#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);
}
+static inline unsigned
+build_mask_linear(int32_t c, int32_t dcdx, int32_t dcdy)
+{
+ unsigned mask = 0;
-/**
- * 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 )
+ int32_t c0 = c;
+ int32_t c1 = c0 + dcdy;
+ int32_t c2 = c1 + dcdy;
+ int32_t 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(int32_t c,
+ int32_t cdiff,
+ int32_t dcdx,
+ int32_t dcdy,
+ unsigned *outmask,
+ unsigned *partmask)
{
- assert(x >= 0);
- assert(y >= 0);
-
- lp_rast_shade_quads(rast_task->rast,
- rast_task->thread_index,
- &tri->inputs,
- x, y,
- -c1, -c2, -c3);
+ *outmask |= build_mask_linear(c, dcdx, dcdy);
+ *partmask |= build_mask_linear(c + cdiff, dcdx, dcdy);
}
+void
+lp_rast_triangle_3_16(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
+{
+ union lp_rast_cmd_arg arg2;
+ arg2.triangle.tri = arg.triangle.tri;
+ arg2.triangle.plane_mask = (1<<3)-1;
+ lp_rast_triangle_3(task, arg2);
+}
-/**
- * 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 c0,
- int c1,
- int c2 )
+void
+lp_rast_triangle_3_4(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
{
- unsigned mask = 0;
- int eo[3];
- int c[3];
- int i, j;
+ lp_rast_triangle_3_16(task, arg);
+}
- assert(x >= 0);
- assert(y >= 0);
- assert(x % 16 == 0);
- assert(y % 16 == 0);
+void
+lp_rast_triangle_4_16(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
+{
+ union lp_rast_cmd_arg arg2;
+ arg2.triangle.tri = arg.triangle.tri;
+ arg2.triangle.plane_mask = (1<<4)-1;
+ lp_rast_triangle_4(task, arg2);
+}
+
+void
+lp_rast_triangle_ms_3_16(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
+{
+ union lp_rast_cmd_arg arg2;
+ arg2.triangle.tri = arg.triangle.tri;
+ arg2.triangle.plane_mask = (1<<3)-1;
+ lp_rast_triangle_ms_3(task, arg2);
+}
+
+void
+lp_rast_triangle_ms_3_4(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
+{
+ lp_rast_triangle_ms_3_16(task, arg);
+}
+
+void
+lp_rast_triangle_ms_4_16(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
+{
+ union lp_rast_cmd_arg arg2;
+ arg2.triangle.tri = arg.triangle.tri;
+ arg2.triangle.plane_mask = (1<<4)-1;
+ lp_rast_triangle_ms_4(task, arg2);
+}
- eo[0] = tri->eo1 * 4;
- eo[1] = tri->eo2 * 4;
- eo[2] = tri->eo3 * 4;
+#if defined(PIPE_ARCH_SSE)
- c[0] = c0;
- c[1] = c1;
- c[2] = c2;
+#include <emmintrin.h>
+#include "util/u_sse.h"
- 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);
- }
+static inline void
+build_masks_sse(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);
+
+ {
+ __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);
}
- 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(rast_task, tri, px, py, 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,
- unsigned thread_index,
- const union lp_rast_cmd_arg arg )
+static inline unsigned
+build_mask_linear_sse(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
+sign_bits4(const __m128i *cstep, int cdiff)
{
