* 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
-/* Render a 4x4 unmasked block:
+/**
+ * Shade all pixels in a 4x4 block.
*/
-static void block_full_4( struct lp_rasterizer *rast,
- int x, int y )
+static void
+block_full_4(struct lp_rasterizer_task *task,
+ const struct lp_rast_triangle *tri,
+ int x, int y)
{
- int i = rast->nr_blocks;
- 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);
}
-static void block_full_16( struct lp_rasterizer *rast,
- int x, int y )
+/**
+ * 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)
{
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);
+}
+
+#if !defined(PIPE_ARCH_SSE)
+static INLINE unsigned
+build_mask(int c, int dcdx, int dcdy)
+{
+ int mask = 0;
- for (iy = 0; iy < 16; iy+=4)
- for (ix = 0; ix < 16; ix+=4)
- block_full_4(rast, x + ix, y + iy);
+ int c0 = c;
+ int c1 = c0 + dcdx;
+ int c2 = c1 + dcdx;
+ int c3 = c2 + dcdx;
+
+ mask |= ((c0 + 0 * dcdy) >> 31) & (1 << 0);
+ mask |= ((c0 + 1 * dcdy) >> 31) & (1 << 2);
+ mask |= ((c0 + 2 * dcdy) >> 31) & (1 << 8);
+ mask |= ((c0 + 3 * dcdy) >> 31) & (1 << 10);
+ mask |= ((c1 + 0 * dcdy) >> 31) & (1 << 1);
+ mask |= ((c1 + 1 * dcdy) >> 31) & (1 << 3);
+ mask |= ((c1 + 2 * dcdy) >> 31) & (1 << 9);
+ mask |= ((c1 + 3 * dcdy) >> 31) & (1 << 11);
+ mask |= ((c2 + 0 * dcdy) >> 31) & (1 << 4);
+ mask |= ((c2 + 1 * dcdy) >> 31) & (1 << 6);
+ mask |= ((c2 + 2 * dcdy) >> 31) & (1 << 12);
+ mask |= ((c2 + 3 * dcdy) >> 31) & (1 << 14);
+ mask |= ((c3 + 0 * dcdy) >> 31) & (1 << 5);
+ mask |= ((c3 + 1 * dcdy) >> 31) & (1 << 7);
+ mask |= ((c3 + 2 * dcdy) >> 31) & (1 << 13);
+ mask |= ((c3 + 3 * dcdy) >> 31) & (1 << 15);
+
+ return mask;
}
+static INLINE unsigned
+build_mask_linear(int c, int dcdx, int dcdy)
+{
+ int mask = 0;
-/* Evaluate each pixel in a block, generate a mask and possibly render
- * the quad:
- */
-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 )
+ int c0 = c;
+ int c1 = c0 + dcdy;
+ int c2 = c1 + dcdy;
+ int c3 = c2 + dcdy;
+
+ mask |= ((c0 + 0 * dcdx) >> 31) & (1 << 0);
+ mask |= ((c0 + 1 * dcdx) >> 31) & (1 << 1);
+ mask |= ((c0 + 2 * dcdx) >> 31) & (1 << 2);
+ mask |= ((c0 + 3 * dcdx) >> 31) & (1 << 3);
+ mask |= ((c1 + 0 * dcdx) >> 31) & (1 << 4);
+ mask |= ((c1 + 1 * dcdx) >> 31) & (1 << 5);
+ mask |= ((c1 + 2 * dcdx) >> 31) & (1 << 6);
+ mask |= ((c1 + 3 * dcdx) >> 31) & (1 << 7);
+ mask |= ((c2 + 0 * dcdx) >> 31) & (1 << 8);
+ mask |= ((c2 + 1 * dcdx) >> 31) & (1 << 9);
+ mask |= ((c2 + 2 * dcdx) >> 31) & (1 << 10);
+ mask |= ((c2 + 3 * dcdx) >> 31) & (1 << 11);
+ mask |= ((c3 + 0 * dcdx) >> 31) & (1 << 12);
+ mask |= ((c3 + 1 * dcdx) >> 31) & (1 << 13);
+ mask |= ((c3 + 2 * dcdx) >> 31) & (1 << 14);
+ mask |= ((c3 + 3 * dcdx) >> 31) & (1 << 15);
+
+ return mask;
+}
+
+
+static INLINE void
+build_masks(int c,
+ int cdiff,
+ int dcdx,
+ int dcdy,
+ unsigned *outmask,
+ unsigned *partmask)
{
- int i;
- unsigned mask = 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) {
- int i = rast->nr_blocks;
- rast->blocks[i].x = x;
- rast->blocks[i].y = y;
- rast->blocks[i].mask = mask;
- rast->nr_blocks++;
- }
+ *outmask |= build_mask_linear(c, dcdx, dcdy);
+ *partmask |= build_mask_linear(c + cdiff, dcdx, dcdy);
}
-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 )
