projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
swr: [rasterizer core] Replace all naked OSALIGN macro uses with OSALIGNSIMD / OSALIG...
[mesa.git] / src / gallium / drivers / swr / rasterizer / core / clip.h
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 clip.h
24 *
25 * @brief Definitions for clipping
26 *
27 ******************************************************************************/
28 #pragma once
29
30 #include "common/simdintrin.h"
31 #include "core/context.h"
32 #include "core/pa.h"
33 #include "rdtsc_core.h"
34
35 // Temp storage used by the clipper
36 extern THREAD simdvertex tlsTempVertices[7];
37
38 enum SWR_CLIPCODES
39 {
40 // Shift clip codes out of the mantissa to prevent denormalized values when used in float compare.
41 // Guardband is able to use a single high-bit with 4 separate LSBs, because it computes a union, rather than intersection, of clipcodes.
42 #define CLIPCODE_SHIFT 23
43 FRUSTUM_LEFT = (0x01 << CLIPCODE_SHIFT),
44 FRUSTUM_TOP = (0x02 << CLIPCODE_SHIFT),
45 FRUSTUM_RIGHT = (0x04 << CLIPCODE_SHIFT),
46 FRUSTUM_BOTTOM = (0x08 << CLIPCODE_SHIFT),
47
48 FRUSTUM_NEAR = (0x10 << CLIPCODE_SHIFT),
49 FRUSTUM_FAR = (0x20 << CLIPCODE_SHIFT),
50
51 NEGW = (0x40 << CLIPCODE_SHIFT),
52
53 GUARDBAND_LEFT = (0x80 << CLIPCODE_SHIFT | 0x1),
54 GUARDBAND_TOP = (0x80 << CLIPCODE_SHIFT | 0x2),
55 GUARDBAND_RIGHT = (0x80 << CLIPCODE_SHIFT | 0x4),
56 GUARDBAND_BOTTOM = (0x80 << CLIPCODE_SHIFT | 0x8)
57 };
58
59 #define FRUSTUM_CLIP_MASK (FRUSTUM_LEFT|FRUSTUM_TOP|FRUSTUM_RIGHT|FRUSTUM_BOTTOM|FRUSTUM_NEAR|FRUSTUM_FAR)
60 #define GUARDBAND_CLIP_MASK (FRUSTUM_NEAR|FRUSTUM_FAR|GUARDBAND_LEFT|GUARDBAND_TOP|GUARDBAND_RIGHT|GUARDBAND_BOTTOM|NEGW)
61
62 void Clip(const float *pTriangle, const float *pAttribs, int numAttribs, float *pOutTriangles,
63 int *numVerts, float *pOutAttribs);
64
65 INLINE
66 void ComputeClipCodes(DRIVER_TYPE type, const API_STATE& state, const simdvector& vertex, simdscalar& clipCodes)
67 {
68 clipCodes = _simd_setzero_ps();
69
70 // -w
71 simdscalar vNegW = _simd_mul_ps(vertex.w, _simd_set1_ps(-1.0f));
72
73 // FRUSTUM_LEFT
74 simdscalar vRes = _simd_cmplt_ps(vertex.x, vNegW);
75 clipCodes = _simd_and_ps(vRes, _simd_castsi_ps(_simd_set1_epi32(FRUSTUM_LEFT)));
76
77 // FRUSTUM_TOP
78 vRes = _simd_cmplt_ps(vertex.y, vNegW);
79 clipCodes = _simd_or_ps(clipCodes, _simd_and_ps(vRes, _simd_castsi_ps(_simd_set1_epi32(FRUSTUM_TOP))));
80
81 // FRUSTUM_RIGHT
82 vRes = _simd_cmpgt_ps(vertex.x, vertex.w);
83 clipCodes = _simd_or_ps(clipCodes, _simd_and_ps(vRes, _simd_castsi_ps(_simd_set1_epi32(FRUSTUM_RIGHT))));
84
85 // FRUSTUM_BOTTOM
86 vRes = _simd_cmpgt_ps(vertex.y, vertex.w);
87 clipCodes = _simd_or_ps(clipCodes, _simd_and_ps(vRes, _simd_castsi_ps(_simd_set1_epi32(FRUSTUM_BOTTOM))));
88
89 if (state.rastState.depthClipEnable)
90 {
91 // FRUSTUM_NEAR
92 // DX clips depth [0..w], GL clips [-w..w]
93 if (type == DX)
94 {
95 vRes = _simd_cmplt_ps(vertex.z, _simd_setzero_ps());
96 }
97 else
98 {
99 vRes = _simd_cmplt_ps(vertex.z, vNegW);
100 }
101 clipCodes = _simd_or_ps(clipCodes, _simd_and_ps(vRes, _simd_castsi_ps(_simd_set1_epi32(FRUSTUM_NEAR))));
102
103 // FRUSTUM_FAR
104 vRes = _simd_cmpgt_ps(vertex.z, vertex.w);
105 clipCodes = _simd_or_ps(clipCodes, _simd_and_ps(vRes, _simd_castsi_ps(_simd_set1_epi32(FRUSTUM_FAR))));
106 }
107
108 // NEGW
109 vRes = _simd_cmple_ps(vertex.w, _simd_setzero_ps());
110 clipCodes = _simd_or_ps(clipCodes, _simd_and_ps(vRes, _simd_castsi_ps(_simd_set1_epi32(NEGW))));
111
112 // GUARDBAND_LEFT
113 simdscalar gbMult = _simd_mul_ps(vNegW, _simd_set1_ps(state.gbState.left));
114 vRes = _simd_cmplt_ps(vertex.x, gbMult);
115 clipCodes = _simd_or_ps(clipCodes, _simd_and_ps(vRes, _simd_castsi_ps(_simd_set1_epi32(GUARDBAND_LEFT))));
116
117 // GUARDBAND_TOP
118 gbMult = _simd_mul_ps(vNegW, _simd_set1_ps(state.gbState.top));
119 vRes = _simd_cmplt_ps(vertex.y, gbMult);
120 clipCodes = _simd_or_ps(clipCodes, _simd_and_ps(vRes, _simd_castsi_ps(_simd_set1_epi32(GUARDBAND_TOP))));
121
122 // GUARDBAND_RIGHT
123 gbMult = _simd_mul_ps(vertex.w, _simd_set1_ps(state.gbState.right));
124 vRes = _simd_cmpgt_ps(vertex.x, gbMult);
125 clipCodes = _simd_or_ps(clipCodes, _simd_and_ps(vRes, _simd_castsi_ps(_simd_set1_epi32(GUARDBAND_RIGHT))));
