• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /****************************************************************************
2 * Copyright (C) 2014-2015 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 fetch_jit.cpp
24 *
25 * @brief Implementation of the fetch jitter
26 *
27 * Notes:
28 *
29 ******************************************************************************/
30 #include "jit_pch.hpp"
31 #include "builder.h"
32 #include "jit_api.h"
33 #include "fetch_jit.h"
34 #include "gen_state_llvm.h"
35 
36 //#define FETCH_DUMP_VERTEX 1
37 using namespace llvm;
38 using namespace SwrJit;
39 
40 bool isComponentEnabled(ComponentEnable enableMask, uint8_t component);
41 
42 enum ConversionType
43 {
44     CONVERT_NONE,
45     CONVERT_NORMALIZED,
46     CONVERT_USCALED,
47     CONVERT_SSCALED,
48     CONVERT_SFIXED,
49 };
50 
51 #if USE_SIMD16_SHADERS
52 #define USE_SIMD16_GATHERS 0
53 #endif
54 
55 //////////////////////////////////////////////////////////////////////////
56 /// Interface to Jitting a fetch shader
57 //////////////////////////////////////////////////////////////////////////
58 struct FetchJit : public Builder
59 {
FetchJitFetchJit60     FetchJit(JitManager* pJitMgr) : Builder(pJitMgr){};
61 
62     Function* Create(const FETCH_COMPILE_STATE& fetchState);
63 
64     Value* GetSimdValid32bitIndices(Value* vIndices, Value* pLastIndex);
65     Value* GetSimdValid16bitIndices(Value* vIndices, Value* pLastIndex);
66     Value* GetSimdValid8bitIndices(Value* vIndices, Value* pLastIndex);
67 
68     // package up Shuffle*bpcGatherd args into a tuple for convenience
69     typedef std::tuple<Value*&, Value*, const Instruction::CastOps, const ConversionType,
70         uint32_t&, uint32_t&, const ComponentEnable, const ComponentControl(&)[4], Value*(&)[4],
71         const uint32_t(&)[4]> Shuffle8bpcArgs;
72 
73 #if USE_SIMD16_SHADERS
74 #if USE_SIMD16_GATHERS
75     void Shuffle8bpcGatherd16(Shuffle8bpcArgs &args);
76 #else
77     void Shuffle8bpcGatherd(Shuffle8bpcArgs &args, bool useVertexID2);
78 #endif
79 #else
80     void Shuffle8bpcGatherd(Shuffle8bpcArgs &args);
81 #endif
82 
83     typedef std::tuple<Value*(&)[2], Value*, const Instruction::CastOps, const ConversionType,
84         uint32_t&, uint32_t&, const ComponentEnable, const ComponentControl(&)[4], Value*(&)[4]> Shuffle16bpcArgs;
85 
86 #if USE_SIMD16_SHADERS
87 #if USE_SIMD16_GATHERS
88     void Shuffle16bpcGather16(Shuffle16bpcArgs &args);
89 #else
90     void Shuffle16bpcGather(Shuffle16bpcArgs &args, bool useVertexID2);
91 #endif
92 #else
93     void Shuffle16bpcGather(Shuffle16bpcArgs &args);
94 #endif
95 
96 #if USE_SIMD16_GATHERS
97     void StoreVertexElements16(Value* pVtxOut, const uint32_t outputElt, const uint32_t numEltsToStore, Value* (&vVertexElements)[4]);
98 #else
99     void StoreVertexElements(Value* pVtxOut, const uint32_t outputElt, const uint32_t numEltsToStore, Value* (&vVertexElements)[4]);
100 #endif
101 
102 #if USE_SIMD16_SHADERS
103 #if USE_SIMD16_GATHERS
104     Value *GenerateCompCtrlVector16(const ComponentControl ctrl);
105 #else
106     Value *GenerateCompCtrlVector(const ComponentControl ctrl, bool useVertexID2);
107 #endif
108 #else
109     Value *GenerateCompCtrlVector(const ComponentControl ctrl);
110 #endif
111 
112     void JitLoadVertices(const FETCH_COMPILE_STATE &fetchState, Value* streams, Value* vIndices, Value* pVtxOut);
113 
114 #if USE_SIMD16_SHADERS
115 #if USE_SIMD16_GATHERS
116     void JitGatherVertices(const FETCH_COMPILE_STATE &fetchState, Value *streams, Value *vIndices, Value *vIndices2, Value *pVtxOut, bool useVertexID2);
117 #else
118     void JitGatherVertices(const FETCH_COMPILE_STATE &fetchState, Value* streams, Value* vIndices, Value* pVtxOut, bool useVertexID2);
119 #endif
120 #else
121     void JitGatherVertices(const FETCH_COMPILE_STATE &fetchState, Value* streams, Value* vIndices, Value* pVtxOut);
122 #endif
123 
124     bool IsOddFormat(SWR_FORMAT format);
125     bool IsUniformFormat(SWR_FORMAT format);
126     void UnpackComponents(SWR_FORMAT format, Value* vInput, Value* result[4]);
127     void CreateGatherOddFormats(SWR_FORMAT format, Value* pMask, Value* pBase, Value* offsets, Value* result[4]);
128     void ConvertFormat(SWR_FORMAT format, Value *texels[4]);
129 
130     Value* mpPrivateContext;
131     Value* mpFetchInfo;
132 };
133 
Create(const FETCH_COMPILE_STATE & fetchState)134 Function* FetchJit::Create(const FETCH_COMPILE_STATE& fetchState)
135 {
136     std::stringstream fnName("FCH_", std::ios_base::in | std::ios_base::out | std::ios_base::ate);
137     fnName << ComputeCRC(0, &fetchState, sizeof(fetchState));
138 
139     Function*    fetch = Function::Create(JM()->mFetchShaderTy, GlobalValue::ExternalLinkage, fnName.str(), JM()->mpCurrentModule);
140     BasicBlock*    entry = BasicBlock::Create(JM()->mContext, "entry", fetch);
141 
142     fetch->getParent()->setModuleIdentifier(fetch->getName());
143 
144     IRB()->SetInsertPoint(entry);
145 
146     auto    argitr = fetch->arg_begin();
147 
148     // Fetch shader arguments
149     mpPrivateContext = &*argitr; ++argitr;
150     mpPrivateContext->setName("privateContext");
151 
152     mpFetchInfo = &*argitr; ++argitr;
153     mpFetchInfo->setName("fetchInfo");
154     Value*    pVtxOut = &*argitr;
155     pVtxOut->setName("vtxOutput");
156     // this is just shorthand to tell LLVM to get a pointer to the base address of simdvertex
157     // index 0(just the pointer to the simdvertex structure
158     // index 1(which element of the simdvertex structure to offset to(in this case 0)
159     // so the indices being i32's doesn't matter
160     // TODO: generated this GEP with a VECTOR structure type so this makes sense
161     std::vector<Value*>    vtxInputIndices(2, C(0));
162     // GEP
163     pVtxOut = GEP(pVtxOut, C(0));
164 #if USE_SIMD16_SHADERS
165 #if 0// USE_SIMD16_BUILDER
166     pVtxOut = BITCAST(pVtxOut, PointerType::get(VectorType::get(mFP32Ty, mVWidth16), 0));
167 #else
168     pVtxOut = BITCAST(pVtxOut, PointerType::get(VectorType::get(mFP32Ty, mVWidth), 0));
169 #endif
170 #else
171     pVtxOut = BITCAST(pVtxOut, PointerType::get(VectorType::get(mFP32Ty, mVWidth), 0));
172 #endif
173 
174     // SWR_FETCH_CONTEXT::pStreams
175     Value*    streams = LOAD(mpFetchInfo,{0, SWR_FETCH_CONTEXT_pStreams});
176     streams->setName("pStreams");
177 
178     // SWR_FETCH_CONTEXT::pIndices
179     Value*    indices = LOAD(mpFetchInfo,{0, SWR_FETCH_CONTEXT_pIndices});
180     indices->setName("pIndices");
181 
182     // SWR_FETCH_CONTEXT::pLastIndex
183     Value*    pLastIndex = LOAD(mpFetchInfo,{0, SWR_FETCH_CONTEXT_pLastIndex});
184     pLastIndex->setName("pLastIndex");
185 
186 
187     Value* vIndices;
188 #if USE_SIMD16_SHADERS
189     Value* indices2;
190     Value* vIndices2;
191 #endif
192     switch(fetchState.indexType)
193     {
194         case R8_UINT:
195             indices = BITCAST(indices, Type::getInt8PtrTy(JM()->mContext, 0));
196 #if USE_SIMD16_SHADERS
197             indices2 = GEP(indices, C(8));
198 #endif
199             if(fetchState.bDisableIndexOOBCheck)
200             {
201                 vIndices = LOAD(BITCAST(indices, PointerType::get(VectorType::get(mInt8Ty, mpJitMgr->mVWidth), 0)), {(uint32_t)0});
202                 vIndices = Z_EXT(vIndices, mSimdInt32Ty);
203 #if USE_SIMD16_SHADERS
204                 vIndices2 = LOAD(BITCAST(indices2, PointerType::get(VectorType::get(mInt8Ty, mpJitMgr->mVWidth), 0)), { (uint32_t)0 });
205                 vIndices2 = Z_EXT(vIndices2, mSimdInt32Ty);
206 #endif
207             }
208             else
209             {
210                 pLastIndex = BITCAST(pLastIndex, Type::getInt8PtrTy(JM()->mContext, 0));
211                 vIndices = GetSimdValid8bitIndices(indices, pLastIndex);
212 #if USE_SIMD16_SHADERS
213                 pLastIndex = BITCAST(pLastIndex, Type::getInt8PtrTy(JM()->mContext, 0));
214                 vIndices2 = GetSimdValid8bitIndices(indices2, pLastIndex);
215 #endif
216             }
217             break;
218         case R16_UINT:
219             indices = BITCAST(indices, Type::getInt16PtrTy(JM()->mContext, 0));
220 #if USE_SIMD16_SHADERS
221             indices2 = GEP(indices, C(8));
222 #endif
223             if(fetchState.bDisableIndexOOBCheck)
224             {
225                 vIndices = LOAD(BITCAST(indices, PointerType::get(VectorType::get(mInt16Ty, mpJitMgr->mVWidth), 0)), {(uint32_t)0});
226                 vIndices = Z_EXT(vIndices, mSimdInt32Ty);
227 #if USE_SIMD16_SHADERS
228                 vIndices2 = LOAD(BITCAST(indices2, PointerType::get(VectorType::get(mInt16Ty, mpJitMgr->mVWidth), 0)), { (uint32_t)0 });
229                 vIndices2 = Z_EXT(vIndices2, mSimdInt32Ty);
230 #endif
231             }
232             else
233             {
234                 pLastIndex = BITCAST(pLastIndex, Type::getInt16PtrTy(JM()->mContext, 0));
235                 vIndices = GetSimdValid16bitIndices(indices, pLastIndex);
236 #if USE_SIMD16_SHADERS
237                 pLastIndex = BITCAST(pLastIndex, Type::getInt16PtrTy(JM()->mContext, 0));
238                 vIndices2 = GetSimdValid16bitIndices(indices2, pLastIndex);
239 #endif
240             }
241             break;
242         case R32_UINT:
243 #if USE_SIMD16_SHADERS
244             indices2 = GEP(indices, C(8));
245 #endif
246             (fetchState.bDisableIndexOOBCheck) ? vIndices = LOAD(BITCAST(indices, PointerType::get(mSimdInt32Ty,0)),{(uint32_t)0})
247                                                : vIndices = GetSimdValid32bitIndices(indices, pLastIndex);
248 #if USE_SIMD16_SHADERS
249             (fetchState.bDisableIndexOOBCheck) ? vIndices2 = LOAD(BITCAST(indices2, PointerType::get(mSimdInt32Ty, 0)), { (uint32_t)0 })
250                                                : vIndices2 = GetSimdValid32bitIndices(indices2, pLastIndex);
251 #endif
252             break; // incoming type is already 32bit int
253         default: SWR_INVALID("Unsupported index type"); vIndices = nullptr; break;
254     }
255 
256     if(fetchState.bForceSequentialAccessEnable)
257     {
258         Value* pOffsets = C({ 0, 1, 2, 3, 4, 5, 6, 7 });
259 
260         // VertexData buffers are accessed sequentially, the index is equal to the vertex number
261         vIndices = VBROADCAST(LOAD(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_StartVertex }));
262         vIndices = ADD(vIndices, pOffsets);
263 #if USE_SIMD16_SHADERS
264         vIndices2 = ADD(vIndices, VIMMED1(8));
265 #endif
266     }
267 
268     Value* vVertexId = vIndices;
269 #if USE_SIMD16_SHADERS
270     Value* vVertexId2 = vIndices2;
271 #endif
272     if (fetchState.bVertexIDOffsetEnable)
273     {
274         // Assuming one of baseVertex or startVertex is 0, so adding both should be functionally correct
275         Value* vBaseVertex = VBROADCAST(LOAD(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_BaseVertex }));
276         Value* vStartVertex = VBROADCAST(LOAD(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_StartVertex }));
277         vVertexId = ADD(vIndices, vBaseVertex);
278         vVertexId = ADD(vVertexId, vStartVertex);
279 #if USE_SIMD16_SHADERS
280         vVertexId2 = ADD(vIndices2, vBaseVertex);
281         vVertexId2 = ADD(vVertexId2, vStartVertex);
282 #endif
283     }
284 
285     // store out vertex IDs
286     STORE(vVertexId, GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_VertexID }));
287 #if USE_SIMD16_SHADERS
288     STORE(vVertexId2, GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_VertexID2 }));
289 #endif
290 
291     // store out cut mask if enabled
292     if (fetchState.bEnableCutIndex)
293     {
294         Value* vCutIndex = VIMMED1(fetchState.cutIndex);
295         Value* cutMask = VMASK(ICMP_EQ(vIndices, vCutIndex));
296         STORE(cutMask, GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_CutMask }));
297 #if USE_SIMD16_SHADERS
298         Value* cutMask2 = VMASK(ICMP_EQ(vIndices2, vCutIndex));
299         STORE(cutMask2, GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_CutMask2 }));
300 #endif
301     }
302 
303     // Fetch attributes from memory and output to a simdvertex struct
304     // since VGATHER has a perf penalty on HSW vs BDW, allow client to choose which fetch method to use
305 #if USE_SIMD16_SHADERS
306     if (fetchState.bDisableVGATHER)
307     {
308         JitLoadVertices(fetchState, streams, vIndices, pVtxOut);
309         JitLoadVertices(fetchState, streams, vIndices2, GEP(pVtxOut, C(1)));
310     }
311     else
312     {
313 #if USE_SIMD16_GATHERS
314         JitGatherVertices(fetchState, streams, vIndices, vIndices2, pVtxOut, false);
315 #else
316         JitGatherVertices(fetchState, streams, vIndices, pVtxOut, false);
317         JitGatherVertices(fetchState, streams, vIndices2, GEP(pVtxOut, C(1)), true);
318 #endif
319     }
320 #else
321     (fetchState.bDisableVGATHER) ? JitLoadVertices(fetchState, streams, vIndices, pVtxOut)
322                                  : JitGatherVertices(fetchState, streams, vIndices, pVtxOut);
323 #endif
324 
325     RET_VOID();
326 
327     JitManager::DumpToFile(fetch, "src");
328 
329 #if defined(_DEBUG)
330     verifyFunction(*fetch);
331 #endif
332 
333     ::FunctionPassManager setupPasses(JM()->mpCurrentModule);
334 
335     ///@todo We don't need the CFG passes for fetch. (e.g. BreakCriticalEdges and CFGSimplification)
336     setupPasses.add(createBreakCriticalEdgesPass());
337     setupPasses.add(createCFGSimplificationPass());
338     setupPasses.add(createEarlyCSEPass());
339     setupPasses.add(createPromoteMemoryToRegisterPass());
340 
341     setupPasses.run(*fetch);
342 
343     JitManager::DumpToFile(fetch, "se");
344 
345     ::FunctionPassManager optPasses(JM()->mpCurrentModule);
346 
347     ///@todo Haven't touched these either. Need to remove some of these and add others.
348     optPasses.add(createCFGSimplificationPass());
349     optPasses.add(createEarlyCSEPass());
350     optPasses.add(createInstructionCombiningPass());
351     optPasses.add(createInstructionSimplifierPass());
352     optPasses.add(createConstantPropagationPass());
353     optPasses.add(createSCCPPass());
354     optPasses.add(createAggressiveDCEPass());
355 
356     optPasses.run(*fetch);
357     optPasses.run(*fetch);
358 
359     JitManager::DumpToFile(fetch, "opt");
360 
361 
362     return fetch;
363 }
364 
365 //////////////////////////////////////////////////////////////////////////
366 /// @brief Loads attributes from memory using LOADs, shuffling the
367 /// components into SOA form.
368 /// *Note* currently does not support component control,
369 /// component packing, instancing
370 /// @param fetchState - info about attributes to be fetched from memory
371 /// @param streams - value pointer to the current vertex stream
372 /// @param vIndices - vector value of indices to load
373 /// @param pVtxOut - value pointer to output simdvertex struct
JitLoadVertices(const FETCH_COMPILE_STATE & fetchState,Value * streams,Value * vIndices,Value * pVtxOut)374 void FetchJit::JitLoadVertices(const FETCH_COMPILE_STATE &fetchState, Value* streams, Value* vIndices, Value* pVtxOut)
375 {
376     // Zack shuffles; a variant of the Charleston.