+#else
+#include <emmintrin.h>
+#include "util/u_sse.h"
+
+
+static INLINE void
+build_masks(int c,
+ int cdiff,
+ int dcdx,
+ int dcdy,
+ unsigned *outmask,
+ unsigned *partmask)
{
- int ix,iy,i = 0;
+ __m128i cstep0 = _mm_setr_epi32(c, c+dcdx, c+dcdx*2, c+dcdx*3);
+ __m128i xdcdy = _mm_set1_epi32(dcdy);
+
+ /* Get values across the quad
+ */
+ __m128i cstep1 = _mm_add_epi32(cstep0, xdcdy);
+ __m128i cstep2 = _mm_add_epi32(cstep1, xdcdy);
+ __m128i cstep3 = _mm_add_epi32(cstep2, xdcdy);
+
+ {
+ __m128i cstep01, cstep23, result;
+
+ cstep01 = _mm_packs_epi32(cstep0, cstep1);
+ cstep23 = _mm_packs_epi32(cstep2, cstep3);
+ result = _mm_packs_epi16(cstep01, cstep23);
- int ei1 = tri->ei1 << 2;
- int ei2 = tri->ei2 << 2;
- int ei3 = tri->ei3 << 2;
+ *outmask |= _mm_movemask_epi8(result);
+ }
- int eo1 = tri->eo1 << 2;
- int eo2 = tri->eo2 << 2;
- int eo3 = tri->eo3 << 2;
- for (iy = 0; iy < 16; iy+=4)
{
- for (ix = 0; ix < 16; ix+=4, i++)
- {
- int cx1 = c1 + (tri->step[0][i] << 2);
- int cx2 = c2 + (tri->step[1][i] << 2);
- int cx3 = c3 + (tri->step[2][i] << 2);
-
- if (cx1 + eo1 < 0 ||
- cx2 + eo2 < 0 ||
- cx3 + eo3 < 0)
- {
- }
- else if (cx1 + ei1 > 0 &&
- cx2 + ei2 > 0 &&
- cx3 + ei3 > 0)
- {
- block_full_4(rast, x+ix, y+iy); /* trivial accept */
- }
- else
- {
- do_block_4(rast, tri, x+ix, y+iy, cx1, cx2, cx3);
- }
- }
+ __m128i cio4 = _mm_set1_epi32(cdiff);
+ __m128i cstep01, cstep23, result;
+
+ cstep0 = _mm_add_epi32(cstep0, cio4);
+ cstep1 = _mm_add_epi32(cstep1, cio4);
+ cstep2 = _mm_add_epi32(cstep2, cio4);
+ cstep3 = _mm_add_epi32(cstep3, cio4);
+
+ cstep01 = _mm_packs_epi32(cstep0, cstep1);
+ cstep23 = _mm_packs_epi32(cstep2, cstep3);
+ result = _mm_packs_epi16(cstep01, cstep23);
+
+ *partmask |= _mm_movemask_epi8(result);
}
}
-/* 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 )
+
+static INLINE unsigned
+build_mask_linear(int c, int dcdx, int dcdy)
{
- const struct lp_rast_triangle *tri = arg.triangle;
+ __m128i cstep0 = _mm_setr_epi32(c, c+dcdx, c+dcdx*2, c+dcdx*3);
+ __m128i xdcdy = _mm_set1_epi32(dcdy);
+
+ /* Get values across the quad
+ */
+ __m128i cstep1 = _mm_add_epi32(cstep0, xdcdy);
+ __m128i cstep2 = _mm_add_epi32(cstep1, xdcdy);
+ __m128i cstep3 = _mm_add_epi32(cstep2, xdcdy);
- int x = rast->x;
- int y = rast->y;
- int ix,iy,i = 0;
+ /* pack pairs of results into epi16
+ */
+ __m128i cstep01 = _mm_packs_epi32(cstep0, cstep1);
+ __m128i cstep23 = _mm_packs_epi32(cstep2, cstep3);
- 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;
+ /* pack into epi8, preserving sign bits
+ */
+ __m128i result = _mm_packs_epi16(cstep01, cstep23);
- int ei1 = tri->ei1 << 4;
- int ei2 = tri->ei2 << 4;
- int ei3 = tri->ei3 << 4;
+ /* extract sign bits to create mask
+ */
+ return _mm_movemask_epi8(result);
+}
- int eo1 = tri->eo1 << 4;
- int eo2 = tri->eo2 << 4;
- int eo3 = tri->eo3 << 4;
+static INLINE unsigned
+build_mask(int c, int dcdx, int dcdy)
+{
+ __m128i step = _mm_setr_epi32(0, dcdx, dcdy, dcdx + dcdy);
+ __m128i c0 = _mm_set1_epi32(c);
- debug_printf("%s\n", __FUNCTION__);
+ /* Get values across the quad
+ */
+ __m128i cstep0 = _mm_add_epi32(c0, step);
- rast->nr_blocks = 0;
+ /* Scale up step for moving between quads.