126
127 // GUARDBAND_BOTTOM
128 gbMult = _simd_mul_ps(vertex.w, _simd_set1_ps(state.gbState.bottom));
129 vRes = _simd_cmpgt_ps(vertex.y, gbMult);
130 clipCodes = _simd_or_ps(clipCodes, _simd_and_ps(vRes, _simd_castsi_ps(_simd_set1_epi32(GUARDBAND_BOTTOM))));
131 }
132
133 template<uint32_t NumVertsPerPrim>
134 class Clipper
135 {
136 public:
137 Clipper(uint32_t in_workerId, DRAW_CONTEXT* in_pDC) :
138 workerId(in_workerId), driverType(in_pDC->pContext->driverType), pDC(in_pDC), state(GetApiState(in_pDC))
139 {
140 static_assert(NumVertsPerPrim >= 1 && NumVertsPerPrim <= 3, "Invalid NumVertsPerPrim");
141 }
142
143 void ComputeClipCodes(simdvector vertex[])
144 {
145 for (uint32_t i = 0; i < NumVertsPerPrim; ++i)
146 {
147 ::ComputeClipCodes(this->driverType, this->state, vertex[i], this->clipCodes[i]);
148 }
149 }
150
151 simdscalar ComputeClipCodeIntersection()
152 {
153 simdscalar result = this->clipCodes[0];
154 for (uint32_t i = 1; i < NumVertsPerPrim; ++i)
155 {
156 result = _simd_and_ps(result, this->clipCodes[i]);
157 }
158 return result;
159 }
160
161 simdscalar ComputeClipCodeUnion()
162 {
163 simdscalar result = this->clipCodes[0];
164 for (uint32_t i = 1; i < NumVertsPerPrim; ++i)
165 {
166 result = _simd_or_ps(result, this->clipCodes[i]);
167 }
168 return result;
169 }
170
171 int ComputeNegWMask()
172 {
173 simdscalar clipCodeUnion = ComputeClipCodeUnion();
174 clipCodeUnion = _simd_and_ps(clipCodeUnion, _simd_castsi_ps(_simd_set1_epi32(NEGW)));
175 return _simd_movemask_ps(_simd_cmpneq_ps(clipCodeUnion, _simd_setzero_ps()));
176 }
177
178 int ComputeClipMask()
179 {
180 simdscalar clipUnion = ComputeClipCodeUnion();
181 clipUnion = _simd_and_ps(clipUnion, _simd_castsi_ps(_simd_set1_epi32(GUARDBAND_CLIP_MASK)));
182 return _simd_movemask_ps(_simd_cmpneq_ps(clipUnion, _simd_setzero_ps()));
183 }
184
185 // clipper is responsible for culling any prims with NAN coordinates
186 int ComputeNaNMask(simdvector prim[])
187 {
188 simdscalar vNanMask = _simd_setzero_ps();
189 for (uint32_t e = 0; e < NumVertsPerPrim; ++e)
190 {
191 simdscalar vNan01 = _simd_cmp_ps(prim[e].v[0], prim[e].v[1], _CMP_UNORD_Q);
192 vNanMask = _simd_or_ps(vNanMask, vNan01);
193 simdscalar vNan23 = _simd_cmp_ps(prim[e].v[2], prim[e].v[3], _CMP_UNORD_Q);
194 vNanMask = _simd_or_ps(vNanMask, vNan23);
195 }
196
197 return _simd_movemask_ps(vNanMask);
198 }
199
200 int ComputeUserClipCullMask(PA_STATE& pa, simdvector prim[])
201 {
202 uint8_t cullMask = this->state.rastState.cullDistanceMask;
203 simdscalar vClipCullMask = _simd_setzero_ps();
204 DWORD index;
205
206 simdvector vClipCullDistLo[3];
207 simdvector vClipCullDistHi[3];
208
209 pa.Assemble(VERTEX_CLIPCULL_DIST_LO_SLOT, vClipCullDistLo);
210 pa.Assemble(VERTEX_CLIPCULL_DIST_HI_SLOT, vClipCullDistHi);
211 while (_BitScanForward(&index, cullMask))
212 {
213 cullMask &= ~(1 << index);
214 uint32_t slot = index >> 2;
215 uint32_t component = index & 0x3;
216
217 simdscalar vCullMaskElem = _simd_set1_ps(-1.0f);
218 for (uint32_t e = 0; e < NumVertsPerPrim; ++e)
219 {
220 simdscalar vCullComp;
221 if (slot == 0)
222 {
223 vCullComp = vClipCullDistLo[e][component];
224 }
225 else
226 {
227 vCullComp = vClipCullDistHi[e][component];
228 }
229
230 // cull if cull distance < 0 || NAN
231 simdscalar vCull = _simd_cmp_ps(_mm256_setzero_ps(), vCullComp, _CMP_NLE_UQ);
232 vCullMaskElem = _simd_and_ps(vCullMaskElem, vCull);
233 }
234 vClipCullMask = _simd_or_ps(vClipCullMask, vCullMaskElem);
235 }
236
237 // clipper should also discard any primitive with NAN clip distance
238 uint8_t clipMask = this->state.rastState.clipDistanceMask;
239 while (_BitScanForward(&index, clipMask))
240 {
241 clipMask &= ~(1 << index);
242 uint32_t slot = index >> 2;
243 uint32_t component = index & 0x3;
244
245 for (uint32_t e = 0; e < NumVertsPerPrim; ++e)
246 {
247 simdscalar vClipComp;
248 if (slot == 0)
249 {
250 vClipComp = vClipCullDistLo[e][component];
251 }
252 else
253 {
254 vClipComp = vClipCullDistHi[e][component];
255 }
256
257 simdscalar vClip = _simd_cmp_ps(vClipComp, vClipComp, _CMP_UNORD_Q);
258 vClipCullMask = _simd_or_ps(vClipCullMask, vClip);
259 }
260 }
261
262 return _simd_movemask_ps(vClipCullMask);
263 }
264
