src/gallium/drivers/swr/rasterizer/jitter/streamout_jit.cpp

   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 streamout_jit.cpp
  24 *
  25 * @brief Implementation of the streamout jitter
  26 *
  27 * Notes:
  28 *
  29 ******************************************************************************/
  30 #include "jit_api.h"
  31 #include "streamout_jit.h"
  32 #include "builder.h"
  33 #include "state_llvm.h"
  34 #include "llvm/IR/DataLayout.h"
  35
  36 #include <sstream>
  37 #include <unordered_set>
  38
  39 //////////////////////////////////////////////////////////////////////////
  40 /// Interface to Jitting a fetch shader
  41 //////////////////////////////////////////////////////////////////////////
  42 struct StreamOutJit : public Builder
  43 {
  44     StreamOutJit(JitManager* pJitMgr) : Builder(pJitMgr){};
  45
  46     // returns pointer to SWR_STREAMOUT_BUFFER
  47     Value* getSOBuffer(Value* pSoCtx, uint32_t buffer)
  48     {
  49         return LOAD(pSoCtx, { 0, SWR_STREAMOUT_CONTEXT_pBuffer, buffer });
  50     }
  51
  52
  53     //////////////////////////////////////////////////////////////////////////
  54     // @brief checks if streamout buffer is oob
  55     // @return <i1> true/false
  56     Value* oob(const STREAMOUT_COMPILE_STATE& state, Value* pSoCtx, uint32_t buffer)
  57     {
  58         Value* returnMask = C(false);
  59
  60         Value* pBuf = getSOBuffer(pSoCtx, buffer);
  61
  62         // load enable
  63         // @todo bool data types should generate <i1> llvm type
  64         Value* enabled = TRUNC(LOAD(pBuf, { 0, SWR_STREAMOUT_BUFFER_enable }), IRB()->getInt1Ty());
  65
  66         // load buffer size
  67         Value* bufferSize = LOAD(pBuf, { 0, SWR_STREAMOUT_BUFFER_bufferSize });
  68
  69         // load current streamOffset
  70         Value* streamOffset = LOAD(pBuf, { 0, SWR_STREAMOUT_BUFFER_streamOffset });
  71
  72         // load buffer pitch
  73         Value* pitch = LOAD(pBuf, { 0, SWR_STREAMOUT_BUFFER_pitch });
  74
  75         // buffer is considered oob if in use in a decl but not enabled
  76         returnMask = OR(returnMask, NOT(enabled));
  77
  78         // buffer is oob if cannot fit a prims worth of verts
  79         Value* newOffset = ADD(streamOffset, MUL(pitch, C(state.numVertsPerPrim)));
  80         returnMask = OR(returnMask, ICMP_SGT(newOffset, bufferSize));
  81
  82         return returnMask;
  83     }
  84
  85
  86     //////////////////////////////////////////////////////////////////////////
  87     // @brief converts scalar bitmask to <4 x i32> suitable for shuffle vector,
  88     //        packing the active mask bits
  89     //        ex. bitmask 0011 -> (0, 1, 0, 0)
  90     //            bitmask 1000 -> (3, 0, 0, 0)
  91     //            bitmask 1100 -> (2, 3, 0, 0)
  92     Value* PackMask(uint32_t bitmask)
  93     {
  94         std::vector<Constant*> indices(4, C(0));
  95         DWORD index;
  96         uint32_t elem = 0;
  97         while (_BitScanForward(&index, bitmask))
  98         {
  99             indices[elem++] = C((int)index);
 100             bitmask &= ~(1 << index);
 101         }
 102
 103         return ConstantVector::get(indices);
 104     }
 105
 106     //////////////////////////////////////////////////////////////////////////
 107     // @brief convert scalar bitmask to <4xfloat> bitmask
 108     Value* ToMask(uint32_t bitmask)
 109     {
 110         std::vector<Constant*> indices;
 111         for (uint32_t i = 0; i < 4; ++i)
 112         {
 113             if (bitmask & (1 << i))
 114             {
