projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
swr/rast: Rework attribute layout
[mesa.git] / src / gallium / drivers / swr / rasterizer / core / frontend.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 frontend.cpp
24 *
25 * @brief Implementation for Frontend which handles vertex processing,
26 * primitive assembly, clipping, binning, etc.
27 *
28 ******************************************************************************/
29
30 #include "api.h"
31 #include "frontend.h"
32 #include "backend.h"
33 #include "context.h"
34 #include "rdtsc_core.h"
35 #include "utils.h"
36 #include "threads.h"
37 #include "pa.h"
38 #include "clip.h"
39 #include "tilemgr.h"
40 #include "tessellator.h"
41 #include <limits>
42
43 //////////////////////////////////////////////////////////////////////////
44 /// @brief Helper macro to generate a bitmask
45 static INLINE uint32_t GenMask(uint32_t numBits)
46 {
47 SWR_ASSERT(numBits <= (sizeof(uint32_t) * 8), "Too many bits (%d) for %s", numBits, __FUNCTION__);
48 return ((1U << numBits) - 1);
49 }
50
51 //////////////////////////////////////////////////////////////////////////
52 /// @brief FE handler for SwrSync.
53 /// @param pContext - pointer to SWR context.
54 /// @param pDC - pointer to draw context.
55 /// @param workerId - thread's worker id. Even thread has a unique id.
56 /// @param pUserData - Pointer to user data passed back to sync callback.
57 /// @todo This should go away when we switch this to use compute threading.
58 void ProcessSync(
59 SWR_CONTEXT *pContext,
60 DRAW_CONTEXT *pDC,
61 uint32_t workerId,
62 void *pUserData)
63 {
64 BE_WORK work;
65 work.type = SYNC;
66 work.pfnWork = ProcessSyncBE;
67
68 MacroTileMgr *pTileMgr = pDC->pTileMgr;
69 pTileMgr->enqueue(0, 0, &work);
70 }
71
72 //////////////////////////////////////////////////////////////////////////
73 /// @brief FE handler for SwrDestroyContext.
74 /// @param pContext - pointer to SWR context.
75 /// @param pDC - pointer to draw context.
76 /// @param workerId - thread's worker id. Even thread has a unique id.
77 /// @param pUserData - Pointer to user data passed back to sync callback.
78 void ProcessShutdown(
79 SWR_CONTEXT *pContext,
80 DRAW_CONTEXT *pDC,
81 uint32_t workerId,
82 void *pUserData)
83 {
84 BE_WORK work;
85 work.type = SHUTDOWN;
86 work.pfnWork = ProcessShutdownBE;
87
88 MacroTileMgr *pTileMgr = pDC->pTileMgr;
89 // Enqueue at least 1 work item for each worker thread
90 // account for number of numa nodes
91 uint32_t numNumaNodes = pContext->threadPool.numaMask + 1;
92
93 for (uint32_t i = 0; i < pContext->threadPool.numThreads; ++i)
94 {
95 for (uint32_t n = 0; n < numNumaNodes; ++n)
96 {
97 pTileMgr->enqueue(i, n, &work);
98 }
99 }
100 }
101
102 //////////////////////////////////////////////////////////////////////////
103 /// @brief FE handler for SwrClearRenderTarget.
104 /// @param pContext - pointer to SWR context.
105 /// @param pDC - pointer to draw context.
106 /// @param workerId - thread's worker id. Even thread has a unique id.
107 /// @param pUserData - Pointer to user data passed back to clear callback.
108 /// @todo This should go away when we switch this to use compute threading.
109 void ProcessClear(
110 SWR_CONTEXT *pContext,
111 DRAW_CONTEXT *pDC,
112 uint32_t workerId,
113 void *pUserData)
114 {
115 CLEAR_DESC *pDesc = (CLEAR_DESC*)pUserData;
116 MacroTileMgr *pTileMgr = pDC->pTileMgr;
117
118 // queue a clear to each macro tile
119 // compute macro tile bounds for the specified rect
120 uint32_t macroTileXMin = pDesc->rect.xmin / KNOB_MACROTILE_X_DIM;
121 uint32_t macroTileXMax = (pDesc->rect.xmax - 1) / KNOB_MACROTILE_X_DIM;
122 uint32_t macroTileYMin = pDesc->rect.ymin / KNOB_MACROTILE_Y_DIM;
123 uint32_t macroTileYMax = (pDesc->rect.ymax - 1) / KNOB_MACROTILE_Y_DIM;
124
125 BE_WORK work;
126 work.type = CLEAR;
127 work.pfnWork = ProcessClearBE;
128 work.desc.clear = *pDesc;
129
130 for (uint32_t y = macroTileYMin; y <= macroTileYMax; ++y)
131 {
132 for (uint32_t x = macroTileXMin; x <= macroTileXMax; ++x)
133 {
134 pTileMgr->enqueue(x, y, &work);
135 }
136 }
137 }
138
139 //////////////////////////////////////////////////////////////////////////
140 /// @brief FE handler for SwrStoreTiles.
141 /// @param pContext - pointer to SWR context.
142 /// @param pDC - pointer to draw context.
143 /// @param workerId - thread's worker id. Even thread has a unique id.
144 /// @param pUserData - Pointer to user data passed back to callback.
145 /// @todo This should go away when we switch this to use compute threading.
146 void ProcessStoreTiles(
147 SWR_CONTEXT *pContext,
148 DRAW_CONTEXT *pDC,
149 uint32_t workerId,
150 void *pUserData)
151 {
152 AR_BEGIN(FEProcessStoreTiles, pDC->drawId);
153 MacroTileMgr *pTileMgr = pDC->pTileMgr;
154 STORE_TILES_DESC* pDesc = (STORE_TILES_DESC*)pUserData;
155
156 // queue a store to each macro tile
157 // compute macro tile bounds for the specified rect
158 uint32_t macroTileXMin = pDesc->rect.xmin / KNOB_MACROTILE_X_DIM;
159 uint32_t macroTileXMax = (pDesc->rect.xmax - 1) / KNOB_MACROTILE_X_DIM;
160 uint32_t macroTileYMin = pDesc->rect.ymin / KNOB_MACROTILE_Y_DIM;
161 uint32_t macroTileYMax = (pDesc->rect.ymax - 1) / KNOB_MACROTILE_Y_DIM;
162
163 // store tiles
164 BE_WORK work;
165 work.type = STORETILES;
166 work.pfnWork = ProcessStoreTilesBE;
167 work.desc.storeTiles = *pDesc;
168
169 for (uint32_t y = macroTileYMin; y <= macroTileYMax; ++y)
170 {
171 for (uint32_t x = macroTileXMin; x <= macroTileXMax; ++x)
172 {
173 pTileMgr->enqueue(x, y, &work);
174 }
175 }
176
177 AR_END(FEProcessStoreTiles, 0);
178 }
179
180 //////////////////////////////////////////////////////////////////////////
181 /// @brief FE handler for SwrInvalidateTiles.
182 /// @param pContext - pointer to SWR context.
183 /// @param pDC - pointer to draw context.
184 /// @param workerId - thread's worker id. Even thread has a unique id.
185 /// @param pUserData - Pointer to user data passed back to callback.
186 /// @todo This should go away when we switch this to use compute threading.
187 void ProcessDiscardInvalidateTiles(
188 SWR_CONTEXT *pContext,
189 DRAW_CONTEXT *pDC,
190 uint32_t workerId,
191 void *pUserData)
192 {
193 AR_BEGIN(FEProcessInvalidateTiles, pDC->drawId);
194 DISCARD_INVALIDATE_TILES_DESC *pDesc = (DISCARD_INVALIDATE_TILES_DESC*)pUserData;
195 MacroTileMgr *pTileMgr = pDC->pTileMgr;
196
197 // compute macro tile bounds for the specified rect
198 uint32_t macroTileXMin = (pDesc->rect.xmin + KNOB_MACROTILE_X_DIM - 1) / KNOB_MACROTILE_X_DIM;
199 uint32_t macroTileXMax = (pDesc->rect.xmax / KNOB_MACROTILE_X_DIM) - 1;
200 uint32_t macroTileYMin = (pDesc->rect.ymin + KNOB_MACROTILE_Y_DIM - 1) / KNOB_MACROTILE_Y_DIM;
201 uint32_t macroTileYMax = (pDesc->rect.ymax / KNOB_MACROTILE_Y_DIM) - 1;
202
203 if (pDesc->fullTilesOnly == false)
204 {
205 // include partial tiles
206 macroTileXMin = pDesc->rect.xmin / KNOB_MACROTILE_X_DIM;
207 macroTileXMax = (pDesc->rect.xmax - 1) / KNOB_MACROTILE_X_DIM;
208 macroTileYMin = pDesc->rect.ymin / KNOB_MACROTILE_Y_DIM;
209 macroTileYMax = (pDesc->rect.ymax - 1) / KNOB_MACROTILE_Y_DIM;
210 }
211
212 SWR_ASSERT(macroTileXMax <= KNOB_NUM_HOT_TILES_X);
213 SWR_ASSERT(macroTileYMax <= KNOB_NUM_HOT_TILES_Y);
214
215 macroTileXMax = std::min<int32_t>(macroTileXMax, KNOB_NUM_HOT_TILES_X);
216 macroTileYMax = std::min<int32_t>(macroTileYMax, KNOB_NUM_HOT_TILES_Y);
217
218 // load tiles
219 BE_WORK work;
220 work.type = DISCARDINVALIDATETILES;
221 work.pfnWork = ProcessDiscardInvalidateTilesBE;
222 work.desc.discardInvalidateTiles = *pDesc;
223
224 for (uint32_t x = macroTileXMin; x <= macroTileXMax; ++x)
225 {
226 for (uint32_t y = macroTileYMin; y <= macroTileYMax; ++y)
227 {
228 pTileMgr->enqueue(x, y, &work);
229 }
230 }
231
232 AR_END(FEProcessInvalidateTiles, 0);
233 }
234
235 //////////////////////////////////////////////////////////////////////////
236 /// @brief Computes the number of primitives given the number of verts.
237 /// @param mode - primitive topology for draw operation.
238 /// @param numPrims - number of vertices or indices for draw.
239 /// @todo Frontend needs to be refactored. This will go in appropriate place then.
