1 /****************************************************************************
2 * Copyright (C) 2014-2015 Intel Corporation. All Rights Reserved.
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:
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
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
23 * @file streamout_jit.cpp
25 * @brief Implementation of the streamout jitter
29 ******************************************************************************/
32 #include "streamout_jit.h"
33 #include "gen_state_llvm.h"
34 #include "llvm/IR/DataLayout.h"
37 #include <unordered_set>
40 using namespace SwrJit
;
42 //////////////////////////////////////////////////////////////////////////
43 /// Interface to Jitting a fetch shader
44 //////////////////////////////////////////////////////////////////////////
45 struct StreamOutJit
: public Builder
47 StreamOutJit(JitManager
* pJitMgr
) : Builder(pJitMgr
){};
49 // returns pointer to SWR_STREAMOUT_BUFFER
50 Value
* getSOBuffer(Value
* pSoCtx
, uint32_t buffer
)
52 return LOAD(pSoCtx
, { 0, SWR_STREAMOUT_CONTEXT_pBuffer
, buffer
});
56 //////////////////////////////////////////////////////////////////////////
57 // @brief checks if streamout buffer is oob
58 // @return <i1> true/false
59 Value
* oob(const STREAMOUT_COMPILE_STATE
& state
, Value
* pSoCtx
, uint32_t buffer
)
61 Value
* returnMask
= C(false);
63 Value
* pBuf
= getSOBuffer(pSoCtx
, buffer
);
66 // @todo bool data types should generate <i1> llvm type
67 Value
* enabled
= TRUNC(LOAD(pBuf
, { 0, SWR_STREAMOUT_BUFFER_enable
}), IRB()->getInt1Ty());
70 Value
* bufferSize
= LOAD(pBuf
, { 0, SWR_STREAMOUT_BUFFER_bufferSize
});
72 // load current streamOffset
73 Value
* streamOffset
= LOAD(pBuf
, { 0, SWR_STREAMOUT_BUFFER_streamOffset
});
76 Value
* pitch
= LOAD(pBuf
, { 0, SWR_STREAMOUT_BUFFER_pitch
});
78 // buffer is considered oob if in use in a decl but not enabled
79 returnMask
= OR(returnMask
, NOT(enabled
));
81 // buffer is oob if cannot fit a prims worth of verts
82 Value
* newOffset
= ADD(streamOffset
, MUL(pitch
, C(state
.numVertsPerPrim
)));
83 returnMask
= OR(returnMask
, ICMP_SGT(newOffset
, bufferSize
));
89 //////////////////////////////////////////////////////////////////////////
90 // @brief converts scalar bitmask to <4 x i32> suitable for shuffle vector,
91 // packing the active mask bits
92 // ex. bitmask 0011 -> (0, 1, 0, 0)
93 // bitmask 1000 -> (3, 0, 0, 0)
94 // bitmask 1100 -> (2, 3, 0, 0)
95 Value
* PackMask(uint32_t bitmask
)
97 std::vector
<Constant
*> indices(4, C(0));
100 while (_BitScanForward(&index
, bitmask
))
102 indices
[elem
++] = C((int)index
);
103 bitmask
&= ~(1 << index
);
106 return ConstantVector::get(indices
);
109 //////////////////////////////////////////////////////////////////////////
110 // @brief convert scalar bitmask to <4xfloat> bitmask
111 Value
* ToMask(uint32_t bitmask
)
113 std::vector
<Constant
*> indices
;
114 for (uint32_t i
= 0; i
< 4; ++i
)
116 if (bitmask
& (1 << i
))
118 indices
.push_back(C(-1.0f
));
122 indices
.push_back(C(0.0f
));
125 return ConstantVector::get(indices
);
128 //////////////////////////////////////////////////////////////////////////
129 // @brief processes a single decl from the streamout stream. Reads 4 components from the input
130 // stream and writes N components to the output buffer given the componentMask or if
131 // a hole, just increments the buffer pointer
132 // @param pStream - pointer to current attribute
133 // @param pOutBuffers - pointers to the current location of each output buffer
134 // @param decl - input decl
135 void buildDecl(Value
* pStream
, Value
* pOutBuffers
[4], const STREAMOUT_DECL
& decl
)
137 // @todo add this to x86 macros
138 Function
* maskStore
= Intrinsic::getDeclaration(JM()->mpCurrentModule
, Intrinsic::x86_avx_maskstore_ps
);
140 uint32_t numComponents
= _mm_popcnt_u32(decl
.componentMask
);
141 uint32_t packedMask
= (1 << numComponents