377 
378     std::vector<Value*> vectors(16);
379     std::vector<Constant*>    pMask(mVWidth);
380     for(uint32_t i = 0; i < mVWidth; ++i)
381     {
382         pMask[i] = (C(i < 4 ? i : 4));
383     }
384     Constant* promoteMask = ConstantVector::get(pMask);
385     Constant* uwvec = UndefValue::get(VectorType::get(mFP32Ty, 4));
386 
387     Value* startVertex = LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_StartVertex});
388     Value* startInstance = LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_StartInstance});
389     Value* curInstance = LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_CurInstance});
390     Value* vBaseVertex = VBROADCAST(LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_BaseVertex}));
391     curInstance->setName("curInstance");
392 
393     for(uint32_t nelt = 0; nelt < fetchState.numAttribs; ++nelt)
394     {
395         Value*    elements[4] = {0};
396         const INPUT_ELEMENT_DESC& ied = fetchState.layout[nelt];
397         const SWR_FORMAT_INFO &info = GetFormatInfo((SWR_FORMAT)ied.Format);
398         SWR_ASSERT((info.bpp != 0), "Unsupported format in JitLoadVertices.");
399         uint32_t    numComponents = info.numComps;
400         uint32_t bpc = info.bpp / info.numComps;  ///@todo Code below assumes all components are same size. Need to fix.
401 
402         // load path doesn't support component packing
403         SWR_ASSERT(ied.ComponentPacking == ComponentEnable::XYZW, "Fetch load path doesn't support component packing.");
404 
405         vectors.clear();
406 
407         if (fetchState.bInstanceIDOffsetEnable)
408         {
409             SWR_ASSERT((0), "TODO: Fill out more once driver sends this down");
410         }
411 
412         Value *vCurIndices;
413         Value *startOffset;
414         if(ied.InstanceEnable)
415         {
416             Value* stepRate = C(ied.InstanceAdvancementState);
417 
418             // prevent a div by 0 for 0 step rate
419             Value* isNonZeroStep = ICMP_UGT(stepRate, C(0));
420             stepRate = SELECT(isNonZeroStep, stepRate, C(1));
421 
422             // calc the current offset into instanced data buffer
423             Value* calcInstance = UDIV(curInstance, stepRate);
424 
425             // if step rate is 0, every instance gets instance 0
426             calcInstance = SELECT(isNonZeroStep, calcInstance, C(0));
427 
428             vCurIndices = VBROADCAST(calcInstance);
429 
430             startOffset = startInstance;
431         }
432         else if (ied.InstanceStrideEnable)
433         {
434             SWR_ASSERT((0), "TODO: Fill out more once driver sends this down.");
435         }
436         else
437         {
438             // offset indices by baseVertex
439             vCurIndices = ADD(vIndices, vBaseVertex);
440 
441             startOffset = startVertex;
442         }
443 
444         // load SWR_VERTEX_BUFFER_STATE::pData
445         Value *stream = LOAD(streams, {ied.StreamIndex, SWR_VERTEX_BUFFER_STATE_pData});
446 
447         // load SWR_VERTEX_BUFFER_STATE::pitch
448         Value *stride = LOAD(streams, {ied.StreamIndex, SWR_VERTEX_BUFFER_STATE_pitch});
449         stride = Z_EXT(stride, mInt64Ty);
450 
451         // load SWR_VERTEX_BUFFER_STATE::size
452         Value *size = LOAD(streams, {ied.StreamIndex, SWR_VERTEX_BUFFER_STATE_size});
453         size = Z_EXT(size, mInt64Ty);
454 
455         Value* startVertexOffset = MUL(Z_EXT(startOffset, mInt64Ty), stride);
456 
457         Value *minVertex = NULL;
458         Value *minVertexOffset = NULL;
459         if (fetchState.bPartialVertexBuffer) {
460             // fetch min index for low bounds checking
461             minVertex = GEP(streams, {C(ied.StreamIndex), C(SWR_VERTEX_BUFFER_STATE_minVertex)});
462             minVertex = LOAD(minVertex);
463             if (!fetchState.bDisableIndexOOBCheck) {
464                 minVertexOffset = MUL(Z_EXT(minVertex, mInt64Ty), stride);
465             }
466         }
467 
468         // Load from the stream.
469         for(uint32_t lane = 0; lane < mVWidth; ++lane)
470         {
471             // Get index
472             Value* index = VEXTRACT(vCurIndices, C(lane));
473 
474             if (fetchState.bPartialVertexBuffer) {
475                 // clamp below minvertex
476                 Value *isBelowMin = ICMP_SLT(index, minVertex);
477                 index = SELECT(isBelowMin, minVertex, index);
478             }
479 
480             index = Z_EXT(index, mInt64Ty);
481 
482             Value*    offset = MUL(index, stride);
483             offset = ADD(offset, C((int64_t)ied.AlignedByteOffset));
484             offset = ADD(offset, startVertexOffset);
485 
486             if (!fetchState.bDisableIndexOOBCheck) {
487                 // check for out of bound access, including partial OOB, and replace them with minVertex
488                 Value *endOffset = ADD(offset, C((int64_t)info.Bpp));
489                 Value *oob = ICMP_ULE(endOffset, size);
490                 if (fetchState.bPartialVertexBuffer) {
491                     offset = SELECT(oob, offset, minVertexOffset);
492                 } else {
493                     offset = SELECT(oob, offset, ConstantInt::get(mInt64Ty, 0));
494                 }
495             }
496 
497             Value*    pointer = GEP(stream, offset);
498             // We use a full-lane, but don't actually care.
499             Value*    vptr = 0;
500 
501             // get a pointer to a 4 component attrib in default address space
502             switch(bpc)
503             {
504                 case 8: vptr = BITCAST(pointer, PointerType::get(VectorType::get(mInt8Ty, 4), 0)); break;
505                 case 16: vptr = BITCAST(pointer, PointerType::get(VectorType::get(mInt16Ty, 4), 0)); break;
506                 case 32: vptr = BITCAST(pointer, PointerType::get(VectorType::get(mFP32Ty, 4), 0)); break;
507                 default: SWR_INVALID("Unsupported underlying bpp!");
508             }
509 
510             // load 4 components of attribute
511             Value*    vec = ALIGNED_LOAD(vptr, 1, false);
512 
513             // Convert To FP32 internally
514             switch(info.type[0])
515             {
516                 case SWR_TYPE_UNORM:
517                     switch(bpc)
518                     {
519                         case 8:
520                             vec = UI_TO_FP(vec, VectorType::get(mFP32Ty, 4));
521                             vec = FMUL(vec, ConstantVector::get(std::vector<Constant*>(4, ConstantFP::get(mFP32Ty, 1.0 / 255.0))));
522                             break;
523                         case 16:
524                             vec = UI_TO_FP(vec, VectorType::get(mFP32Ty, 4));
525                             vec = FMUL(vec, ConstantVector::get(std::vector<Constant*>(4, ConstantFP::get(mFP32Ty, 1.0 / 65535.0))));
526                             break;
527                         default:
528                             SWR_INVALID("Unsupported underlying type!");
529                             break;
530                     }
531                     break;
532                 case SWR_TYPE_SNORM:
533                     switch(bpc)
534                     {
535                         case 8:
536                             vec = SI_TO_FP(vec, VectorType::get(mFP32Ty, 4));
537                             vec = FMUL(vec, ConstantVector::get(std::vector<Constant*>(4, ConstantFP::get(mFP32Ty, 1.0 / 128.0))));
538                             break;
539                         case 16:
540                             vec = SI_TO_FP(vec, VectorType::get(mFP32Ty, 4));
541                             vec = FMUL(vec, ConstantVector::get(std::vector<Constant*>(4, ConstantFP::get(mFP32Ty, 1.0 / 32768.0))));
542                             break;
543                         default:
544                             SWR_INVALID("Unsupported underlying type!");
545                             break;
546                     }
547                     break;
548                 case SWR_TYPE_UINT:
549                     // Zero extend uint32_t types.
550                     switch(bpc)
551                     {
552                         case 8:
553                         case 16:
554                             vec = Z_EXT(vec, VectorType::get(mInt32Ty, 4));
555                             vec = BITCAST(vec, VectorType::get(mFP32Ty, 4));
556                             break;
557                         case 32:
558                             break; // Pass through unchanged.
559                         default:
560                             SWR_INVALID("Unsupported underlying type!");
561                             break;
562                     }
563                     break;
564                 case SWR_TYPE_SINT:
565                     // Sign extend SINT types.
566                     switch(bpc)
567                     {
568                         case 8:
569                         case 16:
570                             vec = S_EXT(vec, VectorType::get(mInt32Ty, 4));
571                             vec = BITCAST(vec, VectorType::get(mFP32Ty, 4));
572                             break;
573                         case 32:
574                             break; // Pass through unchanged.
575                         default:
576                             SWR_INVALID("Unsupported underlying type!");
577                             break;
578                     }
579                     break;
580                 case SWR_TYPE_FLOAT:
581                     switch(bpc)
582                     {
583                         case 32:
584                             break; // Pass through unchanged.
585                         default:
586                             SWR_INVALID("Unsupported underlying type!");
587                     }
588                     break;
589                 case SWR_TYPE_USCALED:
590                     vec = UI_TO_FP(vec, VectorType::get(mFP32Ty, 4));
591                     break;
592                 case SWR_TYPE_SSCALED:
593                     vec = SI_TO_FP(vec, VectorType::get(mFP32Ty, 4));
594                     break;
595                 case SWR_TYPE_SFIXED:
596                     vec = FMUL(SI_TO_FP(vec, VectorType::get(mFP32Ty, 4)), VBROADCAST(C(1/65536.0f)));
597                     break;
598                 case SWR_TYPE_UNKNOWN:
599                 case SWR_TYPE_UNUSED:
600                     SWR_INVALID("Unsupported type %d!", info.type[0]);
601             }
602 
603             // promote mask: sse(0,1,2,3) | avx(0,1,2,3,4,4,4,4)
604             // uwvec: 4 x F32, undef value
605             Value*    wvec = VSHUFFLE(vec, uwvec, promoteMask);
606             vectors.push_back(wvec);
607         }
608 
609         std::vector<Constant*>        v01Mask(mVWidth);
610         std::vector<Constant*>        v23Mask(mVWidth);
611         std::vector<Constant*>        v02Mask(mVWidth);
612         std::vector<Constant*>        v13Mask(mVWidth);
613 
614         // Concatenate the vectors together.
615         elements[0] = VUNDEF_F();
616         elements[1] = VUNDEF_F();
617         elements[2] = VUNDEF_F();
618         elements[3] = VUNDEF_F();
619         for(uint32_t b = 0, num4Wide = mVWidth / 4; b < num4Wide; ++b)
620         {
621             v01Mask[4 * b + 0] = C(0 + 4 * b);
622             v01Mask[4 * b + 1] = C(1 + 4 * b);
623             v01Mask[4 * b + 2] = C(0 + 4 * b + mVWidth);
624             v01Mask[4 * b + 3] = C(1 + 4 * b + mVWidth);
625 
626             v23Mask[4 * b + 0] = C(2 + 4 * b);
627             v23Mask[4 * b + 1] = C(3 + 4 * b);
628             v23Mask[4 * b + 2] = C(2 + 4 * b + mVWidth);
629             v23Mask[4 * b + 3] = C(3 + 4 * b + mVWidth);
630 
631             v02Mask[4 * b + 0] = C(0 + 4 * b);
632             v02Mask[4 * b + 1] = C(2 + 4 * b);
633             v02Mask[4 * b + 2] = C(0 + 4 * b + mVWidth);
634             v02Mask[4 * b + 3] = C(2 + 4 * b + mVWidth);
635 
636             v13Mask[4 * b + 0] = C(1 + 4 * b);
637             v13Mask[4 * b + 1] = C(3 + 4 * b);
638             v13Mask[4 * b + 2] = C(1 + 4 * b + mVWidth);
639             v13Mask[4 * b + 3] = C(3 + 4 * b + mVWidth);
640 
641             std::vector<Constant*>    iMask(mVWidth);
642             for(uint32_t i = 0; i < mVWidth; ++i)
643             {
644                 if(((4 * b) <= i) && (i < (4 * (b + 1))))
645                 {
646                     iMask[i] = C(i % 4 + mVWidth);
647                 }
648                 else
649                 {
650                     iMask[i] = C(i);
651                 }
652             }
653             Constant* insertMask = ConstantVector::get(iMask);
654             elements[0] = VSHUFFLE(elements[0], vectors[4 * b + 0], insertMask);
655             elements[1] = VSHUFFLE(elements[1], vectors[4 * b + 1], insertMask);
656             elements[2] = VSHUFFLE(elements[2], vectors[4 * b + 2], insertMask);
657             elements[3] = VSHUFFLE(elements[3], vectors[4 * b + 3], insertMask);
658         }
659 
660         Value* x0y0x1y1 = VSHUFFLE(elements[0], elements[1], ConstantVector::get(v01Mask));
661         Value* x2y2x3y3 = VSHUFFLE(elements[2], elements[3], ConstantVector::get(v01Mask));
662         Value* z0w0z1w1 = VSHUFFLE(elements[0], elements[1], ConstantVector::get(v23Mask));
663         Value* z2w3z2w3 = VSHUFFLE(elements[2], elements[3], ConstantVector::get(v23Mask));
664         elements[0] = VSHUFFLE(x0y0x1y1, x2y2x3y3, ConstantVector::get(v02Mask));
665         elements[1] = VSHUFFLE(x0y0x1y1, x2y2x3y3, ConstantVector::get(v13Mask));
666         elements[2] = VSHUFFLE(z0w0z1w1, z2w3z2w3, ConstantVector::get(v02Mask));
667         elements[3] = VSHUFFLE(z0w0z1w1, z2w3z2w3, ConstantVector::get(v13Mask));
668 
669         switch(numComponents + 1)
670         {
671             case    1: elements[0] = VIMMED1(0.0f);
672             case    2: elements[1] = VIMMED1(0.0f);
673             case    3: elements[2] = VIMMED1(0.0f);
674             case    4: elements[3] = VIMMED1(1.0f);
675         }
676 
677         for(uint32_t c = 0; c < 4; ++c)
678         {
679 #if USE_SIMD16_SHADERS
680             Value* dest = GEP(pVtxOut, C(nelt * 8 + c * 2), "destGEP");
681 #else
682             Value* dest = GEP(pVtxOut, C(nelt * 4 + c), "destGEP");
683 #endif
684             STORE(elements[c], dest);
685         }
686     }
687 }
688 
689 // returns true for odd formats that require special state.gather handling
IsOddFormat(SWR_FORMAT format)690 bool FetchJit::IsOddFormat(SWR_FORMAT format)
691 {
692     const SWR_FORMAT_INFO& info = GetFormatInfo(format);
693     if (info.bpc[0] != 8 && info.bpc[0] != 16 && info.bpc[0] != 32 && info.bpc[0] != 64)
694     {
695         return true;
696     }
697     return false;
698 }
699 
700 // format is uniform if all components are the same size and type
IsUniformFormat(SWR_FORMAT format)701 bool FetchJit::IsUniformFormat(SWR_FORMAT format)
702 {
703     const SWR_FORMAT_INFO& info = GetFormatInfo(format);
704     uint32_t bpc0 = info.bpc[0];
705     uint32_t type0 = info.type[0];
706 
707     for (uint32_t c = 1; c < info.numComps; ++c)
708     {
709         if (bpc0 != info.bpc[c] || type0 != info.type[c])
710         {
711             return false;
712         }
713     }
714     return true;
715 }
716 
717 // unpacks components based on format
718 // foreach component in the pixel
719 //   mask off everything but this component
720 //   shift component to LSB
UnpackComponents(SWR_FORMAT format,Value * vInput,Value * result[4])721 void FetchJit::UnpackComponents(SWR_FORMAT format, Value* vInput, Value* result[4])
722 {
723     const SWR_FORMAT_INFO& info = GetFormatInfo(format);
724 
725     uint32_t bitOffset = 0;
726     for (uint32_t c = 0; c < info.numComps; ++c)
727     {
728         uint32_t swizzledIndex = info.swizzle[c];
729         uint32_t compBits = info.bpc[c];
730         uint32_t bitmask = ((1 << compBits) - 1) << bitOffset;
731         Value* comp = AND(vInput, bitmask);
732         comp = LSHR(comp, bitOffset);
733 
734         result[swizzledIndex] = comp;
735         bitOffset += compBits;
736     }
737 }
738 
739 // gather for odd component size formats
740 // gather SIMD full pixels per lane then shift/mask to move each component to their
741 // own vector
CreateGatherOddFormats(SWR_FORMAT format,Value * pMask,Value * pBase,Value * pOffsets,Value * pResult[4])742 void FetchJit::CreateGatherOddFormats(SWR_FORMAT format, Value* pMask, Value* pBase, Value* pOffsets, Value* pResult[4])
743 {
744     const SWR_FORMAT_INFO &info = GetFormatInfo(format);
745 
746     // only works if pixel size is <= 32bits
747     SWR_ASSERT(info.bpp <= 32);
748 
749     Value *pGather;
750     if (info.bpp == 32)
751     {
752         pGather = GATHERDD(VIMMED1(0), pBase, pOffsets, pMask);
753     }
754     else
755     {
756         // Can't use 32-bit gather for items less than 32-bits, could cause page faults.