265 // clip a single primitive
266 int ClipScalar(PA_STATE& pa, uint32_t primIndex, float* pOutPos, float* pOutAttribs)
267 {
268 OSALIGNSIMD(float) inVerts[3 * 4];
269 OSALIGNSIMD(float) inAttribs[3 * KNOB_NUM_ATTRIBUTES * 4];
270
271 // transpose primitive position
272 __m128 verts[3];
273 pa.AssembleSingle(VERTEX_POSITION_SLOT, primIndex, verts);
274 _mm_store_ps(&inVerts[0], verts[0]);
275 _mm_store_ps(&inVerts[4], verts[1]);
276 _mm_store_ps(&inVerts[8], verts[2]);
277
278 // transpose attribs
279 uint32_t numScalarAttribs = this->state.linkageCount * 4;
280
281 int idx = 0;
282 DWORD slot = 0;
283 uint32_t mapIdx = 0;
284 uint32_t tmpLinkage = uint32_t(this->state.linkageMask);
285 while (_BitScanForward(&slot, tmpLinkage))
286 {
287 tmpLinkage &= ~(1 << slot);
288 // Compute absolute attrib slot in vertex array
289 uint32_t inputSlot = VERTEX_ATTRIB_START_SLOT + this->state.linkageMap[mapIdx++];
290 __m128 attrib[3]; // triangle attribs (always 4 wide)
291 pa.AssembleSingle(inputSlot, primIndex, attrib);
292 _mm_store_ps(&inAttribs[idx], attrib[0]);
293 _mm_store_ps(&inAttribs[idx + numScalarAttribs], attrib[1]);
294 _mm_store_ps(&inAttribs[idx + numScalarAttribs * 2], attrib[2]);
295 idx += 4;
296 }
297
298 int numVerts;
299 Clip(inVerts, inAttribs, numScalarAttribs, pOutPos, &numVerts, pOutAttribs);
300
301 return numVerts;
302 }
303
304 // clip SIMD primitives
305 void ClipSimd(const simdscalar& vPrimMask, const simdscalar& vClipMask, PA_STATE& pa, const simdscalari& vPrimId)
306 {
307 // input/output vertex store for clipper
308 simdvertex vertices[7]; // maximum 7 verts generated per triangle
309
310 LONG constantInterpMask = this->state.backendState.constantInterpolationMask;
311 uint32_t provokingVertex = 0;
312 if(pa.binTopology == TOP_TRIANGLE_FAN)
313 {
314 provokingVertex = this->state.frontendState.provokingVertex.triFan;
315 }
316 ///@todo: line topology for wireframe?
317
318 // assemble pos
319 simdvector tmpVector[NumVertsPerPrim];
320 pa.Assemble(VERTEX_POSITION_SLOT, tmpVector);
321 for (uint32_t i = 0; i < NumVertsPerPrim; ++i)
322 {
323 vertices[i].attrib[VERTEX_POSITION_SLOT] = tmpVector[i];
324 }
325
326 // assemble attribs
327 DWORD slot = 0;
328 uint32_t mapIdx = 0;
329 uint32_t tmpLinkage = this->state.linkageMask;
330
331 int32_t maxSlot = -1;
332 while (_BitScanForward(&slot, tmpLinkage))
333 {
334 tmpLinkage &= ~(1 << slot);
335 // Compute absolute attrib slot in vertex array
336 uint32_t mapSlot = this->state.linkageMap[mapIdx++];
337 maxSlot = std::max<int32_t>(maxSlot, mapSlot);
338 uint32_t inputSlot = VERTEX_ATTRIB_START_SLOT + mapSlot;
339
340 pa.Assemble(inputSlot, tmpVector);
341
342 // if constant interpolation enabled for this attribute, assign the provoking
343 // vertex values to all edges
344 if (_bittest(&constantInterpMask, slot))
345 {
346 for (uint32_t i = 0; i < NumVertsPerPrim; ++i)
347 {
348 vertices[i].attrib[inputSlot] = tmpVector[provokingVertex];
349 }
350 }
351 else
352 {
353 for (uint32_t i = 0; i < NumVertsPerPrim; ++i)
354 {
355 vertices[i].attrib[inputSlot] = tmpVector[i];
356 }
357 }
358 }
359
360 // assemble user clip distances if enabled
361 if (this->state.rastState.clipDistanceMask & 0xf)
362 {
363 pa.Assemble(VERTEX_CLIPCULL_DIST_LO_SLOT, tmpVector);
364 for (uint32_t i = 0; i < NumVertsPerPrim; ++i)
365 {
366 vertices[i].attrib[VERTEX_CLIPCULL_DIST_LO_SLOT] = tmpVector[i];
367 }
368 }
369
370 if (this->state.rastState.clipDistanceMask & 0xf0)
371 {
372 pa.Assemble(VERTEX_CLIPCULL_DIST_HI_SLOT, tmpVector);
373 for (uint32_t i = 0; i < NumVertsPerPrim; ++i)
374 {
375 vertices[i].attrib[VERTEX_CLIPCULL_DIST_HI_SLOT] = tmpVector[i];
376 }
377 }
378
379 uint32_t numAttribs = maxSlot + 1;
380
381 simdscalari vNumClippedVerts = ClipPrims((float*)&vertices[0], vPrimMask, vClipMask, numAttribs);
382
383 // set up new PA for binning clipped primitives
384 PFN_PROCESS_PRIMS pfnBinFunc = nullptr;
385 PRIMITIVE_TOPOLOGY clipTopology = TOP_UNKNOWN;
386 if (NumVertsPerPrim == 3)
387 {
388 pfnBinFunc = BinTriangles;
389 clipTopology = TOP_TRIANGLE_FAN;
390
391 // so that the binner knows to bloat wide points later
392 if (pa.binTopology == TOP_POINT_LIST)
393 clipTopology = TOP_POINT_LIST;
394 }