240 uint32_t GetNumPrims(
241 PRIMITIVE_TOPOLOGY mode,
242 uint32_t numPrims)
243 {
244 switch (mode)
245 {
246 case TOP_POINT_LIST: return numPrims;
247 case TOP_TRIANGLE_LIST: return numPrims / 3;
248 case TOP_TRIANGLE_STRIP: return numPrims < 3 ? 0 : numPrims - 2;
249 case TOP_TRIANGLE_FAN: return numPrims < 3 ? 0 : numPrims - 2;
250 case TOP_TRIANGLE_DISC: return numPrims < 2 ? 0 : numPrims - 1;
251 case TOP_QUAD_LIST: return numPrims / 4;
252 case TOP_QUAD_STRIP: return numPrims < 4 ? 0 : (numPrims - 2) / 2;
253 case TOP_LINE_STRIP: return numPrims < 2 ? 0 : numPrims - 1;
254 case TOP_LINE_LIST: return numPrims / 2;
255 case TOP_LINE_LOOP: return numPrims;
256 case TOP_RECT_LIST: return numPrims / 3;
257 case TOP_LINE_LIST_ADJ: return numPrims / 4;
258 case TOP_LISTSTRIP_ADJ: return numPrims < 3 ? 0 : numPrims - 3;
259 case TOP_TRI_LIST_ADJ: return numPrims / 6;
260 case TOP_TRI_STRIP_ADJ: return numPrims < 4 ? 0 : (numPrims / 2) - 2;
261
262 case TOP_PATCHLIST_1:
263 case TOP_PATCHLIST_2:
264 case TOP_PATCHLIST_3:
265 case TOP_PATCHLIST_4:
266 case TOP_PATCHLIST_5:
267 case TOP_PATCHLIST_6:
268 case TOP_PATCHLIST_7:
269 case TOP_PATCHLIST_8:
270 case TOP_PATCHLIST_9:
271 case TOP_PATCHLIST_10:
272 case TOP_PATCHLIST_11:
273 case TOP_PATCHLIST_12:
274 case TOP_PATCHLIST_13:
275 case TOP_PATCHLIST_14:
276 case TOP_PATCHLIST_15:
277 case TOP_PATCHLIST_16:
278 case TOP_PATCHLIST_17:
279 case TOP_PATCHLIST_18:
280 case TOP_PATCHLIST_19:
281 case TOP_PATCHLIST_20:
282 case TOP_PATCHLIST_21:
283 case TOP_PATCHLIST_22:
284 case TOP_PATCHLIST_23:
285 case TOP_PATCHLIST_24:
286 case TOP_PATCHLIST_25:
287 case TOP_PATCHLIST_26:
288 case TOP_PATCHLIST_27:
289 case TOP_PATCHLIST_28:
290 case TOP_PATCHLIST_29:
291 case TOP_PATCHLIST_30:
292 case TOP_PATCHLIST_31:
293 case TOP_PATCHLIST_32:
294 return numPrims / (mode - TOP_PATCHLIST_BASE);
295
296 case TOP_POLYGON:
297 case TOP_POINT_LIST_BF:
298 case TOP_LINE_STRIP_CONT:
299 case TOP_LINE_STRIP_BF:
300 case TOP_LINE_STRIP_CONT_BF:
301 case TOP_TRIANGLE_FAN_NOSTIPPLE:
302 case TOP_TRI_STRIP_REVERSE:
303 case TOP_PATCHLIST_BASE:
304 case TOP_UNKNOWN:
305 SWR_INVALID("Unsupported topology: %d", mode);
306 return 0;
307 }
308
309 return 0;
310 }
311
312 //////////////////////////////////////////////////////////////////////////
313 /// @brief Computes the number of verts given the number of primitives.
314 /// @param mode - primitive topology for draw operation.
315 /// @param numPrims - number of primitives for draw.
316 uint32_t GetNumVerts(
317 PRIMITIVE_TOPOLOGY mode,
318 uint32_t numPrims)
319 {
320 switch (mode)
321 {
322 case TOP_POINT_LIST: return numPrims;
323 case TOP_TRIANGLE_LIST: return numPrims * 3;
324 case TOP_TRIANGLE_STRIP: return numPrims ? numPrims + 2 : 0;
325 case TOP_TRIANGLE_FAN: return numPrims ? numPrims + 2 : 0;
326 case TOP_TRIANGLE_DISC: return numPrims ? numPrims + 1 : 0;
327 case TOP_QUAD_LIST: return numPrims * 4;
328 case TOP_QUAD_STRIP: return numPrims ? numPrims * 2 + 2 : 0;
329 case TOP_LINE_STRIP: return numPrims ? numPrims + 1 : 0;
330 case TOP_LINE_LIST: return numPrims * 2;
331 case TOP_LINE_LOOP: return numPrims;
332 case TOP_RECT_LIST: return numPrims * 3;
333 case TOP_LINE_LIST_ADJ: return numPrims * 4;
334 case TOP_LISTSTRIP_ADJ: return numPrims ? numPrims + 3 : 0;
335 case TOP_TRI_LIST_ADJ: return numPrims * 6;
336 case TOP_TRI_STRIP_ADJ: return numPrims ? (numPrims + 2) * 2 : 0;
337
338 case TOP_PATCHLIST_1:
339 case TOP_PATCHLIST_2:
340 case TOP_PATCHLIST_3:
341 case TOP_PATCHLIST_4:
342 case TOP_PATCHLIST_5:
343 case TOP_PATCHLIST_6:
344 case TOP_PATCHLIST_7:
345 case TOP_PATCHLIST_8:
346 case TOP_PATCHLIST_9:
347 case TOP_PATCHLIST_10:
348 case TOP_PATCHLIST_11:
349 case TOP_PATCHLIST_12:
350 case TOP_PATCHLIST_13:
351 case TOP_PATCHLIST_14:
352 case TOP_PATCHLIST_15:
353 case TOP_PATCHLIST_16:
354 case TOP_PATCHLIST_17:
355 case TOP_PATCHLIST_18:
356 case TOP_PATCHLIST_19:
357 case TOP_PATCHLIST_20:
358 case TOP_PATCHLIST_21:
359 case TOP_PATCHLIST_22:
360 case TOP_PATCHLIST_23:
361 case TOP_PATCHLIST_24:
362 case TOP_PATCHLIST_25:
363 case TOP_PATCHLIST_26:
364 case TOP_PATCHLIST_27:
365 case TOP_PATCHLIST_28:
366 case TOP_PATCHLIST_29:
367 case TOP_PATCHLIST_30:
368 case TOP_PATCHLIST_31:
369 case TOP_PATCHLIST_32:
370 return numPrims * (mode - TOP_PATCHLIST_BASE);
371
372 case TOP_POLYGON:
373 case TOP_POINT_LIST_BF:
374 case TOP_LINE_STRIP_CONT:
375 case TOP_LINE_STRIP_BF:
376 case TOP_LINE_STRIP_CONT_BF:
377 case TOP_TRIANGLE_FAN_NOSTIPPLE:
378 case TOP_TRI_STRIP_REVERSE:
379 case TOP_PATCHLIST_BASE:
380 case TOP_UNKNOWN:
381 SWR_INVALID("Unsupported topology: %d", mode);
382 return 0;
383 }
384
385 return 0;
386 }
387
388 //////////////////////////////////////////////////////////////////////////
389 /// @brief Return number of verts per primitive.
390 /// @param topology - topology
391 /// @param includeAdjVerts - include adjacent verts in primitive vertices
392 INLINE uint32_t NumVertsPerPrim(PRIMITIVE_TOPOLOGY topology, bool includeAdjVerts)
393 {
394 uint32_t numVerts = 0;
395 switch (topology)
396 {
397 case TOP_POINT_LIST:
398 case TOP_POINT_LIST_BF:
399 numVerts = 1;
400 break;
401 case TOP_LINE_LIST:
402 case TOP_LINE_STRIP:
403 case TOP_LINE_LIST_ADJ:
404 case TOP_LINE_LOOP:
405 case TOP_LINE_STRIP_CONT:
406 case TOP_LINE_STRIP_BF:
407 case TOP_LISTSTRIP_ADJ:
408 numVerts = 2;
409 break;
410 case TOP_TRIANGLE_LIST:
411 case TOP_TRIANGLE_STRIP:
412 case TOP_TRIANGLE_FAN:
413 case TOP_TRI_LIST_ADJ:
414 case TOP_TRI_STRIP_ADJ:
415 case TOP_TRI_STRIP_REVERSE:
416 case TOP_RECT_LIST:
417 numVerts = 3;
418 break;
419 case TOP_QUAD_LIST:
420 case TOP_QUAD_STRIP:
421 numVerts = 4;
422 break;
423 case TOP_PATCHLIST_1:
424 case TOP_PATCHLIST_2:
425 case TOP_PATCHLIST_3:
426 case TOP_PATCHLIST_4:
427 case TOP_PATCHLIST_5:
428 case TOP_PATCHLIST_6:
429 case TOP_PATCHLIST_7:
430 case TOP_PATCHLIST_8:
431 case TOP_PATCHLIST_9:
432 case TOP_PATCHLIST_10:
433 case TOP_PATCHLIST_11:
434 case TOP_PATCHLIST_12:
435 case TOP_PATCHLIST_13:
436 case TOP_PATCHLIST_14:
437 case TOP_PATCHLIST_15:
438 case TOP_PATCHLIST_16:
439 case TOP_PATCHLIST_17:
440 case TOP_PATCHLIST_18:
441 case TOP_PATCHLIST_19:
442 case TOP_PATCHLIST_20:
443 case TOP_PATCHLIST_21:
444 case TOP_PATCHLIST_22:
445 case TOP_PATCHLIST_23:
446 case TOP_PATCHLIST_24:
447 case TOP_PATCHLIST_25:
448 case TOP_PATCHLIST_26:
449 case TOP_PATCHLIST_27:
450 case TOP_PATCHLIST_28:
451 case TOP_PATCHLIST_29:
452 case TOP_PATCHLIST_30:
453 case TOP_PATCHLIST_31:
454 case TOP_PATCHLIST_32:
455 numVerts = topology - TOP_PATCHLIST_BASE;
456 break;
457 default:
458 SWR_INVALID("Unsupported topology: %d", topology);
459 break;
460 }
461
462 if (includeAdjVerts)
463 {
464 switch (topology)
465 {
466 case TOP_LISTSTRIP_ADJ:
467 case TOP_LINE_LIST_ADJ: numVerts = 4; break;
468 case TOP_TRI_STRIP_ADJ:
469 case TOP_TRI_LIST_ADJ: numVerts = 6; break;
470 default: break;
471 }
472 }
473
474 return numVerts;
475 }
476
477 //////////////////////////////////////////////////////////////////////////
478 /// @brief Generate mask from remaining work.
479 /// @param numWorkItems - Number of items being worked on by a SIMD.
480 static INLINE simdscalari GenerateMask(uint32_t numItemsRemaining)
481 {
482 uint32_t numActive = (numItemsRemaining >= KNOB_SIMD_WIDTH) ? KNOB_SIMD_WIDTH : numItemsRemaining;
483 uint32_t mask = (numActive > 0) ? ((1 << numActive) - 1) : 0;
484 return _simd_castps_si(vMask(mask));
485 }
486
487 //////////////////////////////////////////////////////////////////////////
488 /// @brief StreamOut - Streams vertex data out to SO buffers.