757         Value *pMem = ALLOCA(mSimdInt32Ty);
758         STORE(VIMMED1(0u), pMem);
759 
760         pBase = BITCAST(pBase, PointerType::get(mInt8Ty, 0));
761         Value* pDstMem = BITCAST(pMem, mInt32PtrTy);
762 
763         for (uint32_t lane = 0; lane < mVWidth; ++lane)
764         {
765             // Get index
766             Value* index = VEXTRACT(pOffsets, C(lane));
767             Value* mask = VEXTRACT(pMask, C(lane));
768             switch (info.bpp)
769             {
770             case 8:
771             {
772                 Value* pDst = BITCAST(GEP(pDstMem, C(lane)), PointerType::get(mInt8Ty, 0));
773                 Value* pSrc = BITCAST(GEP(pBase, index), PointerType::get(mInt8Ty, 0));
774                 STORE(LOAD(SELECT(mask, pSrc, pDst)), pDst);
775                 break;
776             }
777 
778             case 16:
779             {
780                 Value* pDst = BITCAST(GEP(pDstMem, C(lane)), PointerType::get(mInt16Ty, 0));
781                 Value* pSrc = BITCAST(GEP(pBase, index), PointerType::get(mInt16Ty, 0));
782                 STORE(LOAD(SELECT(mask, pSrc, pDst)), pDst);
783                 break;
784             }
785             break;
786 
787             case 24:
788             {
789                 // First 16-bits of data
790                 Value* pDst = BITCAST(GEP(pDstMem, C(lane)), PointerType::get(mInt16Ty, 0));
791                 Value* pSrc = BITCAST(GEP(pBase, index), PointerType::get(mInt16Ty, 0));
792                 STORE(LOAD(SELECT(mask, pSrc, pDst)), pDst);
793 
794                 // Last 8-bits of data
795                 pDst = BITCAST(GEP(pDst, C(1)), PointerType::get(mInt8Ty, 0));
796                 pSrc = BITCAST(GEP(pSrc, C(1)), PointerType::get(mInt8Ty, 0));
797                 STORE(LOAD(SELECT(mask, pSrc, pDst)), pDst);
798                 break;
799             }
800 
801             default:
802                 SWR_INVALID("Shouldn't have BPP = %d now", info.bpp);
803                 break;
804             }
805         }
806 
807         pGather = LOAD(pMem);
808     }
809 
810     for (uint32_t comp = 0; comp < 4; ++comp)
811     {
812         pResult[comp] = VIMMED1((int)info.defaults[comp]);
813     }
814 
815     UnpackComponents(format, pGather, pResult);
816 
817     // cast to fp32
818     pResult[0] = BITCAST(pResult[0], mSimdFP32Ty);
819     pResult[1] = BITCAST(pResult[1], mSimdFP32Ty);
820     pResult[2] = BITCAST(pResult[2], mSimdFP32Ty);
821     pResult[3] = BITCAST(pResult[3], mSimdFP32Ty);
822 }
823 
ConvertFormat(SWR_FORMAT format,Value * texels[4])824 void FetchJit::ConvertFormat(SWR_FORMAT format, Value *texels[4])
825 {
826     const SWR_FORMAT_INFO &info = GetFormatInfo(format);
827 
828     for (uint32_t c = 0; c < info.numComps; ++c)
829     {
830         uint32_t compIndex = info.swizzle[c];
831 
832         // skip any conversion on UNUSED components
833         if (info.type[c] == SWR_TYPE_UNUSED)
834         {
835             continue;
836         }
837 
838         if (info.isNormalized[c])
839         {
840             if (info.type[c] == SWR_TYPE_SNORM)
841             {
842                 /// @todo The most-negative value maps to -1.0f. e.g. the 5-bit value 10000 maps to -1.0f.
843 
844                 /// result = c * (1.0f / (2^(n-1) - 1);
845                 uint32_t n = info.bpc[c];
846                 uint32_t pow2 = 1 << (n - 1);
847                 float scale = 1.0f / (float)(pow2 - 1);
848                 Value *vScale = VIMMED1(scale);
849                 texels[compIndex] = BITCAST(texels[compIndex], mSimdInt32Ty);
850                 texels[compIndex] = SI_TO_FP(texels[compIndex], mSimdFP32Ty);
851                 texels[compIndex] = FMUL(texels[compIndex], vScale);
852             }
853             else
854             {
855                 SWR_ASSERT(info.type[c] == SWR_TYPE_UNORM);
856 
857                 /// result = c * (1.0f / (2^n - 1))
858                 uint32_t n = info.bpc[c];
859                 uint32_t pow2 = 1 << n;
860                 // special case 24bit unorm format, which requires a full divide to meet ULP requirement
861                 if (n == 24)
862                 {
863                     float scale = (float)(pow2 - 1);
864                     Value* vScale = VIMMED1(scale);
865                     texels[compIndex] = BITCAST(texels[compIndex], mSimdInt32Ty);
866                     texels[compIndex] = SI_TO_FP(texels[compIndex], mSimdFP32Ty);
867                     texels[compIndex] = FDIV(texels[compIndex], vScale);
868                 }
869                 else
870                 {
871                     float scale = 1.0f / (float)(pow2 - 1);
872                     Value *vScale = VIMMED1(scale);
873                     texels[compIndex] = BITCAST(texels[compIndex], mSimdInt32Ty);
874                     texels[compIndex] = UI_TO_FP(texels[compIndex], mSimdFP32Ty);
875                     texels[compIndex] = FMUL(texels[compIndex], vScale);
876                 }
877             }
878             continue;
879         }
880     }
881 }
882 
883 //////////////////////////////////////////////////////////////////////////
884 /// @brief Loads attributes from memory using AVX2 GATHER(s)
885 /// @param fetchState - info about attributes to be fetched from memory
886 /// @param streams - value pointer to the current vertex stream
887 /// @param vIndices - vector value of indices to gather
888 /// @param pVtxOut - value pointer to output simdvertex struct
889 #if USE_SIMD16_SHADERS
890 #if USE_SIMD16_GATHERS
JitGatherVertices(const FETCH_COMPILE_STATE & fetchState,Value * streams,Value * vIndices,Value * vIndices2,Value * pVtxOut,bool useVertexID2)891 void FetchJit::JitGatherVertices(const FETCH_COMPILE_STATE &fetchState,
892     Value *streams, Value *vIndices, Value *vIndices2, Value *pVtxOut, bool useVertexID2)
893 #else
894 void FetchJit::JitGatherVertices(const FETCH_COMPILE_STATE &fetchState,
895     Value* streams, Value* vIndices, Value* pVtxOut, bool useVertexID2)
896 #endif
897 #else
898 void FetchJit::JitGatherVertices(const FETCH_COMPILE_STATE &fetchState,
899     Value* streams, Value* vIndices, Value* pVtxOut)
900 #endif
901 {
902     uint32_t currentVertexElement = 0;
903     uint32_t outputElt = 0;
904     Value* vVertexElements[4];
905 #if USE_SIMD16_GATHERS
906     Value *pVtxSrc2[4];
907 #endif
908 
909     Value* startVertex = LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_StartVertex});
910     Value* startInstance = LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_StartInstance});
911     Value* curInstance = LOAD(mpFetchInfo, {0, SWR_FETCH_CONTEXT_CurInstance});
912 #if USE_SIMD16_GATHERS
913     Value* vBaseVertex16 = VBROADCAST_16(LOAD(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_BaseVertex }));
914 #else
915     Value* vBaseVertex = VBROADCAST(LOAD(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_BaseVertex }));
916 #endif
917     curInstance->setName("curInstance");
918 
919     for (uint32_t nInputElt = 0; nInputElt < fetchState.numAttribs; nInputElt += 1)
920     {
921         const INPUT_ELEMENT_DESC& ied = fetchState.layout[nInputElt];
922 
923         // skip element if all components are disabled
924         if (ied.ComponentPacking == ComponentEnable::NONE)
925         {
926             continue;
927         }
928 
929         const SWR_FORMAT_INFO &info = GetFormatInfo((SWR_FORMAT)ied.Format);
930         SWR_ASSERT((info.bpp != 0), "Unsupported format in JitGatherVertices.");
931         uint32_t bpc = info.bpp / info.numComps;  ///@todo Code below assumes all components are same size. Need to fix.
932 
933         Value *stream = LOAD(streams, {ied.StreamIndex, SWR_VERTEX_BUFFER_STATE_pData});
934 
935         // VGATHER* takes an *i8 src pointer
936         Value *pStreamBase = BITCAST(stream, PointerType::get(mInt8Ty, 0));
937 
938         Value *stride = LOAD(streams, {ied.StreamIndex, SWR_VERTEX_BUFFER_STATE_pitch});
939 #if USE_SIMD16_GATHERS
940         Value *vStride16 = VBROADCAST_16(stride);
941 #else
942         Value *vStride = VBROADCAST(stride);
943 #endif
944 
945         // max vertex index that is fully in bounds
946         Value *maxVertex = GEP(streams, {C(ied.StreamIndex), C(SWR_VERTEX_BUFFER_STATE_maxVertex)});
947         maxVertex = LOAD(maxVertex);
948 
949         Value *minVertex = NULL;
950         if (fetchState.bPartialVertexBuffer)
951         {
952             // min vertex index for low bounds OOB checking
953             minVertex = GEP(streams, {C(ied.StreamIndex), C(SWR_VERTEX_BUFFER_STATE_minVertex)});
954             minVertex = LOAD(minVertex);
955         }
956 
957         if (fetchState.bInstanceIDOffsetEnable)
958         {
959             // the InstanceID (curInstance) value is offset by StartInstanceLocation
960             curInstance = ADD(curInstance, startInstance);
961         }
962 
963 #if USE_SIMD16_GATHERS
964         Value *vCurIndices16;
965 #else
966         Value *vCurIndices;
967 #endif
968         Value *startOffset;
969 #if USE_SIMD16_GATHERS
970         Value *vInstanceStride16 = VIMMED1_16(0);
971 #else
972         Value *vInstanceStride = VIMMED1(0);
973 #endif
974 
975         if (ied.InstanceEnable)
976         {
977             Value* stepRate = C(ied.InstanceAdvancementState);
978 
979             // prevent a div by 0 for 0 step rate
980             Value* isNonZeroStep = ICMP_UGT(stepRate, C(0));
981             stepRate = SELECT(isNonZeroStep, stepRate, C(1));
982 
983             // calc the current offset into instanced data buffer
984             Value* calcInstance = UDIV(curInstance, stepRate);
985 
986             // if step rate is 0, every instance gets instance 0
987             calcInstance = SELECT(isNonZeroStep, calcInstance, C(0));
988 
989 #if USE_SIMD16_GATHERS
990             vCurIndices16 = VBROADCAST_16(calcInstance);
991 #else
992             vCurIndices = VBROADCAST(calcInstance);
993 #endif
994 
995             startOffset = startInstance;
996         }
997         else if (ied.InstanceStrideEnable)
998         {
999             // grab the instance advancement state, determines stride in bytes from one instance to the next
1000             Value* stepRate = C(ied.InstanceAdvancementState);
1001 #if USE_SIMD16_GATHERS
1002             vInstanceStride16 = VBROADCAST_16(MUL(curInstance, stepRate));
1003 #else
1004             vInstanceStride = VBROADCAST(MUL(curInstance, stepRate));
1005 #endif
1006 
1007             // offset indices by baseVertex
1008 #if USE_SIMD16_GATHERS
1009             Value *vIndices16 = JOIN_16(vIndices, vIndices2);
1010 
1011             vCurIndices16 = ADD(vIndices16, vBaseVertex16);
1012 #else
1013             vCurIndices = ADD(vIndices, vBaseVertex);
1014 #endif
1015 
1016             startOffset = startVertex;
1017             SWR_ASSERT((0), "TODO: Fill out more once driver sends this down.");
1018         }
1019         else
1020         {
1021             // offset indices by baseVertex
1022 #if USE_SIMD16_GATHERS
1023             Value *vIndices16 = JOIN_16(vIndices, vIndices2);
1024 
1025             vCurIndices16 = ADD(vIndices16, vBaseVertex16);
1026 #else
1027             vCurIndices = ADD(vIndices, vBaseVertex);
1028 #endif
1029 
1030             startOffset = startVertex;
1031         }
1032 
1033         // All of the OOB calculations are in vertices, not VB offsets, to prevent having to
1034         // do 64bit address offset calculations.
1035 
1036         // calculate byte offset to the start of the VB
1037         Value* baseOffset = MUL(Z_EXT(startOffset, mInt64Ty), Z_EXT(stride, mInt64Ty));
1038         pStreamBase = GEP(pStreamBase, baseOffset);
1039 
1040         // if we have a start offset, subtract from max vertex. Used for OOB check
1041         maxVertex = SUB(Z_EXT(maxVertex, mInt64Ty), Z_EXT(startOffset, mInt64Ty));
1042         Value* maxNeg = ICMP_SLT(maxVertex, C((int64_t)0));
1043         // if we have a negative value, we're already OOB. clamp at 0.
1044         maxVertex = SELECT(maxNeg, C(0), TRUNC(maxVertex, mInt32Ty));
1045 
1046         if (fetchState.bPartialVertexBuffer)
1047         {
1048             // similary for min vertex
1049             minVertex = SUB(Z_EXT(minVertex, mInt64Ty), Z_EXT(startOffset, mInt64Ty));
1050             Value *minNeg = ICMP_SLT(minVertex, C((int64_t)0));
1051             minVertex = SELECT(minNeg, C(0), TRUNC(minVertex, mInt32Ty));
1052         }
1053 
1054         // Load the in bounds size of a partially valid vertex
1055         Value *partialInboundsSize = GEP(streams, {C(ied.StreamIndex), C(SWR_VERTEX_BUFFER_STATE_partialInboundsSize)});
1056         partialInboundsSize = LOAD(partialInboundsSize);
1057 #if USE_SIMD16_GATHERS
1058         Value *vPartialVertexSize = VBROADCAST_16(partialInboundsSize);
1059         Value *vBpp = VBROADCAST_16(C(info.Bpp));
1060         Value *vAlignmentOffsets = VBROADCAST_16(C(ied.AlignedByteOffset));
1061 #else
1062         Value *vPartialVertexSize = VBROADCAST(partialInboundsSize);
1063         Value *vBpp = VBROADCAST(C(info.Bpp));
1064         Value *vAlignmentOffsets = VBROADCAST(C(ied.AlignedByteOffset));
1065 #endif
1066 
1067         // is the element is <= the partially valid size
1068         Value *vElementInBoundsMask = ICMP_SLE(vBpp, SUB(vPartialVertexSize, vAlignmentOffsets));
1069 
1070 #if USE_SIMD16_GATHERS
1071         // override cur indices with 0 if pitch is 0
1072         Value *pZeroPitchMask16 = ICMP_EQ(vStride16, VIMMED1_16(0));
1073         vCurIndices16 = SELECT(pZeroPitchMask16, VIMMED1_16(0), vCurIndices16);
1074 
1075         // are vertices partially OOB?
1076         Value *vMaxVertex16 = VBROADCAST_16(maxVertex);
1077         Value *vPartialOOBMask = ICMP_EQ(vCurIndices16, vMaxVertex16);
1078 
1079         // are vertices fully in bounds?
1080         Value *vMaxGatherMask16 = ICMP_ULT(vCurIndices16, vMaxVertex16);
1081 
1082         Value *vGatherMask16;
1083 
1084         if (fetchState.bPartialVertexBuffer)
1085         {
1086             // are vertices below minVertex limit?
1087             Value *vMinVertex16 = VBROADCAST_16(minVertex);
1088             Value *vMinGatherMask16 = ICMP_UGE(vCurIndices16, vMinVertex16);
1089 
1090             // only fetch lanes that pass both tests
1091             vGatherMask16 = AND(vMaxGatherMask16, vMinGatherMask16);
1092         }
1093         else
1094         {
1095             vGatherMask16 = vMaxGatherMask16;
1096         }
1097 
1098         // blend in any partially OOB indices that have valid elements
1099         vGatherMask16 = SELECT(vPartialOOBMask, vElementInBoundsMask, vGatherMask16);
1100 
1101         // calculate the actual offsets into the VB
1102         Value *vOffsets16 = MUL(vCurIndices16, vStride16);
1103         vOffsets16 = ADD(vOffsets16, vAlignmentOffsets);
1104 
1105         // if instance stride enable is:
1106         //  true  - add product of the instanceID and advancement state to the offst into the VB
1107         //  false - value of vInstanceStride has been initialialized to zero
1108         vOffsets16 = ADD(vOffsets16, vInstanceStride16);
1109 
1110         // TODO: remove the following simd8 interop stuff once all code paths are fully widened to SIMD16..
1111 
1112         Value *vGatherMask  = EXTRACT_16(vGatherMask16, 0);
1113         Value *vGatherMask2 = EXTRACT_16(vGatherMask16, 1);
1114 
1115         Value *vOffsets  = EXTRACT_16(vOffsets16, 0);
1116         Value *vOffsets2 = EXTRACT_16(vOffsets16, 1);
1117 #else
1118         // override cur indices with 0 if pitch is 0
1119         Value* pZeroPitchMask = ICMP_EQ(vStride, VIMMED1(0));
1120         vCurIndices = SELECT(pZeroPitchMask, VIMMED1(0), vCurIndices);
1121 
1122         // are vertices partially OOB?
1123         Value* vMaxVertex = VBROADCAST(maxVertex);
1124         Value* vPartialOOBMask = ICMP_EQ(vCurIndices, vMaxVertex);
1125 
1126         // are vertices fully in bounds?
1127         Value* vMaxGatherMask = ICMP_ULT(vCurIndices, vMaxVertex);
1128 
1129         Value *vGatherMask;
1130         if (fetchState.bPartialVertexBuffer)
1131         {
1132             // are vertices below minVertex limit?