- struct lp_rasterizer_task *rast_task = &rast->tasks[thread_index];
- const struct lp_rast_triangle *tri = arg.triangle;
- int x = rast_task->x;
- int y = rast_task->y;
- int ei[3], eo[3], c[3];
- unsigned outmask, inmask, partial_mask;
+ /* Adjust the step values
+ */
+ __m128i cio4 = _mm_set1_epi32(cdiff);
+ __m128i cstep0 = _mm_add_epi32(cstep[0], cio4);
+ __m128i cstep1 = _mm_add_epi32(cstep[1], cio4);
+ __m128i cstep2 = _mm_add_epi32(cstep[2], cio4);
+ __m128i cstep3 = _mm_add_epi32(cstep[3], cio4);
+
+ /* Pack down to epi8
+ */
+ __m128i cstep01 = _mm_packs_epi32(cstep0, cstep1);
+ __m128i cstep23 = _mm_packs_epi32(cstep2, cstep3);
+ __m128i result = _mm_packs_epi16(cstep01, cstep23);
+
+ /* Extract the sign bits
+ */
+ return _mm_movemask_epi8(result);
+}
+
+
+#define NR_PLANES 3
+
+void
+lp_rast_triangle_32_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 = GET_PLANES(tri);
+ int x = (arg.triangle.plane_mask & 0xff) + task->x;
+ int y = (arg.triangle.plane_mask >> 8) + task->y;
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;
-
- eo[0] = tri->eo1 * 16;
- eo[1] = tri->eo2 * 16;
- eo[2] = tri->eo3 * 16;
-
- ei[0] = tri->ei1 * 16;
- ei[1] = tri->ei2 * 16;
- ei[2] = tri->ei3 * 16;
-
- outmask = 0;
- inmask = 0xffff;
-
- 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];
-
- /* 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));
+ struct { unsigned mask:16; unsigned i:8; unsigned j:8; } out[16];
+ unsigned nr = 0;
+
+ /* p0 and p2 are aligned, p1 is not (plane size 24 bytes). */
+ __m128i p0 = _mm_load_si128((__m128i *)&plane[0]); /* clo, chi, dcdx, dcdy */
+ __m128i p1 = _mm_loadu_si128((__m128i *)&plane[1]);
+ __m128i p2 = _mm_load_si128((__m128i *)&plane[2]);
+ __m128i zero = _mm_setzero_si128();
+
+ __m128i c, dcdx, dcdy, rej4;
+ __m128i dcdx_neg_mask, dcdy_neg_mask;
+ __m128i dcdx2, dcdx3;
+
+ __m128i span_0; /* 0,dcdx,2dcdx,3dcdx for plane 0 */
+ __m128i span_1; /* 0,dcdx,2dcdx,3dcdx for plane 1 */
+ __m128i span_2; /* 0,dcdx,2dcdx,3dcdx for plane 2 */
+ __m128i unused;
+
+ transpose4_epi32(&p0, &p1, &p2, &zero,
+ &c, &unused, &dcdx, &dcdy);
+
+ /* recalc eo - easier than trying to load as scalars / shuffle... */
+ dcdx_neg_mask = _mm_srai_epi32(dcdx, 31);
+ dcdy_neg_mask = _mm_srai_epi32(dcdy, 31);
+ rej4 = _mm_sub_epi32(_mm_andnot_si128(dcdy_neg_mask, dcdy),
+ _mm_and_si128(dcdx_neg_mask, dcdx));
+
+ /* Adjust dcdx;
+ */
+ dcdx = _mm_sub_epi32(zero, dcdx);
+
+ c = _mm_add_epi32(c, mm_mullo_epi32(dcdx, _mm_set1_epi32(x)));
+ c = _mm_add_epi32(c, mm_mullo_epi32(dcdy, _mm_set1_epi32(y)));
+ rej4 = _mm_slli_epi32(rej4, 2);
+
+ /* Adjust so we can just check the sign bit (< 0 comparison), instead of having to do a less efficient <= 0 comparison */
+ c = _mm_sub_epi32(c, _mm_set1_epi32(1));
+ rej4 = _mm_add_epi32(rej4, _mm_set1_epi32(1));
+
+ dcdx2 = _mm_add_epi32(dcdx, dcdx);
+ dcdx3 = _mm_add_epi32(dcdx2, dcdx);
+
