* 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.
+/**
+ * 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);
+
+/**
+ * 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);
}
-#endif
+#if !defined(PIPE_ARCH_SSE)
static INLINE unsigned
-do_quad( const struct lp_rast_triangle *tri,
- int x, int y,
- float c1, float c2, float c3 )
+build_mask(int c, int dcdx, int dcdy)
{
- 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;
+ int mask = 0;
+
+ 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;
+
+ 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;
}
-/* Evaluate each pixel in a block, generate a mask and possibly render
- * the quad:
- */
-static void
-do_block( struct lp_rasterizer *rast,
- const struct lp_rast_triangle *tri,
- int x, int y,
- float c1,
- float c2,
- float c3 )
+
+static INLINE void
+build_masks(int c,
+ int cdiff,
+ int dcdx,
+ int dcdy,
+ unsigned *outmask,
+ unsigned *partmask)
{
- const int step = 2;
+ *outmask |= build_mask_linear(c, dcdx, dcdy);
+ *partmask |= build_mask_linear(c + cdiff, dcdx, dcdy);
+}
- float xstep1 = -step * tri->dy12;
- float xstep2 = -step * tri->dy23;
- float xstep3 = -step * tri->dy31;
+#else
+#include <emmintrin.h>
+#include "util/u_sse.h"
- float ystep1 = step * tri->dx12;
- float ystep2 = step * tri->dx23;
- float ystep3 = step * tri->dx31;
- int ix, iy;
+static INLINE void
+build_masks(int c,
+ int cdiff,
+ int dcdx,
+ int dcdy,
+ unsigned *outmask,
+ unsigned *partmask)
+{
+ __m128i cstep0 = _mm_setr_epi32(c, c+dcdx, c+dcdx*2, c+dcdx*3);
+ __m128i xdcdy = _mm_set1_epi32(dcdy);
- for (iy = 0; iy < BLOCKSIZE; iy += 2) {
- float cx1 = c1;
- float cx2 = c2;
- float cx3 = c3;
+ /* 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);
- unsigned masks[4] = {0, 0, 0, 0};
+ {
+ __m128i cstep01, cstep23, result;
- for (ix = 0; ix < BLOCKSIZE; ix += 2) {
+ cstep01 = _mm_packs_epi32(cstep0, cstep1);
+ cstep23 = _mm_packs_epi32(cstep2, cstep3);
+ result = _mm_packs_epi16(cstep01, cstep23);
- masks[ix >> 1] = do_quad(tri, x + ix, y + iy, cx1, cx2, cx3);
+ *outmask |= _mm_movemask_epi8(result);
+ }
- cx1 += xstep1;
- cx2 += xstep2;
- cx3 += xstep3;
- }
- if(masks[0] || masks[1] || masks[2] || masks[3])
- lp_rast_shade_quads(rast, &tri->inputs, x, y + iy, masks);
+ {
+ __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);
- c1 += ystep1;
- c2 += ystep2;
- c3 += ystep3;
- }
+ 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);
+
+ /* 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);
+}
- const int step = BLOCKSIZE;
+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);
- float ei1 = tri->ei1 * step;
- float ei2 = tri->ei2 * step;
- float ei3 = tri->ei3 * step;
+ /* Get values across the quad
+ */
+ __m128i cstep0 = _mm_add_epi32(c0, step);
- float eo1 = tri->eo1 * step;
- float eo2 = tri->eo2 * step;
- float eo3 = tri->eo3 * step;
+ /* Scale up step for moving between quads.
+ */
+ __m128i step4 = _mm_add_epi32(step, step);
- float xstep1 = -step * tri->dy12;
- float xstep2 = -step * tri->dy23;
- float xstep3 = -step * tri->dy31;
+ /* 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);
- float ystep1 = step * tri->dx12;
- float ystep2 = step * tri->dx23;
- float ystep3 = step * tri->dx31;
+ /* pack into epi8, preserving sign bits
+ */
+ __m128i result = _mm_packs_epi16(cstep01, cstep23);
- /* Clamp to tile dimensions:
+ /* extract sign bits to create mask
*/
- 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);
+ return _mm_movemask_epi8(result);
+}
+
+#endif
- int x, y;
- float x0, y0;
- float c1, c2, c3;
- debug_printf("%s\n", __FUNCTION__);
- if (miny == maxy || minx == maxx) {
- debug_printf("%s: non-intersecting triangle in bin\n", __FUNCTION__);
- return;
- }
- minx &= ~(step-1);
- miny &= ~(step-1);
+#define TAG(x) x##_1
+#define NR_PLANES 1
+#include "lp_rast_tri_tmp.h"
- x0 = (float)minx;
- y0 = (float)miny;
+#define TAG(x) x##_2
+#define NR_PLANES 2
+#include "lp_rast_tri_tmp.h"
- 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;
+#define TAG(x) x##_3
+#define NR_PLANES 3
+#include "lp_rast_tri_tmp.h"
- for (y = miny; y < maxy; y += step)
- {
- float cx1 = c1;
- float cx2 = c2;
- float cx3 = c3;
+#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 (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;
+ 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 */
}
+ }
- /* Iterate c values down the region:
- */
- c1 += ystep1;
- c2 += ystep2;
- c3 += ystep3;
+ 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);
+ }
+}