489 /// Generally, we are only streaming out a SIMDs worth of triangles.
490 /// @param pDC - pointer to draw context.
491 /// @param workerId - thread's worker id. Even thread has a unique id.
492 /// @param numPrims - Number of prims to streamout (e.g. points, lines, tris)
493 static void StreamOut(
494 DRAW_CONTEXT* pDC,
495 PA_STATE& pa,
496 uint32_t workerId,
497 uint32_t* pPrimData,
498 uint32_t streamIndex)
499 {
500 SWR_CONTEXT *pContext = pDC->pContext;
501
502 AR_BEGIN(FEStreamout, pDC->drawId);
503
504 const API_STATE& state = GetApiState(pDC);
505 const SWR_STREAMOUT_STATE &soState = state.soState;
506
507 uint32_t soVertsPerPrim = NumVertsPerPrim(pa.binTopology, false);
508
509 // The pPrimData buffer is sparse in that we allocate memory for all 32 attributes for each vertex.
510 uint32_t primDataDwordVertexStride = (SWR_VTX_NUM_SLOTS * sizeof(float) * 4) / sizeof(uint32_t);
511
512 SWR_STREAMOUT_CONTEXT soContext = { 0 };
513
514 // Setup buffer state pointers.
515 for (uint32_t i = 0; i < 4; ++i)
516 {
517 soContext.pBuffer[i] = &state.soBuffer[i];
518 }
519
520 uint32_t numPrims = pa.NumPrims();
521
522 for (uint32_t primIndex = 0; primIndex < numPrims; ++primIndex)
523 {
524 DWORD slot = 0;
525 uint32_t soMask = soState.streamMasks[streamIndex];
526
527 // Write all entries into primitive data buffer for SOS.
528 while (_BitScanForward(&slot, soMask))
529 {
530 __m128 attrib[MAX_NUM_VERTS_PER_PRIM]; // prim attribs (always 4 wide)
531 uint32_t paSlot = slot + VERTEX_ATTRIB_START_SLOT;
532 pa.AssembleSingle(paSlot, primIndex, attrib);
533
534 // Attribute offset is relative offset from start of vertex.
535 // Note that attributes start at slot 1 in the PA buffer. We need to write this
536 // to prim data starting at slot 0. Which is why we do (slot - 1).
537 // Also note: GL works slightly differently, and needs slot 0
538 uint32_t primDataAttribOffset = slot * sizeof(float) * 4 / sizeof(uint32_t);
539
540 // Store each vertex's attrib at appropriate locations in pPrimData buffer.
541 for (uint32_t v = 0; v < soVertsPerPrim; ++v)
542 {
543 uint32_t* pPrimDataAttrib = pPrimData + primDataAttribOffset + (v * primDataDwordVertexStride);
544
545 _mm_store_ps((float*)pPrimDataAttrib, attrib[v]);
546 }
547
548 soMask &= ~(1 << slot);
549 }
550
551 // Update pPrimData pointer
552 soContext.pPrimData = pPrimData;
553
554 // Call SOS
555 SWR_ASSERT(state.pfnSoFunc[streamIndex] != nullptr, "Trying to execute uninitialized streamout jit function.");
556 state.pfnSoFunc[streamIndex](soContext);
557 }
558
559 // Update SO write offset. The driver provides memory for the update.
560 for (uint32_t i = 0; i < 4; ++i)
561 {
562 if (state.soBuffer[i].pWriteOffset)
563 {
564 *state.soBuffer[i].pWriteOffset = soContext.pBuffer[i]->streamOffset * sizeof(uint32_t);
565 }
566
567 if (state.soBuffer[i].soWriteEnable)
568 {
569 pDC->dynState.SoWriteOffset[i] = soContext.pBuffer[i]->streamOffset * sizeof(uint32_t);
570 pDC->dynState.SoWriteOffsetDirty[i] = true;
571 }
572 }
573
574 UPDATE_STAT_FE(SoPrimStorageNeeded[streamIndex], soContext.numPrimStorageNeeded);
575 UPDATE_STAT_FE(SoNumPrimsWritten[streamIndex], soContext.numPrimsWritten);
576
577 AR_END(FEStreamout, 1);
578 }
579
580 #if USE_SIMD16_FRONTEND
581 //////////////////////////////////////////////////////////////////////////
582 /// Is value an even number (a multiple of two)
583 ///
584 template <typename T>
585 INLINE static bool IsEven(T value)
586 {
587 return (value & 1) == 0;
588 }
589
590 //////////////////////////////////////////////////////////////////////////
591 /// Round up value to an even number (a multiple of two)
592 ///
593 template <typename T>
594 INLINE static T RoundUpEven(T value)
595 {
596 return (value + 1) & ~1;
597 }
598
599 //////////////////////////////////////////////////////////////////////////
600 /// Round down value to an even number (a multiple of two)
601 ///
602 template <typename T>
603 INLINE static T RoundDownEven(T value)
604 {
605 return value & ~1;
606 }
607
608 //////////////////////////////////////////////////////////////////////////
609 /// Pack pairs of simdvertexes into simd16vertexes, assume non-overlapping
610 ///
611 /// vertexCount is in terms of the source simdvertexes and must be even
612 ///
613 /// attribCount will limit the vector copies to those attribs specified
614 ///
615 /// note: the stride between vertexes is determinded by SWR_VTX_NUM_SLOTS
616 ///
617 void PackPairsOfSimdVertexIntoSimd16Vertex(simd16vertex *vertex_simd16, const simdvertex *vertex, uint32_t vertexCount, uint32_t attribCount)
618 {
619 SWR_ASSERT(vertex);
620 SWR_ASSERT(vertex_simd16);
621 SWR_ASSERT(attribCount <= SWR_VTX_NUM_SLOTS);
622
623 simd16vertex temp;
624
625 for (uint32_t i = 0; i < vertexCount; i += 2)
626 {
627 for (uint32_t j = 0; j < attribCount; j += 1)
628 {
629 for (uint32_t k = 0; k < 4; k += 1)
630 {
631 temp.attrib[j][k] = _simd16_insert_ps(_simd16_setzero_ps(), vertex[i].attrib[j][k], 0);
632
633 if ((i + 1) < vertexCount)
634 {
635 temp.attrib[j][k] = _simd16_insert_ps(temp.attrib[j][k], vertex[i + 1].attrib[j][k], 1);
636 }
637 }
638 }
639
640 for (uint32_t j = 0; j < attribCount; j += 1)
641 {
642 vertex_simd16[i >> 1].attrib[j] = temp.attrib[j];
643 }
644 }
645 }
646
647 #endif
648 //////////////////////////////////////////////////////////////////////////
649 /// @brief Computes number of invocations. The current index represents
650 /// the start of the SIMD. The max index represents how much work
651 /// items are remaining. If there is less then a SIMD's xmin of work
652 /// then return the remaining amount of work.
653 /// @param curIndex - The start index for the SIMD.
654 /// @param maxIndex - The last index for all work items.
655 static INLINE uint32_t GetNumInvocations(
656 uint32_t curIndex,
657 uint32_t maxIndex)
658 {
659 uint32_t remainder = (maxIndex - curIndex);
660 #if USE_SIMD16_FRONTEND
661 return (remainder >= KNOB_SIMD16_WIDTH) ? KNOB_SIMD16_WIDTH : remainder;
662 #else
663 return (remainder >= KNOB_SIMD_WIDTH) ? KNOB_SIMD_WIDTH : remainder;
664 #endif
665 }
666
667 //////////////////////////////////////////////////////////////////////////
668 /// @brief Converts a streamId buffer to a cut buffer for the given stream id.
669 /// The geometry shader will loop over each active streamout buffer, assembling
670 /// primitives for the downstream stages. When multistream output is enabled,
671 /// the generated stream ID buffer from the GS needs to be converted to a cut
672 /// buffer for the primitive assembler.
673 /// @param stream - stream id to generate the cut buffer for
674 /// @param pStreamIdBase - pointer to the stream ID buffer
675 /// @param numEmittedVerts - Number of total verts emitted by the GS
676 /// @param pCutBuffer - output buffer to write cuts to
677 void ProcessStreamIdBuffer(uint32_t stream, uint8_t* pStreamIdBase, uint32_t numEmittedVerts, uint8_t *pCutBuffer)
678 {
679 SWR_ASSERT(stream < MAX_SO_STREAMS);
680
681 uint32_t numInputBytes = (numEmittedVerts * 2 + 7) / 8;
682 uint32_t numOutputBytes = std::max(numInputBytes / 2, 1U);
683
684 for (uint32_t b = 0; b < numOutputBytes; ++b)
685 {
686 uint8_t curInputByte = pStreamIdBase[2*b];
687 uint8_t outByte = 0;
688 for (uint32_t i = 0; i < 4; ++i)
689 {
690 if ((curInputByte & 0x3) != stream)
691 {
692 outByte |= (1 << i);
693 }
694 curInputByte >>= 2;
695 }
696
697 curInputByte = pStreamIdBase[2 * b + 1];
698 for (uint32_t i = 0; i < 4; ++i)
699 {
700 if ((curInputByte & 0x3) != stream)
701 {
702 outByte |= (1 << (i + 4));
703 }
704 curInputByte >>= 2;
705 }
706
707 *pCutBuffer++ = outByte;
708 }
709 }
710
711 THREAD SWR_GS_CONTEXT tlsGsContext;
712
713 template<typename SIMDVERTEX, uint32_t SIMD_WIDTH>
714 struct GsBufferInfo
715 {
716 GsBufferInfo(const SWR_GS_STATE &gsState)
717 {
718 const uint32_t vertexCount = gsState.maxNumVerts;
719 const uint32_t vertexStride = sizeof(SIMDVERTEX);
720 const uint32_t numSimdBatches = (vertexCount + SIMD_WIDTH - 1) / SIMD_WIDTH;
721
722 vertexPrimitiveStride = vertexStride * numSimdBatches;
723 vertexInstanceStride = vertexPrimitiveStride * SIMD_WIDTH;
724
725 if (gsState.isSingleStream)
726 {
727 cutPrimitiveStride = (vertexCount + 7) / 8;
728 cutInstanceStride = cutPrimitiveStride * SIMD_WIDTH;
729
730 streamCutPrimitiveStride = 0;
731 streamCutInstanceStride = 0;
732 }
733 else
734 {
735 cutPrimitiveStride = AlignUp(vertexCount * 2 / 8, 4);
736 cutInstanceStride = cutPrimitiveStride * SIMD_WIDTH;
737
738 streamCutPrimitiveStride = (vertexCount + 7) / 8;
739 streamCutInstanceStride = streamCutPrimitiveStride * SIMD_WIDTH;
740 }
741 }
742
743 uint32_t vertexPrimitiveStride;
744 uint32_t vertexInstanceStride;
745
746 uint32_t cutPrimitiveStride;
747 uint32_t cutInstanceStride;
748
749 uint32_t streamCutPrimitiveStride;
750 uint32_t streamCutInstanceStride;
751 };
752
753 //////////////////////////////////////////////////////////////////////////
754 /// @brief Implements GS stage.