+ transpose4_epi32(&zero, &dcdx, &dcdx2, &dcdx3,
+ &span_0, &span_1, &span_2, &unused);
+
+ for (i = 0; i < 4; i++) {
+ __m128i cx = c;
+
+ for (j = 0; j < 4; j++) {
+ __m128i c4rej = _mm_add_epi32(cx, rej4);
+ __m128i rej_masks = _mm_srai_epi32(c4rej, 31);
+
+ /* if (is_zero(rej_masks)) */
+ if (_mm_movemask_epi8(rej_masks) == 0) {
+ __m128i c0_0 = _mm_add_epi32(SCALAR_EPI32(cx, 0), span_0);
+ __m128i c1_0 = _mm_add_epi32(SCALAR_EPI32(cx, 1), span_1);
+ __m128i c2_0 = _mm_add_epi32(SCALAR_EPI32(cx, 2), span_2);
+
+ __m128i c_0 = _mm_or_si128(_mm_or_si128(c0_0, c1_0), c2_0);
+
+ __m128i c0_1 = _mm_add_epi32(c0_0, SCALAR_EPI32(dcdy, 0));
+ __m128i c1_1 = _mm_add_epi32(c1_0, SCALAR_EPI32(dcdy, 1));
+ __m128i c2_1 = _mm_add_epi32(c2_0, SCALAR_EPI32(dcdy, 2));
+
+ __m128i c_1 = _mm_or_si128(_mm_or_si128(c0_1, c1_1), c2_1);
+ __m128i c_01 = _mm_packs_epi32(c_0, c_1);
+
+ __m128i c0_2 = _mm_add_epi32(c0_1, SCALAR_EPI32(dcdy, 0));
+ __m128i c1_2 = _mm_add_epi32(c1_1, SCALAR_EPI32(dcdy, 1));
+ __m128i c2_2 = _mm_add_epi32(c2_1, SCALAR_EPI32(dcdy, 2));
+
+ __m128i c_2 = _mm_or_si128(_mm_or_si128(c0_2, c1_2), c2_2);
+
+ __m128i c0_3 = _mm_add_epi32(c0_2, SCALAR_EPI32(dcdy, 0));
+ __m128i c1_3 = _mm_add_epi32(c1_2, SCALAR_EPI32(dcdy, 1));
+ __m128i c2_3 = _mm_add_epi32(c2_2, SCALAR_EPI32(dcdy, 2));
+
+ __m128i c_3 = _mm_or_si128(_mm_or_si128(c0_3, c1_3), c2_3);
+ __m128i c_23 = _mm_packs_epi32(c_2, c_3);
+ __m128i c_0123 = _mm_packs_epi16(c_01, c_23);
+
+ unsigned mask = _mm_movemask_epi8(c_0123);
+
+ out[nr].i = i;
+ out[nr].j = j;
+ out[nr].mask = mask;
+ if (mask != 0xffff)
+ nr++;
+ }
+ cx = _mm_add_epi32(cx, _mm_slli_epi32(dcdx, 2));
}
+
+ c = _mm_add_epi32(c, _mm_slli_epi32(dcdy, 2));
}
- assert((outmask & inmask) == 0);
+ for (i = 0; i < nr; i++)
+ lp_rast_shade_quads_mask(task,
+ &tri->inputs,
+ x + 4 * out[i].j,
+ y + 4 * out[i].i,
+ 0xffff & ~out[i].mask);
+}
+
+void
+lp_rast_triangle_32_3_4(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 = GET_PLANES(tri);
+ unsigned x = (arg.triangle.plane_mask & 0xff) + task->x;
+ unsigned y = (arg.triangle.plane_mask >> 8) + task->y;
+
+ /* p0 and p2 are aligned, p1 is not (plane size 24 bytes). */
+ __m128i p0 = _mm_load_si128((__m128i *)&plane[0]); /* clo, chi, dcdx, dcdy */
+ __m128i p1 = _mm_loadu_si128((__m128i *)&plane[1]);
+ __m128i p2 = _mm_load_si128((__m128i *)&plane[2]);
+ __m128i zero = _mm_setzero_si128();
+
+ __m128i c, dcdx, dcdy;
+ __m128i dcdx2, dcdx3;
+
+ __m128i span_0; /* 0,dcdx,2dcdx,3dcdx for plane 0 */
+ __m128i span_1; /* 0,dcdx,2dcdx,3dcdx for plane 1 */
+ __m128i span_2; /* 0,dcdx,2dcdx,3dcdx for plane 2 */
+ __m128i unused;
- if (outmask == 0xffff)
- return;
+ transpose4_epi32(&p0, &p1, &p2, &zero,
+ &c, &unused, &dcdx, &dcdy);
- /* Invert mask, so that bits are set whenever we are at least
- * partially inside all of the edges:
+ /* Adjust dcdx;
*/
- partial_mask = ~inmask & ~outmask & 0xffff;