 115                 indices.push_back(C(-1.0f));
 116             }
 117             else
 118             {
 119                 indices.push_back(C(0.0f));
 120             }
 121         }
 122         return ConstantVector::get(indices);
 123     }
 124
 125     //////////////////////////////////////////////////////////////////////////
 126     // @brief processes a single decl from the streamout stream. Reads 4 components from the input
 127     //        stream and writes N components to the output buffer given the componentMask or if
 128     //        a hole, just increments the buffer pointer
 129     // @param pStream - pointer to current attribute
 130     // @param pOutBuffers - pointers to the current location of each output buffer
 131     // @param decl - input decl
 132     void buildDecl(Value* pStream, Value* pOutBuffers[4], const STREAMOUT_DECL& decl)
 133     {
 134         // @todo add this to x86 macros
 135         Function* maskStore = Intrinsic::getDeclaration(JM()->mpCurrentModule, Intrinsic::x86_avx_maskstore_ps);
 136
 137         uint32_t numComponents = _mm_popcnt_u32(decl.componentMask);
 138         uint32_t packedMask = (1 << numComponents) - 1;
 139         if (!decl.hole)
 140         {
 141             // increment stream pointer to correct slot
 142             Value* pAttrib = GEP(pStream, C(4 * decl.attribSlot));
 143
 144             // load 4 components from stream
 145             Type* simd4Ty = VectorType::get(IRB()->getFloatTy(), 4);
 146             Type* simd4PtrTy = PointerType::get(simd4Ty, 0);
 147             pAttrib = BITCAST(pAttrib, simd4PtrTy);
 148             Value *vattrib = LOAD(pAttrib);
 149
 150             // shuffle/pack enabled components
 151             Value* vpackedAttrib = VSHUFFLE(vattrib, vattrib, PackMask(decl.componentMask));
 152
 153             // store to output buffer
 154             // cast SO buffer to i8*, needed by maskstore
 155             Value* pOut = BITCAST(pOutBuffers[decl.bufferIndex], PointerType::get(mInt8Ty, 0));
 156
 157             // cast input to <4xfloat>
 158             Value* src = BITCAST(vpackedAttrib, simd4Ty);
 159             CALL(maskStore, {pOut, ToMask(packedMask), src});
 160         }
 161
 162         // increment SO buffer
 163         pOutBuffers[decl.bufferIndex] = GEP(pOutBuffers[decl.bufferIndex], C(numComponents));
 164     }
 165
 166     //////////////////////////////////////////////////////////////////////////
 167     // @brief builds a single vertex worth of data for the given stream
 168     // @param streamState - state for this stream
 169     // @param pCurVertex - pointer to src stream vertex data
 170     // @param pOutBuffer - pointers to up to 4 SO buffers
 171     void buildVertex(const STREAMOUT_STREAM& streamState, Value* pCurVertex, Value* pOutBuffer[4])
 172     {
 173         for (uint32_t d = 0; d < streamState.numDecls; ++d)
 174         {
 175             const STREAMOUT_DECL& decl = streamState.decl[d];
 176             buildDecl(pCurVertex, pOutBuffer, decl);
 177         }
 178     }
 179
 180     void buildStream(const STREAMOUT_COMPILE_STATE& state, const STREAMOUT_STREAM& streamState, Value* pSoCtx, BasicBlock* returnBB, Function* soFunc)
 181     {
 182         // get list of active SO buffers
 183         std::unordered_set<uint32_t> activeSOBuffers;
 184         for (uint32_t d = 0; d < streamState.numDecls; ++d)
 185         {
 186             const STREAMOUT_DECL& decl = streamState.decl[d];
 187             activeSOBuffers.insert(decl.bufferIndex);
 188         }
 189
 190         // always increment numPrimStorageNeeded