) - 1;
144 // increment stream pointer to correct slot
145 Value
* pAttrib
= GEP(pStream
, C(4 * decl
.attribSlot
));
147 // load 4 components from stream
148 Type
* simd4Ty
= VectorType::get(IRB()->getFloatTy(), 4);
149 Type
* simd4PtrTy
= PointerType::get(simd4Ty
, 0);
150 pAttrib
= BITCAST(pAttrib
, simd4PtrTy
);
151 Value
*vattrib
= LOAD(pAttrib
);
153 // shuffle/pack enabled components
154 Value
* vpackedAttrib
= VSHUFFLE(vattrib
, vattrib
, PackMask(decl
.componentMask
));
156 // store to output buffer
157 // cast SO buffer to i8*, needed by maskstore
158 Value
* pOut
= BITCAST(pOutBuffers
[decl
.bufferIndex
], PointerType::get(mInt8Ty
, 0));
160 // cast input to <4xfloat>
161 Value
* src
= BITCAST(vpackedAttrib
, simd4Ty
);
163 // cast mask to <4xint>
164 Value
* mask
= ToMask(packedMask
);
165 mask
= BITCAST(mask
, VectorType::get(IRB()->getInt32Ty(), 4));
166 CALL(maskStore
, {pOut
, mask
, src
});
169 // increment SO buffer
170 pOutBuffers
[decl
.bufferIndex
] = GEP(pOutBuffers
[decl
.bufferIndex
], C(numComponents
));
173 //////////////////////////////////////////////////////////////////////////
174 // @brief builds a single vertex worth of data for the given stream
175 // @param streamState - state for this stream
176 // @param pCurVertex - pointer to src stream vertex data
177 // @param pOutBuffer - pointers to up to 4 SO buffers
178 void buildVertex(const STREAMOUT_STREAM
& streamState
, Value
* pCurVertex
, Value
* pOutBuffer
[4])
180 for (uint32_t d
= 0; d
< streamState
.numDecls
; ++d
)
182 const STREAMOUT_DECL
& decl
= streamState
.decl
[d
];
183 buildDecl(pCurVertex
, pOutBuffer
, decl
);
187 void buildStream(const STREAMOUT_COMPILE_STATE
& state
, const STREAMOUT_STREAM
& streamState
, Value
* pSoCtx
, BasicBlock
* returnBB
, Function
* soFunc
)
189 // get list of active SO buffers
190 std::unordered_set
<uint32_t> activeSOBuffers
;
191 for (uint32_t d
= 0; d
< streamState
.numDecls
; ++d
)
193 const STREAMOUT_DECL
& decl
= streamState
.decl
[d
];
194 activeSOBuffers
.insert(decl
.bufferIndex
);
197 // always increment numPrimStorageNeeded
198 Value
*numPrimStorageNeeded
= LOAD(pSoCtx
, { 0, SWR_STREAMOUT_CONTEXT_numPrimStorageNeeded
});
199 numPrimStorageNeeded
= ADD(numPrimStorageNeeded
, C(1));
200 STORE(numPrimStorageNeeded
, pSoCtx
, { 0, SWR_STREAMOUT_CONTEXT_numPrimStorageNeeded
});
202 // check OOB on active SO buffers. If any buffer is out of bound, don't write
203 // the primitive to any buffer
204 Value
* oobMask
= C(false);
205 for (uint32_t buffer
: activeSOBuffers
)
207 oobMask
= OR(oobMask
, oob(state
, pSoCtx
, buffer
));
210 BasicBlock
* validBB
= BasicBlock::Create(JM()->mContext
, "valid", soFunc
);
213 COND_BR(oobMask
, returnBB
, validBB
);
215 IRB()->SetInsertPoint(validBB
);
217 Value
* numPrimsWritten
= LOAD(pSoCtx
, { 0, SWR_STREAMOUT_CONTEXT_numPrimsWritten
});
218 numPrimsWritten
= ADD(numPrimsWritten
, C(1));
219 STORE(numPrimsWritten
, pSoCtx
, { 0, SWR_STREAMOUT_CONTEXT_numPrimsWritten
});
221 // compute start pointer for each output buffer
222 Value
* pOutBuffer
[4];
223 Value
* pOutBufferStartVertex
[4];
224 Value
* outBufferPitch
[4];
225 for (uint32_t b
: activeSOBuffers
)
227 Value
* pBuf
= getSOBuffer(pSoCtx
, b
);
228 Value
* pData
= LOAD(pBuf
, { 0, SWR_STREAMOUT_BUFFER_pBuffer
});
229 Value
* streamOffset
= LOAD(pBuf
, { 0, SWR_STREAMOUT_BUFFER_streamOffset
});
230 pOutBuffer
[b
] = GEP(pData
, streamOffset
);
231 pOutBufferStartVertex
[b
] = pOutBuffer
[b
];
233 outBufferPitch
[b
] = LOAD(pBuf
, { 0, SWR_STREAMOUT_BUFFER_pitch
});
236 // loop over the vertices of the prim
237 Value
* pStreamData
= LOAD(pSoCtx
, { 0, SWR_STREAMOUT_CONTEXT_pPrimData
});
238 for (uint32_t v
= 0; v
< state
.numVertsPerPrim
; ++v
)
240 buildVertex(streamState
, pStreamData
, pOutBuffer
);
242 // increment stream and output buffer pointers
243 // stream verts are always 32*4 dwords apart
244 pStreamData
= GEP(pStreamData
, C(KNOB_NUM_ATTRIBUTES