+ dcdx = _mm_sub_epi32(zero, dcdx);
+
+ c = _mm_add_epi32(c, mm_mullo_epi32(dcdx, _mm_set1_epi32(x)));
+ c = _mm_add_epi32(c, mm_mullo_epi32(dcdy, _mm_set1_epi32(y)));
+
+ /* Adjust so we can just check the sign bit (< 0 comparison), instead of having to do a less efficient <= 0 comparison */
+ c = _mm_sub_epi32(c, _mm_set1_epi32(1));
+
+ dcdx2 = _mm_add_epi32(dcdx, dcdx);
+ dcdx3 = _mm_add_epi32(dcdx2, dcdx);
- /* Iterate over partials:
+ transpose4_epi32(&zero, &dcdx, &dcdx2, &dcdx3,
+ &span_0, &span_1, &span_2, &unused);
+
+
+ {
+ __m128i c0_0 = _mm_add_epi32(SCALAR_EPI32(c, 0), span_0);
+ __m128i c1_0 = _mm_add_epi32(SCALAR_EPI32(c, 1), span_1);
+ __m128i c2_0 = _mm_add_epi32(SCALAR_EPI32(c, 2), span_2);
+
+ __m128i c_0 = _mm_or_si128(_mm_or_si128(c0_0, c1_0), c2_0);
+
+ __m128i c0_1 = _mm_add_epi32(c0_0, SCALAR_EPI32(dcdy, 0));
+ __m128i c1_1 = _mm_add_epi32(c1_0, SCALAR_EPI32(dcdy, 1));
+ __m128i c2_1 = _mm_add_epi32(c2_0, SCALAR_EPI32(dcdy, 2));
+
+ __m128i c_1 = _mm_or_si128(_mm_or_si128(c0_1, c1_1), c2_1);
+ __m128i c_01 = _mm_packs_epi32(c_0, c_1);
+
+ __m128i c0_2 = _mm_add_epi32(c0_1, SCALAR_EPI32(dcdy, 0));
+ __m128i c1_2 = _mm_add_epi32(c1_1, SCALAR_EPI32(dcdy, 1));
+ __m128i c2_2 = _mm_add_epi32(c2_1, SCALAR_EPI32(dcdy, 2));
+
+ __m128i c_2 = _mm_or_si128(_mm_or_si128(c0_2, c1_2), c2_2);
+
+ __m128i c0_3 = _mm_add_epi32(c0_2, SCALAR_EPI32(dcdy, 0));
+ __m128i c1_3 = _mm_add_epi32(c1_2, SCALAR_EPI32(dcdy, 1));
+ __m128i c2_3 = _mm_add_epi32(c2_2, SCALAR_EPI32(dcdy, 2));
+
+ __m128i c_3 = _mm_or_si128(_mm_or_si128(c0_3, c1_3), c2_3);
+ __m128i c_23 = _mm_packs_epi32(c_2, c_3);
+ __m128i c_0123 = _mm_packs_epi16(c_01, c_23);
+
+ unsigned mask = _mm_movemask_epi8(c_0123);
+
+ if (mask != 0xffff)
+ lp_rast_shade_quads_mask(task,
+ &tri->inputs,
+ x,
+ y,
+ 0xffff & ~mask);
+ }
+}
+
+#undef NR_PLANES
+
+#else
+
+#if defined(_ARCH_PWR8) && UTIL_ARCH_LITTLE_ENDIAN
+
+#include <altivec.h>
+#include "util/u_pwr8.h"
+
+static inline void
+build_masks_ppc(int c,
+ int cdiff,
+ int dcdx,
+ int dcdy,
+ unsigned *outmask,
+ unsigned *partmask)
+{
+ __m128i cstep0 = vec_setr_epi32(c, c+dcdx, c+dcdx*2, c+dcdx*3);
+ __m128i xdcdy = (__m128i) vec_splats(dcdy);
+
+ /* Get values across the quad
*/
- 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(rast_task, tri, px, py, cx1, cx2, cx3);
+ __m128i cstep1 = vec_add_epi32(cstep0, xdcdy);
+ __m128i cstep2 = vec_add_epi32(cstep1, xdcdy);
+ __m128i cstep3 = vec_add_epi32(cstep2, xdcdy);
+
+ {
+ __m128i cstep01, cstep23, result;
+
+ cstep01 = vec_packs_epi32(cstep0, cstep1);
+ cstep23 = vec_packs_epi32(cstep2, cstep3);
+ result = vec_packs_epi16(cstep01, cstep23);
+
+ *outmask |= vec_movemask_epi8(result);
}
- /* Iterate over fulls:
+
+ {
+ __m128i cio4 = (__m128i) vec_splats(cdiff);
+ __m128i cstep01, cstep23, result;
+
+ cstep0 = vec_add_epi32(cstep0, cio4);
+ cstep1 = vec_add_epi32(cstep1, cio4);
+ cstep2 = vec_add_epi32(cstep2, cio4);
