* 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 4
+/**
+ * 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 }
+};
/**
- * Add a 4x4 block of pixels to the block list.
- * All pixels are known to be inside the triangle's bounds.
+ * Shade all pixels in a 4x4 block.
*/
static void
-block_full_4( struct lp_rasterizer *rast, int x, int y )
+block_full_4(struct lp_rasterizer_task *task,
+ const struct lp_rast_triangle *tri,
+ int x, int y)
{
- const unsigned i = rast->nr_blocks;
- assert(x % 4 == 0);
- assert(y % 4 == 0);
- rast->blocks[i].x = x;
- rast->blocks[i].y = y;
- rast->blocks[i].mask = ~0;
- rast->nr_blocks++;
+ lp_rast_shade_quads_all(task, &tri->inputs, x, y);
}
/**
- * Add a 16x16 block of pixels to the block list.
- * All pixels are known to be inside the triangle's bounds.
+ * Shade all pixels in a 16x16 block.
*/
static void
-block_full_16( struct lp_rasterizer *rast, int x, int y )
+block_full_16(struct lp_rasterizer_task *task,
+ const struct lp_rast_triangle *tri,
+ int x, int y)
{
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(rast, x + ix, y + iy);
+ block_full_4(task, tri, x + ix, y + iy);
}
/**
- * Evaluate each pixel in a 4x4 block to determine if it lies within
- * the triangle's bounds.
- * Generate a mask of in/out flags and add the block to the blocks list.
+ * 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_4( struct lp_rasterizer *rast,
- const struct lp_rast_triangle *tri,
- int x, int y,
- int c1,
- int c2,
- int 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)
{
- int i;
- unsigned mask = 0;
-
- assert(x % 4 == 0);
- assert(y % 4 == 0);
+ assert(x >= 0);
+ assert(y >= 0);
- for (i = 0; i < 16; i++)
- mask |= (~(((c1 + tri->step[0][i]) |
- (c2 + tri->step[1][i]) |
- (c3 + tri->step[2][i])) >> 31)) & (1 << i);
-
- /* As we do trivial reject already, masks should rarely be all zero:
- */
- if (mask) {
- const unsigned i = rast->nr_blocks;
- rast->blocks[i].x = x;
- rast->blocks[i].y = y;
- rast->blocks[i].mask = mask;
- rast->nr_blocks++;
- }
+ lp_rast_shade_quads(task, &tri->inputs, x, y, -c1, -c2, -c3);
}
* of the triangle's bounds.
*/
static void
-do_block_16( struct lp_rasterizer *rast,
- const struct lp_rast_triangle *tri,
- int x, int y,
- int c1,
- int c2,
- int c3 )
+do_block_16(struct lp_rasterizer_task *task,
+ const struct lp_rast_triangle *tri,
+ int x, int y,
+ int c0, int c1, int c2)
{
- int ix, iy, i = 0;
-
- int ei1 = tri->ei1 * 4;
- int ei2 = tri->ei2 * 4;
- int ei3 = tri->ei3 * 4;
-
- int eo1 = tri->eo1 * 4;
- int eo2 = tri->eo2 * 4;
- int eo3 = tri->eo3 * 4;
+ unsigned mask = 0;
+ int eo[3];
+ int c[3];
+ int i, j;
+ assert(x >= 0);
+ assert(y >= 0);
assert(x % 16 == 0);
assert(y % 16 == 0);
- for (iy = 0; iy < 16; iy+=4) {
- for (ix = 0; ix < 16; ix+=4, i++) {
- int cx1 = c1 + (tri->step[0][i] * 4);
- int cx2 = c2 + (tri->step[1][i] * 4);
- int cx3 = c3 + (tri->step[2][i] * 4);
-
- if (cx1 + eo1 < 0 ||
- cx2 + eo2 < 0 ||
- cx3 + eo3 < 0) {
- /* the block is completely outside the triangle - nop */
- }
- else if (cx1 + ei1 > 0 &&
- cx2 + ei2 > 0 &&
- cx3 + ei3 > 0) {
- /* the block is completely inside the triangle */
- block_full_4(rast, x+ix, y+iy);
- }
- else {
- /* the block is partially in/out of the triangle */
- do_block_4(rast, tri, x+ix, y+iy, cx1, cx2, cx3);
- }
+ eo[0] = tri->eo1 * 4;
+ eo[1] = tri->eo2 * 4;
+ eo[2] = tri->eo3 * 4;
+
+ c[0] = c0;
+ c[1] = c1;
+ c[2] = c2;
+
+ for (j = 0; j < 3; j++) {
+ const int *step = tri->inputs.step[j];
+ const int cx = c[j] + eo[j];
+
+ /* 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);
}
}
+
+ 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;
+
+ 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);
+ }
}
* for this triangle.