395 else if (NumVertsPerPrim == 2)
396 {
397 pfnBinFunc = BinLines;
398 clipTopology = TOP_LINE_LIST;
399 }
400 else
401 {
402 SWR_ASSERT(0 && "Unexpected points in clipper.");
403 }
404
405
406 uint32_t* pVertexCount = (uint32_t*)&vNumClippedVerts;
407 uint32_t* pPrimitiveId = (uint32_t*)&vPrimId;
408
409 const simdscalari vOffsets = _mm256_set_epi32(
410 0 * sizeof(simdvertex), // unused lane
411 6 * sizeof(simdvertex),
412 5 * sizeof(simdvertex),
413 4 * sizeof(simdvertex),
414 3 * sizeof(simdvertex),
415 2 * sizeof(simdvertex),
416 1 * sizeof(simdvertex),
417 0 * sizeof(simdvertex));
418
419 // only need to gather 7 verts
420 // @todo dynamic mask based on actual # of verts generated per lane
421 const simdscalar vMask = _mm256_set_ps(0, -1, -1, -1, -1, -1, -1, -1);
422
423 uint32_t numClippedPrims = 0;
424 for (uint32_t inputPrim = 0; inputPrim < pa.NumPrims(); ++inputPrim)
425 {
426 uint32_t numEmittedVerts = pVertexCount[inputPrim];
427 if (numEmittedVerts < NumVertsPerPrim)
428 {
429 continue;
430 }
431 SWR_ASSERT(numEmittedVerts <= 7, "Unexpected vertex count from clipper.");
432
433 uint32_t numEmittedPrims = GetNumPrims(clipTopology, numEmittedVerts);
434 numClippedPrims += numEmittedPrims;
435
436 // tranpose clipper output so that each lane's vertices are in SIMD order
437 // set aside space for 2 vertices, as the PA will try to read up to 16 verts
438 // for triangle fan
439 simdvertex transposedPrims[2];
440
441 // transpose pos
442 uint8_t* pBase = (uint8_t*)(&vertices[0].attrib[VERTEX_POSITION_SLOT]) + sizeof(float) * inputPrim;
443 for (uint32_t c = 0; c < 4; ++c)
444 {
445 transposedPrims[0].attrib[VERTEX_POSITION_SLOT][c] = _simd_mask_i32gather_ps(_mm256_undefined_ps(), (const float*)pBase, vOffsets, vMask, 1);
446 pBase += sizeof(simdscalar);
447 }
448
449 // transpose attribs
450 pBase = (uint8_t*)(&vertices[0].attrib[VERTEX_ATTRIB_START_SLOT]) + sizeof(float) * inputPrim;
451 for (uint32_t attrib = 0; attrib < numAttribs; ++attrib)
452 {
453 uint32_t attribSlot = VERTEX_ATTRIB_START_SLOT + attrib;
454 for (uint32_t c = 0; c < 4; ++c)
455 {
456 transposedPrims[0].attrib[attribSlot][c] = _simd_mask_i32gather_ps(_mm256_undefined_ps(), (const float*)pBase, vOffsets, vMask, 1);
457 pBase += sizeof(simdscalar);
458 }
459 }
460
461 // transpose user clip distances if enabled
462 if (this->state.rastState.clipDistanceMask & 0xf)
463 {
464 pBase = (uint8_t*)(&vertices[0].attrib[VERTEX_CLIPCULL_DIST_LO_SLOT]) + sizeof(float) * inputPrim;
465 for (uint32_t c = 0; c < 4; ++c)
466 {
467 transposedPrims[0].attrib[VERTEX_CLIPCULL_DIST_LO_SLOT][c] = _simd_mask_i32gather_ps(_mm256_undefined_ps(), (const float*)pBase, vOffsets, vMask, 1);
468 pBase += sizeof(simdscalar);
469 }
470 }
471
472 if (this->state.rastState.clipDistanceMask & 0xf0)
473 {
474 pBase = (uint8_t*)(&vertices[0].attrib[VERTEX_CLIPCULL_DIST_HI_SLOT]) + sizeof(float) * inputPrim;
475 for (uint32_t c = 0; c < 4; ++c)
476 {
477 transposedPrims[0].attrib[VERTEX_CLIPCULL_DIST_HI_SLOT][c] = _simd_mask_i32gather_ps(_mm256_undefined_ps(), (const float*)pBase, vOffsets, vMask, 1);
478 pBase += sizeof(simdscalar);
479 }
480 }
481
482 PA_STATE_OPT clipPa(this->pDC, numEmittedPrims, (uint8_t*)&transposedPrims[0], numEmittedVerts, true, clipTopology);
483
484 while (clipPa.GetNextStreamOutput())
485 {
486 do
487 {
488 simdvector attrib[NumVertsPerPrim];
489 bool assemble = clipPa.Assemble(VERTEX_POSITION_SLOT, attrib);
490 if (assemble)
491 {
492 static const uint32_t primMaskMap[] = { 0x0, 0x1, 0x3, 0x7, 0xf, 0x1f, 0x3f, 0x7f, 0xff };
493 pfnBinFunc(this->pDC, clipPa, this->workerId, attrib, primMaskMap[numEmittedPrims], _simd_set1_epi32(pPrimitiveId[inputPrim]));
494 }
495 } while (clipPa.NextPrim());
496 }
497 }
498
499 // update global pipeline stat
500 SWR_CONTEXT* pContext = this->pDC->pContext;
501 UPDATE_STAT(CPrimitives, numClippedPrims);
502 }
503
504 // execute the clipper stage
505 void ExecuteStage(PA_STATE& pa, simdvector prim[], uint32_t primMask, simdscalari primId)
506 {
507 // set up binner based on PA state
508 PFN_PROCESS_PRIMS pfnBinner;
509 switch (pa.binTopology)
510 {
511 case TOP_POINT_LIST:
512 pfnBinner = BinPoints;
513 break;
514 case TOP_LINE_LIST:
515 case TOP_LINE_STRIP:
516 case TOP_LINE_LOOP:
517 case TOP_LINE_LIST_ADJ:
518 case TOP_LISTSTRIP_ADJ:
519 pfnBinner = BinLines;
520 break;
521 default:
522 pfnBinner = BinTriangles;
523 break;
524 };
525
526 // update clipper invocations pipeline stat
527 SWR_CONTEXT* pContext = this->pDC->pContext;
528 uint32_t numInvoc = _mm_popcnt_u32(primMask);
529 UPDATE_STAT(CInvocations, numInvoc);
530
531 ComputeClipCodes(prim);
532
533 // cull prims with NAN coords
534 primMask &= ~ComputeNaNMask(prim);
535
536 // user cull distance cull
537 if (this->state.rastState.cullDistanceMask)
538 {
539 primMask &= ~ComputeUserClipCullMask(pa, prim);
540 }
541
542 // cull prims outside view frustum
543 simdscalar clipIntersection = ComputeClipCodeIntersection();
544 int validMask = primMask & _simd_movemask_ps(_simd_cmpeq_ps(clipIntersection, _simd_setzero_ps()));
545
546 // skip clipping for points
547 uint32_t clipMask = 0;
548 if (NumVertsPerPrim != 1)
549 {
550 clipMask = primMask & ComputeClipMask();
551 }
552
553 if (clipMask)
554 {
555 RDTSC_START(FEGuardbandClip);
556 // we have to clip tris, execute the clipper, which will also
557 // call the binner
558 ClipSimd(vMask(primMask), vMask(clipMask), pa, primId);
559 RDTSC_STOP(FEGuardbandClip, 1, 0);
560 }
561 else if (validMask)
562 {
563 // update CPrimitives pipeline state
564 SWR_CONTEXT* pContext = this->pDC->pContext;
565 UPDATE_STAT(CPrimitives, _mm_popcnt_u32(validMask));
566
567 // forward valid prims directly to binner
568 pfnBinner(this->pDC, pa, this->workerId, prim, validMask, primId);
569 }
570 }
571
572 private:
573 inline simdscalar ComputeInterpFactor(simdscalar boundaryCoord0, simdscalar boundaryCoord1)
574 {
575 return _simd_div_ps(boundaryCoord0, _simd_sub_ps(boundaryCoord0, boundaryCoord1));
576 }
577
578 inline simdscalari ComputeOffsets(uint32_t attrib, simdscalari vIndices, uint32_t component)
579 {
580 const uint32_t simdVertexStride = sizeof(simdvertex);
581 const uint32_t componentStride = sizeof(simdscalar);
582 const uint32_t attribStride = sizeof(simdvector);
583 const __m256i vElemOffset = _mm256_set_epi32(7 * sizeof(float), 6 * sizeof(float), 5 * sizeof(float), 4 * sizeof(float),
584 3 * sizeof(float), 2 * sizeof(float), 1 * sizeof(float), 0 * sizeof(float));
585
586 // step to the simdvertex
587 simdscalari vOffsets = _simd_mullo_epi32(vIndices, _simd_set1_epi32(simdVertexStride));
588
589 // step to the attribute and component
590 vOffsets = _simd_add_epi32(vOffsets, _simd_set1_epi32(attribStride * attrib + componentStride * component));
591
592 // step to the lane
593 vOffsets = _simd_add_epi32(vOffsets, vElemOffset);
594
595 return vOffsets;
596 }
597
598 // gathers a single component for a given attribute for each SIMD lane
599 inline simdscalar GatherComponent(const float* pBuffer, uint32_t attrib, simdscalar vMask, simdscalari vIndices, uint32_t component)
600 {
601 simdscalari vOffsets = ComputeOffsets(attrib, vIndices, component);
602 simdscalar vSrc = _mm256_undefined_ps();
603 return _simd_mask_i32gather_ps(vSrc, pBuffer, vOffsets, vMask, 1);
604 }
605
606 inline void ScatterComponent(const float* pBuffer, uint32_t attrib, simdscalar vMask, simdscalari vIndices, uint32_t component, simdscalar vSrc)
607 {
608 simdscalari vOffsets = ComputeOffsets(attrib, vIndices, component);
609
610 uint32_t* pOffsets = (uint32_t*)&vOffsets;
611 float* pSrc = (float*)&vSrc;
612 uint32_t mask = _simd_movemask_ps(vMask);
613 DWORD lane;
614 while (_BitScanForward(&lane, mask))
615 {
616 mask &= ~(1 << lane);
617 uint8_t* pBuf = (uint8_t*)pBuffer + pOffsets[lane];
618 *(float*)pBuf = pSrc[lane];
619 }
620 }
621
622 template<SWR_CLIPCODES ClippingPlane>
623 inline void intersect(
624 const simdscalar& vActiveMask, // active lanes to operate on
625 const simdscalari& s, // index to first edge vertex v0 in pInPts.
626 const simdscalari& p, // index to second edge vertex v1 in pInPts.
627 const simdvector& v1, // vertex 0 position
628 const simdvector& v2, // vertex 1 position
629 simdscalari& outIndex, // output index.
630 const float *pInVerts, // array of all the input positions.
631 uint32_t numInAttribs, // number of attributes per vertex.
632 float *pOutVerts) // array of output positions. We'll write our new intersection point at i*4.