755 /// @param pDC - pointer to draw context.
756 /// @param workerId - thread's worker id. Even thread has a unique id.
757 /// @param pa - The primitive assembly object.
758 /// @param pGsOut - output stream for GS
759 template <
760 typename HasStreamOutT,
761 typename HasRastT>
762 static void GeometryShaderStage(
763 DRAW_CONTEXT *pDC,
764 uint32_t workerId,
765 PA_STATE& pa,
766 void* pGsOut,
767 void* pCutBuffer,
768 void* pStreamCutBuffer,
769 uint32_t* pSoPrimData,
770 #if USE_SIMD16_FRONTEND
771 uint32_t numPrims_simd8,
772 #endif
773 simdscalari primID)
774 {
775 SWR_CONTEXT *pContext = pDC->pContext;
776
777 AR_BEGIN(FEGeometryShader, pDC->drawId);
778
779 const API_STATE& state = GetApiState(pDC);
780 const SWR_GS_STATE* pState = &state.gsState;
781
782 SWR_ASSERT(pGsOut != nullptr, "GS output buffer should be initialized");
783 SWR_ASSERT(pCutBuffer != nullptr, "GS output cut buffer should be initialized");
784
785 tlsGsContext.pStream = (uint8_t*)pGsOut;
786 tlsGsContext.pCutOrStreamIdBuffer = (uint8_t*)pCutBuffer;
787 tlsGsContext.PrimitiveID = primID;
788
789 uint32_t numVertsPerPrim = NumVertsPerPrim(pa.binTopology, true);
790 simdvector attrib[MAX_ATTRIBUTES];
791
792 // assemble all attributes for the input primitive
793 for (uint32_t slot = 0; slot < pState->numInputAttribs; ++slot)
794 {
795 uint32_t attribSlot = VERTEX_ATTRIB_START_SLOT + slot;
796 pa.Assemble(attribSlot, attrib);
797
798 for (uint32_t i = 0; i < numVertsPerPrim; ++i)
799 {
800 tlsGsContext.vert[i].attrib[attribSlot] = attrib[i];
801 }
802 }
803
804 // assemble position
805 pa.Assemble(VERTEX_POSITION_SLOT, attrib);
806 for (uint32_t i = 0; i < numVertsPerPrim; ++i)
807 {
808 tlsGsContext.vert[i].attrib[VERTEX_POSITION_SLOT] = attrib[i];
809 }
810
811 #if USE_SIMD16_FRONTEND
812 const GsBufferInfo<simd16vertex, KNOB_SIMD16_WIDTH> bufferInfo(state.gsState);
813 #else
814 const GsBufferInfo<simdvertex, KNOB_SIMD_WIDTH> bufferInfo(state.gsState);
815 #endif
816
817 // record valid prims from the frontend to avoid over binning the newly generated
818 // prims from the GS
819 #if USE_SIMD16_FRONTEND
820 uint32_t numInputPrims = numPrims_simd8;
821 #else
822 uint32_t numInputPrims = pa.NumPrims();
823 #endif
824
825 for (uint32_t instance = 0; instance < pState->instanceCount; ++instance)
826 {
827 tlsGsContext.InstanceID = instance;
828 tlsGsContext.mask = GenerateMask(numInputPrims);
829
830 // execute the geometry shader
831 state.pfnGsFunc(GetPrivateState(pDC), &tlsGsContext);
832
833 tlsGsContext.pStream += bufferInfo.vertexInstanceStride;
834 tlsGsContext.pCutOrStreamIdBuffer += bufferInfo.cutInstanceStride;
835 }
836
837 // set up new binner and state for the GS output topology
838 #if USE_SIMD16_FRONTEND
839 PFN_PROCESS_PRIMS_SIMD16 pfnClipFunc = nullptr;
840 if (HasRastT::value)
841 {
842 switch (pState->outputTopology)
843 {
844 case TOP_TRIANGLE_STRIP: pfnClipFunc = ClipTriangles_simd16; break;
845 case TOP_LINE_STRIP: pfnClipFunc = ClipLines_simd16; break;
846 case TOP_POINT_LIST: pfnClipFunc = ClipPoints_simd16; break;
847 default: SWR_INVALID("Unexpected GS output topology: %d", pState->outputTopology);
848 }
849 }
850
851 #else
852 PFN_PROCESS_PRIMS pfnClipFunc = nullptr;
853 if (HasRastT::value)
854 {
855 switch (pState->outputTopology)
856 {
857 case TOP_TRIANGLE_STRIP: pfnClipFunc = ClipTriangles; break;
858 case TOP_LINE_STRIP: pfnClipFunc = ClipLines; break;
859 case TOP_POINT_LIST: pfnClipFunc = ClipPoints; break;
860 default: SWR_INVALID("Unexpected GS output topology: %d", pState->outputTopology);
861 }
862 }
863
864 #endif
865 // foreach input prim:
866 // - setup a new PA based on the emitted verts for that prim
867 // - loop over the new verts, calling PA to assemble each prim
868 uint32_t* pVertexCount = (uint32_t*)&tlsGsContext.vertexCount;
869 uint32_t* pPrimitiveId = (uint32_t*)&primID;
870
871 uint32_t totalPrimsGenerated = 0;
872 for (uint32_t inputPrim = 0; inputPrim < numInputPrims; ++inputPrim)
873 {
874 uint8_t* pInstanceBase = (uint8_t*)pGsOut + inputPrim * bufferInfo.vertexPrimitiveStride;
875 uint8_t* pCutBufferBase = (uint8_t*)pCutBuffer + inputPrim * bufferInfo.cutPrimitiveStride;
876
877 for (uint32_t instance = 0; instance < pState->instanceCount; ++instance)
878 {
879 uint32_t numEmittedVerts = pVertexCount[inputPrim];
880 if (numEmittedVerts == 0)
881 {
882 continue;
883 }
884
885 uint8_t* pBase = pInstanceBase + instance * bufferInfo.vertexInstanceStride;
886 uint8_t* pCutBase = pCutBufferBase + instance * bufferInfo.cutInstanceStride;
887
888 uint32_t numAttribs = state.feNumAttributes;
889
890 for (uint32_t stream = 0; stream < MAX_SO_STREAMS; ++stream)
891 {
892 bool processCutVerts = false;
893
894 uint8_t* pCutBuffer = pCutBase;
895
896 // assign default stream ID, only relevant when GS is outputting a single stream
897 uint32_t streamID = 0;
898 if (pState->isSingleStream)
899 {
900 processCutVerts = true;
901 streamID = pState->singleStreamID;
902 if (streamID != stream) continue;
903 }
904 else
905 {
906 // early exit if this stream is not enabled for streamout
907 if (HasStreamOutT::value && !state.soState.streamEnable[stream])
908 {
909 continue;
910 }
911
912 // multi-stream output, need to translate StreamID buffer to a cut buffer
913 ProcessStreamIdBuffer(stream, pCutBase, numEmittedVerts, (uint8_t*)pStreamCutBuffer);
914 pCutBuffer = (uint8_t*)pStreamCutBuffer;
915 processCutVerts = false;
916 }
917
918 #if USE_SIMD16_FRONTEND
919 PA_STATE_CUT gsPa(pDC, pBase, numEmittedVerts, reinterpret_cast<simd16mask *>(pCutBuffer), numEmittedVerts, numAttribs, pState->outputTopology, processCutVerts);
920
921 #else
922 PA_STATE_CUT gsPa(pDC, pBase, numEmittedVerts, pCutBuffer, numEmittedVerts, numAttribs, pState->outputTopology, processCutVerts);
923
924 #endif
925 while (gsPa.GetNextStreamOutput())
926 {
927 do
928 {
929 #if USE_SIMD16_FRONTEND
930 simd16vector attrib_simd16[3];
931
932 bool assemble = gsPa.Assemble_simd16(VERTEX_POSITION_SLOT, attrib_simd16);
933
934 #else
935 bool assemble = gsPa.Assemble(VERTEX_POSITION_SLOT, attrib);
936
937 #endif
938 if (assemble)
939 {
940 totalPrimsGenerated += gsPa.NumPrims();
941
942 if (HasStreamOutT::value)
943 {
944 gsPa.useAlternateOffset = false;
945 StreamOut(pDC, gsPa, workerId, pSoPrimData, stream);
946 }
947
948 if (HasRastT::value && state.soState.streamToRasterizer == stream)
949 {
950 #if USE_SIMD16_FRONTEND
951 simd16scalari vPrimId = _simd16_set1_epi32(pPrimitiveId[inputPrim]);
952
953 // use viewport array index if GS declares it as an output attribute. Otherwise use index 0.
954 simd16scalari vViewPortIdx;
955 if (state.gsState.emitsViewportArrayIndex)
956 {
957 simd16vector vpiAttrib[3];
958 gsPa.Assemble_simd16(VERTEX_SGV_SLOT, vpiAttrib);
959
960 // OOB indices => forced to zero.
961 simd16scalari vpai = _simd16_castps_si(vpiAttrib[0][VERTEX_SGV_VAI_COMP]);
962 simd16scalari vNumViewports = _simd16_set1_epi32(KNOB_NUM_VIEWPORTS_SCISSORS);
963 simd16scalari vClearMask = _simd16_cmplt_epi32(vpai, vNumViewports);
964 vViewPortIdx = _simd16_and_si(vClearMask, vpai);
965 }
966 else
967 {
968 vViewPortIdx = _simd16_set1_epi32(0);
969 }
970
971 gsPa.useAlternateOffset = false;
972 pfnClipFunc(pDC, gsPa, workerId, attrib_simd16, GenMask(gsPa.NumPrims()), vPrimId, vViewPortIdx);
973 #else
974 simdscalari vPrimId = _simd_set1_epi32(pPrimitiveId[inputPrim]);
975
976 // use viewport array index if GS declares it as an output attribute. Otherwise use index 0.
977 simdscalari vViewPortIdx;
978 if (state.gsState.emitsViewportArrayIndex)
979 {
980 simdvector vpiAttrib[3];
981 gsPa.Assemble(VERTEX_SGV_SLOT, vpiAttrib);
982 simdscalari vpai = _simd_castps_si(vpiAttrib[0][VERTEX_SGV_VAI_COMP]);
983
984 // OOB indices => forced to zero.