 191         Value *numPrimStorageNeeded = LOAD(pSoCtx, { 0, SWR_STREAMOUT_CONTEXT_numPrimStorageNeeded });
 192         numPrimStorageNeeded = ADD(numPrimStorageNeeded, C(1));
 193         STORE(numPrimStorageNeeded, pSoCtx, { 0, SWR_STREAMOUT_CONTEXT_numPrimStorageNeeded });
 194
 195         // check OOB on active SO buffers.  If any buffer is out of bound, don't write
 196         // the primitive to any buffer
 197         Value* oobMask = C(false);
 198         for (uint32_t buffer : activeSOBuffers)
 199         {
 200             oobMask = OR(oobMask, oob(state, pSoCtx, buffer));
 201         }
 202
 203         BasicBlock* validBB = BasicBlock::Create(JM()->mContext, "valid", soFunc);
 204
 205         // early out if OOB
 206         COND_BR(oobMask, returnBB, validBB);
 207
 208         IRB()->SetInsertPoint(validBB);
 209
 210         Value* numPrimsWritten = LOAD(pSoCtx, { 0, SWR_STREAMOUT_CONTEXT_numPrimsWritten });
 211         numPrimsWritten = ADD(numPrimsWritten, C(1));
 212         STORE(numPrimsWritten, pSoCtx, { 0, SWR_STREAMOUT_CONTEXT_numPrimsWritten });
 213
 214         // compute start pointer for each output buffer
 215         Value* pOutBuffer[4];
 216         Value* pOutBufferStartVertex[4];
 217         Value* outBufferPitch[4];
 218         for (uint32_t b: activeSOBuffers)
 219         {
 220             Value* pBuf = getSOBuffer(pSoCtx, b);
 221             Value* pData = LOAD(pBuf, { 0, SWR_STREAMOUT_BUFFER_pBuffer });
 222             Value* streamOffset = LOAD(pBuf, { 0, SWR_STREAMOUT_BUFFER_streamOffset });
 223             pOutBuffer[b] = GEP(pData, streamOffset);
 224             pOutBufferStartVertex[b] = pOutBuffer[b];
 225
 226             outBufferPitch[b] = LOAD(pBuf, { 0, SWR_STREAMOUT_BUFFER_pitch });
 227         }
 228
 229         // loop over the vertices of the prim
 230         Value* pStreamData = LOAD(pSoCtx, { 0, SWR_STREAMOUT_CONTEXT_pPrimData });
 231         for (uint32_t v = 0; v < state.numVertsPerPrim; ++v)
 232         {
 233             buildVertex(streamState, pStreamData, pOutBuffer);
 234
 235             // increment stream and output buffer pointers
 236             // stream verts are always 32*4 dwords apart
 237             pStreamData = GEP(pStreamData, C(KNOB_NUM_ATTRIBUTES * 4));
 238
 239             // output buffers offset using pitch in buffer state
 240             for (uint32_t b : activeSOBuffers)
 241             {
 242                 pOutBufferStartVertex[b] = GEP(pOutBufferStartVertex[b], outBufferPitch[b]);
 243                 pOutBuffer[b] = pOutBufferStartVertex[b];
 244             }
 245         }
 246
 247         // update each active buffer's streamOffset
 248         for (uint32_t b : activeSOBuffers)
 249         {
 250             Value* pBuf = getSOBuffer(pSoCtx, b);
 251             Value* streamOffset = LOAD(pBuf, { 0, SWR_STREAMOUT_BUFFER_streamOffset });
 252             streamOffset = ADD(streamOffset, MUL(C(state.numVertsPerPrim), outBufferPitch[b]));
 253             STORE(streamOffset, pBuf, { 0, SWR_STREAMOUT_BUFFER_streamOffset });
 254         }
 255     }
 256
 257     Function* Create(const STREAMOUT_COMPILE_STATE& state)
 258     {
 259         static std::size_t soNum = 0;
 260
 261         std::stringstream fnName("SOShader", std::ios_base::in | std::ios_base::out | std::ios_base::ate);
 262         fnName << soNum++;
 263
 264         // SO function signature
 265         // typedef void(__cdecl *PFN_SO_FUNC)(SWR_STREAMOUT_CONTEXT*)
 266
 267         std::vector<Type*> args{