633 {
634 // compute interpolation factor
635 simdscalar t;
636 switch (ClippingPlane)
637 {
638 case FRUSTUM_LEFT: t = ComputeInterpFactor(_simd_add_ps(v1[3], v1[0]), _simd_add_ps(v2[3], v2[0])); break;
639 case FRUSTUM_RIGHT: t = ComputeInterpFactor(_simd_sub_ps(v1[3], v1[0]), _simd_sub_ps(v2[3], v2[0])); break;
640 case FRUSTUM_TOP: t = ComputeInterpFactor(_simd_add_ps(v1[3], v1[1]), _simd_add_ps(v2[3], v2[1])); break;
641 case FRUSTUM_BOTTOM: t = ComputeInterpFactor(_simd_sub_ps(v1[3], v1[1]), _simd_sub_ps(v2[3], v2[1])); break;
642 case FRUSTUM_NEAR:
643 // DX Znear plane is 0, GL is -w
644 if (this->driverType == DX)
645 {
646 t = ComputeInterpFactor(v1[2], v2[2]);
647 }
648 else
649 {
650 t = ComputeInterpFactor(_simd_add_ps(v1[3], v1[2]), _simd_add_ps(v2[3], v2[2]));
651 }
652 break;
653 case FRUSTUM_FAR: t = ComputeInterpFactor(_simd_sub_ps(v1[3], v1[2]), _simd_sub_ps(v2[3], v2[2])); break;
654 default: SWR_ASSERT(false, "invalid clipping plane: %d", ClippingPlane);
655 };
656
657 // interpolate position and store
658 for (uint32_t c = 0; c < 4; ++c)
659 {
660 simdscalar vOutPos = _simd_fmadd_ps(_simd_sub_ps(v2[c], v1[c]), t, v1[c]);
661 ScatterComponent(pOutVerts, VERTEX_POSITION_SLOT, vActiveMask, outIndex, c, vOutPos);
662 }
663
664 // interpolate attributes and store
665 for (uint32_t a = 0; a < numInAttribs; ++a)
666 {
667 uint32_t attribSlot = VERTEX_ATTRIB_START_SLOT + a;
668 for (uint32_t c = 0; c < 4; ++c)
669 {
670 simdscalar vAttrib0 = GatherComponent(pInVerts, attribSlot, vActiveMask, s, c);
671 simdscalar vAttrib1 = GatherComponent(pInVerts, attribSlot, vActiveMask, p, c);
672 simdscalar vOutAttrib = _simd_fmadd_ps(_simd_sub_ps(vAttrib1, vAttrib0), t, vAttrib0);
673 ScatterComponent(pOutVerts, attribSlot, vActiveMask, outIndex, c, vOutAttrib);
674 }
675 }
676
677 // interpolate clip distance if enabled
678 if (this->state.rastState.clipDistanceMask & 0xf)
679 {
680 uint32_t attribSlot = VERTEX_CLIPCULL_DIST_LO_SLOT;
681 for (uint32_t c = 0; c < 4; ++c)
682 {
683 simdscalar vAttrib0 = GatherComponent(pInVerts, attribSlot, vActiveMask, s, c);
684 simdscalar vAttrib1 = GatherComponent(pInVerts, attribSlot, vActiveMask, p, c);
685 simdscalar vOutAttrib = _simd_fmadd_ps(_simd_sub_ps(vAttrib1, vAttrib0), t, vAttrib0);
686 ScatterComponent(pOutVerts, attribSlot, vActiveMask, outIndex, c, vOutAttrib);
687 }
688 }
689
690 if (this->state.rastState.clipDistanceMask & 0xf0)
691 {
692 uint32_t attribSlot = VERTEX_CLIPCULL_DIST_HI_SLOT;
693 for (uint32_t c = 0; c < 4; ++c)
694 {
695 simdscalar vAttrib0 = GatherComponent(pInVerts, attribSlot, vActiveMask, s, c);
696 simdscalar vAttrib1 = GatherComponent(pInVerts, attribSlot, vActiveMask, p, c);
697 simdscalar vOutAttrib = _simd_fmadd_ps(_simd_sub_ps(vAttrib1, vAttrib0), t, vAttrib0);
698 ScatterComponent(pOutVerts, attribSlot, vActiveMask, outIndex, c, vOutAttrib);
699 }
700 }
701 }
702
703 template<SWR_CLIPCODES ClippingPlane>
704 inline simdscalar inside(const simdvector& v)
705 {
706 switch (ClippingPlane)
707 {
708 case FRUSTUM_LEFT: return _simd_cmpge_ps(v[0], _simd_mul_ps(v[3], _simd_set1_ps(-1.0f)));
709 case FRUSTUM_RIGHT: return _simd_cmple_ps(v[0], v[3]);
710 case FRUSTUM_TOP: return _simd_cmpge_ps(v[1], _simd_mul_ps(v[3], _simd_set1_ps(-1.0f)));
711 case FRUSTUM_BOTTOM: return _simd_cmple_ps(v[1], v[3]);
712 case FRUSTUM_NEAR: return _simd_cmpge_ps(v[2], this->driverType == DX ? _simd_setzero_ps() : _simd_mul_ps(v[3], _simd_set1_ps(-1.0f)));
713 case FRUSTUM_FAR: return _simd_cmple_ps(v[2], v[3]);
714 default:
715 SWR_ASSERT(false, "invalid clipping plane: %d", ClippingPlane);
716 return _simd_setzero_ps();
717 }
718 }
719
720 template<SWR_CLIPCODES ClippingPlane>
721 simdscalari ClipTriToPlane(const float* pInVerts, const simdscalari& vNumInPts, uint32_t numInAttribs, float* pOutVerts)
722 {
723 simdscalari vCurIndex = _simd_setzero_si();
724 simdscalari vOutIndex = _simd_setzero_si();
725 simdscalar vActiveMask = _simd_castsi_ps(_simd_cmplt_epi32(vCurIndex, vNumInPts));
726
727 while (!_simd_testz_ps(vActiveMask, vActiveMask)) // loop until activeMask is empty
728 {
729 simdscalari s = vCurIndex;
730 simdscalari p = _simd_add_epi32(s, _simd_set1_epi32(1));
731 simdscalari underFlowMask = _simd_cmpgt_epi32(vNumInPts, p);
732 p = _simd_castps_si(_simd_blendv_ps(_simd_setzero_ps(), _simd_castsi_ps(p), _simd_castsi_ps(underFlowMask)));