1133             Value *vMinVertex = VBROADCAST(minVertex);
1134             Value *vMinGatherMask = ICMP_UGE(vCurIndices, vMinVertex);
1135 
1136             // only fetch lanes that pass both tests
1137             vGatherMask = AND(vMaxGatherMask, vMinGatherMask);
1138         }
1139         else
1140         {
1141             vGatherMask = vMaxGatherMask;
1142         }
1143 
1144         // blend in any partially OOB indices that have valid elements
1145         vGatherMask = SELECT(vPartialOOBMask, vElementInBoundsMask, vGatherMask);
1146 
1147         // calculate the actual offsets into the VB
1148         Value* vOffsets = MUL(vCurIndices, vStride);
1149         vOffsets = ADD(vOffsets, vAlignmentOffsets);
1150 
1151         // if instance stride enable is:
1152         //  true  - add product of the instanceID and advancement state to the offst into the VB
1153         //  false - value of vInstanceStride has been initialialized to zero
1154         vOffsets = ADD(vOffsets, vInstanceStride);
1155 
1156 #endif
1157         // Packing and component control
1158         ComponentEnable compMask = (ComponentEnable)ied.ComponentPacking;
1159         const ComponentControl compCtrl[4] { (ComponentControl)ied.ComponentControl0, (ComponentControl)ied.ComponentControl1,
1160                                              (ComponentControl)ied.ComponentControl2, (ComponentControl)ied.ComponentControl3};
1161 
1162         // Special gather/conversion for formats without equal component sizes
1163         if (IsOddFormat((SWR_FORMAT)ied.Format))
1164         {
1165 #if USE_SIMD16_GATHERS
1166             Value *pResults[4];
1167             Value *pResults2[4];
1168             CreateGatherOddFormats((SWR_FORMAT)ied.Format, vGatherMask,  pStreamBase, vOffsets,  pResults);
1169             CreateGatherOddFormats((SWR_FORMAT)ied.Format, vGatherMask2, pStreamBase, vOffsets2, pResults2);
1170             ConvertFormat((SWR_FORMAT)ied.Format, pResults);
1171             ConvertFormat((SWR_FORMAT)ied.Format, pResults2);
1172 
1173             for (uint32_t c = 0; c < 4; c += 1)
1174             {
1175                 if (isComponentEnabled(compMask, c))
1176                 {
1177                     // pack adjacent pairs of SIMD8s into SIMD16s
1178                     pVtxSrc2[currentVertexElement++] = JOIN_16(pResults[c], pResults2[c]);
1179 
1180                     if (currentVertexElement > 3)
1181                     {
1182                         // store SIMD16s
1183                         Value *pVtxOut2 = BITCAST(pVtxOut, PointerType::get(VectorType::get(mFP32Ty, mVWidth16), 0));
1184 
1185                         StoreVertexElements16(pVtxOut2, outputElt++, 4, pVtxSrc2);
1186                         // reset to the next vVertexElement to output
1187                         currentVertexElement = 0;
1188                     }
1189                 }
1190             }
1191 #else
1192             Value *pResults[4];
1193             CreateGatherOddFormats((SWR_FORMAT)ied.Format, vGatherMask, pStreamBase, vOffsets, pResults);
1194             ConvertFormat((SWR_FORMAT)ied.Format, pResults);
1195 
1196             for (uint32_t c = 0; c < 4; c += 1)
1197             {
1198                 if (isComponentEnabled(compMask, c))
1199                 {
1200                     vVertexElements[currentVertexElement++] = pResults[c];
1201                     if (currentVertexElement > 3)
1202                     {
1203                         StoreVertexElements(pVtxOut, outputElt++, 4, vVertexElements);
1204                         // reset to the next vVertexElement to output
1205                         currentVertexElement = 0;
1206                     }
1207                 }
1208             }
1209 #endif
1210         }
1211         else if(info.type[0] == SWR_TYPE_FLOAT)
1212         {
1213             ///@todo: support 64 bit vb accesses
1214             Value *gatherSrc = VIMMED1(0.0f);
1215 #if USE_SIMD16_GATHERS
1216             Value *gatherSrc16 = VIMMED1_16(0.0f);
1217 #endif
1218 
1219             SWR_ASSERT(IsUniformFormat((SWR_FORMAT)ied.Format),
1220                 "Unsupported format for standard gather fetch.");
1221 
1222             // Gather components from memory to store in a simdvertex structure
1223             switch (bpc)
1224             {
1225                 case 16:
1226                 {
1227 #if USE_SIMD16_GATHERS
1228                     Value *gatherResult[2];
1229 
1230                     // if we have at least one component out of x or y to fetch
1231                     if (isComponentEnabled(compMask, 0) || isComponentEnabled(compMask, 1))
1232                     {
1233                         gatherResult[0] = GATHERPS_16(gatherSrc16, pStreamBase, vOffsets16, vGatherMask16);
1234 
1235                         // e.g. result of first 8x32bit integer gather for 16bit components
1236                         // 256i - 0    1    2    3    4    5    6    7
1237                         //        xyxy xyxy xyxy xyxy xyxy xyxy xyxy xyxy
1238                         //
1239                     }
1240                     else
1241                     {
1242                         gatherResult[0] = VUNDEF_I_16();
1243                     }
1244 
1245                     // if we have at least one component out of z or w to fetch
1246                     if (isComponentEnabled(compMask, 2) || isComponentEnabled(compMask, 3))
1247                     {
1248                         // offset base to the next components(zw) in the vertex to gather
1249                         pStreamBase = GEP(pStreamBase, C((char)4));
1250 
1251                         gatherResult[1] = GATHERPS_16(gatherSrc16, pStreamBase, vOffsets16, vGatherMask16);
1252 
1253                         // e.g. result of second 8x32bit integer gather for 16bit components
1254                         // 256i - 0    1    2    3    4    5    6    7
1255                         //        zwzw zwzw zwzw zwzw zwzw zwzw zwzw zwzw
1256                         //
1257                     }
1258                     else
1259                     {
1260                         gatherResult[1] = VUNDEF_I_16();
1261                     }
1262 
1263                     // if we have at least one component to shuffle into place
1264                     if (compMask)
1265                     {
1266                         Value *pVtxOut2 = BITCAST(pVtxOut, PointerType::get(VectorType::get(mFP32Ty, mVWidth16), 0));
1267 
1268                         Shuffle16bpcArgs args = std::forward_as_tuple(gatherResult, pVtxOut2, Instruction::CastOps::FPExt, CONVERT_NONE,
1269                             currentVertexElement, outputElt, compMask, compCtrl, pVtxSrc2);
1270 
1271                         // Shuffle gathered components into place in simdvertex struct
1272                         Shuffle16bpcGather16(args);  // outputs to vVertexElements ref
1273                     }
1274 #else
1275                     Value *vGatherResult[2];
1276 
1277                     // if we have at least one component out of x or y to fetch
1278                     if (isComponentEnabled(compMask, 0) || isComponentEnabled(compMask, 1))
1279                     {
1280                         vGatherResult[0] = GATHERPS(gatherSrc, pStreamBase, vOffsets, vGatherMask);
1281                         // e.g. result of first 8x32bit integer gather for 16bit components
1282                         // 256i - 0    1    2    3    4    5    6    7
1283                         //        xyxy xyxy xyxy xyxy xyxy xyxy xyxy xyxy
1284                         //
1285                     }
1286 
1287                     // if we have at least one component out of z or w to fetch
1288                     if (isComponentEnabled(compMask, 2) || isComponentEnabled(compMask, 3))
1289                     {
1290                         // offset base to the next components(zw) in the vertex to gather
1291                         pStreamBase = GEP(pStreamBase, C((char)4));
1292 
1293                         vGatherResult[1] = GATHERPS(gatherSrc, pStreamBase, vOffsets, vGatherMask);
1294                         // e.g. result of second 8x32bit integer gather for 16bit components
1295                         // 256i - 0    1    2    3    4    5    6    7
1296                         //        zwzw zwzw zwzw zwzw zwzw zwzw zwzw zwzw
1297                         //
1298                     }
1299 
1300                     // if we have at least one component to shuffle into place
1301                     if (compMask)
1302                     {
1303                         Shuffle16bpcArgs args = std::forward_as_tuple(vGatherResult, pVtxOut, Instruction::CastOps::FPExt, CONVERT_NONE,
1304                             currentVertexElement, outputElt, compMask, compCtrl, vVertexElements);
1305 
1306                         // Shuffle gathered components into place in simdvertex struct
1307 #if USE_SIMD16_SHADERS
1308                         Shuffle16bpcGather(args, useVertexID2);  // outputs to vVertexElements ref
1309 #else
1310                         Shuffle16bpcGather(args);  // outputs to vVertexElements ref
1311 #endif
1312                     }
1313 #endif
1314                 }
1315                     break;
1316                 case 32:
1317                 {
1318                     for (uint32_t i = 0; i < 4; i += 1)
1319                     {
1320 #if USE_SIMD16_GATHERS
1321                         if (isComponentEnabled(compMask, i))
1322                         {
1323                             // if we need to gather the component
1324                             if (compCtrl[i] == StoreSrc)
1325                             {
1326                                 // Gather a SIMD of vertices
1327                                 // APIs allow a 4GB range for offsets
1328                                 // However, GATHERPS uses signed 32-bit offsets, so only a 2GB range :(
1329                                 // But, we know that elements must be aligned for FETCH. :)
1330                                 // Right shift the offset by a bit and then scale by 2 to remove the sign extension.
1331                                 Value *shiftedOffsets16 = LSHR(vOffsets16, 1);
1332                                 pVtxSrc2[currentVertexElement++] = GATHERPS_16(gatherSrc16, pStreamBase, shiftedOffsets16, vGatherMask16, 2);
1333                             }
1334                             else
1335                             {
1336                                 pVtxSrc2[currentVertexElement++] = GenerateCompCtrlVector16(compCtrl[i]);
1337                             }
1338 
1339                             if (currentVertexElement > 3)
1340                             {
1341                                 // store SIMD16s
1342                                 Value *pVtxOut2 = BITCAST(pVtxOut, PointerType::get(VectorType::get(mFP32Ty, mVWidth16), 0));
1343 
1344                                 StoreVertexElements16(pVtxOut2, outputElt++, 4, pVtxSrc2);
1345                                 // reset to the next vVertexElement to output
1346                                 currentVertexElement = 0;
1347                             }
1348                         }
1349 
1350                         // offset base to the next component in the vertex to gather
1351                         pStreamBase = GEP(pStreamBase, C((char)4));
1352 #else
1353                         if (isComponentEnabled(compMask, i))
1354                         {
1355                             // if we need to gather the component
1356                             if (compCtrl[i] == StoreSrc)
1357                             {
1358                                 // Gather a SIMD of vertices
1359                                 // APIs allow a 4GB range for offsets
1360                                 // However, GATHERPS uses signed 32-bit offsets, so only a 2GB range :(
1361                                 // But, we know that elements must be aligned for FETCH. :)
1362                                 // Right shift the offset by a bit and then scale by 2 to remove the sign extension.
1363                                 Value *vShiftedOffsets = LSHR(vOffsets, 1);
1364                                 vVertexElements[currentVertexElement++] = GATHERPS(gatherSrc, pStreamBase, vShiftedOffsets, vGatherMask, 2);
1365                             }
1366                             else
1367                             {
1368 #if USE_SIMD16_SHADERS
1369                                 vVertexElements[currentVertexElement++] = GenerateCompCtrlVector(compCtrl[i], useVertexID2);
1370 #else
1371                                 vVertexElements[currentVertexElement++] = GenerateCompCtrlVector(compCtrl[i]);
1372 #endif
1373                             }
1374 
1375                             if (currentVertexElement > 3)
1376                             {
1377                                 StoreVertexElements(pVtxOut, outputElt++, 4, vVertexElements);
1378                                 // reset to the next vVertexElement to output
1379                                 currentVertexElement = 0;
1380                             }
1381                         }
1382 
1383                         // offset base to the next component in the vertex to gather
1384                         pStreamBase = GEP(pStreamBase, C((char)4));
1385 #endif
1386                     }
1387                 }
1388                     break;
1389                 case 64:
1390                 {
1391                     for (uint32_t i = 0; i < 4; i += 1)
1392                     {
1393 #if USE_SIMD16_GATHERS
1394                         if (isComponentEnabled(compMask, i))
1395                         {
1396                             // if we need to gather the component
1397                             if (compCtrl[i] == StoreSrc)
1398                             {
1399                                 Value *vMaskLo  = VSHUFFLE(vGatherMask,  VUNDEF(mInt1Ty, 8), C({ 0, 1, 2, 3 }));
1400                                 Value *vMaskLo2 = VSHUFFLE(vGatherMask2, VUNDEF(mInt1Ty, 8), C({ 0, 1, 2, 3 }));
1401                                 Value *vMaskHi  = VSHUFFLE(vGatherMask,  VUNDEF(mInt1Ty, 8), C({ 4, 5, 6, 7 }));
1402                                 Value *vMaskHi2 = VSHUFFLE(vGatherMask2, VUNDEF(mInt1Ty, 8), C({ 4, 5, 6, 7 }));
1403 
1404                                 Value *vOffsetsLo  = VEXTRACTI128(vOffsets,  C(0));
1405                                 Value *vOffsetsLo2 = VEXTRACTI128(vOffsets2, C(0));
1406                                 Value *vOffsetsHi  = VEXTRACTI128(vOffsets,  C(1));
1407                                 Value *vOffsetsHi2 = VEXTRACTI128(vOffsets2, C(1));
1408 
1409                                 Value *vZeroDouble = VECTOR_SPLAT(4, ConstantFP::get(IRB()->getDoubleTy(), 0.0f));
1410 
1411                                 Value* pGatherLo  = GATHERPD(vZeroDouble, pStreamBase, vOffsetsLo,  vMaskLo);
1412                                 Value* pGatherLo2 = GATHERPD(vZeroDouble, pStreamBase, vOffsetsLo2, vMaskLo2);
1413                                 Value* pGatherHi  = GATHERPD(vZeroDouble, pStreamBase, vOffsetsHi,  vMaskHi);
1414                                 Value* pGatherHi2 = GATHERPD(vZeroDouble, pStreamBase, vOffsetsHi2, vMaskHi2);
1415 
1416                                 pGatherLo  = VCVTPD2PS(pGatherLo);
1417                                 pGatherLo2 = VCVTPD2PS(pGatherLo2);
1418                                 pGatherHi  = VCVTPD2PS(pGatherHi);
1419                                 pGatherHi2 = VCVTPD2PS(pGatherHi2);
1420 
1421                                 Value *pGather  = VSHUFFLE(pGatherLo,  pGatherHi,  C({ 0, 1, 2, 3, 4, 5, 6, 7 }));
1422                                 Value *pGather2 = VSHUFFLE(pGatherLo2, pGatherHi2, C({ 0, 1, 2, 3, 4, 5, 6, 7 }));
1423 
1424                                 // pack adjacent pairs of SIMD8s into SIMD16s
1425                                 pVtxSrc2[currentVertexElement++] = JOIN_16(pGather, pGather2);
1426                             }
1427                             else
1428                             {
1429                                 pVtxSrc2[currentVertexElement++] = GenerateCompCtrlVector16(compCtrl[i]);
1430                             }
1431 
1432                             if (currentVertexElement > 3)
1433                             {
1434                                 // store SIMD16s
1435                                 Value *pVtxOut2 = BITCAST(pVtxOut, PointerType::get(VectorType::get(mFP32Ty, mVWidth16), 0));
1436 
1437                                 StoreVertexElements16(pVtxOut2, outputElt++, 4, pVtxSrc2);
1438                                 // reset to the next vVertexElement to output
1439                                 currentVertexElement = 0;
1440                             }
1441                         }
1442 
1443                         // offset base to the next component  in the vertex to gather
1444                         pStreamBase = GEP(pStreamBase, C((char)8));
1445 #else
1446                         if (isComponentEnabled(compMask, i))
1447                         {
1448                             // if we need to gather the component
1449                             if (compCtrl[i] == StoreSrc)
1450                             {
1451                                 Value *vMaskLo = VSHUFFLE(vGatherMask, VUNDEF(mInt1Ty, 8), C({0, 1, 2, 3}));
1452                                 Value *vMaskHi = VSHUFFLE(vGatherMask, VUNDEF(mInt1Ty, 8), C({4, 5, 6, 7}));
1453 
1454                                 Value *vOffsetsLo = VEXTRACTI128(vOffsets, C(0));
1455                                 Value *vOffsetsHi = VEXTRACTI128(vOffsets, C(1));
1456 
1457                                 Value *vZeroDouble = VECTOR_SPLAT(4, ConstantFP::get(IRB()->getDoubleTy(), 0.0f));
1458 
1459                                 Value* pGatherLo = GATHERPD(vZeroDouble, pStreamBase, vOffsetsLo, vMaskLo);
1460                                 Value* pGatherHi = GATHERPD(vZeroDouble, pStreamBase, vOffsetsHi, vMaskHi);
1461 
1462                                 pGatherLo = VCVTPD2PS(pGatherLo);
1463                                 pGatherHi = VCVTPD2PS(pGatherHi);
1464 
1465                                 Value *pGather = VSHUFFLE(pGatherLo, pGatherHi, C({0, 1, 2, 3, 4, 5, 6, 7}));
1466 
1467                                 vVertexElements[currentVertexElement++] = pGather;
1468                             }
1469                             else
1470                             {
1471 #if USE_SIMD16_SHADERS
1472                                 vVertexElements[currentVertexElement++] = GenerateCompCtrlVector(compCtrl[i], useVertexID2);
1473 #else
1474                                 vVertexElements[currentVertexElement++] = GenerateCompCtrlVector(compCtrl[i]);
1475 #endif
1476                             }
1477 
1478                             if (currentVertexElement > 3)
1479                             {
1480                                 StoreVertexElements(pVtxOut, outputElt++, 4, vVertexElements);
1481                                 // reset to the next vVertexElement to output
1482                                 currentVertexElement = 0;
1483                             }
1484                         }
1485 
1486                         // offset base to the next component  in the vertex to gather
1487                         pStreamBase = GEP(pStreamBase, C((char)8));
1488 #endif
1489                     }
1490                 }
1491                     break;
1492                 default:
1493                     SWR_INVALID("Tried to fetch invalid FP format");
1494                     break;
1495             }
1496         }
1497         else
1498         {
1499             Instruction::CastOps extendCastType = Instruction::CastOps::CastOpsEnd;