+ */
+ __m128i step4 = _mm_add_epi32(step, step);
+
+ /* Get values for the remaining quads:
+ */
+ __m128i cstep1 = _mm_add_epi32(cstep0,
+ _mm_shuffle_epi32(step4, _MM_SHUFFLE(1,1,1,1)));
+ __m128i cstep2 = _mm_add_epi32(cstep0,
+ _mm_shuffle_epi32(step4, _MM_SHUFFLE(2,2,2,2)));
+ __m128i cstep3 = _mm_add_epi32(cstep2,
+ _mm_shuffle_epi32(step4, _MM_SHUFFLE(1,1,1,1)));
+
+ /* pack pairs of results into epi16
+ */
+ __m128i cstep01 = _mm_packs_epi32(cstep0, cstep1);
+ __m128i cstep23 = _mm_packs_epi32(cstep2, cstep3);
+
+ /* pack into epi8, preserving sign bits
+ */
+ __m128i result = _mm_packs_epi16(cstep01, cstep23);
+
+ /* extract sign bits to create mask
+ */
+ return _mm_movemask_epi8(result);
+}
+
+#endif
+
+
+
+
+#define TAG(x) x##_1
+#define NR_PLANES 1
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_2
+#define NR_PLANES 2
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_3
+#define NR_PLANES 3
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_4
+#define NR_PLANES 4
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_5
+#define NR_PLANES 5
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_6
+#define NR_PLANES 6
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_7
+#define NR_PLANES 7
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_8
+#define NR_PLANES 8
+#include "lp_rast_tri_tmp.h"
+
+
+/* Special case for 3 plane triangle which is contained entirely
+ * within a 16x16 block.
+ */
+void
+lp_rast_triangle_3_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 = tri->plane;
+ unsigned mask = arg.triangle.plane_mask;
+ const int x = task->x + (mask & 0xf) * 16;
+ const int y = task->y + (mask >> 4) * 16;
+ unsigned outmask, inmask, partmask, partial_mask;
+ unsigned j;
+ int c[3];
+
+ outmask = 0; /* outside one or more trivial reject planes */
+ partmask = 0; /* outside one or more trivial accept planes */
+
+ for (j = 0; j < 3; j++) {
+ c[j] = plane[j].c + plane[j].dcdy * y - plane[j].dcdx * x;
- for (iy = 0; iy < 64; iy+=16)
- {
- for (ix = 0; ix < 64; ix+=16, 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)
- {
- }
- else if (cx1 + ei1 > 0 &&
- cx2 + ei2 > 0 &&
- cx3 + ei3 > 0)
- {
- block_full_16(rast, x+ix, y+iy); /* trivial accept */
- }
- else
- {
- do_block_16(rast, tri, x+ix, y+iy, cx1, cx2, cx3);
- }
+ 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 */
}
}
- 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);
-}
+ if (outmask == 0xffff)
+ return;
+ /* Mask of sub-blocks which are inside all trivial accept planes:
+ */
+ inmask = ~partmask & 0xffff;
+
+ /* Mask of sub-blocks which are inside all trivial reject planes,
+ * but outside at least one trivial accept plane:
+ */
+ partial_mask = partmask & ~outmask;
+
+ assert((partial_mask & inmask) == 0);
+
+ /* 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;
+ int cx[3];
+
+ partial_mask &= ~(1 << i);
+
+ for (j = 0; j < 3; j++)
+ cx[j] = (c[j]
+ - plane[j].dcdx * ix
+ + plane[j].dcdy * iy);
+
+ do_block_4_3(task, tri, plane, px, py, cx);
+ }
+
+ /* Iterate over fulls:
+ */
+ while (inmask) {
+ int i = ffs(inmask) - 1;
+ int ix = (i & 3) * 4;
+ int iy = (i >> 2) * 4;
+ int px = x + ix;
+ int py = y + iy;
+
+ inmask &= ~(1 << i);
+
+ block_full_4(task, tri, px, py);
+ }
+}