733
734 // gather position
735 simdvector vInPos0, vInPos1;
736 for (uint32_t c = 0; c < 4; ++c)
737 {
738 vInPos0[c] = GatherComponent(pInVerts, VERTEX_POSITION_SLOT, vActiveMask, s, c);
739 vInPos1[c] = GatherComponent(pInVerts, VERTEX_POSITION_SLOT, vActiveMask, p, c);
740 }
741
742 // compute inside mask
743 simdscalar s_in = inside<ClippingPlane>(vInPos0);
744 simdscalar p_in = inside<ClippingPlane>(vInPos1);
745
746 // compute intersection mask (s_in != p_in)
747 simdscalar intersectMask = _simd_xor_ps(s_in, p_in);
748 intersectMask = _simd_and_ps(intersectMask, vActiveMask);
749
750 // store s if inside
751 s_in = _simd_and_ps(s_in, vActiveMask);
752 if (!_simd_testz_ps(s_in, s_in))
753 {
754 // store position
755 for (uint32_t c = 0; c < 4; ++c)
756 {
757 ScatterComponent(pOutVerts, VERTEX_POSITION_SLOT, s_in, vOutIndex, c, vInPos0[c]);
758 }
759
760 // store attribs
761 for (uint32_t a = 0; a < numInAttribs; ++a)
762 {
763 uint32_t attribSlot = VERTEX_ATTRIB_START_SLOT + a;
764 for (uint32_t c = 0; c < 4; ++c)
765 {
766 simdscalar vAttrib = GatherComponent(pInVerts, attribSlot, s_in, s, c);
767 ScatterComponent(pOutVerts, attribSlot, s_in, vOutIndex, c, vAttrib);
768 }
769 }
770
771 // store clip distance if enabled
772 if (this->state.rastState.clipDistanceMask & 0xf)
773 {
774 uint32_t attribSlot = VERTEX_CLIPCULL_DIST_LO_SLOT;
775 for (uint32_t c = 0; c < 4; ++c)
776 {
777 simdscalar vAttrib = GatherComponent(pInVerts, attribSlot, s_in, s, c);
778 ScatterComponent(pOutVerts, attribSlot, s_in, vOutIndex, c, vAttrib);
779 }
780 }
781
782 if (this->state.rastState.clipDistanceMask & 0xf0)
783 {
784 uint32_t attribSlot = VERTEX_CLIPCULL_DIST_HI_SLOT;
785 for (uint32_t c = 0; c < 4; ++c)
786 {
787 simdscalar vAttrib = GatherComponent(pInVerts, attribSlot, s_in, s, c);
788 ScatterComponent(pOutVerts, attribSlot, s_in, vOutIndex, c, vAttrib);
789 }
790 }
791
792 // increment outIndex
793 vOutIndex = _simd_blendv_epi32(vOutIndex, _simd_add_epi32(vOutIndex, _simd_set1_epi32(1)), s_in);
794 }
795
796 // compute and store intersection
797 if (!_simd_testz_ps(intersectMask, intersectMask))
798 {
799 intersect<ClippingPlane>(intersectMask, s, p, vInPos0, vInPos1, vOutIndex, pInVerts, numInAttribs, pOutVerts);
800
801 // increment outIndex for active lanes
802 vOutIndex = _simd_blendv_epi32(vOutIndex, _simd_add_epi32(vOutIndex, _simd_set1_epi32(1)), intersectMask);
803 }
804
805 // increment loop index and update active mask
806 vCurIndex = _simd_add_epi32(vCurIndex, _simd_set1_epi32(1));
807 vActiveMask = _simd_castsi_ps(_simd_cmplt_epi32(vCurIndex, vNumInPts));
808 }
809
810 return vOutIndex;
811 }
812
813 template<SWR_CLIPCODES ClippingPlane>
814 simdscalari ClipLineToPlane(const float* pInVerts, const simdscalari& vNumInPts, uint32_t numInAttribs, float* pOutVerts)
815 {
816 simdscalari vCurIndex = _simd_setzero_si();
817 simdscalari vOutIndex = _simd_setzero_si();
818 simdscalar vActiveMask = _simd_castsi_ps(_simd_cmplt_epi32(vCurIndex, vNumInPts));
819
820 if (!_simd_testz_ps(vActiveMask, vActiveMask))
821 {
822 simdscalari s = vCurIndex;
823 simdscalari p = _simd_add_epi32(s, _simd_set1_epi32(1));
824
825 // gather position
826 simdvector vInPos0, vInPos1;
827 for (uint32_t c = 0; c < 4; ++c)
828 {
829 vInPos0[c] = GatherComponent(pInVerts, VERTEX_POSITION_SLOT, vActiveMask, s, c);
830 vInPos1[c] = GatherComponent(pInVerts, VERTEX_POSITION_SLOT, vActiveMask, p, c);
831 }
832
833 // compute inside mask
834 simdscalar s_in = inside<ClippingPlane>(vInPos0);
835 simdscalar p_in = inside<ClippingPlane>(vInPos1);
836
837 // compute intersection mask (s_in != p_in)
838 simdscalar intersectMask = _simd_xor_ps(s_in, p_in);
839 intersectMask = _simd_and_ps(intersectMask, vActiveMask);
840
841 // store s if inside
842 s_in = _simd_and_ps(s_in, vActiveMask);
843 if (!_simd_testz_ps(s_in, s_in))
844 {
845 for (uint32_t c = 0; c < 4; ++c)
846 {
847 ScatterComponent(pOutVerts, VERTEX_POSITION_SLOT, s_in, vOutIndex, c, vInPos0[c]);
848 }
849
850 // interpolate attributes and store
851 for (uint32_t a = 0; a < numInAttribs; ++a)
852 {
853 uint32_t attribSlot = VERTEX_ATTRIB_START_SLOT + a;
854 for (uint32_t c = 0; c < 4; ++c)
855 {
856 simdscalar vAttrib = GatherComponent(pInVerts, attribSlot, s_in, s, c);
857 ScatterComponent(pOutVerts, attribSlot, s_in, vOutIndex, c, vAttrib);
