#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"
-/**
- * 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.
*/
static void
-block_full_4( struct lp_rasterizer_task *rast_task,
- const struct lp_rast_triangle *tri,
- int x, int y )
+block_full_4(struct lp_rasterizer_task *task,
+ const struct lp_rast_triangle *tri,
+ int x, int y)
{
- lp_rast_shade_quads_all(rast_task->rast,
- rast_task->thread_index,
- &tri->inputs,
- x, y);
+ lp_rast_shade_quads_all(task, &tri->inputs, x, y);
}
* Shade all pixels in a 16x16 block.
*/
static void
-block_full_16( struct lp_rasterizer_task *rast_task,
- const struct lp_rast_triangle *tri,
- 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_task, tri, x + ix, y + iy);
+ 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;
+
+ int c0 = c;
+ int c1 = c0 + dcdx;
+ int c2 = c1 + dcdx;
+ int c3 = c2 + dcdx;
-/**
- * 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_task *rast_task,
- const struct lp_rast_triangle *tri,
- int x, int y,
- int c1,
- int c2,
- int c3 )
+ 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)
{
- lp_rast_shade_quads(rast_task->rast,
- rast_task->thread_index,
- &tri->inputs,
- x, y,
- -c1, -c2, -c3);
+ 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 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 *rast_task,
- const struct lp_rast_triangle *tri,
- int x, int y,
- int c1,
- int c2,
- int c3 )
+static INLINE void
+build_masks(int c,
+ int cdiff,
+ int dcdx,
+ int dcdy,
+ unsigned *outmask,
+ unsigned *partmask)
{
- const int eo1 = tri->eo1 * 4;
- const int eo2 = tri->eo2 * 4;
- const int eo3 = tri->eo3 * 4;
+ *outmask |= build_mask_linear(c, dcdx, dcdy);
+ *partmask |= build_mask_linear(c + cdiff, dcdx, dcdy);
+}
- int i;
+#else
+#include <emmintrin.h>
+#include "util/u_sse.h"
- assert(x % 16 == 0);
- assert(y % 16 == 0);
- for (i = 0; i < 16; i++) {
- int cx1 = c1 + (tri->inputs.step[0][i] * 4);
- int cx2 = c2 + (tri->inputs.step[1][i] * 4);
- int cx3 = c3 + (tri->inputs.step[2][i] * 4);
+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);
+
+ /* 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);
- if (cx1 + eo1 < 0 ||
- cx2 + eo2 < 0 ||
- cx3 + eo3 < 0) {
- /* the block is completely outside the triangle - nop */
- }
- else {
- int px = x + pos_table4[i][0];
- int py = y + pos_table4[i][1];
- /* 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.
- */
- do_block_4(rast_task, tri, px, py, cx1, cx2, cx3);
- }
+ {
+ __m128i cstep01, cstep23, result;
+
+ cstep01 = _mm_packs_epi32(cstep0, cstep1);
+ cstep23 = _mm_packs_epi32(cstep2, cstep3);
+ result = _mm_packs_epi16(cstep01, cstep23);
+
+ *outmask |= _mm_movemask_epi8(result);
+ }
+
+
+ {
+ __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.
+static INLINE unsigned
+build_mask_linear(int c, int dcdx, int dcdy)
+{
+ __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);
+}
+
+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);
+
+ /* Get values across the quad
+ */
+ __m128i cstep0 = _mm_add_epi32(c0, step);
+
+ /* 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( struct lp_rasterizer *rast,
- unsigned thread_index,
- const union lp_rast_cmd_arg arg )
+lp_rast_triangle_3_16(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
{
- struct lp_rasterizer_task *rast_task = &rast->tasks[thread_index];
- const struct lp_rast_triangle *tri = arg.triangle;
+ 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];
- int x = rast_task->x;
- int y = rast_task->y;
- unsigned i;
+ outmask = 0; /* outside one or more trivial reject planes */
+ partmask = 0; /* outside one or more trivial accept planes */
- 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;
+ for (j = 0; j < 3; j++) {
+ c[j] = plane[j].c + plane[j].dcdy * y - plane[j].dcdx * x;
- int ei1 = tri->ei1 * 16;
- int ei2 = tri->ei2 * 16;
- int ei3 = tri->ei3 * 16;
+ {
+ 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;
- int eo1 = tri->eo1 * 16;
- int eo2 = tri->eo2 * 16;
- int eo3 = tri->eo3 * 16;
+ 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 */
+ }
+ }
- LP_DBG(DEBUG_RAST, "lp_rast_triangle\n");
+ if (outmask == 0xffff)
+ return;
- /* 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.
+ /* Mask of sub-blocks which are inside all trivial accept planes:
*/
- for (i = 0; i < 16; i++) {
- int cx1 = c1 + (tri->inputs.step[0][i] * 16);
- int cx2 = c2 + (tri->inputs.step[1][i] * 16);
- int cx3 = c3 + (tri->inputs.step[2][i] * 16);
-
- if (cx1 + eo1 < 0 ||
- cx2 + eo2 < 0 ||
- cx3 + eo3 < 0) {
- /* the block is completely outside the triangle - nop */
- }
- else {
- int px = x + pos_table16[i][0];
- int py = y + pos_table16[i][1];
-
- if (cx1 + ei1 > 0 &&
- cx2 + ei2 > 0 &&
- cx3 + ei3 > 0) {
- /* the block is completely inside the triangle */
- block_full_16(rast_task, tri, px, py);
- }
- else {
- /* the block is partially in/out of the triangle */
- do_block_16(rast_task, tri, px, py, cx1, cx2, cx3);
- }
- }
+ 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);
}
}