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 = 0xffff;
int j;
+#ifndef MULTISAMPLE
+ unsigned mask = 0xffff;
+#else
+ uint64_t mask = UINT64_MAX;
+#endif
for (j = 0; j < NR_PLANES; j++) {
- mask &= ~build_mask_linear(c[j] - 1,
- -plane[j].dcdx,
- plane[j].dcdy);
+#ifndef MULTISAMPLE
+#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
+#else
+ for (unsigned s = 0; s < 4; s++) {
+ int64_t new_c = (c[j]) + ((IMUL64(task->scene->fixed_sample_pos[s][1], plane[j].dcdy) + IMUL64(task->scene->fixed_sample_pos[s][0], -plane[j].dcdx)) >> FIXED_ORDER);
+ uint32_t build_mask;
+#ifdef RASTER_64
+ build_mask = BUILD_MASK_LINEAR((new_c - 1) >> (int64_t)FIXED_ORDER,
+ -plane[j].dcdx >> FIXED_ORDER,
+ plane[j].dcdy >> FIXED_ORDER);
+#else
+ build_mask = BUILD_MASK_LINEAR((new_c - 1),
+ -plane[j].dcdx,
+ plane[j].dcdy);
+#endif
+ mask &= ~((uint64_t)build_mask << (s * 16));
+ }
+#endif
}
/* Now pass to the shader:
*/
if (mask)
- lp_rast_shade_quads_mask(task, &tri->inputs, x, y, mask);
+ lp_rast_shade_quads_mask_sample(task, &tri->inputs, x, y, mask);
}
/**
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 j;
partmask = 0; /* outside one or more trivial accept planes */
for (j = 0; j < NR_PLANES; j++) {
- const int dcdx = -plane[j].dcdx * 4;
- const int dcdy = plane[j].dcdy * 4;
- const int cox = plane[j].eo * 4;
- const int cio = plane[j].ei * 4 - 1;
-
- build_masks(c[j] + cox,
- cio - cox,
- dcdx, dcdy,
- &outmask, /* sign bits from c[i][0..15] + cox */
- &partmask); /* sign bits from c[i][0..15] + cio */
+#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)
int iy = (i >> 2) * 4;
int px = x + ix;
int py = y + iy;
- int cx[NR_PLANES];
+ int64_t cx[NR_PLANES];
partial_mask &= ~(1 << i);
for (j = 0; j < NR_PLANES; j++)
cx[j] = (c[j]
- - plane[j].dcdx * ix
- + plane[j].dcdy * iy);
+ - IMUL64(plane[j].dcdx, ix)
+ + IMUL64(plane[j].dcdy, iy));
TAG(do_block_4)(task, tri, plane, px, py, cx);
}
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 j = 0;
int i = ffs(plane_mask) - 1;
plane[j] = tri_plane[i];
plane_mask &= ~(1 << i);
- c[j] = plane[j].c + plane[j].dcdy * y - plane[j].dcdx * x;
+ c[j] = plane[j].c + IMUL64(plane[j].dcdy, y) - IMUL64(plane[j].dcdx, x);
{
- const int dcdx = -plane[j].dcdx * 16;
- const int dcdy = plane[j].dcdy * 16;
- const int cox = plane[j].eo * 16;
- const int cio = plane[j].ei * 16 - 1;
-
- build_masks(c[j] + cox,
- cio - cox,
- dcdx, dcdy,
- &outmask, /* sign bits from c[i][0..15] + cox */
- &partmask); /* sign bits from c[i][0..15] + cio */
+#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++;
int iy = (i >> 2) * 16;
int px = x + ix;
int py = y + iy;
- int cx[NR_PLANES];
+ int64_t cx[NR_PLANES];
for (j = 0; j < NR_PLANES; j++)
cx[j] = (c[j]
- - plane[j].dcdx * ix
- + plane[j].dcdy * iy);
+ - IMUL64(plane[j].dcdx, ix)
+ + IMUL64(plane[j].dcdy, iy));
partial_mask &= ~(1 << i);