*/
void
-lp_rast_triangle( struct lp_rasterizer *rast,
- const union lp_rast_cmd_arg arg )
+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;
+
+ 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;
- int x = rast->x;
- int y = rast->y;
- int ix, iy;
- unsigned i = 0;
+ eo[0] = tri->eo1 * 16;
+ eo[1] = tri->eo2 * 16;
+ eo[2] = tri->eo3 * 16;
- int c1 = tri->c1 + tri->dx12 * y - tri->dy12 * x;
- int c2 = tri->c2 + tri->dx23 * y - tri->dy23 * x;
- int c3 = tri->c3 + tri->dx31 * y - tri->dy31 * x;
+ ei[0] = tri->ei1 * 16;
+ ei[1] = tri->ei2 * 16;
+ ei[2] = tri->ei3 * 16;
- int ei1 = tri->ei1 * 16;
- int ei2 = tri->ei2 * 16;
- int ei3 = tri->ei3 * 16;
+ outmask = 0;
+ inmask = 0xffff;
- int eo1 = tri->eo1 * 16;
- int eo2 = tri->eo2 * 16;
- int eo3 = tri->eo3 * 16;
+ 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];
- debug_printf("%s\n", __FUNCTION__);
+ /* 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));
+ }
+ }
- rast->nr_blocks = 0;
+ assert((outmask & inmask) == 0);
- /* Walk over the tile to build a list of 4x4 pixel blocks which will
- * be filled/shaded. We do this at two granularities: 16x16 blocks
- * and then 4x4 blocks.
+ if (outmask == 0xffff)
+ return;
+
+ /* Invert mask, so that bits are set whenever we are at least
+ * partially inside all of the edges:
*/
- for (iy = 0; iy < TILE_SIZE; iy += 16) {
- for (ix = 0; ix < TILE_SIZE; ix += 16, i++) {
- int cx1 = c1 + (tri->step[0][i] * 16);
- int cx2 = c2 + (tri->step[1][i] * 16);
- int cx3 = c3 + (tri->step[2][i] * 16);
-
- if (cx1 + eo1 < 0 ||
- cx2 + eo2 < 0 ||
- cx3 + eo3 < 0) {
- /* the block is completely outside the triangle - nop */
- }
- else if (cx1 + ei1 > 0 &&
- cx2 + ei2 > 0 &&
- cx3 + ei3 > 0) {
- /* the block is completely inside the triangle */
- block_full_16(rast, x+ix, y+iy);
- }
- else {
- /* the block is partially in/out of the triangle */
- do_block_16(rast, tri, x+ix, y+iy, cx1, cx2, cx3);
- }
- }
+ 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);
}
- /* Shade the 4x4 pixel blocks */
- for (i = 0; i < rast->nr_blocks; i++)
- lp_rast_shade_quads(rast, &tri->inputs,
- rast->blocks[i].x,
- rast->blocks[i].y,
- rast->blocks[i].mask);
+ /* Iterate over fulls:
+ */
+ while (inmask) {
+ int i = ffs(inmask) - 1;
+ int px = x + pos_table16[i][0];
+ int py = y + pos_table16[i][1];
+
+ inmask &= ~(1 << i);
+
+ LP_COUNT(nr_fully_covered_16);
+ block_full_16(task, tri, px, py);
+ }
}