1500             ConversionType conversionType = CONVERT_NONE;
1501 
1502             SWR_ASSERT(IsUniformFormat((SWR_FORMAT)ied.Format),
1503                 "Unsupported format for standard gather fetch.");
1504 
1505             switch(info.type[0])
1506             {
1507                 case SWR_TYPE_UNORM:
1508                     conversionType = CONVERT_NORMALIZED;
1509                 case SWR_TYPE_UINT:
1510                     extendCastType = Instruction::CastOps::ZExt;
1511                     break;
1512                 case SWR_TYPE_SNORM:
1513                     conversionType = CONVERT_NORMALIZED;
1514                 case SWR_TYPE_SINT:
1515                     extendCastType = Instruction::CastOps::SExt;
1516                     break;
1517                 case SWR_TYPE_USCALED:
1518                     conversionType = CONVERT_USCALED;
1519                     extendCastType = Instruction::CastOps::UIToFP;
1520                     break;
1521                 case SWR_TYPE_SSCALED:
1522                     conversionType = CONVERT_SSCALED;
1523                     extendCastType = Instruction::CastOps::SIToFP;
1524                     break;
1525                 case SWR_TYPE_SFIXED:
1526                     conversionType = CONVERT_SFIXED;
1527                     extendCastType = Instruction::CastOps::SExt;
1528                     break;
1529                 default:
1530                     break;
1531             }
1532 
1533             // value substituted when component of gather is masked
1534             Value* gatherSrc = VIMMED1(0);
1535 #if USE_SIMD16_GATHERS
1536             Value *gatherSrc16 = VIMMED1_16(0);
1537 #endif
1538 
1539             // Gather components from memory to store in a simdvertex structure
1540             switch (bpc)
1541             {
1542                 case 8:
1543                 {
1544                     // if we have at least one component to fetch
1545                     if (compMask)
1546                     {
1547 #if USE_SIMD16_GATHERS
1548                         Value *gatherResult = GATHERDD_16(gatherSrc16, pStreamBase, vOffsets16, vGatherMask16);
1549 
1550                         // e.g. result of an 8x32bit integer gather for 8bit components
1551                         // 256i - 0    1    2    3    4    5    6    7
1552                         //        xyzw xyzw xyzw xyzw xyzw xyzw xyzw xyzw
1553 
1554                         Value *pVtxOut2 = BITCAST(pVtxOut, PointerType::get(VectorType::get(mFP32Ty, mVWidth16), 0));
1555 
1556                         Shuffle8bpcArgs args = std::forward_as_tuple(gatherResult, pVtxOut2, extendCastType, conversionType,
1557                             currentVertexElement, outputElt, compMask, compCtrl, pVtxSrc2, info.swizzle);
1558 
1559                         // Shuffle gathered components into place in simdvertex struct
1560                         Shuffle8bpcGatherd16(args);  // outputs to vVertexElements ref
1561 #else
1562                         Value *vGatherResult = GATHERDD(gatherSrc, pStreamBase, vOffsets, vGatherMask);
1563                         // e.g. result of an 8x32bit integer gather for 8bit components
1564                         // 256i - 0    1    2    3    4    5    6    7
1565                         //        xyzw xyzw xyzw xyzw xyzw xyzw xyzw xyzw
1566 
1567                         Shuffle8bpcArgs args = std::forward_as_tuple(vGatherResult, pVtxOut, extendCastType, conversionType,
1568                             currentVertexElement, outputElt, compMask, compCtrl, vVertexElements, info.swizzle);
1569 
1570                         // Shuffle gathered components into place in simdvertex struct
1571 #if USE_SIMD16_SHADERS
1572                         Shuffle8bpcGatherd(args, useVertexID2); // outputs to vVertexElements ref
1573 #else
1574                         Shuffle8bpcGatherd(args); // outputs to vVertexElements ref
1575 #endif
1576 #endif
1577                     }
1578                 }
1579                 break;
1580                 case 16:
1581                 {
1582 #if USE_SIMD16_GATHERS
1583                     Value *gatherResult[2];
1584 
1585                     // if we have at least one component out of x or y to fetch
1586                     if (isComponentEnabled(compMask, 0) || isComponentEnabled(compMask, 1))
1587                     {
1588                         gatherResult[0] = GATHERDD_16(gatherSrc16, pStreamBase, vOffsets16, vGatherMask16);
1589 
1590                         // e.g. result of first 8x32bit integer gather for 16bit components
1591                         // 256i - 0    1    2    3    4    5    6    7
1592                         //        xyxy xyxy xyxy xyxy xyxy xyxy xyxy xyxy
1593                         //
1594                     }
1595                     else
1596                     {
1597                         gatherResult[0] = VUNDEF_I_16();
1598                     }
1599 
1600                     // if we have at least one component out of z or w to fetch
1601                     if (isComponentEnabled(compMask, 2) || isComponentEnabled(compMask, 3))
1602                     {
1603                         // offset base to the next components(zw) in the vertex to gather
1604                         pStreamBase = GEP(pStreamBase, C((char)4));
1605 
1606                         gatherResult[1] = GATHERDD_16(gatherSrc16, pStreamBase, vOffsets16, vGatherMask16);
1607 
1608                         // e.g. result of second 8x32bit integer gather for 16bit components
1609                         // 256i - 0    1    2    3    4    5    6    7
1610                         //        zwzw zwzw zwzw zwzw zwzw zwzw zwzw zwzw
1611                         //
1612                     }
1613                     else
1614                     {
1615                         gatherResult[1] = VUNDEF_I_16();
1616                     }
1617 
1618                     // if we have at least one component to shuffle into place
1619                     if (compMask)
1620                     {
1621                         Value *pVtxOut2 = BITCAST(pVtxOut, PointerType::get(VectorType::get(mFP32Ty, mVWidth16), 0));
1622 
1623                         Shuffle16bpcArgs args = std::forward_as_tuple(gatherResult, pVtxOut2, extendCastType, conversionType,
1624                             currentVertexElement, outputElt, compMask, compCtrl, pVtxSrc2);
1625 
1626                         // Shuffle gathered components into place in simdvertex struct
1627                         Shuffle16bpcGather16(args);  // outputs to vVertexElements ref
1628                     }
1629 #else
1630                     Value *vGatherResult[2];
1631 
1632                     // if we have at least one component out of x or y to fetch
1633                     if (isComponentEnabled(compMask, 0) || isComponentEnabled(compMask, 1))
1634                     {
1635                         vGatherResult[0] = GATHERDD(gatherSrc, pStreamBase, vOffsets, vGatherMask);
1636                         // e.g. result of first 8x32bit integer gather for 16bit components
1637                         // 256i - 0    1    2    3    4    5    6    7
1638                         //        xyxy xyxy xyxy xyxy xyxy xyxy xyxy xyxy
1639                         //
1640                     }
1641 
1642                     // if we have at least one component out of z or w to fetch
1643                     if (isComponentEnabled(compMask, 2) || isComponentEnabled(compMask, 3))
1644                     {
1645                         // offset base to the next components(zw) in the vertex to gather
1646                         pStreamBase = GEP(pStreamBase, C((char)4));
1647 
1648                         vGatherResult[1] = GATHERDD(gatherSrc, pStreamBase, vOffsets, vGatherMask);
1649                         // e.g. result of second 8x32bit integer gather for 16bit components
1650                         // 256i - 0    1    2    3    4    5    6    7
1651                         //        zwzw zwzw zwzw zwzw zwzw zwzw zwzw zwzw
1652                         //
1653                     }
1654 
1655                     // if we have at least one component to shuffle into place
1656                     if (compMask)
1657                     {
1658                         Shuffle16bpcArgs args = std::forward_as_tuple(vGatherResult, pVtxOut, extendCastType, conversionType,
1659                             currentVertexElement, outputElt, compMask, compCtrl, vVertexElements);
1660 
1661                         // Shuffle gathered components into place in simdvertex struct
1662 #if USE_SIMD16_SHADERS
1663                         Shuffle16bpcGather(args, useVertexID2);  // outputs to vVertexElements ref
1664 #else
1665                         Shuffle16bpcGather(args);  // outputs to vVertexElements ref
1666 #endif
1667                     }
1668 #endif
1669                 }
1670                 break;
1671                 case 32:
1672                 {
1673                     // Gathered components into place in simdvertex struct
1674                     for (uint32_t i = 0; i < 4; i++)
1675                     {
1676                         if (isComponentEnabled(compMask, i))
1677                         {
1678                             // if we need to gather the component
1679                             if (compCtrl[i] == StoreSrc)
1680                             {
1681 #if USE_SIMD16_GATHERS
1682                                 Value *pGather = GATHERDD_16(gatherSrc16, pStreamBase, vOffsets16, vGatherMask16);
1683 
1684                                 if (conversionType == CONVERT_USCALED)
1685                                 {
1686                                     pGather = UI_TO_FP(pGather, mSimd16FP32Ty);
1687                                 }
1688                                 else if (conversionType == CONVERT_SSCALED)
1689                                 {
1690                                     pGather = SI_TO_FP(pGather, mSimd16FP32Ty);
1691                                 }
1692                                 else if (conversionType == CONVERT_SFIXED)
1693                                 {
1694                                     pGather = FMUL(SI_TO_FP(pGather, mSimd16FP32Ty), VBROADCAST_16(C(1 / 65536.0f)));
1695                                 }
1696 
1697                                 pVtxSrc2[currentVertexElement++] = pGather;
1698 
1699                                 // e.g. result of a single 8x32bit integer gather for 32bit components
1700                                 // 256i - 0    1    2    3    4    5    6    7
1701                                 //        xxxx xxxx xxxx xxxx xxxx xxxx xxxx xxxx
1702 #else
1703                                 Value* pGather = GATHERDD(gatherSrc, pStreamBase, vOffsets, vGatherMask);
1704 
1705                                 if (conversionType == CONVERT_USCALED)
1706                                 {
1707                                     pGather = UI_TO_FP(pGather, mSimdFP32Ty);
1708                                 }
1709                                 else if (conversionType == CONVERT_SSCALED)
1710                                 {
1711                                     pGather = SI_TO_FP(pGather, mSimdFP32Ty);
1712                                 }
1713                                 else if (conversionType == CONVERT_SFIXED)
1714                                 {
1715                                     pGather = FMUL(SI_TO_FP(pGather, mSimdFP32Ty), VBROADCAST(C(1/65536.0f)));
1716                                 }
1717 
1718                                 vVertexElements[currentVertexElement++] = pGather;
1719 
1720                                 // e.g. result of a single 8x32bit integer gather for 32bit components
1721                                 // 256i - 0    1    2    3    4    5    6    7
1722                                 //        xxxx xxxx xxxx xxxx xxxx xxxx xxxx xxxx
1723 #endif
1724                             }
1725                             else
1726                             {
1727 #if USE_SIMD16_GATHERS
1728                                 pVtxSrc2[currentVertexElement++] = GenerateCompCtrlVector16(compCtrl[i]);
1729 #else
1730 #if USE_SIMD16_SHADERS
1731                                 vVertexElements[currentVertexElement++] = GenerateCompCtrlVector(compCtrl[i], useVertexID2);
1732 #else
1733                                 vVertexElements[currentVertexElement++] = GenerateCompCtrlVector(compCtrl[i]);
1734 #endif
1735 #endif
1736                             }
1737 
1738                             if (currentVertexElement > 3)
1739                             {
1740 #if USE_SIMD16_GATHERS
1741                                 // store SIMD16s
1742                                 Value *pVtxOut2 = BITCAST(pVtxOut, PointerType::get(VectorType::get(mFP32Ty, mVWidth16), 0));
1743 
1744                                 StoreVertexElements16(pVtxOut2, outputElt++, 4, pVtxSrc2);
1745 #else
1746                                 StoreVertexElements(pVtxOut, outputElt++, 4, vVertexElements);
1747 #endif
1748 
1749                                 // reset to the next vVertexElement to output
1750                                 currentVertexElement = 0;
1751                             }
1752 
1753                         }
1754 
1755                         // offset base to the next component  in the vertex to gather
1756                         pStreamBase = GEP(pStreamBase, C((char)4));
1757                     }
1758                 }
1759                 break;
1760             }
1761         }
1762     }
1763 
1764     // if we have a partially filled vVertexElement struct, output it
1765     if (currentVertexElement > 0)
1766     {
1767 #if USE_SIMD16_GATHERS
1768         // store SIMD16s
1769         Value *pVtxOut2 = BITCAST(pVtxOut, PointerType::get(VectorType::get(mFP32Ty, mVWidth16), 0));
1770 
1771         StoreVertexElements16(pVtxOut2, outputElt++, currentVertexElement, pVtxSrc2);
1772 #else
1773         StoreVertexElements(pVtxOut, outputElt++, currentVertexElement, vVertexElements);
1774 #endif
1775     }
1776 }
1777 
1778 //////////////////////////////////////////////////////////////////////////
1779 /// @brief Loads a simd of valid indices. OOB indices are set to 0
1780 /// *Note* have to do 16bit index checking in scalar until we have AVX-512
1781 /// support
1782 /// @param pIndices - pointer to 8 bit indices
1783 /// @param pLastIndex - pointer to last valid index
GetSimdValid8bitIndices(Value * pIndices,Value * pLastIndex)1784 Value* FetchJit::GetSimdValid8bitIndices(Value* pIndices, Value* pLastIndex)
1785 {
1786     // can fit 2 16 bit integers per vWidth lane
1787     Value* vIndices =  VUNDEF_I();
1788 
1789     // store 0 index on stack to be used to conditionally load from if index address is OOB
1790     Value* pZeroIndex = ALLOCA(mInt8Ty);
1791     STORE(C((uint8_t)0), pZeroIndex);
1792 
1793     // Load a SIMD of index pointers
1794     for(int64_t lane = 0; lane < mVWidth; lane++)
1795     {
1796         // Calculate the address of the requested index
1797         Value *pIndex = GEP(pIndices, C(lane));
1798 
1799         // check if the address is less than the max index,
1800         Value* mask = ICMP_ULT(pIndex, pLastIndex);
1801 
1802         // if valid, load the index. if not, load 0 from the stack
1803         Value* pValid = SELECT(mask, pIndex, pZeroIndex);
1804         Value *index = LOAD(pValid, "valid index");
1805 
1806         // zero extended index to 32 bits and insert into the correct simd lane
1807         index = Z_EXT(index, mInt32Ty);
1808         vIndices = VINSERT(vIndices, index, lane);
1809     }
1810     return vIndices;
1811 }
1812 
1813 //////////////////////////////////////////////////////////////////////////
1814 /// @brief Loads a simd of valid indices. OOB indices are set to 0
1815 /// *Note* have to do 16bit index checking in scalar until we have AVX-512
1816 /// support
1817 /// @param pIndices - pointer to 16 bit indices
1818 /// @param pLastIndex - pointer to last valid index
GetSimdValid16bitIndices(Value * pIndices,Value * pLastIndex)1819 Value* FetchJit::GetSimdValid16bitIndices(Value* pIndices, Value* pLastIndex)
1820 {
1821     // can fit 2 16 bit integers per vWidth lane
1822     Value* vIndices =  VUNDEF_I();
1823 
1824     // store 0 index on stack to be used to conditionally load from if index address is OOB
1825     Value* pZeroIndex = ALLOCA(mInt16Ty);
1826     STORE(C((uint16_t)0), pZeroIndex);
1827 
1828     // Load a SIMD of index pointers
1829     for(int64_t lane = 0; lane < mVWidth; lane++)
1830     {
1831         // Calculate the address of the requested index
1832         Value *pIndex = GEP(pIndices, C(lane));
1833 
1834         // check if the address is less than the max index,
1835         Value* mask = ICMP_ULT(pIndex, pLastIndex);
1836 
1837         // if valid, load the index. if not, load 0 from the stack
1838         Value* pValid = SELECT(mask, pIndex, pZeroIndex);
1839         Value *index = LOAD(pValid, "valid index");
1840 
1841         // zero extended index to 32 bits and insert into the correct simd lane
1842         index = Z_EXT(index, mInt32Ty);
1843         vIndices = VINSERT(vIndices, index, lane);
1844     }
1845     return vIndices;
1846 }
1847 
1848 //////////////////////////////////////////////////////////////////////////
1849 /// @brief Loads a simd of valid indices. OOB indices are set to 0
1850 /// @param pIndices - pointer to 32 bit indices
1851 /// @param pLastIndex - pointer to last valid index
GetSimdValid32bitIndices(Value * pIndices,Value * pLastIndex)1852 Value* FetchJit::GetSimdValid32bitIndices(Value* pIndices, Value* pLastIndex)
1853 {
1854     DataLayout dL(JM()->mpCurrentModule);
1855     unsigned int ptrSize = dL.getPointerSize() * 8;  // ptr size in bits
1856     Value* iLastIndex = PTR_TO_INT(pLastIndex, Type::getIntNTy(JM()->mContext, ptrSize));
1857     Value* iIndices = PTR_TO_INT(pIndices, Type::getIntNTy(JM()->mContext, ptrSize));
1858 
1859     // get the number of indices left in the buffer (endPtr - curPtr) / sizeof(index)
1860     Value* numIndicesLeft = SUB(iLastIndex,iIndices);
1861     numIndicesLeft = TRUNC(numIndicesLeft, mInt32Ty);
1862     numIndicesLeft = SDIV(numIndicesLeft, C(4));
1863 
1864     // create a vector of index counts from the base index ptr passed into the fetch
1865     const std::vector<Constant*> vecIndices {C(0), C(1), C(2), C(3), C(4), C(5), C(6), C(7)};
1866     Constant* vIndexOffsets = ConstantVector::get(vecIndices);
1867 
1868     // compare index count to the max valid index
1869     // e.g vMaxIndex      4 4 4 4 4 4 4 4 : 4 indices left to load
1870     //     vIndexOffsets  0 1 2 3 4 5 6 7
1871     //     ------------------------------
1872     //     vIndexMask    -1-1-1-1 0 0 0 0 : offsets < max pass
1873     //     vLoadedIndices 0 1 2 3 0 0 0 0 : offsets >= max masked to 0
1874     Value* vMaxIndex = VBROADCAST(numIndicesLeft);
1875     Value* vIndexMask = VPCMPGTD(vMaxIndex,vIndexOffsets);
1876 
1877     // VMASKLOAD takes an *i8 src pointer
1878     pIndices = BITCAST(pIndices,PointerType::get(mInt8Ty,0));
1879 
1880     // Load the indices; OOB loads 0
1881     return MASKLOADD(pIndices,vIndexMask);
1882 }
1883 
1884 //////////////////////////////////////////////////////////////////////////
1885 /// @brief Takes a SIMD of gathered 8bpc verts, zero or sign extends,
1886 /// denormalizes if needed, converts to F32 if needed, and positions in
1887 //  the proper SIMD rows to be output to the simdvertex structure
1888 /// @param args: (tuple of args, listed below)
1889 ///   @param vGatherResult - 8 gathered 8bpc vertices
1890 ///   @param pVtxOut - base pointer to output simdvertex struct
1891 ///   @param extendType - sign extend or zero extend
1892 ///   @param bNormalized - do we need to denormalize?
