for (uint32_t i = 0; i < 4; i++)
+ {
+ if (isComponentEnabled(compMask, i))
+ {
+ // if we need to gather the component
+ if (compCtrl[i] == StoreSrc)
+ {
+ Value *vMaskLo = VSHUFFLE(pMask, VUNDEF(mInt1Ty, 8), C({0, 1, 2, 3}));
+ Value *vMaskHi = VSHUFFLE(pMask, VUNDEF(mInt1Ty, 8), C({4, 5, 6, 7}));
+ vMaskLo = S_EXT(vMaskLo, VectorType::get(mInt64Ty, 4));
+ vMaskHi = S_EXT(vMaskHi, VectorType::get(mInt64Ty, 4));
+ vMaskLo = BITCAST(vMaskLo, VectorType::get(mDoubleTy, 4));
+ vMaskHi = BITCAST(vMaskHi, VectorType::get(mDoubleTy, 4));
+
+ Value *vOffsetsLo = VEXTRACTI128(vOffsets, C(0));
+ Value *vOffsetsHi = VEXTRACTI128(vOffsets, C(1));
+
+ Value *vZeroDouble = VECTOR_SPLAT(4, ConstantFP::get(IRB()->getDoubleTy(), 0.0f));
+
+ Value* pGatherLo = GATHERPD(vZeroDouble,
+ pStreamBase, vOffsetsLo, vMaskLo, C((char)1));
+ Value* pGatherHi = GATHERPD(vZeroDouble,
+ pStreamBase, vOffsetsHi, vMaskHi, C((char)1));
+
+ pGatherLo = VCVTPD2PS(pGatherLo);
+ pGatherHi = VCVTPD2PS(pGatherHi);
+
+ Value *pGather = VSHUFFLE(pGatherLo, pGatherHi, C({0, 1, 2, 3, 4, 5, 6, 7}));
+
+ vVertexElements[currentVertexElement++] = pGather;
+ }
+ else
+ {
+ vVertexElements[currentVertexElement++] = GenerateCompCtrlVector(compCtrl[i]);
+ }
+
+ if (currentVertexElement > 3)
+ {
+ StoreVertexElements(pVtxOut, outputElt++, 4, vVertexElements);
+ // reset to the next vVertexElement to output
+ currentVertexElement = 0;
+ }
+
+ }
+
+ // offset base to the next component in the vertex to gather
+ pStreamBase = GEP(pStreamBase, C((char)8));
+ }
+ }
+ break;
default:
SWR_ASSERT(0, "Tried to fetch invalid FP format");
break;
conversionType = CONVERT_SSCALED;
extendCastType = Instruction::CastOps::SIToFP;
break;
+ case SWR_TYPE_SFIXED:
+ conversionType = CONVERT_SFIXED;
+ extendCastType = Instruction::CastOps::SExt;
+ break;
default:
break;
}
break;
case 32:
{
- SWR_ASSERT(conversionType == CONVERT_NONE);
-
// Gathered components into place in simdvertex struct
for (uint32_t i = 0; i < 4; i++)
{
// save mask as it is zero'd out after each gather
Value *vMask = vGatherMask;
- vVertexElements[currentVertexElement++] = GATHERDD(gatherSrc, pStreamBase, vOffsets, vMask, C((char)1));
-
+ Value* pGather = GATHERDD(gatherSrc, pStreamBase, vOffsets, vMask, C((char)1));
+
+ if (conversionType == CONVERT_USCALED)
+ {
+ pGather = UI_TO_FP(pGather, mSimdFP32Ty);
+ }
+ else if (conversionType == CONVERT_SSCALED)
+ {
+ pGather = SI_TO_FP(pGather, mSimdFP32Ty);
+ }
+ else if (conversionType == CONVERT_SFIXED)
+ {
+ pGather = FMUL(SI_TO_FP(pGather, mSimdFP32Ty), VBROADCAST(C(1/65536.0f)));
+ }
+
+ vVertexElements[currentVertexElement++] = pGather;
// e.g. result of a single 8x32bit integer gather for 32bit components
// 256i - 0 1 2 3 4 5 6 7
// xxxx xxxx xxxx xxxx xxxx xxxx xxxx xxxx
case Store0: return VIMMED1(0);
case Store1Fp: return VIMMED1(1.0f);
case Store1Int: return VIMMED1(1);
+ case StoreVertexId:
+ {
+ Value* pId = BITCAST(LOAD(GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_VertexID })), mSimdFP32Ty);
+ return VBROADCAST(pId);
+ }
+ case StoreInstanceId:
+ {
+ Value* pId = BITCAST(LOAD(GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_CurInstance })), mFP32Ty);
+ return VBROADCAST(pId);
+ }
case StoreSrc:
default: SWR_ASSERT(0, "Invalid component control"); return VUNDEF_I();
}
fclose(fd);
#endif
+ pJitMgr->DumpAsm(const_cast<llvm::Function*>(func), "final");
+
return pfnFetch;
}