* Rasterization for binned triangles within a tile
*/
+#include <limits.h>
#include "util/u_math.h"
+#include "lp_debug.h"
+#include "lp_perf.h"
#include "lp_rast_priv.h"
#include "lp_tile_soa.h"
-#define BLOCKSIZE 8
-
-
-/* Convert 8x8 block into four runs of quads and render each in turn.
+/**
+ * Map an index in [0,15] to an x,y position, multiplied by 4.
+ * This is used to get the position of each subtile in a 4x4
+ * grid of edge step values.
+ * Note: we can use some bit twiddling to compute these values instead
+ * of using a look-up table, but there's no measurable performance
+ * difference.
+ */
+static const int pos_table4[16][2] = {
+ { 0, 0 },
+ { 4, 0 },
+ { 0, 4 },
+ { 4, 4 },
+ { 8, 0 },
+ { 12, 0 },
+ { 8, 4 },
+ { 12, 4 },
+ { 0, 8 },
+ { 4, 8 },
+ { 0, 12 },
+ { 4, 12 },
+ { 8, 8 },
+ { 12, 8 },
+ { 8, 12 },
+ { 12, 12 }
+};
+
+
+static const int pos_table16[16][2] = {
+ { 0, 0 },
+ { 16, 0 },
+ { 0, 16 },
+ { 16, 16 },
+ { 32, 0 },
+ { 48, 0 },
+ { 32, 16 },
+ { 48, 16 },
+ { 0, 32 },
+ { 16, 32 },
+ { 0, 48 },
+ { 16, 48 },
+ { 32, 32 },
+ { 48, 32 },
+ { 32, 48 },
+ { 48, 48 }
+};
+
+
+/**
+ * Shade all pixels in a 4x4 block.
*/
-#if (BLOCKSIZE == 8)
-static void block_full( struct lp_rasterizer *rast,
- const struct lp_rast_triangle *tri,
- int x, int y )
+static void
+block_full_4(struct lp_rasterizer_task *task,
+ const struct lp_rast_triangle *tri,
+ int x, int y)
{
- const unsigned masks[4] = {~0, ~0, ~0, ~0};
- int iy;
-
- for (iy = 0; iy < 8; iy += 2)
- lp_rast_shade_quads(rast, &tri->inputs, x, y + iy, masks);
+ lp_rast_shade_quads_all(task, &tri->inputs, x, y);
}
-#else
-static void block_full( struct lp_rasterizer *rast,
- const struct lp_rast_triangle *tri,
- int x, int y )
-{
- const unsigned masks[4] = {~0, ~0, 0, 0}; /* FIXME: Wasting quads!!! */
- int iy;
- for (iy = 0; iy < 4; iy += 2)
- lp_rast_shade_quads(rast, &tri->inputs, x, y + iy, masks);
-}
-#endif
-static INLINE unsigned
-do_quad( const struct lp_rast_triangle *tri,
- int x, int y,
- float c1, float c2, float c3 )
+/**
+ * Shade all pixels in a 16x16 block.
+ */
+static void
+block_full_16(struct lp_rasterizer_task *task,
+ const struct lp_rast_triangle *tri,
+ int x, int y)
{
- float xstep1 = -tri->dy12;
- float xstep2 = -tri->dy23;
- float xstep3 = -tri->dy31;
-
- float ystep1 = tri->dx12;
- float ystep2 = tri->dx23;
- float ystep3 = tri->dx31;
-
- unsigned mask = 0;
-
- if (c1 > 0 &&
- c2 > 0 &&
- c3 > 0)
- mask |= 1;
-
- if (c1 + xstep1 > 0 &&
- c2 + xstep2 > 0 &&
- c3 + xstep3 > 0)
- mask |= 2;
-
- if (c1 + ystep1 > 0 &&
- c2 + ystep2 > 0 &&
- c3 + ystep3 > 0)
- mask |= 4;
-
- if (c1 + ystep1 + xstep1 > 0 &&
- c2 + ystep2 + xstep2 > 0 &&
- c3 + ystep3 + xstep3 > 0)
- mask |= 8;
-
- return mask;
+ unsigned ix, iy;
+ assert(x % 16 == 0);
+ assert(y % 16 == 0);
+ for (iy = 0; iy < 16; iy += 4)
+ for (ix = 0; ix < 16; ix += 4)
+ block_full_4(task, tri, x + ix, y + iy);
}
-/* Evaluate each pixel in a block, generate a mask and possibly render
- * the quad:
+
+/**
+ * Pass the 4x4 pixel block to the shader function.