+ cstep3 = vec_add_epi32(cstep3, cio4);
+
+ cstep01 = vec_packs_epi32(cstep0, cstep1);
+ cstep23 = vec_packs_epi32(cstep2, cstep3);
+ result = vec_packs_epi16(cstep01, cstep23);
+
+ *partmask |= vec_movemask_epi8(result);
+ }
+}
+
+static inline unsigned
+build_mask_linear_ppc(int c, int dcdx, int dcdy)
+{
+ __m128i cstep0 = vec_setr_epi32(c, c+dcdx, c+dcdx*2, c+dcdx*3);
+ __m128i xdcdy = (__m128i) vec_splats(dcdy);
+
+ /* Get values across the quad
+ */
+ __m128i cstep1 = vec_add_epi32(cstep0, xdcdy);
+ __m128i cstep2 = vec_add_epi32(cstep1, xdcdy);
+ __m128i cstep3 = vec_add_epi32(cstep2, xdcdy);
+
+ /* pack pairs of results into epi16
+ */
+ __m128i cstep01 = vec_packs_epi32(cstep0, cstep1);
+ __m128i cstep23 = vec_packs_epi32(cstep2, cstep3);
+
+ /* pack into epi8, preserving sign bits
*/
- while (inmask) {
- int i = ffs(inmask) - 1;
- int px = x + pos_table16[i][0];
- int py = y + pos_table16[i][1];
+ __m128i result = vec_packs_epi16(cstep01, cstep23);
- inmask &= ~(1 << i);
+ /* extract sign bits to create mask
+ */
+ return vec_movemask_epi8(result);
+}
+
+static inline __m128i
+lp_plane_to_m128i(const struct lp_rast_plane *plane)
+{
+ return vec_setr_epi32((int32_t)plane->c, (int32_t)plane->dcdx,
+ (int32_t)plane->dcdy, (int32_t)plane->eo);
+}
+
+#define NR_PLANES 3
+
+void
+lp_rast_triangle_32_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 = GET_PLANES(tri);
+ int x = (arg.triangle.plane_mask & 0xff) + task->x;
+ int y = (arg.triangle.plane_mask >> 8) + task->y;
+ unsigned i, j;
+
+ struct { unsigned mask:16; unsigned i:8; unsigned j:8; } out[16];
+ unsigned nr = 0;
+
+ __m128i p0 = lp_plane_to_m128i(&plane[0]); /* c, dcdx, dcdy, eo */
+ __m128i p1 = lp_plane_to_m128i(&plane[1]); /* c, dcdx, dcdy, eo */
+ __m128i p2 = lp_plane_to_m128i(&plane[2]); /* c, dcdx, dcdy, eo */
+ __m128i zero = vec_splats((unsigned char) 0);
+
+ __m128i c;
+ __m128i dcdx;
+ __m128i dcdy;
+ __m128i rej4;
+
+ __m128i dcdx2;
+ __m128i dcdx3;
+
+ __m128i span_0; /* 0,dcdx,2dcdx,3dcdx for plane 0 */
+ __m128i span_1; /* 0,dcdx,2dcdx,3dcdx for plane 1 */
+ __m128i span_2; /* 0,dcdx,2dcdx,3dcdx for plane 2 */
+ __m128i unused;
+
+ __m128i vshuf_mask0;
+ __m128i vshuf_mask1;
+ __m128i vshuf_mask2;
+
+#if UTIL_ARCH_LITTLE_ENDIAN
+ vshuf_mask0 = (__m128i) vec_splats((unsigned int) 0x03020100);
+ vshuf_mask1 = (__m128i) vec_splats((unsigned int) 0x07060504);
+ vshuf_mask2 = (__m128i) vec_splats((unsigned int) 0x0B0A0908);
+#else
+ vshuf_mask0 = (__m128i) vec_splats((unsigned int) 0x0C0D0E0F);
+ vshuf_mask1 = (__m128i) vec_splats((unsigned int) 0x08090A0B);
+ vshuf_mask2 = (__m128i) vec_splats((unsigned int) 0x04050607);
+#endif
+
+ transpose4_epi32(&p0, &p1, &p2, &zero,
+ &c, &dcdx, &dcdy, &rej4);
+
+ /* Adjust dcdx;
+ */
+ dcdx = vec_sub_epi32(zero, dcdx);
+
+ c = vec_add_epi32(c, vec_mullo_epi32(dcdx, (__m128i) vec_splats(x)));