1893 ///   @param currentVertexElement - reference to the current vVertexElement
1894 ///   @param outputElt - reference to the current offset from simdvertex we're o
1895 ///   @param compMask - component packing mask
1896 ///   @param compCtrl - component control val
1897 ///   @param vVertexElements[4] - vertex components to output
1898 ///   @param swizzle[4] - component swizzle location
1899 #if USE_SIMD16_GATHERS
Shuffle8bpcGatherd16(Shuffle8bpcArgs & args)1900 void FetchJit::Shuffle8bpcGatherd16(Shuffle8bpcArgs &args)
1901 {
1902     // Unpack tuple args
1903     Value*& vGatherResult = std::get<0>(args);
1904     Value* pVtxOut = std::get<1>(args);
1905     const Instruction::CastOps extendType = std::get<2>(args);
1906     const ConversionType conversionType = std::get<3>(args);
1907     uint32_t &currentVertexElement = std::get<4>(args);
1908     uint32_t &outputElt = std::get<5>(args);
1909     const ComponentEnable compMask = std::get<6>(args);
1910     const ComponentControl(&compCtrl)[4] = std::get<7>(args);
1911     Value* (&vVertexElements)[4] = std::get<8>(args);
1912     const uint32_t(&swizzle)[4] = std::get<9>(args);
1913 
1914     // cast types
1915     Type *vGatherTy = mSimdInt32Ty;
1916     Type *v32x8Ty = VectorType::get(mInt8Ty, mVWidth * 4); // vwidth is units of 32 bits
1917 
1918     // have to do extra work for sign extending
1919     if ((extendType == Instruction::CastOps::SExt) || (extendType == Instruction::CastOps::SIToFP))
1920     {
1921         Type *v16x8Ty = VectorType::get(mInt8Ty, mVWidth * 2); // 8x16bit ints in a 128bit lane
1922         Type *v128Ty = VectorType::get(IntegerType::getIntNTy(JM()->mContext, 128), mVWidth / 4); // vwidth is units of 32 bits
1923 
1924         // shuffle mask, including any swizzling
1925         const char x = (char)swizzle[0]; const char y = (char)swizzle[1];
1926         const char z = (char)swizzle[2]; const char w = (char)swizzle[3];
1927         Value *vConstMask = C<char>({ char(x), char(x + 4), char(x + 8), char(x + 12),
1928             char(y), char(y + 4), char(y + 8), char(y + 12),
1929             char(z), char(z + 4), char(z + 8), char(z + 12),
1930             char(w), char(w + 4), char(w + 8), char(w + 12),
1931             char(x), char(x + 4), char(x + 8), char(x + 12),
1932             char(y), char(y + 4), char(y + 8), char(y + 12),
1933             char(z), char(z + 4), char(z + 8), char(z + 12),
1934             char(w), char(w + 4), char(w + 8), char(w + 12) });
1935 
1936         // SIMD16 PSHUFB isnt part of AVX-512F, so split into SIMD8 for the sake of KNL, for now..
1937 
1938         Value *vGatherResult_lo = EXTRACT_16(vGatherResult, 0);
1939         Value *vGatherResult_hi = EXTRACT_16(vGatherResult, 1);
1940 
1941         Value *vShufResult_lo = BITCAST(PSHUFB(BITCAST(vGatherResult_lo, v32x8Ty), vConstMask), vGatherTy);
1942         Value *vShufResult_hi = BITCAST(PSHUFB(BITCAST(vGatherResult_hi, v32x8Ty), vConstMask), vGatherTy);
1943 
1944         // after pshufb: group components together in each 128bit lane
1945         // 256i - 0    1    2    3    4    5    6    7
1946         //        xxxx yyyy zzzz wwww xxxx yyyy zzzz wwww
1947 
1948         Value *vi128XY_lo = nullptr;
1949         Value *vi128XY_hi = nullptr;
1950         if (isComponentEnabled(compMask, 0) || isComponentEnabled(compMask, 1))
1951         {
1952             vi128XY_lo = BITCAST(PERMD(vShufResult_lo, C<int32_t>({ 0, 4, 0, 0, 1, 5, 0, 0 })), v128Ty);
1953             vi128XY_hi = BITCAST(PERMD(vShufResult_hi, C<int32_t>({ 0, 4, 0, 0, 1, 5, 0, 0 })), v128Ty);
1954 
1955             // after PERMD: move and pack xy and zw components in low 64 bits of each 128bit lane
1956             // 256i - 0    1    2    3    4    5    6    7
1957             //        xxxx xxxx dcdc dcdc yyyy yyyy dcdc dcdc (dc - don't care)
1958         }
1959 
1960         // do the same for zw components
1961         Value *vi128ZW_lo = nullptr;
1962         Value *vi128ZW_hi = nullptr;
1963         if (isComponentEnabled(compMask, 2) || isComponentEnabled(compMask, 3))
1964         {
1965             vi128ZW_lo = BITCAST(PERMD(vShufResult_lo, C<int32_t>({ 2, 6, 0, 0, 3, 7, 0, 0 })), v128Ty);
1966             vi128ZW_hi = BITCAST(PERMD(vShufResult_hi, C<int32_t>({ 2, 6, 0, 0, 3, 7, 0, 0 })), v128Ty);
1967         }
1968 
1969         // init denormalize variables if needed
1970         Instruction::CastOps fpCast;
1971         Value *conversionFactor;
1972 
1973         switch (conversionType)
1974         {
1975         case CONVERT_NORMALIZED:
1976             fpCast = Instruction::CastOps::SIToFP;
1977             conversionFactor = VIMMED1((float)(1.0 / 127.0));
1978             break;
1979         case CONVERT_SSCALED:
1980             fpCast = Instruction::CastOps::SIToFP;
1981             conversionFactor = VIMMED1((float)(1.0));
1982             break;
1983         case CONVERT_USCALED:
1984             SWR_INVALID("Type should not be sign extended!");
1985             conversionFactor = nullptr;
1986             break;
1987         default:
1988             SWR_ASSERT(conversionType == CONVERT_NONE);
1989             conversionFactor = nullptr;
1990             break;
1991         }
1992 
1993         // sign extend all enabled components. If we have a fill vVertexElements, output to current simdvertex
1994         for (uint32_t i = 0; i < 4; i++)
1995         {
1996             if (isComponentEnabled(compMask, i))
1997             {
1998                 if (compCtrl[i] == ComponentControl::StoreSrc)
1999                 {
2000                     // if x or z, extract 128bits from lane 0, else for y or w, extract from lane 1
2001                     uint32_t lane = ((i == 0) || (i == 2)) ? 0 : 1;
2002                     // if x or y, use vi128XY permute result, else use vi128ZW
2003                     Value *selectedPermute_lo = (i < 2) ? vi128XY_lo : vi128ZW_lo;
2004                     Value *selectedPermute_hi = (i < 2) ? vi128XY_hi : vi128ZW_hi;
2005 
2006                     // sign extend
2007                     Value *temp_lo = PMOVSXBD(BITCAST(VEXTRACT(selectedPermute_lo, C(lane)), v16x8Ty));
2008                     Value *temp_hi = PMOVSXBD(BITCAST(VEXTRACT(selectedPermute_hi, C(lane)), v16x8Ty));
2009 
2010                     // denormalize if needed
2011                     if (conversionType != CONVERT_NONE)
2012                     {
2013                         temp_lo = FMUL(CAST(fpCast, temp_lo, mSimdFP32Ty), conversionFactor);
2014                         temp_hi = FMUL(CAST(fpCast, temp_hi, mSimdFP32Ty), conversionFactor);
2015                     }
2016 
2017                     vVertexElements[currentVertexElement] = JOIN_16(temp_lo, temp_hi);
2018 
2019                     currentVertexElement += 1;
2020                 }
2021                 else
2022                 {
2023                     vVertexElements[currentVertexElement++] = GenerateCompCtrlVector16(compCtrl[i]);
2024                 }
2025 
2026                 if (currentVertexElement > 3)
2027                 {
2028                     StoreVertexElements16(pVtxOut, outputElt++, 4, vVertexElements);
2029                     // reset to the next vVertexElement to output
2030                     currentVertexElement = 0;
2031                 }
2032             }
2033         }
2034     }
2035     // else zero extend
2036     else if ((extendType == Instruction::CastOps::ZExt) || (extendType == Instruction::CastOps::UIToFP))
2037     {
2038         // init denormalize variables if needed
2039         Instruction::CastOps fpCast;
2040         Value *conversionFactor;
2041 
2042         switch (conversionType)
2043         {
2044         case CONVERT_NORMALIZED:
2045             fpCast = Instruction::CastOps::UIToFP;
2046             conversionFactor = VIMMED1((float)(1.0 / 255.0));
2047             break;
2048         case CONVERT_USCALED:
2049             fpCast = Instruction::CastOps::UIToFP;
2050             conversionFactor = VIMMED1((float)(1.0));
2051             break;
2052         case CONVERT_SSCALED:
2053             SWR_INVALID("Type should not be zero extended!");
2054             conversionFactor = nullptr;
2055             break;
2056         default:
2057             SWR_ASSERT(conversionType == CONVERT_NONE);
2058             conversionFactor = nullptr;
2059             break;
2060         }
2061 
2062         // shuffle enabled components into lower byte of each 32bit lane, 0 extending to 32 bits
2063         for (uint32_t i = 0; i < 4; i++)
2064         {
2065             if (isComponentEnabled(compMask, i))
2066             {
2067                 if (compCtrl[i] == ComponentControl::StoreSrc)
2068                 {
2069                     // pshufb masks for each component
2070                     Value *vConstMask;
2071                     switch (swizzle[i])
2072                     {
2073                     case 0:
2074                         // x shuffle mask
2075                         vConstMask = C<char>({ 0, -1, -1, -1, 4, -1, -1, -1, 8, -1, -1, -1, 12, -1, -1, -1,
2076                             0, -1, -1, -1, 4, -1, -1, -1, 8, -1, -1, -1, 12, -1, -1, -1 });
2077                         break;
2078                     case 1:
2079                         // y shuffle mask
2080                         vConstMask = C<char>({ 1, -1, -1, -1, 5, -1, -1, -1, 9, -1, -1, -1, 13, -1, -1, -1,
2081                             1, -1, -1, -1, 5, -1, -1, -1, 9, -1, -1, -1, 13, -1, -1, -1 });
2082                         break;
2083                     case 2:
2084                         // z shuffle mask
2085                         vConstMask = C<char>({ 2, -1, -1, -1, 6, -1, -1, -1, 10, -1, -1, -1, 14, -1, -1, -1,
2086                             2, -1, -1, -1, 6, -1, -1, -1, 10, -1, -1, -1, 14, -1, -1, -1 });
2087                         break;
2088                     case 3:
2089                         // w shuffle mask
2090                         vConstMask = C<char>({ 3, -1, -1, -1, 7, -1, -1, -1, 11, -1, -1, -1, 15, -1, -1, -1,
2091                             3, -1, -1, -1, 7, -1, -1, -1, 11, -1, -1, -1, 15, -1, -1, -1 });
2092                         break;
2093                     default:
2094                         vConstMask = nullptr;
2095                         break;
2096                     }
2097 
2098                     Value *vGatherResult_lo = EXTRACT_16(vGatherResult, 0);
2099                     Value *vGatherResult_hi = EXTRACT_16(vGatherResult, 1);
2100 
2101                     Value *temp_lo = BITCAST(PSHUFB(BITCAST(vGatherResult_lo, v32x8Ty), vConstMask), vGatherTy);
2102                     Value *temp_hi = BITCAST(PSHUFB(BITCAST(vGatherResult_hi, v32x8Ty), vConstMask), vGatherTy);
2103 
2104                     // after pshufb for x channel
2105                     // 256i - 0    1    2    3    4    5    6    7
2106                     //        x000 x000 x000 x000 x000 x000 x000 x000
2107 
2108                     // denormalize if needed
2109                     if (conversionType != CONVERT_NONE)
2110                     {
2111                         temp_lo = FMUL(CAST(fpCast, temp_lo, mSimdFP32Ty), conversionFactor);
2112                         temp_hi = FMUL(CAST(fpCast, temp_hi, mSimdFP32Ty), conversionFactor);
2113                     }
2114 
2115                     vVertexElements[currentVertexElement] = JOIN_16(temp_lo, temp_hi);
2116 
2117                     currentVertexElement += 1;
2118                 }
2119                 else
2120                 {
2121                     vVertexElements[currentVertexElement++] = GenerateCompCtrlVector16(compCtrl[i]);
2122                 }
2123 
2124                 if (currentVertexElement > 3)
2125                 {
2126                     StoreVertexElements16(pVtxOut, outputElt++, 4, vVertexElements);
2127                     // reset to the next vVertexElement to output
2128                     currentVertexElement = 0;
2129                 }
2130             }
2131         }
2132     }
2133     else
2134     {
2135         SWR_INVALID("Unsupported conversion type");
2136     }
2137 }
2138 
2139 #else
2140 #if USE_SIMD16_SHADERS
Shuffle8bpcGatherd(Shuffle8bpcArgs & args,bool useVertexID2)2141 void FetchJit::Shuffle8bpcGatherd(Shuffle8bpcArgs &args, bool useVertexID2)
2142 #else
2143 void FetchJit::Shuffle8bpcGatherd(Shuffle8bpcArgs &args)
2144 #endif
2145 {
2146     // Unpack tuple args
2147     Value*& vGatherResult = std::get<0>(args);
2148     Value* pVtxOut = std::get<1>(args);
2149     const Instruction::CastOps extendType = std::get<2>(args);
2150     const ConversionType conversionType = std::get<3>(args);
2151     uint32_t &currentVertexElement = std::get<4>(args);
2152     uint32_t &outputElt = std::get<5>(args);
2153     const ComponentEnable compMask = std::get<6>(args);
2154     const ComponentControl(&compCtrl)[4] = std::get<7>(args);
2155     Value* (&vVertexElements)[4] = std::get<8>(args);
2156     const uint32_t(&swizzle)[4] = std::get<9>(args);
2157 
2158     // cast types
2159     Type* v32x8Ty = VectorType::get(mInt8Ty, mVWidth * 4); // vwidth is units of 32 bits
2160 
2161     for (uint32_t i = 0; i < 4; i++)
2162     {
2163         if (!isComponentEnabled(compMask, i))
2164             continue;
2165 
2166         if (compCtrl[i] == ComponentControl::StoreSrc)
2167         {
2168             std::vector<uint32_t> vShuffleMasks[4] = {
2169                 { 0, 4,  8, 12, 16, 20, 24, 28 }, // x
2170                 { 1, 5,  9, 13, 17, 21, 25, 29 }, // y
2171                 { 2, 6, 10, 14, 18, 22, 26, 30 }, // z
2172                 { 3, 7, 11, 15, 19, 23, 27, 31 }, // w
2173             };
2174 
2175             Value *val = VSHUFFLE(BITCAST(vGatherResult, v32x8Ty),
2176                 UndefValue::get(v32x8Ty),
2177                 vShuffleMasks[swizzle[i]]);
2178 
2179             if ((extendType == Instruction::CastOps::SExt) ||
2180                 (extendType == Instruction::CastOps::SIToFP)) {
2181                 switch (conversionType)
2182                 {
2183                 case CONVERT_NORMALIZED:
2184                     val = FMUL(SI_TO_FP(val, mSimdFP32Ty), VIMMED1((float)(1.0 / 127.0)));
2185                     break;
2186                 case CONVERT_SSCALED:
2187                     val = SI_TO_FP(val, mSimdFP32Ty);
2188                     break;
2189                 case CONVERT_USCALED:
2190                     SWR_INVALID("Type should not be sign extended!");
2191                     break;
2192                 default:
2193                     SWR_ASSERT(conversionType == CONVERT_NONE);
2194                     val = S_EXT(val, mSimdInt32Ty);
2195                     break;
2196                 }
2197             }
2198             else if ((extendType == Instruction::CastOps::ZExt) ||
2199                 (extendType == Instruction::CastOps::UIToFP)) {
2200                 switch (conversionType)
2201                 {
2202                 case CONVERT_NORMALIZED:
2203                     val = FMUL(UI_TO_FP(val, mSimdFP32Ty), VIMMED1((float)(1.0 / 255.0)));
2204                     break;
2205                 case CONVERT_SSCALED:
2206                     SWR_INVALID("Type should not be zero extended!");
2207                     break;
2208                 case CONVERT_USCALED:
2209                     val = UI_TO_FP(val, mSimdFP32Ty);
2210                     break;
2211                 default:
2212                     SWR_ASSERT(conversionType == CONVERT_NONE);
2213                     val = Z_EXT(val, mSimdInt32Ty);
2214                     break;
2215                 }
2216             }
2217             else
2218             {
2219                 SWR_INVALID("Unsupported conversion type");
2220             }
2221 
2222             vVertexElements[currentVertexElement++] = val;
2223         }
2224         else
2225         {
2226 #if USE_SIMD16_SHADERS
2227             vVertexElements[currentVertexElement++] = GenerateCompCtrlVector(compCtrl[i], useVertexID2);
2228 #else
2229             vVertexElements[currentVertexElement++] = GenerateCompCtrlVector(compCtrl[i]);
2230 #endif
2231         }
2232 
2233         if (currentVertexElement > 3)
2234         {
2235             StoreVertexElements(pVtxOut, outputElt++, 4, vVertexElements);
2236             // reset to the next vVertexElement to output
2237             currentVertexElement = 0;
2238         }
2239     }
2240 }
2241 
2242 #endif
2243 //////////////////////////////////////////////////////////////////////////
2244 /// @brief Takes a SIMD of gathered 16bpc verts, zero or sign extends,
2245 /// denormalizes if needed, converts to F32 if needed, and positions in
2246 //  the proper SIMD rows to be output to the simdvertex structure
2247 /// @param args: (tuple of args, listed below)
2248 ///   @param vGatherResult[2] - array of gathered 16bpc vertices, 4 per index
2249 ///   @param pVtxOut - base pointer to output simdvertex struct
2250 ///   @param extendType - sign extend or zero extend
2251 ///   @param bNormalized - do we need to denormalize?