* 4));
246 // output buffers offset using pitch in buffer state
247 for (uint32_t b
: activeSOBuffers
)
249 pOutBufferStartVertex
[b
] = GEP(pOutBufferStartVertex
[b
], outBufferPitch
[b
]);
250 pOutBuffer
[b
] = pOutBufferStartVertex
[b
];
254 // update each active buffer's streamOffset
255 for (uint32_t b
: activeSOBuffers
)
257 Value
* pBuf
= getSOBuffer(pSoCtx
, b
);
258 Value
* streamOffset
= LOAD(pBuf
, { 0, SWR_STREAMOUT_BUFFER_streamOffset
});
259 streamOffset
= ADD(streamOffset
, MUL(C(state
.numVertsPerPrim
), outBufferPitch
[b
]));
260 STORE(streamOffset
, pBuf
, { 0, SWR_STREAMOUT_BUFFER_streamOffset
});
264 Function
* Create(const STREAMOUT_COMPILE_STATE
& state
)
266 static std::size_t soNum
= 0;
268 std::stringstream
fnName("SOShader", std::ios_base::in
| std::ios_base::out
| std::ios_base::ate
);
271 // SO function signature
272 // typedef void(__cdecl *PFN_SO_FUNC)(SWR_STREAMOUT_CONTEXT*)
274 std::vector
<Type
*> args
{
275 PointerType::get(Gen_SWR_STREAMOUT_CONTEXT(JM()), 0), // SWR_STREAMOUT_CONTEXT*
278 FunctionType
* fTy
= FunctionType::get(IRB()->getVoidTy(), args
, false);
279 Function
* soFunc
= Function::Create(fTy
, GlobalValue::ExternalLinkage
, fnName
.str(), JM()->mpCurrentModule
);
281 // create return basic block
282 BasicBlock
* entry
= BasicBlock::Create(JM()->mContext
, "entry", soFunc
);
283 BasicBlock
* returnBB
= BasicBlock::Create(JM()->mContext
, "return", soFunc
);
285 IRB()->SetInsertPoint(entry
);
288 auto argitr
= soFunc
->arg_begin();
289 Value
* pSoCtx
= &*argitr
++;
290 pSoCtx
->setName("pSoCtx");
292 const STREAMOUT_STREAM
& streamState
= state
.stream
;
293 buildStream(state
, streamState
, pSoCtx
, returnBB
, soFunc
);
297 IRB()->SetInsertPoint(returnBB
);
300 JitManager::DumpToFile(soFunc
, "SoFunc");
302 ::FunctionPassManager
passes(JM()->mpCurrentModule
);
304 passes
.add(createBreakCriticalEdgesPass());
305 passes
.add(createCFGSimplificationPass());
306 passes
.add(createEarlyCSEPass());
307 passes
.add(createPromoteMemoryToRegisterPass());
308 passes
.add(createCFGSimplificationPass());
309 passes
.add(createEarlyCSEPass());
310 passes
.add(createInstructionCombiningPass());
311 passes
.add(createInstructionSimplifierPass());
312 passes
.add(createConstantPropagationPass());
313 passes
.add(createSCCPPass());
314 passes
.add(createAggressiveDCEPass());
318 JitManager::DumpToFile(soFunc
, "SoFunc_optimized");
324 //////////////////////////////////////////////////////////////////////////
325 /// @brief JITs from streamout shader IR
326 /// @param hJitMgr - JitManager handle
327 /// @param func - LLVM function IR
328 /// @return PFN_SO_FUNC - pointer to SOS function
329 PFN_SO_FUNC
JitStreamoutFunc(HANDLE hJitMgr
, const HANDLE hFunc
)
331 const llvm::Function
*func
= (const llvm::Function
*)hFunc
;
332 JitManager
* pJitMgr
= reinterpret_cast<JitManager
*>(hJitMgr
);
333 PFN_SO_FUNC pfnStreamOut
;
334 pfnStreamOut
= (PFN_SO_FUNC
)(pJitMgr
->mpExec
->getFunctionAddress(func
->getName().str()));
335 // MCJIT finalizes modules the first time you JIT code from them. After finalized, you cannot add new IR to the module
336 pJitMgr
->mIsModuleFinalized
= true;
341 //////////////////////////////////////////////////////////////////////////
342 /// @brief JIT compiles streamout shader
343 /// @param hJitMgr - JitManager handle
344 /// @param state - SO state to build function from
345 extern "C" PFN_SO_FUNC JITCALL
JitCompileStreamout(HANDLE hJitMgr
, const STREAMOUT_COMPILE_STATE
& state
)
347 JitManager
* pJitMgr
= reinterpret_cast<JitManager
*>(hJitMgr
);
349 STREAMOUT_COMPILE_STATE soState
= state
;
350 if (soState
.offsetAttribs
)
352 for (uint32_t i
= 0; i
< soState
.stream
.numDecls
; ++i
)
354 soState
.stream
.decl
[i
].attribSlot
-= soState
.offsetAttribs
;
358 pJitMgr
->SetupNewModule();
360 StreamOutJit
theJit(pJitMgr
);
361 HANDLE hFunc
= theJit
.Create(soState
);
363 return JitStreamoutFunc(hJitMgr
, hFunc
);