+ c = vec_add_epi32(c, vec_mullo_epi32(dcdy, (__m128i) vec_splats(y)));
+ rej4 = vec_slli_epi32(rej4, 2);
+
+ /*
+ * Adjust so we can just check the sign bit (< 0 comparison),
+ * instead of having to do a less efficient <= 0 comparison
+ */
+ c = vec_sub_epi32(c, (__m128i) vec_splats((unsigned int) 1));
+ rej4 = vec_add_epi32(rej4, (__m128i) vec_splats((unsigned int) 1));
+
+ dcdx2 = vec_add_epi32(dcdx, dcdx);
+ dcdx3 = vec_add_epi32(dcdx2, dcdx);
+
+ transpose4_epi32(&zero, &dcdx, &dcdx2, &dcdx3,
+ &span_0, &span_1, &span_2, &unused);
+
+ for (i = 0; i < 4; i++) {
+ __m128i cx = c;
+
+ for (j = 0; j < 4; j++) {
+ __m128i c4rej = vec_add_epi32(cx, rej4);
+ __m128i rej_masks = vec_srai_epi32(c4rej, 31);
+
+ /* if (is_zero(rej_masks)) */
+ if (vec_movemask_epi8(rej_masks) == 0) {
+ __m128i c0_0 = vec_add_epi32(vec_perm(cx, cx, vshuf_mask0), span_0);
+ __m128i c1_0 = vec_add_epi32(vec_perm(cx, cx, vshuf_mask1), span_1);
+ __m128i c2_0 = vec_add_epi32(vec_perm(cx, cx, vshuf_mask2), span_2);
+
+ __m128i c_0 = vec_or(vec_or(c0_0, c1_0), c2_0);
+
+ __m128i c0_1 = vec_add_epi32(c0_0, vec_perm(dcdy, dcdy, vshuf_mask0));
+ __m128i c1_1 = vec_add_epi32(c1_0, vec_perm(dcdy, dcdy, vshuf_mask1));
+ __m128i c2_1 = vec_add_epi32(c2_0, vec_perm(dcdy, dcdy, vshuf_mask2));
+
+ __m128i c_1 = vec_or(vec_or(c0_1, c1_1), c2_1);
+ __m128i c_01 = vec_packs_epi32(c_0, c_1);
+
+ __m128i c0_2 = vec_add_epi32(c0_1, vec_perm(dcdy, dcdy, vshuf_mask0));
+ __m128i c1_2 = vec_add_epi32(c1_1, vec_perm(dcdy, dcdy, vshuf_mask1));
+ __m128i c2_2 = vec_add_epi32(c2_1, vec_perm(dcdy, dcdy, vshuf_mask2));
- LP_COUNT(nr_fully_covered_16);
- block_full_16(rast_task, tri, px, py);
+ __m128i c_2 = vec_or(vec_or(c0_2, c1_2), c2_2);
+
+ __m128i c0_3 = vec_add_epi32(c0_2, vec_perm(dcdy, dcdy, vshuf_mask0));
+ __m128i c1_3 = vec_add_epi32(c1_2, vec_perm(dcdy, dcdy, vshuf_mask1));
+ __m128i c2_3 = vec_add_epi32(c2_2, vec_perm(dcdy, dcdy, vshuf_mask2));
+
+ __m128i c_3 = vec_or(vec_or(c0_3, c1_3), c2_3);
+ __m128i c_23 = vec_packs_epi32(c_2, c_3);
+ __m128i c_0123 = vec_packs_epi16(c_01, c_23);
+
+ unsigned mask = vec_movemask_epi8(c_0123);
+
+ out[nr].i = i;
+ out[nr].j = j;
+ out[nr].mask = mask;
+ if (mask != 0xffff)
+ nr++;
+ }
+ cx = vec_add_epi32(cx, vec_slli_epi32(dcdx, 2));
+ }
+
+ c = vec_add_epi32(c, vec_slli_epi32(dcdy, 2));
}
+
+ for (i = 0; i < nr; i++)
+ lp_rast_shade_quads_mask(task,
+ &tri->inputs,
+ x + 4 * out[i].j,
+ y + 4 * out[i].i,
+ 0xffff & ~out[i].mask);
}
+
+#undef NR_PLANES
+
+#else
+
+void
+lp_rast_triangle_32_3_16(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
+{
+ union lp_rast_cmd_arg arg2;
+ arg2.triangle.tri = arg.triangle.tri;
+ arg2.triangle.plane_mask = (1<<3)-1;
+ lp_rast_triangle_32_3(task, arg2);
+}
+
+#endif /* _ARCH_PWR8 && UTIL_ARCH_LITTLE_ENDIAN */
+
+void
+lp_rast_triangle_32_4_16(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
+{
+ union lp_rast_cmd_arg arg2;
