Lines Matching refs:SIMD_WIDTH
40 SIMD_WIDTH = KNOB_SIMD16_WIDTH,
56 SIMD_WIDTH = KNOB_SIMD_WIDTH,
255 const uint32_t numSimdVerts = streamSizeInVerts / SIMD_WIDTH;
296 (SIMD_WIDTH - (this->numPrimsComplete + this->nextNumPrimsIncrement - this->numPrims)) : SIMD_WIDTH;
353 _simd16_set1_epi32(startID + this->primIDIncr * (this->numPrimsComplete / SIMD_WIDTH)));
356 _simd_set1_epi32(startID + this->primIDIncr * (this->numPrimsComplete / SIMD_WIDTH)));
404 OSALIGNSIMD16(uint32_t) indices[MAX_NUM_VERTS_PER_PRIM][SIMD_WIDTH]; // current index buffer for gather
406 OSALIGNSIMD(uint32_t) indices[MAX_NUM_VERTS_PER_PRIM][SIMD_WIDTH]; // current index buffer for gather
490 uint32_t vertexIndex = this->headVertex / SIMD_WIDTH;
491 this->headVertex = (this->headVertex + SIMD_WIDTH) % this->numVerts;
500 uint32_t vertexIndex = this->headVertex / SIMD_WIDTH;
523 this->headVertex += SIMD_WIDTH;
565 return ((this->headVertex + SIMD_WIDTH) % this->numVerts) == this->tailVertex;
577 uint32_t vertexIndex = vertex / SIMD_WIDTH;
578 uint32_t vertexOffset = vertex & (SIMD_WIDTH - 1);
586 while (this->numPrimsAssembled != SIMD_WIDTH &&
617 if (this->numPrimsAssembled != SIMD_WIDTH && this->numRemainingVerts == 0 && this->adjExtraVert != -1)
630 this->vPrimId = _simd16_add_epi32(vPrimId, _simd16_set1_epi32(SIMD_WIDTH));
632 this->vPrimId = _simd_add_epi32(vPrimId, _simd_set1_epi32(SIMD_WIDTH));
639 if (this->numPrimsAssembled == SIMD_WIDTH || this->numRemainingVerts <= 0)
664 const uint32_t simdMask = SIMD_WIDTH - 1;
681 if (this->numPrimsAssembled != SIMD_WIDTH && this->numRemainingVerts > 0)
718 pBase += SIMD_WIDTH;
732 if (this->numPrimsAssembled != SIMD_WIDTH && this->numRemainingVerts > 0)
765 pBase += SIMD_WIDTH;
790 offset += SIMD_WIDTH * sizeof(float);
1211 SWR_ASSERT(numPrims <= SIMD_WIDTH);
1213 static const OSALIGNLINE(int32_t) maskGen[SIMD_WIDTH * 2] =
1219 return _simd16_loadu_si((const SIMDSCALARI*)&maskGen[SIMD_WIDTH - numPrims]);
1221 static const OSALIGNLINE(int32_t) maskGen[SIMD_WIDTH * 2] =
1227 return _simd_loadu_si((const SIMDSCALARI*)&maskGen[SIMD_WIDTH - numPrims]);
1272 pBase += m_attributeStrideInVectors * SIMD_WIDTH;
1320 pBase += m_attributeStrideInVectors * SIMD_WIDTH;
1347 pVertData += m_attributeStrideInVectors * SIMD_WIDTH;
1383 return std::min<uint32_t>(m_numPrims, SIMD_WIDTH);
1440 new (&this->paCut) PA_STATE_CUT(pDC, reinterpret_cast<uint8_t *>(pVertexStore), vertexStoreSize * PA_STATE::SIMD_WIDTH,
1448 new (&this->paOpt) PA_STATE_OPT(pDC, numPrims, reinterpret_cast<uint8_t *>(pVertexStore), vertexStoreSize * PA_STATE::SIMD_WIDTH, vertexStride, false, numVertsPerPrim);