2252 ///   @param currentVertexElement - reference to the current vVertexElement
2253 ///   @param outputElt - reference to the current offset from simdvertex we're o
2254 ///   @param compMask - component packing mask
2255 ///   @param compCtrl - component control val
2256 ///   @param vVertexElements[4] - vertex components to output
2257 #if USE_SIMD16_GATHERS
Shuffle16bpcGather16(Shuffle16bpcArgs & args)2258 void FetchJit::Shuffle16bpcGather16(Shuffle16bpcArgs &args)
2259 {
2260     // Unpack tuple args
2261     Value* (&vGatherResult)[2] = std::get<0>(args);
2262     Value* pVtxOut = std::get<1>(args);
2263     const Instruction::CastOps extendType = std::get<2>(args);
2264     const ConversionType conversionType = std::get<3>(args);
2265     uint32_t &currentVertexElement = std::get<4>(args);
2266     uint32_t &outputElt = std::get<5>(args);
2267     const ComponentEnable compMask = std::get<6>(args);
2268     const ComponentControl(&compCtrl)[4] = std::get<7>(args);
2269     Value* (&vVertexElements)[4] = std::get<8>(args);
2270 
2271     // cast types
2272     Type *vGatherTy = VectorType::get(IntegerType::getInt32Ty(JM()->mContext), mVWidth);
2273     Type *v32x8Ty = VectorType::get(mInt8Ty, mVWidth * 4); // vwidth is units of 32 bits
2274 
2275     // have to do extra work for sign extending
2276     if ((extendType == Instruction::CastOps::SExt) || (extendType == Instruction::CastOps::SIToFP) || (extendType == Instruction::CastOps::FPExt))
2277     {
2278         // is this PP float?
2279         bool bFP = (extendType == Instruction::CastOps::FPExt) ? true : false;
2280 
2281         Type *v8x16Ty = VectorType::get(mInt16Ty, 8); // 8x16bit in a 128bit lane
2282         Type *v128bitTy = VectorType::get(IntegerType::getIntNTy(JM()->mContext, 128), mVWidth / 4); // vwidth is units of 32 bits
2283 
2284         // shuffle mask
2285         Value *vConstMask = C<char>({ 0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15,
2286                                       0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15 });
2287         Value *vi128XY_lo = nullptr;
2288         Value *vi128XY_hi = nullptr;
2289         if (isComponentEnabled(compMask, 0) || isComponentEnabled(compMask, 1))
2290         {
2291             // SIMD16 PSHUFB isnt part of AVX-512F, so split into SIMD8 for the sake of KNL, for now..
2292 
2293             Value *vGatherResult_lo = EXTRACT_16(vGatherResult[0], 0);
2294             Value *vGatherResult_hi = EXTRACT_16(vGatherResult[0], 1);
2295 
2296             Value *vShufResult_lo = BITCAST(PSHUFB(BITCAST(vGatherResult_lo, v32x8Ty), vConstMask), vGatherTy);
2297             Value *vShufResult_hi = BITCAST(PSHUFB(BITCAST(vGatherResult_hi, v32x8Ty), vConstMask), vGatherTy);
2298 
2299             // after pshufb: group components together in each 128bit lane
2300             // 256i - 0    1    2    3    4    5    6    7
2301             //        xxxx xxxx yyyy yyyy xxxx xxxx yyyy yyyy
2302 
2303             vi128XY_lo = BITCAST(PERMD(vShufResult_lo, C<int32_t>({ 0, 1, 4, 5, 2, 3, 6, 7 })), v128bitTy);
2304             vi128XY_hi = BITCAST(PERMD(vShufResult_hi, C<int32_t>({ 0, 1, 4, 5, 2, 3, 6, 7 })), v128bitTy);
2305 
2306             // after PERMD: move and pack xy components into each 128bit lane
2307             // 256i - 0    1    2    3    4    5    6    7
2308             //        xxxx xxxx xxxx xxxx yyyy yyyy yyyy yyyy
2309         }
2310 
2311         // do the same for zw components
2312         Value *vi128ZW_lo = nullptr;
2313         Value *vi128ZW_hi = nullptr;
2314         if (isComponentEnabled(compMask, 2) || isComponentEnabled(compMask, 3))
2315         {
2316             Value *vGatherResult_lo = EXTRACT_16(vGatherResult[1], 0);
2317             Value *vGatherResult_hi = EXTRACT_16(vGatherResult[1], 1);
2318 
2319             Value *vShufResult_lo = BITCAST(PSHUFB(BITCAST(vGatherResult_lo, v32x8Ty), vConstMask), vGatherTy);
2320             Value *vShufResult_hi = BITCAST(PSHUFB(BITCAST(vGatherResult_hi, v32x8Ty), vConstMask), vGatherTy);
2321 
2322             vi128ZW_lo = BITCAST(PERMD(vShufResult_lo, C<int32_t>({ 0, 1, 4, 5, 2, 3, 6, 7 })), v128bitTy);
2323             vi128ZW_hi = BITCAST(PERMD(vShufResult_hi, C<int32_t>({ 0, 1, 4, 5, 2, 3, 6, 7 })), v128bitTy);
2324         }
2325 
2326         // init denormalize variables if needed
2327         Instruction::CastOps IntToFpCast;
2328         Value *conversionFactor;
2329 
2330         switch (conversionType)
2331         {
2332         case CONVERT_NORMALIZED:
2333             IntToFpCast = Instruction::CastOps::SIToFP;
2334             conversionFactor = VIMMED1((float)(1.0 / 32767.0));
2335             break;
2336         case CONVERT_SSCALED:
2337             IntToFpCast = Instruction::CastOps::SIToFP;
2338             conversionFactor = VIMMED1((float)(1.0));
2339             break;
2340         case CONVERT_USCALED:
2341             SWR_INVALID("Type should not be sign extended!");
2342             conversionFactor = nullptr;
2343             break;
2344         default:
2345             SWR_ASSERT(conversionType == CONVERT_NONE);
2346             conversionFactor = nullptr;
2347             break;
2348         }
2349 
2350         // sign extend all enabled components. If we have a fill vVertexElements, output to current simdvertex
2351         for (uint32_t i = 0; i < 4; i++)
2352         {
2353             if (isComponentEnabled(compMask, i))
2354             {
2355                 if (compCtrl[i] == ComponentControl::StoreSrc)
2356                 {
2357                     // if x or z, extract 128bits from lane 0, else for y or w, extract from lane 1
2358                     uint32_t lane = ((i == 0) || (i == 2)) ? 0 : 1;
2359                     // if x or y, use vi128XY permute result, else use vi128ZW
2360                     Value *selectedPermute_lo = (i < 2) ? vi128XY_lo : vi128ZW_lo;
2361                     Value *selectedPermute_hi = (i < 2) ? vi128XY_hi : vi128ZW_hi;
2362 
2363                     if (bFP)
2364                     {
2365                         // extract 128 bit lanes to sign extend each component
2366                         Value *temp_lo = CVTPH2PS(BITCAST(VEXTRACT(selectedPermute_lo, C(lane)), v8x16Ty));
2367                         Value *temp_hi = CVTPH2PS(BITCAST(VEXTRACT(selectedPermute_hi, C(lane)), v8x16Ty));
2368 
2369                         vVertexElements[currentVertexElement] = JOIN_16(temp_lo, temp_hi);
2370                     }
2371                     else
2372                     {
2373                         // extract 128 bit lanes to sign extend each component
2374                         Value *temp_lo = PMOVSXWD(BITCAST(VEXTRACT(selectedPermute_lo, C(lane)), v8x16Ty));
2375                         Value *temp_hi = PMOVSXWD(BITCAST(VEXTRACT(selectedPermute_hi, C(lane)), v8x16Ty));
2376 
2377                         // denormalize if needed
2378                         if (conversionType != CONVERT_NONE)
2379                         {
2380                             temp_lo = FMUL(CAST(IntToFpCast, temp_lo, mSimdFP32Ty), conversionFactor);
2381                             temp_hi = FMUL(CAST(IntToFpCast, temp_hi, mSimdFP32Ty), conversionFactor);
2382                         }
2383 
2384                         vVertexElements[currentVertexElement] = JOIN_16(temp_lo, temp_hi);
2385                     }
2386 
2387                     currentVertexElement += 1;
2388                 }
2389                 else
2390                 {
2391                     vVertexElements[currentVertexElement++] = GenerateCompCtrlVector16(compCtrl[i]);
2392                 }
2393 
2394                 if (currentVertexElement > 3)
2395                 {
2396                     StoreVertexElements16(pVtxOut, outputElt++, 4, vVertexElements);
2397                     // reset to the next vVertexElement to output
2398                     currentVertexElement = 0;
2399                 }
2400             }
2401         }
2402     }
2403     // else zero extend
2404     else if ((extendType == Instruction::CastOps::ZExt) || (extendType == Instruction::CastOps::UIToFP))
2405     {
2406         // pshufb masks for each component
2407         Value *vConstMask[2];
2408 
2409         if (isComponentEnabled(compMask, 0) || isComponentEnabled(compMask, 2))
2410         {
2411             // x/z shuffle mask
2412             vConstMask[0] = C<char>({ 0, 1, -1, -1, 4, 5, -1, -1, 8, 9, -1, -1, 12, 13, -1, -1,
2413                 0, 1, -1, -1, 4, 5, -1, -1, 8, 9, -1, -1, 12, 13, -1, -1, });
2414         }
2415 
2416         if (isComponentEnabled(compMask, 1) || isComponentEnabled(compMask, 3))
2417         {
2418             // y/w shuffle mask
2419             vConstMask[1] = C<char>({ 2, 3, -1, -1, 6, 7, -1, -1, 10, 11, -1, -1, 14, 15, -1, -1,
2420                 2, 3, -1, -1, 6, 7, -1, -1, 10, 11, -1, -1, 14, 15, -1, -1 });
2421         }
2422 
2423         // init denormalize variables if needed
2424         Instruction::CastOps fpCast;
2425         Value* conversionFactor;
2426 
2427         switch (conversionType)
2428         {
2429         case CONVERT_NORMALIZED:
2430             fpCast = Instruction::CastOps::UIToFP;
2431             conversionFactor = VIMMED1((float)(1.0 / 65535.0));
2432             break;
2433         case CONVERT_USCALED:
2434             fpCast = Instruction::CastOps::UIToFP;
2435             conversionFactor = VIMMED1((float)(1.0f));
2436             break;
2437         case CONVERT_SSCALED:
2438             SWR_INVALID("Type should not be zero extended!");
2439             conversionFactor = nullptr;
2440             break;
2441         default:
2442             SWR_ASSERT(conversionType == CONVERT_NONE);
2443             conversionFactor = nullptr;
2444             break;
2445         }
2446 
2447         // shuffle enabled components into lower word of each 32bit lane, 0 extending to 32 bits
2448         for (uint32_t i = 0; i < 4; i++)
2449         {
2450             if (isComponentEnabled(compMask, i))
2451             {
2452                 if (compCtrl[i] == ComponentControl::StoreSrc)
2453                 {
2454                     // select correct constMask for x/z or y/w pshufb
2455                     uint32_t selectedMask = ((i == 0) || (i == 2)) ? 0 : 1;
2456                     // if x or y, use vi128XY permute result, else use vi128ZW
2457                     uint32_t selectedGather = (i < 2) ? 0 : 1;
2458 
2459                     // SIMD16 PSHUFB isnt part of AVX-512F, so split into SIMD8 for the sake of KNL, for now..
2460 
2461                     Value *vGatherResult_lo = EXTRACT_16(vGatherResult[selectedGather], 0);
2462                     Value *vGatherResult_hi = EXTRACT_16(vGatherResult[selectedGather], 1);
2463 
2464                     Value *temp_lo = BITCAST(PSHUFB(BITCAST(vGatherResult_lo, v32x8Ty), vConstMask[selectedMask]), vGatherTy);
2465                     Value *temp_hi = BITCAST(PSHUFB(BITCAST(vGatherResult_hi, v32x8Ty), vConstMask[selectedMask]), vGatherTy);
2466 
2467                     // after pshufb mask for x channel; z uses the same shuffle from the second gather
2468                     // 256i - 0    1    2    3    4    5    6    7
2469                     //        xx00 xx00 xx00 xx00 xx00 xx00 xx00 xx00
2470 
2471                     // denormalize if needed
2472                     if (conversionType != CONVERT_NONE)
2473                     {
2474                         temp_lo = FMUL(CAST(fpCast, temp_lo, mSimdFP32Ty), conversionFactor);
2475                         temp_hi = FMUL(CAST(fpCast, temp_hi, mSimdFP32Ty), conversionFactor);
2476                     }
2477 
2478                     vVertexElements[currentVertexElement] = JOIN_16(temp_lo, temp_hi);
2479 
2480                     currentVertexElement += 1;
2481                 }
2482                 else
2483                 {
2484                     vVertexElements[currentVertexElement++] = GenerateCompCtrlVector16(compCtrl[i]);
2485                 }
2486 
2487                 if (currentVertexElement > 3)
2488                 {
2489                     StoreVertexElements16(pVtxOut, outputElt++, 4, vVertexElements);
2490                     // reset to the next vVertexElement to output
2491                     currentVertexElement = 0;
2492                 }
2493             }
2494         }
2495     }
2496     else
2497     {
2498         SWR_INVALID("Unsupported conversion type");
2499     }
2500 }
2501 
2502 #else
2503 #if USE_SIMD16_SHADERS
Shuffle16bpcGather(Shuffle16bpcArgs & args,bool useVertexID2)2504 void FetchJit::Shuffle16bpcGather(Shuffle16bpcArgs &args, bool useVertexID2)
2505 #else
2506 void FetchJit::Shuffle16bpcGather(Shuffle16bpcArgs &args)
2507 #endif
2508 {
2509     // Unpack tuple args
2510     Value* (&vGatherResult)[2] = std::get<0>(args);
2511     Value* pVtxOut = std::get<1>(args);
2512     const Instruction::CastOps extendType = std::get<2>(args);
2513     const ConversionType conversionType = std::get<3>(args);
2514     uint32_t &currentVertexElement = std::get<4>(args);
2515     uint32_t &outputElt = std::get<5>(args);
2516     const ComponentEnable compMask = std::get<6>(args);
2517     const ComponentControl(&compCtrl)[4] = std::get<7>(args);
2518     Value* (&vVertexElements)[4] = std::get<8>(args);
2519 
2520     // cast types
2521     Type* vGatherTy = VectorType::get(IntegerType::getInt32Ty(JM()->mContext), mVWidth);
2522     Type* v32x8Ty = VectorType::get(mInt8Ty, mVWidth * 4); // vwidth is units of 32 bits
2523 
2524                                                            // have to do extra work for sign extending
2525     if ((extendType == Instruction::CastOps::SExt) || (extendType == Instruction::CastOps::SIToFP) ||
2526         (extendType == Instruction::CastOps::FPExt))
2527     {
2528         // is this PP float?