858 }
859 }
860
861 // increment outIndex
862 vOutIndex = _simd_blendv_epi32(vOutIndex, _simd_add_epi32(vOutIndex, _simd_set1_epi32(1)), s_in);
863 }
864
865 // compute and store intersection
866 if (!_simd_testz_ps(intersectMask, intersectMask))
867 {
868 intersect<ClippingPlane>(intersectMask, s, p, vInPos0, vInPos1, vOutIndex, pInVerts, numInAttribs, pOutVerts);
869
870 // increment outIndex for active lanes
871 vOutIndex = _simd_blendv_epi32(vOutIndex, _simd_add_epi32(vOutIndex, _simd_set1_epi32(1)), intersectMask);
872 }
873
874 // store p if inside
875 p_in = _simd_and_ps(p_in, vActiveMask);
876 if (!_simd_testz_ps(p_in, p_in))
877 {
878 for (uint32_t c = 0; c < 4; ++c)
879 {
880 ScatterComponent(pOutVerts, VERTEX_POSITION_SLOT, p_in, vOutIndex, c, vInPos1[c]);
881 }
882
883 // interpolate attributes and store
884 for (uint32_t a = 0; a < numInAttribs; ++a)
885 {
886 uint32_t attribSlot = VERTEX_ATTRIB_START_SLOT + a;
887 for (uint32_t c = 0; c < 4; ++c)
888 {
889 simdscalar vAttrib = GatherComponent(pInVerts, attribSlot, p_in, p, c);
890 ScatterComponent(pOutVerts, attribSlot, p_in, vOutIndex, c, vAttrib);
891 }
892 }
893
894 // increment outIndex
895 vOutIndex = _simd_blendv_epi32(vOutIndex, _simd_add_epi32(vOutIndex, _simd_set1_epi32(1)), p_in);
896 }
897 }
898
899 return vOutIndex;
900 }
901
902 //////////////////////////////////////////////////////////////////////////
903 /// @brief Vertical clipper. Clips SIMD primitives at a time
904 /// @param pVertices - pointer to vertices in SOA form. Clipper will read input and write results to this buffer
905 /// @param vPrimMask - mask of valid input primitives, including non-clipped prims
906 /// @param numAttribs - number of valid input attribs, including position
907 simdscalari ClipPrims(float* pVertices, const simdscalar& vPrimMask, const simdscalar& vClipMask, int numAttribs)
908 {
909 // temp storage
910 float* pTempVerts = (float*)&tlsTempVertices[0];
911
912 // zero out num input verts for non-active lanes
913 simdscalari vNumInPts = _simd_set1_epi32(NumVertsPerPrim);
914 vNumInPts = _simd_blendv_epi32(_simd_setzero_si(), vNumInPts, vClipMask);
915
916 // clip prims to frustum
917 simdscalari vNumOutPts;
918 if (NumVertsPerPrim == 3)
919 {
920 vNumOutPts = ClipTriToPlane<FRUSTUM_NEAR>(pVertices, vNumInPts, numAttribs, pTempVerts);
921 vNumOutPts = ClipTriToPlane<FRUSTUM_FAR>(pTempVerts, vNumOutPts, numAttribs, pVertices);
922 vNumOutPts = ClipTriToPlane<FRUSTUM_LEFT>(pVertices, vNumOutPts, numAttribs, pTempVerts);
923 vNumOutPts = ClipTriToPlane<FRUSTUM_RIGHT>(pTempVerts, vNumOutPts, numAttribs, pVertices);
924 vNumOutPts = ClipTriToPlane<FRUSTUM_BOTTOM>(pVertices, vNumOutPts, numAttribs, pTempVerts);
925 vNumOutPts = ClipTriToPlane<FRUSTUM_TOP>(pTempVerts, vNumOutPts, numAttribs, pVertices);
926 }
927 else
928 {
929 SWR_ASSERT(NumVertsPerPrim == 2);
930 vNumOutPts = ClipLineToPlane<FRUSTUM_NEAR>(pVertices, vNumInPts, numAttribs, pTempVerts);
931 vNumOutPts = ClipLineToPlane<FRUSTUM_FAR>(pTempVerts, vNumOutPts, numAttribs, pVertices);
932 vNumOutPts = ClipLineToPlane<FRUSTUM_LEFT>(pVertices, vNumOutPts, numAttribs, pTempVerts);
933 vNumOutPts = ClipLineToPlane<FRUSTUM_RIGHT>(pTempVerts, vNumOutPts, numAttribs, pVertices);
934 vNumOutPts = ClipLineToPlane<FRUSTUM_BOTTOM>(pVertices, vNumOutPts, numAttribs, pTempVerts);
935 vNumOutPts = ClipLineToPlane<FRUSTUM_TOP>(pTempVerts, vNumOutPts, numAttribs, pVertices);
936 }
937
938 // restore num verts for non-clipped, active lanes
939 simdscalar vNonClippedMask = _simd_andnot_ps(vClipMask, vPrimMask);
940 vNumOutPts = _simd_blendv_epi32(vNumOutPts, _simd_set1_epi32(NumVertsPerPrim), vNonClippedMask);
941
942 return vNumOutPts;
943 }
944
945 const uint32_t workerId{ 0 };
946 const DRIVER_TYPE driverType{ DX };
947 DRAW_CONTEXT* pDC{ nullptr };
948 const API_STATE& state;
949 simdscalar clipCodes[NumVertsPerPrim];
950 };
951
952
953 // pipeline stage functions
954 void ClipTriangles(DRAW_CONTEXT *pDC, PA_STATE& pa, uint32_t workerId, simdvector prims[], uint32_t primMask, simdscalari primId);
955 void ClipLines(DRAW_CONTEXT *pDC, PA_STATE& pa, uint32_t workerId, simdvector prims[], uint32_t primMask, simdscalari primId);
956 void ClipPoints(DRAW_CONTEXT *pDC, PA_STATE& pa, uint32_t workerId, simdvector prims[], uint32_t primMask, simdscalari primId);