985 simdscalari vNumViewports = _simd_set1_epi32(KNOB_NUM_VIEWPORTS_SCISSORS);
986 simdscalari vClearMask = _simd_cmplt_epi32(vpai, vNumViewports);
987 vViewPortIdx = _simd_and_si(vClearMask, vpai);
988 }
989 else
990 {
991 vViewPortIdx = _simd_set1_epi32(0);
992 }
993
994 pfnClipFunc(pDC, gsPa, workerId, attrib, GenMask(gsPa.NumPrims()), vPrimId, vViewPortIdx);
995 #endif
996 }
997 }
998 } while (gsPa.NextPrim());
999 }
1000 }
1001 }
1002 }
1003
1004 // update GS pipeline stats
1005 UPDATE_STAT_FE(GsInvocations, numInputPrims * pState->instanceCount);
1006 UPDATE_STAT_FE(GsPrimitives, totalPrimsGenerated);
1007 AR_EVENT(GSPrimInfo(numInputPrims, totalPrimsGenerated, numVertsPerPrim*numInputPrims));
1008 AR_END(FEGeometryShader, 1);
1009 }
1010
1011 //////////////////////////////////////////////////////////////////////////
1012 /// @brief Allocate GS buffers
1013 /// @param pDC - pointer to draw context.
1014 /// @param state - API state
1015 /// @param ppGsOut - pointer to GS output buffer allocation
1016 /// @param ppCutBuffer - pointer to GS output cut buffer allocation
1017 template<typename SIMDVERTEX, uint32_t SIMD_WIDTH>
1018 static INLINE void AllocateGsBuffers(DRAW_CONTEXT* pDC, const API_STATE& state, void** ppGsOut, void** ppCutBuffer,
1019 void **ppStreamCutBuffer)
1020 {
1021 auto pArena = pDC->pArena;
1022 SWR_ASSERT(pArena != nullptr);
1023 SWR_ASSERT(state.gsState.gsEnable);
1024
1025 // allocate arena space to hold GS output verts
1026 // @todo pack attribs
1027 // @todo support multiple streams
1028
1029 const GsBufferInfo<SIMDVERTEX, SIMD_WIDTH> bufferInfo(state.gsState);
1030
1031 const uint32_t vertexBufferSize = state.gsState.instanceCount * bufferInfo.vertexInstanceStride;
1032
1033 *ppGsOut = pArena->AllocAligned(vertexBufferSize, SIMD_WIDTH * sizeof(float));
1034
1035 // allocate arena space to hold cut or streamid buffer, which is essentially a bitfield sized to the
1036 // maximum vertex output as defined by the GS state, per SIMD lane, per GS instance
1037
1038 // allocate space for temporary per-stream cut buffer if multi-stream is enabled
1039 if (state.gsState.isSingleStream)
1040 {
1041 const uint32_t cutBufferSize = state.gsState.instanceCount * bufferInfo.cutInstanceStride;
1042
1043 *ppCutBuffer = pArena->AllocAligned(cutBufferSize, SIMD_WIDTH * sizeof(float));
1044 *ppStreamCutBuffer = nullptr;
1045 }
1046 else
1047 {
1048 const uint32_t cutBufferSize = state.gsState.instanceCount * bufferInfo.cutInstanceStride;
1049 const uint32_t streamCutBufferSize = state.gsState.instanceCount * bufferInfo.streamCutInstanceStride;
1050
1051 *ppCutBuffer = pArena->AllocAligned(cutBufferSize, SIMD_WIDTH * sizeof(float));
1052 *ppStreamCutBuffer = pArena->AllocAligned(streamCutBufferSize, SIMD_WIDTH * sizeof(float));
1053 }
1054 }
1055
1056 //////////////////////////////////////////////////////////////////////////
1057 /// @brief Contains all data generated by the HS and passed to the
1058 /// tessellator and DS.
1059 struct TessellationThreadLocalData
1060 {
1061 SWR_HS_CONTEXT hsContext;
1062 ScalarPatch patchData[KNOB_SIMD_WIDTH];
1063 void* pTxCtx;
1064 size_t tsCtxSize;
1065
1066 simdscalar* pDSOutput;
1067 size_t numDSOutputVectors;
1068 };
1069
1070 THREAD TessellationThreadLocalData* gt_pTessellationThreadData = nullptr;
1071
1072 //////////////////////////////////////////////////////////////////////////
1073 /// @brief Allocate tessellation data for this worker thread.
1074 INLINE
1075 static void AllocateTessellationData(SWR_CONTEXT* pContext)
1076 {
1077 /// @TODO - Don't use thread local storage. Use Worker local storage instead.
1078 if (gt_pTessellationThreadData == nullptr)
1079 {
1080 gt_pTessellationThreadData = (TessellationThreadLocalData*)
1081 AlignedMalloc(sizeof(TessellationThreadLocalData), 64);
1082 memset(gt_pTessellationThreadData, 0, sizeof(*gt_pTessellationThreadData));
1083 }
1084 }
1085
1086 //////////////////////////////////////////////////////////////////////////
1087 /// @brief Implements Tessellation Stages.
1088 /// @param pDC - pointer to draw context.
1089 /// @param workerId - thread's worker id. Even thread has a unique id.
1090 /// @param pa - The primitive assembly object.
1091 /// @param pGsOut - output stream for GS
1092 template <
1093 typename HasGeometryShaderT,
1094 typename HasStreamOutT,
1095 typename HasRastT>
1096 static void TessellationStages(
1097 DRAW_CONTEXT *pDC,
1098 uint32_t workerId,
1099 PA_STATE& pa,
1100 void* pGsOut,
1101 void* pCutBuffer,
1102 void* pCutStreamBuffer,
1103 uint32_t* pSoPrimData,
1104 #if USE_SIMD16_FRONTEND
1105 uint32_t numPrims_simd8,
1106 #endif
1107 simdscalari primID)
1108 {
1109 SWR_CONTEXT *pContext = pDC->pContext;
1110 const API_STATE& state = GetApiState(pDC);
1111 const SWR_TS_STATE& tsState = state.tsState;
1112
1113 SWR_ASSERT(gt_pTessellationThreadData);
1114
1115 HANDLE tsCtx = TSInitCtx(
1116 tsState.domain,
1117 tsState.partitioning,
1118 tsState.tsOutputTopology,
1119 gt_pTessellationThreadData->pTxCtx,
1120 gt_pTessellationThreadData->tsCtxSize);
1121 if (tsCtx == nullptr)
1122 {
1123 gt_pTessellationThreadData->pTxCtx = AlignedMalloc(gt_pTessellationThreadData->tsCtxSize, 64);
1124 tsCtx = TSInitCtx(
1125 tsState.domain,
1126 tsState.partitioning,
1127 tsState.tsOutputTopology,
1128 gt_pTessellationThreadData->pTxCtx,
1129 gt_pTessellationThreadData->tsCtxSize);
1130 }
1131 SWR_ASSERT(tsCtx);
1132
1133 #if USE_SIMD16_FRONTEND
1134 PFN_PROCESS_PRIMS_SIMD16 pfnClipFunc = nullptr;
1135 if (HasRastT::value)
1136 {
1137 switch (tsState.postDSTopology)
1138 {
1139 case TOP_TRIANGLE_LIST: pfnClipFunc = ClipTriangles_simd16; break;
1140 case TOP_LINE_LIST: pfnClipFunc = ClipLines_simd16; break;
1141 case TOP_POINT_LIST: pfnClipFunc = ClipPoints_simd16; break;
1142 default: SWR_INVALID("Unexpected DS output topology: %d", tsState.postDSTopology);
1143 }
1144 }
1145
1146 #else
1147 PFN_PROCESS_PRIMS pfnClipFunc = nullptr;
1148 if (HasRastT::value)
1149 {
1150 switch (tsState.postDSTopology)
1151 {
1152 case TOP_TRIANGLE_LIST: pfnClipFunc = ClipTriangles; break;
1153 case TOP_LINE_LIST: pfnClipFunc = ClipLines; break;
1154 case TOP_POINT_LIST: pfnClipFunc = ClipPoints; break;
1155 default: SWR_INVALID("Unexpected DS output topology: %d", tsState.postDSTopology);
1156 }
1157 }
1158
1159 #endif
1160 SWR_HS_CONTEXT& hsContext = gt_pTessellationThreadData->hsContext;
1161 hsContext.pCPout = gt_pTessellationThreadData->patchData;
1162 hsContext.PrimitiveID = primID;
1163
1164 uint32_t numVertsPerPrim = NumVertsPerPrim(pa.binTopology, false);
1165 // Max storage for one attribute for an entire simdprimitive
1166 simdvector simdattrib[MAX_NUM_VERTS_PER_PRIM];
1167
1168 // assemble all attributes for the input primitives
1169 for (uint32_t slot = 0; slot < tsState.numHsInputAttribs; ++slot)
1170 {
1171 uint32_t attribSlot = VERTEX_ATTRIB_START_SLOT + slot;
1172 pa.Assemble(attribSlot, simdattrib);
1173
1174 for (uint32_t i = 0; i < numVertsPerPrim; ++i)
1175 {
1176 hsContext.vert[i].attrib[attribSlot] = simdattrib[i];
1177 }
1178 }
1179
1180 #if defined(_DEBUG)
1181 memset(hsContext.pCPout, 0x90, sizeof(ScalarPatch) * KNOB_SIMD_WIDTH);
1182 #endif
1183
1184 #if USE_SIMD16_FRONTEND
1185 uint32_t numPrims = numPrims_simd8;
1186 #else
1187 uint32_t numPrims = pa.NumPrims();
1188 #endif
1189 hsContext.mask = GenerateMask(numPrims);
1190
1191 // Run the HS
1192 AR_BEGIN(FEHullShader, pDC->drawId);
1193 state.pfnHsFunc(GetPrivateState(pDC), &hsContext);
1194 AR_END(FEHullShader, 0);
1195
1196 UPDATE_STAT_FE(HsInvocations, numPrims);
1197
1198 const uint32_t* pPrimId = (const uint32_t*)&primID;
1199
1200 for (uint32_t p = 0; p < numPrims; ++p)
1201 {
1202 // Run Tessellator
1203 SWR_TS_TESSELLATED_DATA tsData = { 0 };
1204 AR_BEGIN(FETessellation, pDC->drawId);
1205 TSTessellate(tsCtx, hsContext.pCPout[p].tessFactors, tsData);
1206 AR_EVENT(TessPrimCount(1));