 268             PointerType::get(Gen_SWR_STREAMOUT_CONTEXT(JM()), 0), // SWR_STREAMOUT_CONTEXT*
 269         };
 270
 271         FunctionType* fTy = FunctionType::get(IRB()->getVoidTy(), args, false);
 272         Function* soFunc = Function::Create(fTy, GlobalValue::ExternalLinkage, fnName.str(), JM()->mpCurrentModule);
 273
 274         // create return basic block
 275         BasicBlock* entry = BasicBlock::Create(JM()->mContext, "entry", soFunc);
 276         BasicBlock* returnBB = BasicBlock::Create(JM()->mContext, "return", soFunc);
 277
 278         IRB()->SetInsertPoint(entry);
 279
 280         // arguments
 281         auto argitr = soFunc->getArgumentList().begin();
 282         Value* pSoCtx = &*argitr++;
 283         pSoCtx->setName("pSoCtx");
 284
 285         const STREAMOUT_STREAM& streamState = state.stream;
 286         buildStream(state, streamState, pSoCtx, returnBB, soFunc);
 287
 288         BR(returnBB);
 289
 290         IRB()->SetInsertPoint(returnBB);
 291         RET_VOID();
 292
 293         JitManager::DumpToFile(soFunc, "SoFunc");
 294
 295         ::FunctionPassManager passes(JM()->mpCurrentModule);
 296
 297         passes.add(createBreakCriticalEdgesPass());
 298         passes.add(createCFGSimplificationPass());
 299         passes.add(createEarlyCSEPass());
 300         passes.add(createPromoteMemoryToRegisterPass());
 301         passes.add(createCFGSimplificationPass());
 302         passes.add(createEarlyCSEPass());
 303         passes.add(createInstructionCombiningPass());
 304         passes.add(createInstructionSimplifierPass());
 305         passes.add(createConstantPropagationPass());
 306         passes.add(createSCCPPass());
 307         passes.add(createAggressiveDCEPass());
 308
 309         passes.run(*soFunc);
 310
 311         JitManager::DumpToFile(soFunc, "SoFunc_optimized");
 312
 313         return soFunc;
 314     }
 315 };
 316
 317 //////////////////////////////////////////////////////////////////////////
 318 /// @brief JITs from streamout shader IR
 319 /// @param hJitMgr - JitManager handle
 320 /// @param func   - LLVM function IR
 321 /// @return PFN_SO_FUNC - pointer to SOS function
 322 PFN_SO_FUNC JitStreamoutFunc(HANDLE hJitMgr, const HANDLE hFunc)
 323 {
 324     const llvm::Function *func = (const llvm::Function*)hFunc;
 325     JitManager* pJitMgr = reinterpret_cast<JitManager*>(hJitMgr);
 326     PFN_SO_FUNC pfnStreamOut;
 327     pfnStreamOut = (PFN_SO_FUNC)(pJitMgr->mpExec->getFunctionAddress(func->getName().str()));
 328     // MCJIT finalizes modules the first time you JIT code from them. After finalized, you cannot add new IR to the module
 329     pJitMgr->mIsModuleFinalized = true;
 330
 331     return pfnStreamOut;
 332 }
 333
 334 //////////////////////////////////////////////////////////////////////////
 335 /// @brief JIT compiles streamout shader
 336 /// @param hJitMgr - JitManager handle
 337 /// @param state   - SO state to build function from
 338 extern "C" PFN_SO_FUNC JITCALL JitCompileStreamout(HANDLE hJitMgr, const STREAMOUT_COMPILE_STATE& state)
 339 {
 340     JitManager* pJitMgr = reinterpret_cast<JitManager*>(hJitMgr);
 341
 342     STREAMOUT_COMPILE_STATE soState = state;
 343     if (soState.offsetAttribs)
 344     {
 345         for (uint32_t i = 0; i < soState.stream.numDecls; ++i)
 346         {
 347             soState.stream.decl[i].attribSlot -= soState.offsetAttribs;
 348         }
 349     }
 350
 351     pJitMgr->SetupNewModule();
 352
 353     StreamOutJit theJit(pJitMgr);
 354     HANDLE hFunc = theJit.Create(soState);
 355
 356     return JitStreamoutFunc(hJitMgr, hFunc);
 357 }