+ * Determination of which of the 16 pixels lies inside the triangle
+ * will be done as part of the fragment shader.
*/
static void
-do_block( struct lp_rasterizer *rast,
- const struct lp_rast_triangle *tri,
- int x, int y,
- float c1,
- float c2,
- float c3 )
+do_block_4(struct lp_rasterizer_task *task,
+ const struct lp_rast_triangle *tri,
+ int x, int y,
+ int c1, int c2, int c3)
{
- const int step = 2;
+ assert(x >= 0);
+ assert(y >= 0);
- float xstep1 = -step * tri->dy12;
- float xstep2 = -step * tri->dy23;
- float xstep3 = -step * tri->dy31;
+ lp_rast_shade_quads(task, &tri->inputs, x, y, -c1, -c2, -c3);
+}
- float ystep1 = step * tri->dx12;
- float ystep2 = step * tri->dx23;
- float ystep3 = step * tri->dx31;
- int ix, iy;
+/**
+ * Evaluate a 16x16 block of pixels to determine which 4x4 subblocks are in/out
+ * of the triangle's bounds.
+ */
+static void
+do_block_16(struct lp_rasterizer_task *task,
+ const struct lp_rast_triangle *tri,
+ int x, int y,
+ int c0, int c1, int c2)
+{
+ unsigned mask = 0;
+ int eo[3];
+ int c[3];
+ int i, j;
- for (iy = 0; iy < BLOCKSIZE; iy += 2) {
- float cx1 = c1;
- float cx2 = c2;
- float cx3 = c3;
+ assert(x >= 0);
+ assert(y >= 0);
+ assert(x % 16 == 0);
+ assert(y % 16 == 0);
- unsigned masks[4] = {0, 0, 0, 0};
+ eo[0] = tri->eo1 * 4;
+ eo[1] = tri->eo2 * 4;
+ eo[2] = tri->eo3 * 4;
- for (ix = 0; ix < BLOCKSIZE; ix += 2) {
+ c[0] = c0;
+ c[1] = c1;
+ c[2] = c2;
- masks[ix >> 1] = do_quad(tri, x + ix, y + iy, cx1, cx2, cx3);
+ for (j = 0; j < 3; j++) {
+ const int *step = tri->inputs.step[j];
+ const int cx = c[j] + eo[j];
- cx1 += xstep1;
- cx2 += xstep2;
- cx3 += xstep3;
+ /* Mask has bits set whenever we are outside any of the edges.
+ */
+ for (i = 0; i < 16; i++) {
+ int out = cx + step[i] * 4;
+ mask |= (out >> 31) & (1 << i);
}
+ }
- lp_rast_shade_quads(rast, &tri->inputs, x, y + iy, masks);
+ mask = ~mask & 0xffff;
+ while (mask) {
+ int i = ffs(mask) - 1;
+ int px = x + pos_table4[i][0];
+ int py = y + pos_table4[i][1];
+ int cx1 = c0 + tri->inputs.step[0][i] * 4;
+ int cx2 = c1 + tri->inputs.step[1][i] * 4;
+ int cx3 = c2 + tri->inputs.step[2][i] * 4;
- c1 += ystep1;
- c2 += ystep2;
- c3 += ystep3;
- }
+ mask &= ~(1 << i);
+ /* Don't bother testing if the 4x4 block is entirely in/out of
+ * the triangle. It's a little faster to do it in the jit code.
+ */
+ LP_COUNT(nr_non_empty_4);
+ do_block_4(task, tri, px, py, cx1, cx2, cx3);
+ }
}
-
-/* Scan the tile in chunks and figure out which pixels to rasterize
- * for this triangle:
+/**
+ * Scan the tile in chunks and figure out which pixels to rasterize
+ * for this triangle.