1207 AR_END(FETessellation, 0);
1208
1209 if (tsData.NumPrimitives == 0)
1210 {
1211 continue;
1212 }
1213 SWR_ASSERT(tsData.NumDomainPoints);
1214
1215 // Allocate DS Output memory
1216 uint32_t requiredDSVectorInvocations = AlignUp(tsData.NumDomainPoints, KNOB_SIMD_WIDTH) / KNOB_SIMD_WIDTH;
1217 size_t requiredDSOutputVectors = requiredDSVectorInvocations * tsState.numDsOutputAttribs;
1218 #if USE_SIMD16_FRONTEND
1219 size_t requiredAllocSize = sizeof(simdvector) * RoundUpEven(requiredDSVectorInvocations) * tsState.numDsOutputAttribs; // simd8 -> simd16, padding
1220 #else
1221 size_t requiredAllocSize = sizeof(simdvector) * requiredDSOutputVectors;
1222 #endif
1223 if (requiredDSOutputVectors > gt_pTessellationThreadData->numDSOutputVectors)
1224 {
1225 AlignedFree(gt_pTessellationThreadData->pDSOutput);
1226 gt_pTessellationThreadData->pDSOutput = (simdscalar*)AlignedMalloc(requiredAllocSize, 64);
1227 #if USE_SIMD16_FRONTEND
1228 gt_pTessellationThreadData->numDSOutputVectors = RoundUpEven(requiredDSVectorInvocations) * tsState.numDsOutputAttribs; // simd8 -> simd16, padding
1229 #else
1230 gt_pTessellationThreadData->numDSOutputVectors = requiredDSOutputVectors;
1231 #endif
1232 }
1233 SWR_ASSERT(gt_pTessellationThreadData->pDSOutput);
1234 SWR_ASSERT(gt_pTessellationThreadData->numDSOutputVectors >= requiredDSOutputVectors);
1235
1236 #if defined(_DEBUG)
1237 memset(gt_pTessellationThreadData->pDSOutput, 0x90, requiredAllocSize);
1238 #endif
1239
1240 // Run Domain Shader
1241 SWR_DS_CONTEXT dsContext;
1242 dsContext.PrimitiveID = pPrimId[p];
1243 dsContext.pCpIn = &hsContext.pCPout[p];
1244 dsContext.pDomainU = (simdscalar*)tsData.pDomainPointsU;
1245 dsContext.pDomainV = (simdscalar*)tsData.pDomainPointsV;
1246 dsContext.pOutputData = gt_pTessellationThreadData->pDSOutput;
1247 #if USE_SIMD16_FRONTEND
1248 dsContext.vectorStride = RoundUpEven(requiredDSVectorInvocations); // simd8 -> simd16
1249 #else
1250 dsContext.vectorStride = requiredDSVectorInvocations;
1251 #endif
1252
1253 uint32_t dsInvocations = 0;
1254
1255 for (dsContext.vectorOffset = 0; dsContext.vectorOffset < requiredDSVectorInvocations; ++dsContext.vectorOffset)
1256 {
1257 dsContext.mask = GenerateMask(tsData.NumDomainPoints - dsInvocations);
1258
1259 AR_BEGIN(FEDomainShader, pDC->drawId);
1260 state.pfnDsFunc(GetPrivateState(pDC), &dsContext);
1261 AR_END(FEDomainShader, 0);
1262
1263 dsInvocations += KNOB_SIMD_WIDTH;
1264 }
1265 UPDATE_STAT_FE(DsInvocations, tsData.NumDomainPoints);
1266
1267 #if USE_SIMD16_FRONTEND
1268 SWR_ASSERT(IsEven(dsContext.vectorStride)); // simd8 -> simd16
1269
1270 #endif
1271 PA_TESS tessPa(
1272 pDC,
1273 #if USE_SIMD16_FRONTEND
1274 reinterpret_cast<const simd16scalar *>(dsContext.pOutputData), // simd8 -> simd16
1275 dsContext.vectorStride / 2, // simd8 -> simd16
1276 #else
1277 dsContext.pOutputData,
1278 dsContext.vectorStride,
1279 #endif
1280 tsState.numDsOutputAttribs,
1281 tsData.ppIndices,
1282 tsData.NumPrimitives,
1283 tsState.postDSTopology);
1284
1285 while (tessPa.HasWork())
1286 {
1287 #if USE_SIMD16_FRONTEND
1288 const uint32_t numPrims = tessPa.NumPrims();
1289 const uint32_t numPrims_lo = std::min<uint32_t>(numPrims, KNOB_SIMD_WIDTH);
1290 const uint32_t numPrims_hi = std::max<uint32_t>(numPrims, KNOB_SIMD_WIDTH) - KNOB_SIMD_WIDTH;
1291
1292 const simd16scalari primID = _simd16_set1_epi32(dsContext.PrimitiveID);
1293 const simdscalari primID_lo = _simd16_extract_si(primID, 0);
1294 const simdscalari primID_hi = _simd16_extract_si(primID, 1);
1295
1296 #endif
1297 if (HasGeometryShaderT::value)
1298 {
1299 #if USE_SIMD16_FRONTEND
1300 tessPa.useAlternateOffset = false;
1301 GeometryShaderStage<HasStreamOutT, HasRastT>(pDC, workerId, tessPa, pGsOut, pCutBuffer, pCutStreamBuffer, pSoPrimData, numPrims_lo, primID_lo);
1302
1303 if (numPrims_hi)
1304 {
1305 tessPa.useAlternateOffset = true;
1306 GeometryShaderStage<HasStreamOutT, HasRastT>(pDC, workerId, tessPa, pGsOut, pCutBuffer, pCutStreamBuffer, pSoPrimData, numPrims_hi, primID_hi);
1307 }
1308 #else
1309 GeometryShaderStage<HasStreamOutT, HasRastT>(
1310 pDC, workerId, tessPa, pGsOut, pCutBuffer, pCutStreamBuffer, pSoPrimData,
1311 _simd_set1_epi32(dsContext.PrimitiveID));
1312 #endif
1313 }
1314 else
1315 {
1316 if (HasStreamOutT::value)
1317 {
1318 tessPa.useAlternateOffset = false;
1319 StreamOut(pDC, tessPa, workerId, pSoPrimData, 0);
1320 }
1321
1322 if (HasRastT::value)
1323 {
1324 #if USE_SIMD16_FRONTEND
1325 simd16vector prim_simd16[3]; // Only deal with triangles, lines, or points
1326 #else
1327 simdvector prim[3]; // Only deal with triangles, lines, or points
1328 #endif
1329 AR_BEGIN(FEPAAssemble, pDC->drawId);
1330 bool assemble =
1331 #if USE_SIMD16_FRONTEND
1332 tessPa.Assemble_simd16(VERTEX_POSITION_SLOT, prim_simd16);
1333 #else
1334 tessPa.Assemble(VERTEX_POSITION_SLOT, prim);
1335 #endif
1336 AR_END(FEPAAssemble, 1);
1337 SWR_ASSERT(assemble);
1338
1339 SWR_ASSERT(pfnClipFunc);
1340 #if USE_SIMD16_FRONTEND
1341 tessPa.useAlternateOffset = false;
1342 pfnClipFunc(pDC, tessPa, workerId, prim_simd16, GenMask(numPrims), primID, _simd16_set1_epi32(0));
1343 #else
1344 pfnClipFunc(pDC, tessPa, workerId, prim,
1345 GenMask(tessPa.NumPrims()), _simd_set1_epi32(dsContext.PrimitiveID), _simd_set1_epi32(0));
1346 #endif
1347 }
1348 }
1349
1350 tessPa.NextPrim();
1351
1352 } // while (tessPa.HasWork())
1353 } // for (uint32_t p = 0; p < numPrims; ++p)
1354
1355 #if USE_SIMD16_FRONTEND
1356 if (gt_pTessellationThreadData->pDSOutput != nullptr)
1357 {
1358 AlignedFree(gt_pTessellationThreadData->pDSOutput);
1359 gt_pTessellationThreadData->pDSOutput = nullptr;
1360 }
1361 gt_pTessellationThreadData->numDSOutputVectors = 0;
1362
1363 #endif
1364 TSDestroyCtx(tsCtx);
1365 }
1366
1367 THREAD PA_STATE::SIMDVERTEX *pVertexStore = nullptr;
1368 THREAD uint32_t gVertexStoreSize = 0;
1369
1370 //////////////////////////////////////////////////////////////////////////
1371 /// @brief FE handler for SwrDraw.
1372 /// @tparam IsIndexedT - Is indexed drawing enabled
1373 /// @tparam HasTessellationT - Is tessellation enabled
1374 /// @tparam HasGeometryShaderT::value - Is the geometry shader stage enabled
1375 /// @tparam HasStreamOutT - Is stream-out enabled
1376 /// @tparam HasRastT - Is rasterization enabled
1377 /// @param pContext - pointer to SWR context.
1378 /// @param pDC - pointer to draw context.
1379 /// @param workerId - thread's worker id.
1380 /// @param pUserData - Pointer to DRAW_WORK
1381 template <
1382 typename IsIndexedT,
1383 typename IsCutIndexEnabledT,
1384 typename HasTessellationT,
1385 typename HasGeometryShaderT,
1386 typename HasStreamOutT,
1387 typename HasRastT>
1388 void ProcessDraw(
1389 SWR_CONTEXT *pContext,
1390 DRAW_CONTEXT *pDC,
1391 uint32_t workerId,
1392 void *pUserData)
1393 {
1394
1395 #if KNOB_ENABLE_TOSS_POINTS
1396 if (KNOB_TOSS_QUEUE_FE)
1397 {
1398 return;
1399 }
1400 #endif
1401
1402 AR_BEGIN(FEProcessDraw, pDC->drawId);
1403
1404 DRAW_WORK& work = *(DRAW_WORK*)pUserData;
1405 const API_STATE& state = GetApiState(pDC);
1406
1407 uint32_t indexSize = 0;
1408 uint32_t endVertex = work.numVerts;
1409
1410 const int32_t* pLastRequestedIndex = nullptr;
1411 if (IsIndexedT::value)
1412 {
1413 switch (work.type)
1414 {
1415 case R32_UINT:
1416 indexSize = sizeof(uint32_t);
1417 pLastRequestedIndex = &(work.pIB[endVertex]);
1418 break;
1419 case R16_UINT:
1420 indexSize = sizeof(uint16_t);
1421 // nasty address offset to last index
1422 pLastRequestedIndex = (int32_t*)(&(((uint16_t*)work.pIB)[endVertex]));
1423 break;
1424 case R8_UINT:
1425 indexSize = sizeof(uint8_t);
1426 // nasty address offset to last index
1427 pLastRequestedIndex = (int32_t*)(&(((uint8_t*)work.pIB)[endVertex]));
1428 break;
1429 default:
1430 SWR_INVALID("Invalid work.type: %d", work.type);
1431 }
1432 }
1433 else
1434 {
1435 // No cuts, prune partial primitives.