2529         bool bFP = (extendType == Instruction::CastOps::FPExt) ? true : false;
2530 
2531         Type* v8x16Ty = VectorType::get(mInt16Ty, 8); // 8x16bit in a 128bit lane
2532         Type* v128bitTy = VectorType::get(IntegerType::getIntNTy(JM()->mContext, 128), mVWidth / 4); // vwidth is units of 32 bits
2533 
2534                                                                                                      // shuffle mask
2535         Value* vConstMask = C<char>({ 0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15,
2536             0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15 });
2537         Value* vi128XY = nullptr;
2538         if (isComponentEnabled(compMask, 0) || isComponentEnabled(compMask, 1)) {
2539             Value* vShufResult = BITCAST(PSHUFB(BITCAST(vGatherResult[0], v32x8Ty), vConstMask), vGatherTy);
2540             // after pshufb: group components together in each 128bit lane
2541             // 256i - 0    1    2    3    4    5    6    7
2542             //        xxxx xxxx yyyy yyyy xxxx xxxx yyyy yyyy
2543 
2544             vi128XY = BITCAST(PERMD(vShufResult, C<int32_t>({ 0, 1, 4, 5, 2, 3, 6, 7 })), v128bitTy);
2545             // after PERMD: move and pack xy components into each 128bit lane
2546             // 256i - 0    1    2    3    4    5    6    7
2547             //        xxxx xxxx xxxx xxxx yyyy yyyy yyyy yyyy
2548         }
2549 
2550         // do the same for zw components
2551         Value* vi128ZW = nullptr;
2552         if (isComponentEnabled(compMask, 2) || isComponentEnabled(compMask, 3)) {
2553             Value* vShufResult = BITCAST(PSHUFB(BITCAST(vGatherResult[1], v32x8Ty), vConstMask), vGatherTy);
2554             vi128ZW = BITCAST(PERMD(vShufResult, C<int32_t>({ 0, 1, 4, 5, 2, 3, 6, 7 })), v128bitTy);
2555         }
2556 
2557         // init denormalize variables if needed
2558         Instruction::CastOps IntToFpCast;
2559         Value* conversionFactor;
2560 
2561         switch (conversionType)
2562         {
2563         case CONVERT_NORMALIZED:
2564             IntToFpCast = Instruction::CastOps::SIToFP;
2565             conversionFactor = VIMMED1((float)(1.0 / 32767.0));
2566             break;
2567         case CONVERT_SSCALED:
2568             IntToFpCast = Instruction::CastOps::SIToFP;
2569             conversionFactor = VIMMED1((float)(1.0));
2570             break;
2571         case CONVERT_USCALED:
2572             SWR_INVALID("Type should not be sign extended!");
2573             conversionFactor = nullptr;
2574             break;
2575         default:
2576             SWR_ASSERT(conversionType == CONVERT_NONE);
2577             conversionFactor = nullptr;
2578             break;
2579         }
2580 
2581         // sign extend all enabled components. If we have a fill vVertexElements, output to current simdvertex
2582         for (uint32_t i = 0; i < 4; i++)
2583         {
2584             if (isComponentEnabled(compMask, i))
2585             {
2586                 if (compCtrl[i] == ComponentControl::StoreSrc)
2587                 {
2588                     // if x or z, extract 128bits from lane 0, else for y or w, extract from lane 1
2589                     uint32_t lane = ((i == 0) || (i == 2)) ? 0 : 1;
2590                     // if x or y, use vi128XY permute result, else use vi128ZW
2591                     Value* selectedPermute = (i < 2) ? vi128XY : vi128ZW;
2592 
2593                     if (bFP) {
2594                         // extract 128 bit lanes to sign extend each component
2595                         vVertexElements[currentVertexElement] = CVTPH2PS(BITCAST(VEXTRACT(selectedPermute, C(lane)), v8x16Ty));
2596                     }
2597                     else {
2598                         // extract 128 bit lanes to sign extend each component
2599                         vVertexElements[currentVertexElement] = PMOVSXWD(BITCAST(VEXTRACT(selectedPermute, C(lane)), v8x16Ty));
2600 
2601                         // denormalize if needed
2602                         if (conversionType != CONVERT_NONE) {
2603                             vVertexElements[currentVertexElement] = FMUL(CAST(IntToFpCast, vVertexElements[currentVertexElement], mSimdFP32Ty), conversionFactor);
2604                         }
2605                     }
2606                     currentVertexElement++;
2607                 }
2608                 else
2609                 {
2610 #if USE_SIMD16_SHADERS
2611                     vVertexElements[currentVertexElement++] = GenerateCompCtrlVector(compCtrl[i], useVertexID2);
2612 #else
2613                     vVertexElements[currentVertexElement++] = GenerateCompCtrlVector(compCtrl[i]);
2614 #endif
2615                 }
2616 
2617                 if (currentVertexElement > 3)
2618                 {
2619                     StoreVertexElements(pVtxOut, outputElt++, 4, vVertexElements);
2620                     // reset to the next vVertexElement to output
2621                     currentVertexElement = 0;
2622                 }
2623             }
2624         }
2625     }
2626     // else zero extend
2627     else if ((extendType == Instruction::CastOps::ZExt) || (extendType == Instruction::CastOps::UIToFP))
2628     {
2629         // pshufb masks for each component
2630         Value* vConstMask[2];
2631         if (isComponentEnabled(compMask, 0) || isComponentEnabled(compMask, 2)) {
2632             // x/z shuffle mask
2633             vConstMask[0] = C<char>({ 0, 1, -1, -1, 4, 5, -1, -1, 8, 9, -1, -1, 12, 13, -1, -1,
2634                 0, 1, -1, -1, 4, 5, -1, -1, 8, 9, -1, -1, 12, 13, -1, -1, });
2635         }
2636 
2637         if (isComponentEnabled(compMask, 1) || isComponentEnabled(compMask, 3)) {
2638             // y/w shuffle mask
2639             vConstMask[1] = C<char>({ 2, 3, -1, -1, 6, 7, -1, -1, 10, 11, -1, -1, 14, 15, -1, -1,
2640                 2, 3, -1, -1, 6, 7, -1, -1, 10, 11, -1, -1, 14, 15, -1, -1 });
2641         }
2642 
2643         // init denormalize variables if needed
2644         Instruction::CastOps fpCast;
2645         Value* conversionFactor;
2646 
2647         switch (conversionType)
2648         {
2649         case CONVERT_NORMALIZED:
2650             fpCast = Instruction::CastOps::UIToFP;
2651             conversionFactor = VIMMED1((float)(1.0 / 65535.0));
2652             break;
2653         case CONVERT_USCALED:
2654             fpCast = Instruction::CastOps::UIToFP;
2655             conversionFactor = VIMMED1((float)(1.0f));
2656             break;
2657         case CONVERT_SSCALED:
2658             SWR_INVALID("Type should not be zero extended!");
2659             conversionFactor = nullptr;
2660             break;
2661         default:
2662             SWR_ASSERT(conversionType == CONVERT_NONE);
2663             conversionFactor = nullptr;
2664             break;
2665         }
2666 
2667         // shuffle enabled components into lower word of each 32bit lane, 0 extending to 32 bits
2668         for (uint32_t i = 0; i < 4; i++)
2669         {
2670             if (isComponentEnabled(compMask, i))
2671             {
2672                 if (compCtrl[i] == ComponentControl::StoreSrc)
2673                 {
2674                     // select correct constMask for x/z or y/w pshufb
2675                     uint32_t selectedMask = ((i == 0) || (i == 2)) ? 0 : 1;
2676                     // if x or y, use vi128XY permute result, else use vi128ZW
2677                     uint32_t selectedGather = (i < 2) ? 0 : 1;
2678 
2679                     vVertexElements[currentVertexElement] = BITCAST(PSHUFB(BITCAST(vGatherResult[selectedGather], v32x8Ty), vConstMask[selectedMask]), vGatherTy);
2680                     // after pshufb mask for x channel; z uses the same shuffle from the second gather
2681                     // 256i - 0    1    2    3    4    5    6    7
2682                     //        xx00 xx00 xx00 xx00 xx00 xx00 xx00 xx00
2683 
2684                     // denormalize if needed
2685                     if (conversionType != CONVERT_NONE)
2686                     {
2687                         vVertexElements[currentVertexElement] = FMUL(CAST(fpCast, vVertexElements[currentVertexElement], mSimdFP32Ty), conversionFactor);
2688                     }
2689                     currentVertexElement++;
2690                 }
2691                 else
2692                 {
2693 #if USE_SIMD16_SHADERS
2694                     vVertexElements[currentVertexElement++] = GenerateCompCtrlVector(compCtrl[i], useVertexID2);
2695 #else
2696                     vVertexElements[currentVertexElement++] = GenerateCompCtrlVector(compCtrl[i]);
2697 #endif
2698                 }
2699 
2700                 if (currentVertexElement > 3)
2701                 {
2702                     StoreVertexElements(pVtxOut, outputElt++, 4, vVertexElements);
2703                     // reset to the next vVertexElement to output
2704                     currentVertexElement = 0;
2705                 }
2706             }
2707         }
2708     }
2709     else
2710     {
2711         SWR_INVALID("Unsupported conversion type");
2712     }
2713 }
2714 
2715 #endif
2716 //////////////////////////////////////////////////////////////////////////
2717 /// @brief Output a simdvertex worth of elements to the current outputElt
2718 /// @param pVtxOut - base address of VIN output struct
2719 /// @param outputElt - simdvertex offset in VIN to write to
2720 /// @param numEltsToStore - number of simdvertex rows to write out
2721 /// @param vVertexElements - LLVM Value*[] simdvertex to write out
2722 #if USE_SIMD16_GATHERS
StoreVertexElements16(Value * pVtxOut,const uint32_t outputElt,const uint32_t numEltsToStore,Value * (& vVertexElements)[4])2723 void FetchJit::StoreVertexElements16(Value* pVtxOut, const uint32_t outputElt, const uint32_t numEltsToStore, Value* (&vVertexElements)[4])
2724 {
2725     SWR_ASSERT(numEltsToStore <= 4, "Invalid element count.");
2726 
2727     for (uint32_t c = 0; c < numEltsToStore; ++c)
2728     {
2729         // STORE expects FP32 x vWidth type, just bitcast if needed
2730         if (!vVertexElements[c]->getType()->getScalarType()->isFloatTy())
2731         {
2732 #if FETCH_DUMP_VERTEX
2733             PRINT("vVertexElements[%d]: 0x%x\n", { C(c), vVertexElements[c] });
2734 #endif
2735             vVertexElements[c] = BITCAST(vVertexElements[c], mSimd16FP32Ty);
2736         }
2737 #if FETCH_DUMP_VERTEX
2738         else
2739         {
2740             PRINT("vVertexElements[%d]: %f\n", { C(c), vVertexElements[c] });
2741         }
2742 #endif
2743         // outputElt * 4 = offsetting by the size of a simdvertex
2744         // + c offsets to a 32bit x vWidth row within the current vertex
2745         Value* dest = GEP(pVtxOut, C(outputElt * 4 + c), "destGEP");
2746         STORE(vVertexElements[c], dest);
2747     }
2748 }
2749 
2750 #else
StoreVertexElements(Value * pVtxOut,const uint32_t outputElt,const uint32_t numEltsToStore,Value * (& vVertexElements)[4])2751 void FetchJit::StoreVertexElements(Value* pVtxOut, const uint32_t outputElt, const uint32_t numEltsToStore, Value* (&vVertexElements)[4])
2752 {
2753     SWR_ASSERT(numEltsToStore <= 4, "Invalid element count.");
2754 
2755     for (uint32_t c = 0; c < numEltsToStore; ++c)
2756     {
2757         // STORE expects FP32 x vWidth type, just bitcast if needed
2758         if (!vVertexElements[c]->getType()->getScalarType()->isFloatTy())
2759         {
2760 #if FETCH_DUMP_VERTEX
2761             PRINT("vVertexElements[%d]: 0x%x\n", { C(c), vVertexElements[c] });
2762 #endif
2763             vVertexElements[c] = BITCAST(vVertexElements[c], mSimdFP32Ty);
2764         }
2765 #if FETCH_DUMP_VERTEX
2766         else
2767         {
2768             PRINT("vVertexElements[%d]: %f\n", { C(c), vVertexElements[c] });
2769         }
2770 #endif
2771         // outputElt * 4 = offsetting by the size of a simdvertex
2772         // + c offsets to a 32bit x vWidth row within the current vertex
2773 #if USE_SIMD16_SHADERS
2774         Value* dest = GEP(pVtxOut, C(outputElt * 8 + c * 2), "destGEP");
2775 #else
2776         Value* dest = GEP(pVtxOut, C(outputElt * 4 + c), "destGEP");
2777 #endif
2778         STORE(vVertexElements[c], dest);
2779     }
2780 }
2781 
2782 #endif
2783 //////////////////////////////////////////////////////////////////////////
2784 /// @brief Generates a constant vector of values based on the
2785 /// ComponentControl value
2786 /// @param ctrl - ComponentControl value
2787 #if USE_SIMD16_GATHERS
GenerateCompCtrlVector16(const ComponentControl ctrl)2788 Value *FetchJit::GenerateCompCtrlVector16(const ComponentControl ctrl)
2789 {
2790     switch (ctrl)
2791     {
2792         case NoStore:
2793             return VUNDEF_I_16();
2794         case Store0:
2795             return VIMMED1_16(0);
2796         case Store1Fp:
2797             return VIMMED1_16(1.0f);
2798         case Store1Int:
2799             return VIMMED1_16(1);
2800         case StoreVertexId:
2801         {
2802             Value *pId_lo = BITCAST(LOAD(GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_VertexID  })), mSimdFP32Ty);
2803             Value *pId_hi = BITCAST(LOAD(GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_VertexID2 })), mSimdFP32Ty);
2804 
2805             Value *pId = JOIN_16(pId_lo, pId_hi);
2806 
2807             return pId;
2808         }
2809         case StoreInstanceId:
2810         {
2811             Value *pId = BITCAST(LOAD(GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_CurInstance })), mFP32Ty);
2812             return VBROADCAST_16(pId);
2813         }
2814 
2815 
2816         case StoreSrc:
2817         default:
2818             SWR_INVALID("Invalid component control");
2819             return VUNDEF_I_16();
2820     }
2821 }
2822 
2823 #else
2824 #if USE_SIMD16_SHADERS
GenerateCompCtrlVector(const ComponentControl ctrl,bool useVertexID2)2825 Value *FetchJit::GenerateCompCtrlVector(const ComponentControl ctrl, bool useVertexID2)
2826 #else
2827 Value *FetchJit::GenerateCompCtrlVector(const ComponentControl ctrl)
2828 #endif
2829 {
2830     switch (ctrl)
2831     {
2832     case NoStore:
2833         return VUNDEF_I();
2834     case Store0:
2835         return VIMMED1(0);
2836     case Store1Fp:
2837         return VIMMED1(1.0f);
2838     case Store1Int:
2839         return VIMMED1(1);
2840     case StoreVertexId:
2841         {
2842 #if USE_SIMD16_SHADERS
2843             Value *pId;
2844             if (useVertexID2)
2845             {
2846                 pId = BITCAST(LOAD(GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_VertexID2 })), mSimdFP32Ty);
2847             }
2848             else
2849             {
2850                 pId = BITCAST(LOAD(GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_VertexID })), mSimdFP32Ty);
2851             }
2852 #else
2853             Value *pId = BITCAST(LOAD(GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_VertexID })), mSimdFP32Ty);
2854 #endif
2855             return pId;
2856         }
2857     case StoreInstanceId:
2858         {
2859             Value *pId = BITCAST(LOAD(GEP(mpFetchInfo, { 0, SWR_FETCH_CONTEXT_CurInstance })), mFP32Ty);
2860             return VBROADCAST(pId);
2861         }
2862 
2863 
2864     case StoreSrc:
2865     default:
2866         SWR_INVALID("Invalid component control");
2867         return VUNDEF_I();
2868     }
2869 }
2870 
2871 #endif
2872 //////////////////////////////////////////////////////////////////////////
2873 /// @brief Returns the enable mask for the specified component.
2874 /// @param enableMask - enable bits
2875 /// @param component - component to check if enabled.
isComponentEnabled(ComponentEnable enableMask,uint8_t component)2876 bool isComponentEnabled(ComponentEnable enableMask, uint8_t component)
2877 {
2878     switch (component)
2879     {
2880         // X
2881     case 0: return (enableMask & ComponentEnable::X);
2882         // Y
2883     case 1: return (enableMask & ComponentEnable::Y);
2884         // Z
2885     case 2: return (enableMask & ComponentEnable::Z);
2886         // W
2887     case 3: return (enableMask & ComponentEnable::W);
2888 
2889     default: return false;
2890     }
2891 }
2892 
2893 // Don't want two threads compiling the same fetch shader simultaneously
2894 // Has problems in the JIT cache implementation
2895 // This is only a problem for fetch right now.
2896 static std::mutex gFetchCodegenMutex;
2897 
2898 //////////////////////////////////////////////////////////////////////////
2899 /// @brief JITs from fetch shader IR
2900 /// @param hJitMgr - JitManager handle
2901 /// @param func   - LLVM function IR
2902 /// @return PFN_FETCH_FUNC - pointer to fetch code
JitFetchFunc(HANDLE hJitMgr,const HANDLE hFunc)2903 PFN_FETCH_FUNC JitFetchFunc(HANDLE hJitMgr, const HANDLE hFunc)
2904 {
2905     const llvm::Function* func = (const llvm::Function*)hFunc;
2906     JitManager* pJitMgr = reinterpret_cast<JitManager*>(hJitMgr);
2907     PFN_FETCH_FUNC pfnFetch;
2908 
2909     gFetchCodegenMutex.lock();
2910     pfnFetch = (PFN_FETCH_FUNC)(pJitMgr->mpExec->getFunctionAddress(func->getName().str()));
2911     // MCJIT finalizes modules the first time you JIT code from them. After finalized, you cannot add new IR to the module
2912     pJitMgr->mIsModuleFinalized = true;
2913 
2914 #if defined(KNOB_SWRC_TRACING)
2915     char fName[1024];
2916     const char *funcName = func->getName().data();
2917     sprintf(fName, "%s.bin", funcName);
2918     FILE *fd = fopen(fName, "wb");
2919     fwrite((void *)pfnFetch, 1, 2048, fd);
2920     fclose(fd);
2921 #endif
2922 
2923     pJitMgr->DumpAsm(const_cast<llvm::Function*>(func), "final");
2924     gFetchCodegenMutex.unlock();
2925 
2926 
2927 
2928     return pfnFetch;
2929 }
2930 
2931 //////////////////////////////////////////////////////////////////////////
2932 /// @brief JIT compiles fetch shader
2933 /// @param hJitMgr - JitManager handle
2934 /// @param state   - fetch state to build function from
JitCompileFetch(HANDLE hJitMgr,const FETCH_COMPILE_STATE & state)2935 extern "C" PFN_FETCH_FUNC JITCALL JitCompileFetch(HANDLE hJitMgr, const FETCH_COMPILE_STATE& state)
2936 {
2937     JitManager* pJitMgr = reinterpret_cast<JitManager*>(hJitMgr);
2938 
2939     pJitMgr->SetupNewModule();
2940 
2941     FetchJit theJit(pJitMgr);
2942     HANDLE hFunc = theJit.Create(state);
2943 
2944     return JitFetchFunc(hJitMgr, hFunc);
2945 }
2946