*/
-void lp_rast_triangle( struct lp_rasterizer *rast,
- const union lp_rast_cmd_arg arg )
+void
+lp_rast_triangle(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
{
const struct lp_rast_triangle *tri = arg.triangle;
+ const int x = task->x, y = task->y;
+ int ei[3], eo[3], c[3];
+ unsigned outmask, inmask, partial_mask;
+ unsigned i, j;
- const int step = BLOCKSIZE;
+ c[0] = tri->c1 + tri->dx12 * y - tri->dy12 * x;
+ c[1] = tri->c2 + tri->dx23 * y - tri->dy23 * x;
+ c[2] = tri->c3 + tri->dx31 * y - tri->dy31 * x;
- float ei1 = tri->ei1 * step;
- float ei2 = tri->ei2 * step;
- float ei3 = tri->ei3 * step;
+ eo[0] = tri->eo1 * 16;
+ eo[1] = tri->eo2 * 16;
+ eo[2] = tri->eo3 * 16;
- float eo1 = tri->eo1 * step;
- float eo2 = tri->eo2 * step;
- float eo3 = tri->eo3 * step;
+ ei[0] = tri->ei1 * 16;
+ ei[1] = tri->ei2 * 16;
+ ei[2] = tri->ei3 * 16;
- float xstep1 = -step * tri->dy12;
- float xstep2 = -step * tri->dy23;
- float xstep3 = -step * tri->dy31;
+ outmask = 0;
+ inmask = 0xffff;
- float ystep1 = step * tri->dx12;
- float ystep2 = step * tri->dx23;
- float ystep3 = step * tri->dx31;
+ for (j = 0; j < 3; j++) {
+ const int *step = tri->inputs.step[j];
+ const int cox = c[j] + eo[j];
+ const int cio = ei[j]- eo[j];
- /* Clamp to tile dimensions:
- */
- int minx = MAX2(tri->minx, rast->x);
- int miny = MAX2(tri->miny, rast->y);
- int maxx = MIN2(tri->maxx, rast->x + TILE_SIZE);
- int maxy = MIN2(tri->maxy, rast->y + TILE_SIZE);
-
- int x, y;
- float x0, y0;
- float c1, c2, c3;
+ /* Outmask has bits set whenever we are outside any of the
+ * edges.
+ */
+ /* Inmask has bits set whenever we are inside all of the edges.
+ */
+ for (i = 0; i < 16; i++) {
+ int out = cox + step[i] * 16;
+ int in = out + cio;
+ outmask |= (out >> 31) & (1 << i);
+ inmask &= ~((in >> 31) & (1 << i));
+ }
+ }
- debug_printf("%s\n", __FUNCTION__);
+ assert((outmask & inmask) == 0);
- if (miny == maxy || minx == maxx) {
- debug_printf("%s: non-intersecting triangle in bin\n", __FUNCTION__);
+ if (outmask == 0xffff)
return;
+
+ /* Invert mask, so that bits are set whenever we are at least
+ * partially inside all of the edges:
+ */
+ partial_mask = ~inmask & ~outmask & 0xffff;
+
+ /* 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 cx1 = c[0] + tri->inputs.step[0][i] * 16;
+ int cx2 = c[1] + tri->inputs.step[1][i] * 16;
+ int cx3 = c[2] + tri->inputs.step[2][i] * 16;
+
+ partial_mask &= ~(1 << i);
+
+ LP_COUNT(nr_partially_covered_16);
+ do_block_16(task, tri, px, py, cx1, cx2, cx3);
}
- minx &= ~(step-1);
- miny &= ~(step-1);
-
- x0 = (float)minx;
- y0 = (float)miny;
-
- c1 = tri->c1 + tri->dx12 * y0 - tri->dy12 * x0;
- c2 = tri->c2 + tri->dx23 * y0 - tri->dy23 * x0;
- c3 = tri->c3 + tri->dx31 * y0 - tri->dy31 * x0;
-
- for (y = miny; y < maxy; y += step)
- {
- float cx1 = c1;
- float cx2 = c2;
- float cx3 = c3;
-
- for (x = minx; x < maxx; x += step)
- {
- if (cx1 + eo1 < 0 ||
- cx2 + eo2 < 0 ||
- cx3 + eo3 < 0)
- {
- }
- else if (cx1 + ei1 > 0 &&
- cx2 + ei2 > 0 &&
- cx3 + ei3 > 0)
- {
- block_full(rast, tri, x, y); /* trivial accept */
- }
- else
- {
- do_block(rast, tri, x, y, cx1, cx2, cx3);
- }
-
- /* Iterate cx values across the region:
- */
- cx1 += xstep1;
- cx2 += xstep2;
- cx3 += xstep3;
- }
+ /* Iterate over fulls:
+ */
+ while (inmask) {
+ int i = ffs(inmask) - 1;
+ int px = x + pos_table16[i][0];
+ int py = y + pos_table16[i][1];
- /* Iterate c values down the region:
- */
- c1 += ystep1;
- c2 += ystep2;
- c3 += ystep3;
+ inmask &= ~(1 << i);
+
+ LP_COUNT(nr_fully_covered_16);
+ block_full_16(task, tri, px, py);
}
}
-