1436 endVertex = GetNumVerts(state.topology, GetNumPrims(state.topology, work.numVerts));
1437 }
1438
1439 #if defined(KNOB_ENABLE_RDTSC) || defined(KNOB_ENABLE_AR)
1440 uint32_t numPrims = GetNumPrims(state.topology, work.numVerts);
1441 #endif
1442
1443 void* pGsOut = nullptr;
1444 void* pCutBuffer = nullptr;
1445 void* pStreamCutBuffer = nullptr;
1446 if (HasGeometryShaderT::value)
1447 {
1448 #if USE_SIMD16_FRONTEND
1449 AllocateGsBuffers<simd16vertex, KNOB_SIMD16_WIDTH>(pDC, state, &pGsOut, &pCutBuffer, &pStreamCutBuffer);
1450 #else
1451 AllocateGsBuffers<simdvertex, KNOB_SIMD_WIDTH>(pDC, state, &pGsOut, &pCutBuffer, &pStreamCutBuffer);
1452 #endif
1453 }
1454
1455 if (HasTessellationT::value)
1456 {
1457 SWR_ASSERT(state.tsState.tsEnable == true);
1458 SWR_ASSERT(state.pfnHsFunc != nullptr);
1459 SWR_ASSERT(state.pfnDsFunc != nullptr);
1460
1461 AllocateTessellationData(pContext);
1462 }
1463 else
1464 {
1465 SWR_ASSERT(state.tsState.tsEnable == false);
1466 SWR_ASSERT(state.pfnHsFunc == nullptr);
1467 SWR_ASSERT(state.pfnDsFunc == nullptr);
1468 }
1469
1470 // allocate space for streamout input prim data
1471 uint32_t* pSoPrimData = nullptr;
1472 if (HasStreamOutT::value)
1473 {
1474 pSoPrimData = (uint32_t*)pDC->pArena->AllocAligned(4096, 16);
1475 }
1476
1477 const uint32_t vertexCount = NumVertsPerPrim(state.topology, state.gsState.gsEnable);
1478
1479 SWR_ASSERT(vertexCount <= MAX_NUM_VERTS_PER_PRIM);
1480
1481 // grow the vertex store for the PA as necessary
1482 if (gVertexStoreSize < vertexCount)
1483 {
1484 if (pVertexStore != nullptr)
1485 {
1486 AlignedFree(pVertexStore);
1487 }
1488
1489 while (gVertexStoreSize < vertexCount)
1490 {
1491 #if USE_SIMD16_FRONTEND
1492 gVertexStoreSize += 4; // grow in chunks of 4 simd16vertex
1493 #else
1494 gVertexStoreSize += 8; // grow in chunks of 8 simdvertex
1495 #endif
1496 }
1497
1498 SWR_ASSERT(gVertexStoreSize <= MAX_NUM_VERTS_PER_PRIM);
1499
1500 pVertexStore = reinterpret_cast<PA_STATE::SIMDVERTEX *>(AlignedMalloc(gVertexStoreSize * sizeof(pVertexStore[0]), 64));
1501
1502 SWR_ASSERT(pVertexStore != nullptr);
1503 }
1504
1505 // choose primitive assembler
1506 PA_FACTORY<IsIndexedT, IsCutIndexEnabledT> paFactory(pDC, state.topology, work.numVerts, pVertexStore, gVertexStoreSize);
1507 PA_STATE& pa = paFactory.GetPA();
1508
1509 #if USE_SIMD16_FRONTEND
1510 simdvertex vin_lo;
1511 simdvertex vin_hi;
1512 SWR_VS_CONTEXT vsContext_lo;
1513 SWR_VS_CONTEXT vsContext_hi;
1514
1515 vsContext_lo.pVin = &vin_lo;
1516 vsContext_hi.pVin = &vin_hi;
1517 vsContext_lo.AlternateOffset = 0;
1518 vsContext_hi.AlternateOffset = 1;
1519
1520 SWR_FETCH_CONTEXT fetchInfo_lo = { 0 };
1521
1522 fetchInfo_lo.pStreams = &state.vertexBuffers[0];
1523 fetchInfo_lo.StartInstance = work.startInstance;
1524 fetchInfo_lo.StartVertex = 0;
1525
1526 if (IsIndexedT::value)
1527 {
1528 fetchInfo_lo.BaseVertex = work.baseVertex;
1529
1530 // if the entire index buffer isn't being consumed, set the last index
1531 // so that fetches < a SIMD wide will be masked off
1532 fetchInfo_lo.pLastIndex = (const int32_t*)(((uint8_t*)state.indexBuffer.pIndices) + state.indexBuffer.size);
1533 if (pLastRequestedIndex < fetchInfo_lo.pLastIndex)
1534 {
1535 fetchInfo_lo.pLastIndex = pLastRequestedIndex;
1536 }
1537 }
1538 else
1539 {
1540 fetchInfo_lo.StartVertex = work.startVertex;
1541 }
1542
1543 SWR_FETCH_CONTEXT fetchInfo_hi = fetchInfo_lo;
1544
1545 const simd16scalari vScale = _simd16_set_epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
1546
1547 for (uint32_t instanceNum = 0; instanceNum < work.numInstances; instanceNum++)
1548 {
1549 uint32_t i = 0;
1550
1551 simd16scalari vIndex;
1552
1553 if (IsIndexedT::value)
1554 {
1555 fetchInfo_lo.pIndices = work.pIB;
1556 fetchInfo_hi.pIndices = (int32_t *)((uint8_t *)fetchInfo_lo.pIndices + KNOB_SIMD_WIDTH * indexSize); // 1/2 of KNOB_SIMD16_WIDTH
1557 }
1558 else
1559 {
1560 vIndex = _simd16_add_epi32(_simd16_set1_epi32(work.startVertexID), vScale);
1561
1562 fetchInfo_lo.pIndices = (const int32_t *)&vIndex;
1563 fetchInfo_hi.pIndices = (const int32_t *)&vIndex + KNOB_SIMD_WIDTH; // 1/2 of KNOB_SIMD16_WIDTH
1564 }
1565
1566 fetchInfo_lo.CurInstance = instanceNum;
1567 fetchInfo_hi.CurInstance = instanceNum;
1568
1569 vsContext_lo.InstanceID = instanceNum;
1570 vsContext_hi.InstanceID = instanceNum;
1571
1572 while (pa.HasWork())
1573 {
1574 // GetNextVsOutput currently has the side effect of updating some PA state machine state.
1575 // So we need to keep this outside of (i < endVertex) check.
1576
1577 simdmask *pvCutIndices_lo = nullptr;
1578 simdmask *pvCutIndices_hi = nullptr;
1579
1580 if (IsIndexedT::value)
1581 {
1582 // simd16mask <=> simdmask[2]
1583
1584 pvCutIndices_lo = &reinterpret_cast<simdmask *>(&pa.GetNextVsIndices())[0];
1585 pvCutIndices_hi = &reinterpret_cast<simdmask *>(&pa.GetNextVsIndices())[1];
1586 }
1587
1588 simd16vertex &vout = pa.GetNextVsOutput();
1589
1590 vsContext_lo.pVout = reinterpret_cast<simdvertex *>(&vout);
1591 vsContext_hi.pVout = reinterpret_cast<simdvertex *>(&vout);
1592
1593 if (i < endVertex)
1594 {
1595 // 1. Execute FS/VS for a single SIMD.
1596 AR_BEGIN(FEFetchShader, pDC->drawId);
1597 state.pfnFetchFunc(fetchInfo_lo, vin_lo);
1598
1599 if ((i + KNOB_SIMD_WIDTH) < endVertex) // 1/2 of KNOB_SIMD16_WIDTH
1600 {
1601 state.pfnFetchFunc(fetchInfo_hi, vin_hi);
1602 }
1603 AR_END(FEFetchShader, 0);
1604
1605 // forward fetch generated vertex IDs to the vertex shader
1606 vsContext_lo.VertexID = fetchInfo_lo.VertexID;
1607 vsContext_hi.VertexID = fetchInfo_hi.VertexID;
1608
1609 // Setup active mask for vertex shader.
1610 vsContext_lo.mask = GenerateMask(endVertex - i);
1611 vsContext_hi.mask = GenerateMask(endVertex - (i + KNOB_SIMD_WIDTH));
1612
1613 // forward cut mask to the PA
1614 if (IsIndexedT::value)
1615 {
1616 *pvCutIndices_lo = _simd_movemask_ps(_simd_castsi_ps(fetchInfo_lo.CutMask));
1617 *pvCutIndices_hi = _simd_movemask_ps(_simd_castsi_ps(fetchInfo_hi.CutMask));
1618 }
1619
1620 UPDATE_STAT_FE(IaVertices, GetNumInvocations(i, endVertex));
1621
1622 #if KNOB_ENABLE_TOSS_POINTS
1623 if (!KNOB_TOSS_FETCH)
1624 #endif
1625 {
1626 AR_BEGIN(FEVertexShader, pDC->drawId);
1627 state.pfnVertexFunc(GetPrivateState(pDC), &vsContext_lo);
1628
1629 if ((i + KNOB_SIMD_WIDTH) < endVertex) // 1/2 of KNOB_SIMD16_WIDTH
1630 {
1631 state.pfnVertexFunc(GetPrivateState(pDC), &vsContext_hi);
1632 }
1633 AR_END(FEVertexShader, 0);
1634
1635 UPDATE_STAT_FE(VsInvocations, GetNumInvocations(i, endVertex));
1636 }
1637 }
1638
1639 // 2. Assemble primitives given the last two SIMD.
