X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fllvmpipe%2Flp_rast_tri_tmp.h;h=05d2242786feb799d2a9ea690d8a8883d6e91da5;hb=e4ca1d64565b4d665bcaf5d08922bfbe1d920e7a;hp=a410c611a3fa71677d697bfe44373d4ba30c0644;hpb=582132aaef3fa6fa912f27c75c3b3e5bd89d3714;p=mesa.git diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h b/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h index a410c611a3f..05d2242786f 100644 --- a/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h +++ b/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h @@ -32,7 +32,7 @@ /** - * Prototype for a 7 plane rasterizer function. Will codegenerate + * Prototype for a 8 plane rasterizer function. Will codegenerate * several of these. * * XXX: Varients for more/fewer planes. @@ -44,21 +44,21 @@ TAG(do_block_4)(struct lp_rasterizer_task *task, const struct lp_rast_triangle *tri, const struct lp_rast_plane *plane, int x, int y, - const int *c) + const int64_t *c) { - unsigned mask = 0; - int i; + unsigned mask = 0xffff; + int j; - for (i = 0; i < 16; i++) { - int any_negative = 0; - int j; - - for (j = 0; j < NR_PLANES; j++) - any_negative |= (c[j] - 1 + plane[j].step[i]); - - any_negative >>= 31; - - mask |= (~any_negative) & (1 << i); + for (j = 0; j < NR_PLANES; j++) { +#ifdef RASTER_64 + mask &= ~BUILD_MASK_LINEAR(((c[j] - 1) >> (int64_t)FIXED_ORDER), + -plane[j].dcdx >> FIXED_ORDER, + plane[j].dcdy >> FIXED_ORDER); +#else + mask &= ~BUILD_MASK_LINEAR((c[j] - 1), + -plane[j].dcdx, + plane[j].dcdy); +#endif } /* Now pass to the shader: @@ -76,27 +76,42 @@ TAG(do_block_16)(struct lp_rasterizer_task *task, const struct lp_rast_triangle *tri, const struct lp_rast_plane *plane, int x, int y, - const int *c) + const int64_t *c) { unsigned outmask, inmask, partmask, partial_mask; - unsigned i, j; + unsigned j; outmask = 0; /* outside one or more trivial reject planes */ partmask = 0; /* outside one or more trivial accept planes */ for (j = 0; j < NR_PLANES; j++) { - const int *step = plane[j].step; - const int eo = plane[j].eo * 4; - const int ei = plane[j].ei * 4; - const int cox = c[j] + eo; - const int cio = ei - 1 - eo; - - for (i = 0; i < 16; i++) { - int out = cox + step[i] * 4; - int part = out + cio; - outmask |= (out >> 31) & (1 << i); - partmask |= (part >> 31) & (1 << i); - } +#ifdef RASTER_64 + int32_t dcdx = -plane[j].dcdx >> FIXED_ORDER; + int32_t dcdy = plane[j].dcdy >> FIXED_ORDER; + const int32_t cox = plane[j].eo >> FIXED_ORDER; + const int32_t ei = (dcdy + dcdx - cox) << 2; + const int32_t cox_s = cox << 2; + const int32_t co = (int32_t)(c[j] >> (int64_t)FIXED_ORDER) + cox_s; + int32_t cdiff; + cdiff = ei - cox_s + ((int32_t)((c[j] - 1) >> (int64_t)FIXED_ORDER) - + (int32_t)(c[j] >> (int64_t)FIXED_ORDER)); + dcdx <<= 2; + dcdy <<= 2; +#else + const int64_t dcdx = -IMUL64(plane[j].dcdx, 4); + const int64_t dcdy = IMUL64(plane[j].dcdy, 4); + const int64_t cox = IMUL64(plane[j].eo, 4); + const int32_t ei = plane[j].dcdy - plane[j].dcdx - (int64_t)plane[j].eo; + const int64_t cio = IMUL64(ei, 4) - 1; + int32_t co, cdiff; + co = c[j] + cox; + cdiff = cio - cox; +#endif + + BUILD_MASKS(co, cdiff, + dcdx, dcdy, + &outmask, /* sign bits from c[i][0..15] + cox */ + &partmask); /* sign bits from c[i][0..15] + cio */ } if (outmask == 0xffff) @@ -113,19 +128,27 @@ TAG(do_block_16)(struct lp_rasterizer_task *task, assert((partial_mask & inmask) == 0); + LP_COUNT_ADD(nr_empty_4, util_bitcount(0xffff & ~(partial_mask | inmask))); + /* Iterate over partials: */ while (partial_mask) { int i = ffs(partial_mask) - 1; - int px = x + pos_table4[i][0]; - int py = y + pos_table4[i][1]; - int cx[NR_PLANES]; - - for (j = 0; j < NR_PLANES; j++) - cx[j] = c[j] + plane[j].step[i] * 4; + int ix = (i & 3) * 4; + int iy = (i >> 2) * 4; + int px = x + ix; + int py = y + iy; + int64_t cx[NR_PLANES]; partial_mask &= ~(1 << i); + LP_COUNT(nr_partially_covered_4); + + for (j = 0; j < NR_PLANES; j++) + cx[j] = (c[j] + - IMUL64(plane[j].dcdx, ix) + + IMUL64(plane[j].dcdy, iy)); + TAG(do_block_4)(task, tri, plane, px, py, cx); } @@ -133,11 +156,14 @@ TAG(do_block_16)(struct lp_rasterizer_task *task, */ while (inmask) { int i = ffs(inmask) - 1; - int px = x + pos_table4[i][0]; - int py = y + pos_table4[i][1]; + int ix = (i & 3) * 4; + int iy = (i >> 2) * 4; + int px = x + ix; + int py = y + iy; inmask &= ~(1 << i); + LP_COUNT(nr_fully_covered_4); block_full_4(task, tri, px, py); } } @@ -153,39 +179,93 @@ TAG(lp_rast_triangle)(struct lp_rasterizer_task *task, { const struct lp_rast_triangle *tri = arg.triangle.tri; unsigned plane_mask = arg.triangle.plane_mask; + const struct lp_rast_plane *tri_plane = GET_PLANES(tri); const int x = task->x, y = task->y; struct lp_rast_plane plane[NR_PLANES]; - int c[NR_PLANES]; + int64_t c[NR_PLANES]; unsigned outmask, inmask, partmask, partial_mask; - unsigned i, j, nr_planes = 0; + unsigned j = 0; - while (plane_mask) { - int i = ffs(plane_mask) - 1; - plane[nr_planes] = tri->plane[i]; - plane_mask &= ~(1 << i); - nr_planes++; - }; + if (tri->inputs.disable) { + /* This triangle was partially binned and has been disabled */ + return; + } - assert(nr_planes == NR_PLANES); outmask = 0; /* outside one or more trivial reject planes */ partmask = 0; /* outside one or more trivial accept planes */ - for (j = 0; j < NR_PLANES; j++) { - const int *step = plane[j].step; - const int eo = plane[j].eo * 16; - const int ei = plane[j].ei * 16; - int cox, cio; - - c[j] = plane[j].c + plane[j].dcdy * y - plane[j].dcdx * x; - cox = c[j] + eo; - cio = ei - 1 - eo; - - for (i = 0; i < 16; i++) { - int out = cox + step[i] * 16; - int part = out + cio; - outmask |= (out >> 31) & (1 << i); - partmask |= (part >> 31) & (1 << i); + while (plane_mask) { + int i = ffs(plane_mask) - 1; + plane[j] = tri_plane[i]; + plane_mask &= ~(1 << i); + c[j] = plane[j].c + IMUL64(plane[j].dcdy, y) - IMUL64(plane[j].dcdx, x); + + { +#ifdef RASTER_64 + /* + * Strip off lower FIXED_ORDER bits. Note that those bits from + * dcdx, dcdy, eo are always 0 (by definition). + * c values, however, are not. This means that for every + * addition of the form c + n*dcdx the lower FIXED_ORDER bits will + * NOT change. And those bits are not relevant to the sign bit (which + * is only what we need!) that is, + * sign(c + n*dcdx) == sign((c >> FIXED_ORDER) + n*(dcdx >> FIXED_ORDER)) + * This means we can get away with using 32bit math for the most part. + * Only tricky part is the -1 adjustment for cdiff. + */ + int32_t dcdx = -plane[j].dcdx >> FIXED_ORDER; + int32_t dcdy = plane[j].dcdy >> FIXED_ORDER; + const int32_t cox = plane[j].eo >> FIXED_ORDER; + const int32_t ei = (dcdy + dcdx - cox) << 4; + const int32_t cox_s = cox << 4; + const int32_t co = (int32_t)(c[j] >> (int64_t)FIXED_ORDER) + cox_s; + int32_t cdiff; + /* + * Plausibility check to ensure the 32bit math works. + * Note that within a tile, the max we can move the edge function + * is essentially dcdx * TILE_SIZE + dcdy * TILE_SIZE. + * TILE_SIZE is 64, dcdx/dcdy are nominally 21 bit (for 8192 max size + * and 8 subpixel bits), I'd be happy with 2 bits more too (1 for + * increasing fb size to 16384, the required d3d11 value, another one + * because I'm not quite sure we can't be _just_ above the max value + * here). This gives us 30 bits max - hence if c would exceed that here + * that means the plane is either trivial reject for the whole tile + * (in which case the tri will not get binned), or trivial accept for + * the whole tile (in which case plane_mask will not include it). + */ + assert((c[j] >> (int64_t)FIXED_ORDER) > (int32_t)0xb0000000 && + (c[j] >> (int64_t)FIXED_ORDER) < (int32_t)0x3fffffff); + /* + * Note the fixup part is constant throughout the tile - thus could + * just calculate this and avoid _all_ 64bit math in rasterization + * (except exactly this fixup calc). + * In fact theoretically could move that even to setup, albeit that + * seems tricky (pre-bin certainly can have values larger than 32bit, + * and would need to communicate that fixup value through). + * And if we want to support msaa, we'd probably don't want to do the + * downscaling in setup in any case... + */ + cdiff = ei - cox_s + ((int32_t)((c[j] - 1) >> (int64_t)FIXED_ORDER) - + (int32_t)(c[j] >> (int64_t)FIXED_ORDER)); + dcdx <<= 4; + dcdy <<= 4; +#else + const int32_t dcdx = -plane[j].dcdx << 4; + const int32_t dcdy = plane[j].dcdy << 4; + const int32_t cox = plane[j].eo << 4; + const int32_t ei = plane[j].dcdy - plane[j].dcdx - (int32_t)plane[j].eo; + const int32_t cio = (ei << 4) - 1; + int32_t co, cdiff; + co = c[j] + cox; + cdiff = cio - cox; +#endif + BUILD_MASKS(co, cdiff, + dcdx, dcdy, + &outmask, /* sign bits from c[i][0..15] + cox */ + &partmask); /* sign bits from c[i][0..15] + cio */ } + + j++; } if (outmask == 0xffff) @@ -202,16 +282,22 @@ TAG(lp_rast_triangle)(struct lp_rasterizer_task *task, assert((partial_mask & inmask) == 0); + LP_COUNT_ADD(nr_empty_16, util_bitcount(0xffff & ~(partial_mask | inmask))); + /* Iterate over partials: */ while (partial_mask) { int i = ffs(partial_mask) - 1; - int px = x + pos_table16[i][0]; - int py = y + pos_table16[i][1]; - int cx[NR_PLANES]; + int ix = (i & 3) * 16; + int iy = (i >> 2) * 16; + int px = x + ix; + int py = y + iy; + int64_t cx[NR_PLANES]; for (j = 0; j < NR_PLANES; j++) - cx[j] = c[j] + plane[j].step[i] * 16; + cx[j] = (c[j] + - IMUL64(plane[j].dcdx, ix) + + IMUL64(plane[j].dcdy, iy)); partial_mask &= ~(1 << i); @@ -223,8 +309,10 @@ TAG(lp_rast_triangle)(struct lp_rasterizer_task *task, */ while (inmask) { int i = ffs(inmask) - 1; - int px = x + pos_table16[i][0]; - int py = y + pos_table16[i][1]; + int ix = (i & 3) * 16; + int iy = (i >> 2) * 16; + int px = x + ix; + int py = y + iy; inmask &= ~(1 << i); @@ -233,6 +321,133 @@ TAG(lp_rast_triangle)(struct lp_rasterizer_task *task, } } +#if defined(PIPE_ARCH_SSE) && defined(TRI_16) +/* XXX: special case this when intersection is not required. + * - tile completely within bbox, + * - bbox completely within tile. + */ +void +TRI_16(struct lp_rasterizer_task *task, + const union lp_rast_cmd_arg arg) +{ + const struct lp_rast_triangle *tri = arg.triangle.tri; + const struct lp_rast_plane *plane = GET_PLANES(tri); + unsigned mask = arg.triangle.plane_mask; + unsigned outmask, partial_mask; + unsigned j; + __m128i cstep4[NR_PLANES][4]; + + int x = (mask & 0xff); + int y = (mask >> 8); + + outmask = 0; /* outside one or more trivial reject planes */ + + x += task->x; + y += task->y; + + for (j = 0; j < NR_PLANES; j++) { + const int dcdx = -plane[j].dcdx * 4; + const int dcdy = plane[j].dcdy * 4; + __m128i xdcdy = _mm_set1_epi32(dcdy); + + cstep4[j][0] = _mm_setr_epi32(0, dcdx, dcdx*2, dcdx*3); + cstep4[j][1] = _mm_add_epi32(cstep4[j][0], xdcdy); + cstep4[j][2] = _mm_add_epi32(cstep4[j][1], xdcdy); + cstep4[j][3] = _mm_add_epi32(cstep4[j][2], xdcdy); + + { + const int c = plane[j].c + plane[j].dcdy * y - plane[j].dcdx * x; + const int cox = plane[j].eo * 4; + + outmask |= sign_bits4(cstep4[j], c + cox); + } + } + + if (outmask == 0xffff) + return; + + + /* Mask of sub-blocks which are inside all trivial reject planes, + * but outside at least one trivial accept plane: + */ + partial_mask = 0xffff & ~outmask; + + /* Iterate over partials: + */ + while (partial_mask) { + int i = ffs(partial_mask) - 1; + int ix = (i & 3) * 4; + int iy = (i >> 2) * 4; + int px = x + ix; + int py = y + iy; + unsigned mask = 0xffff; + + partial_mask &= ~(1 << i); + + for (j = 0; j < NR_PLANES; j++) { + const int cx = (plane[j].c - 1 + - plane[j].dcdx * px + + plane[j].dcdy * py) * 4; + + mask &= ~sign_bits4(cstep4[j], cx); + } + + if (mask) + lp_rast_shade_quads_mask(task, &tri->inputs, px, py, mask); + } +} +#endif + +#if defined(PIPE_ARCH_SSE) && defined(TRI_4) +void +TRI_4(struct lp_rasterizer_task *task, + const union lp_rast_cmd_arg arg) +{ + const struct lp_rast_triangle *tri = arg.triangle.tri; + const struct lp_rast_plane *plane = GET_PLANES(tri); + unsigned mask = arg.triangle.plane_mask; + const int x = task->x + (mask & 0xff); + const int y = task->y + (mask >> 8); + unsigned j; + + /* Iterate over partials: + */ + { + unsigned mask = 0xffff; + + for (j = 0; j < NR_PLANES; j++) { + const int cx = (plane[j].c + - plane[j].dcdx * x + + plane[j].dcdy * y); + + const int dcdx = -plane[j].dcdx; + const int dcdy = plane[j].dcdy; + __m128i xdcdy = _mm_set1_epi32(dcdy); + + __m128i cstep0 = _mm_setr_epi32(cx, cx + dcdx, cx + dcdx*2, cx + dcdx*3); + __m128i cstep1 = _mm_add_epi32(cstep0, xdcdy); + __m128i cstep2 = _mm_add_epi32(cstep1, xdcdy); + __m128i cstep3 = _mm_add_epi32(cstep2, xdcdy); + + __m128i cstep01 = _mm_packs_epi32(cstep0, cstep1); + __m128i cstep23 = _mm_packs_epi32(cstep2, cstep3); + __m128i result = _mm_packs_epi16(cstep01, cstep23); + + /* Extract the sign bits + */ + mask &= ~_mm_movemask_epi8(result); + } + + if (mask) + lp_rast_shade_quads_mask(task, &tri->inputs, x, y, mask); + } +} +#endif + + + #undef TAG +#undef TRI_4 +#undef TRI_16 #undef NR_PLANES