+ arg2.triangle.tri = arg.triangle.tri;
+ arg2.triangle.plane_mask = (1<<4)-1;
+ lp_rast_triangle_32_4(task, arg2);
+}
+
+void
+lp_rast_triangle_32_3_4(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
+{
+ lp_rast_triangle_32_3_16(task, arg);
+}
+
+#endif
+
+#if defined PIPE_ARCH_SSE
+#define BUILD_MASKS(c, cdiff, dcdx, dcdy, omask, pmask) build_masks_sse((int)c, (int)cdiff, dcdx, dcdy, omask, pmask)
+#define BUILD_MASK_LINEAR(c, dcdx, dcdy) build_mask_linear_sse((int)c, dcdx, dcdy)
+#elif (defined(_ARCH_PWR8) && UTIL_ARCH_LITTLE_ENDIAN)
+#define BUILD_MASKS(c, cdiff, dcdx, dcdy, omask, pmask) build_masks_ppc((int)c, (int)cdiff, dcdx, dcdy, omask, pmask)
+#define BUILD_MASK_LINEAR(c, dcdx, dcdy) build_mask_linear_ppc((int)c, dcdx, dcdy)
+#else
+#define BUILD_MASKS(c, cdiff, dcdx, dcdy, omask, pmask) build_masks(c, cdiff, dcdx, dcdy, omask, pmask)
+#define BUILD_MASK_LINEAR(c, dcdx, dcdy) build_mask_linear(c, dcdx, dcdy)
+#endif
+
+#define RASTER_64 1
+
+#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
+/*#define TRI_4 lp_rast_triangle_3_4*/
+/*#define TRI_16 lp_rast_triangle_3_16*/
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_4
+#define NR_PLANES 4
+/*#define TRI_16 lp_rast_triangle_4_16*/
+#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"
+
+#undef RASTER_64
+
+#define TAG(x) x##_32_1
+#define NR_PLANES 1
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_32_2
+#define NR_PLANES 2
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_32_3
+#define NR_PLANES 3
+/*#define TRI_4 lp_rast_triangle_3_4*/
+/*#define TRI_16 lp_rast_triangle_3_16*/
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_32_4
+#define NR_PLANES 4
+#ifdef PIPE_ARCH_SSE
+#define TRI_16 lp_rast_triangle_32_4_16
+#endif
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_32_5
+#define NR_PLANES 5
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_32_6
+#define NR_PLANES 6
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_32_7
+#define NR_PLANES 7
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_32_8
+#define NR_PLANES 8
+#include "lp_rast_tri_tmp.h"
+
+#define MULTISAMPLE 1
+#define RASTER_64 1
+
+#define TAG(x) x##_ms_1
+#define NR_PLANES 1
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_ms_2
+#define NR_PLANES 2
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_ms_3
+#define NR_PLANES 3
+/*#define TRI_4 lp_rast_triangle_3_4*/
+/*#define TRI_16 lp_rast_triangle_3_16*/
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_ms_4
+#define NR_PLANES 4
+/*#define TRI_16 lp_rast_triangle_4_16*/
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_ms_5
+#define NR_PLANES 5
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_ms_6
+#define NR_PLANES 6
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_ms_7
+#define NR_PLANES 7
+#include "lp_rast_tri_tmp.h"
+
+#define TAG(x) x##_ms_8
+#define NR_PLANES 8
+#include "lp_rast_tri_tmp.h"
+
+#undef RASTER_64