1640 do
1641 {
1642 simd16vector prim_simd16[MAX_NUM_VERTS_PER_PRIM];
1643
1644 RDTSC_START(FEPAAssemble);
1645 bool assemble = pa.Assemble_simd16(VERTEX_POSITION_SLOT, prim_simd16);
1646 RDTSC_STOP(FEPAAssemble, 1, 0);
1647
1648 #if KNOB_ENABLE_TOSS_POINTS
1649 if (!KNOB_TOSS_FETCH)
1650 #endif
1651 {
1652 #if KNOB_ENABLE_TOSS_POINTS
1653 if (!KNOB_TOSS_VS)
1654 #endif
1655 {
1656 if (assemble)
1657 {
1658 UPDATE_STAT_FE(IaPrimitives, pa.NumPrims());
1659
1660 const uint32_t numPrims = pa.NumPrims();
1661 const uint32_t numPrims_lo = std::min<uint32_t>(numPrims, KNOB_SIMD_WIDTH);
1662 const uint32_t numPrims_hi = std::max<uint32_t>(numPrims, KNOB_SIMD_WIDTH) - KNOB_SIMD_WIDTH;
1663
1664 const simd16scalari primID = pa.GetPrimID(work.startPrimID);
1665 const simdscalari primID_lo = _simd16_extract_si(primID, 0);
1666 const simdscalari primID_hi = _simd16_extract_si(primID, 1);
1667
1668 if (HasTessellationT::value)
1669 {
1670 pa.useAlternateOffset = false;
1671 TessellationStages<HasGeometryShaderT, HasStreamOutT, HasRastT>(pDC, workerId, pa, pGsOut, pCutBuffer, pStreamCutBuffer, pSoPrimData, numPrims_lo, primID_lo);
1672
1673 if (numPrims_hi)
1674 {
1675 pa.useAlternateOffset = true;
1676 TessellationStages<HasGeometryShaderT, HasStreamOutT, HasRastT>(pDC, workerId, pa, pGsOut, pCutBuffer, pStreamCutBuffer, pSoPrimData, numPrims_hi, primID_hi);
1677 }
1678 }
1679 else if (HasGeometryShaderT::value)
1680 {
1681 pa.useAlternateOffset = false;
1682 GeometryShaderStage<HasStreamOutT, HasRastT>(pDC, workerId, pa, pGsOut, pCutBuffer, pStreamCutBuffer, pSoPrimData, numPrims_lo, primID_lo);
1683
1684 if (numPrims_hi)
1685 {
1686 pa.useAlternateOffset = true;
1687 GeometryShaderStage<HasStreamOutT, HasRastT>(pDC, workerId, pa, pGsOut, pCutBuffer, pStreamCutBuffer, pSoPrimData, numPrims_hi, primID_hi);
1688 }
1689 }
1690 else
1691 {
1692 // If streamout is enabled then stream vertices out to memory.
1693 if (HasStreamOutT::value)
1694 {
1695 pa.useAlternateOffset = false;
1696 StreamOut(pDC, pa, workerId, pSoPrimData, 0);
1697 }
1698
1699 if (HasRastT::value)
1700 {
1701 SWR_ASSERT(pDC->pState->pfnProcessPrims_simd16);
1702
1703 pa.useAlternateOffset = false;
1704 pDC->pState->pfnProcessPrims_simd16(pDC, pa, workerId, prim_simd16, GenMask(numPrims), primID, _simd16_setzero_si());
1705 }
1706 }
1707 }
1708 }
1709 }
1710 } while (pa.NextPrim());
1711
1712 if (IsIndexedT::value)
1713 {
1714 fetchInfo_lo.pIndices = (int32_t *)((uint8_t*)fetchInfo_lo.pIndices + KNOB_SIMD16_WIDTH * indexSize);
1715 fetchInfo_hi.pIndices = (int32_t *)((uint8_t*)fetchInfo_hi.pIndices + KNOB_SIMD16_WIDTH * indexSize);
1716 }
1717 else
1718 {
1719 vIndex = _simd16_add_epi32(vIndex, _simd16_set1_epi32(KNOB_SIMD16_WIDTH));
1720 }
1721
1722 i += KNOB_SIMD16_WIDTH;
1723 }
1724
1725 pa.Reset();
1726 }
1727
1728 #else
1729 simdvertex vin;
1730 SWR_VS_CONTEXT vsContext;
1731
1732 vsContext.pVin = &vin;
1733
1734 SWR_FETCH_CONTEXT fetchInfo = { 0 };
1735
1736 fetchInfo.pStreams = &state.vertexBuffers[0];
1737 fetchInfo.StartInstance = work.startInstance;
1738 fetchInfo.StartVertex = 0;
1739
1740 if (IsIndexedT::value)
1741 {
1742 fetchInfo.BaseVertex = work.baseVertex;
1743
1744 // if the entire index buffer isn't being consumed, set the last index
1745 // so that fetches < a SIMD wide will be masked off
1746 fetchInfo.pLastIndex = (const int32_t*)(((uint8_t*)state.indexBuffer.pIndices) + state.indexBuffer.size);
1747 if (pLastRequestedIndex < fetchInfo.pLastIndex)
1748 {
1749 fetchInfo.pLastIndex = pLastRequestedIndex;
1750 }
1751 }
1752 else
1753 {
1754 fetchInfo.StartVertex = work.startVertex;
1755 }
1756
1757 const simdscalari vScale = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0);
1758
1759 /// @todo: temporarily move instance loop in the FE to ensure SO ordering
1760 for (uint32_t instanceNum = 0; instanceNum < work.numInstances; instanceNum++)
1761 {
1762 simdscalari vIndex;
1763 uint32_t i = 0;
1764
1765 if (IsIndexedT::value)
1766 {
1767 fetchInfo.pIndices = work.pIB;
1768 }
1769 else
1770 {
1771 vIndex = _simd_add_epi32(_simd_set1_epi32(work.startVertexID), vScale);
1772 fetchInfo.pIndices = (const int32_t*)&vIndex;
1773 }
1774
1775 fetchInfo.CurInstance = instanceNum;
1776 vsContext.InstanceID = instanceNum;
1777
1778 while (pa.HasWork())
1779 {
1780 // GetNextVsOutput currently has the side effect of updating some PA state machine state.
1781 // So we need to keep this outside of (i < endVertex) check.
1782 simdmask* pvCutIndices = nullptr;
1783 if (IsIndexedT::value)
1784 {
1785 pvCutIndices = &pa.GetNextVsIndices();
1786 }
1787
1788 simdvertex& vout = pa.GetNextVsOutput();
1789 vsContext.pVout = &vout;
1790
1791 if (i < endVertex)
1792 {
1793
1794 // 1. Execute FS/VS for a single SIMD.
1795 AR_BEGIN(FEFetchShader, pDC->drawId);
1796 state.pfnFetchFunc(fetchInfo, vin);
1797 AR_END(FEFetchShader, 0);
1798
1799 // forward fetch generated vertex IDs to the vertex shader
1800 vsContext.VertexID = fetchInfo.VertexID;
1801
1802 // Setup active mask for vertex shader.
1803 vsContext.mask = GenerateMask(endVertex - i);
1804
1805 // forward cut mask to the PA
1806 if (IsIndexedT::value)
1807 {
1808 *pvCutIndices = _simd_movemask_ps(_simd_castsi_ps(fetchInfo.CutMask));
1809 }
1810
1811 UPDATE_STAT_FE(IaVertices, GetNumInvocations(i, endVertex));
1812
1813 #if KNOB_ENABLE_TOSS_POINTS
1814 if (!KNOB_TOSS_FETCH)
1815 #endif
1816 {
1817 AR_BEGIN(FEVertexShader, pDC->drawId);
1818 state.pfnVertexFunc(GetPrivateState(pDC), &vsContext);
1819 AR_END(FEVertexShader, 0);
1820
1821 UPDATE_STAT_FE(VsInvocations, GetNumInvocations(i, endVertex));
1822 }
1823 }
1824
1825 // 2. Assemble primitives given the last two SIMD.
1826 do
1827 {
1828 simdvector prim[MAX_NUM_VERTS_PER_PRIM];
1829 // PaAssemble returns false if there is not enough verts to assemble.
1830 AR_BEGIN(FEPAAssemble, pDC->drawId);
1831 bool assemble = pa.Assemble(VERTEX_POSITION_SLOT, prim);
1832 AR_END(FEPAAssemble, 1);
1833
1834 #if KNOB_ENABLE_TOSS_POINTS
1835 if (!KNOB_TOSS_FETCH)
1836 #endif
1837 {
1838 #if KNOB_ENABLE_TOSS_POINTS
1839 if (!KNOB_TOSS_VS)
1840 #endif
1841 {
1842 if (assemble)
1843 {
1844 UPDATE_STAT_FE(IaPrimitives, pa.NumPrims());
1845
1846 if (HasTessellationT::value)
1847 {
1848 TessellationStages<HasGeometryShaderT, HasStreamOutT, HasRastT>(
1849 pDC, workerId, pa, pGsOut, pCutBuffer, pStreamCutBuffer, pSoPrimData, pa.GetPrimID(work.startPrimID));
1850 }
1851 else if (HasGeometryShaderT::value)
1852 {
1853 GeometryShaderStage<HasStreamOutT, HasRastT>(
1854 pDC, workerId, pa, pGsOut, pCutBuffer, pStreamCutBuffer, pSoPrimData, pa.GetPrimID(work.startPrimID));
1855 }
1856 else
1857 {
1858 // If streamout is enabled then stream vertices out to memory.
1859 if (HasStreamOutT::value)
1860 {
1861 StreamOut(pDC, pa, workerId, pSoPrimData, 0);
1862 }
1863
1864 if (HasRastT::value)
1865 {
1866 SWR_ASSERT(pDC->pState->pfnProcessPrims);
1867
1868 pDC->pState->pfnProcessPrims(pDC, pa, workerId, prim,
1869 GenMask(pa.NumPrims()), pa.GetPrimID(work.startPrimID), _simd_set1_epi32(0));
1870 }
1871 }
1872 }
1873 }
1874 }
1875 } while (pa.NextPrim());
1876
1877 if (IsIndexedT::value)
1878 {
1879 fetchInfo.pIndices = (int*)((uint8_t*)fetchInfo.pIndices + KNOB_SIMD_WIDTH * indexSize);
1880 }
1881 else
1882 {
1883 vIndex = _simd_add_epi32(vIndex, _simd_set1_epi32(KNOB_SIMD_WIDTH));
1884 }
1885
1886 i += KNOB_SIMD_WIDTH;
1887 }
1888 pa.Reset();
1889 }
1890
1891 #endif
1892
1893 AR_END(FEProcessDraw, numPrims * work.numInstances);
1894 }
1895
1896 struct FEDrawChooser
1897 {
1898 typedef PFN_FE_WORK_FUNC FuncType;
1899
1900 template <typename... ArgsB>
1901 static FuncType GetFunc()
1902 {
1903 return ProcessDraw<ArgsB...>;
1904 }
1905 };
1906
1907
1908 // Selector for correct templated Draw front-end function
1909 PFN_FE_WORK_FUNC GetProcessDrawFunc(
1910 bool IsIndexed,
1911 bool IsCutIndexEnabled,
1912 bool HasTessellation,
1913 bool HasGeometryShader,
1914 bool HasStreamOut,
1915 bool HasRasterization)
1916 {
1917 return TemplateArgUnroller<FEDrawChooser>::GetFunc(IsIndexed, IsCutIndexEnabled, HasTessellation, HasGeometryShader, HasStreamOut, HasRasterization);
1918 }