* Rasterization for binned triangles within a tile
*/
-#include "lp_context.h"
-#include "lp_quad.h"
-#include "lp_quad_pipe.h"
-#include "lp_setup.h"
-#include "lp_state.h"
-#include "draw/draw_context.h"
-#include "draw/draw_private.h"
-#include "draw/draw_vertex.h"
-#include "pipe/p_shader_tokens.h"
-#include "pipe/p_thread.h"
+#include <limits.h>
#include "util/u_math.h"
-#include "util/u_memory.h"
+#include "lp_debug.h"
+#include "lp_perf.h"
+#include "lp_rast_priv.h"
-#define BLOCKSIZE 4
+/**
+ * Shade all pixels in a 4x4 block.
+ */
+static void
+block_full_4(struct lp_rasterizer_task *task,
+ const struct lp_rast_triangle *tri,
+ int x, int y)
+{
+ lp_rast_shade_quads_all(task, &tri->inputs, x, y);
+}
-/* Convert 8x8 block into four runs of quads and render each in turn.
+/**
+ * Shade all pixels in a 16x16 block.
*/
-#if (BLOCKSIZE == 8)
-static void block_full( struct triangle *tri, int x, int y )
+static void
+block_full_16(struct lp_rasterizer_task *task,
+ const struct lp_rast_triangle *tri,
+ int x, int y)
{
- struct quad_header *ptrs[4];
- int i;
-
- tri->quad[0].input.x0 = x + 0;
- tri->quad[1].input.x0 = x + 2;
- tri->quad[2].input.x0 = x + 4;
- tri->quad[3].input.x0 = x + 6;
-
- for (i = 0; i < 4; i++, y += 2) {
- tri->quad[0].inout.mask = 0xf;
- tri->quad[1].inout.mask = 0xf;
- tri->quad[2].inout.mask = 0xf;
- tri->quad[3].inout.mask = 0xf;
-
- tri->quad[0].input.y0 = y;
- tri->quad[1].input.y0 = y;
- tri->quad[2].input.y0 = y;
- tri->quad[3].input.y0 = y;
-
- /* XXX: don't bother with this ptrs business */
- ptrs[0] = &tri->quad[0];
- ptrs[1] = &tri->quad[1];
- ptrs[2] = &tri->quad[2];
- ptrs[3] = &tri->quad[3];
-
- tri->llvmpipe->quad.first->run( tri->llvmpipe->quad.first, ptrs, 4 );
- }
+ 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);
}
-#else
-static void block_full( struct triangle *tri, int x, int y )
-{
- struct quad_header *ptrs[4];
- int iy;
- tri->quad[0].input.x0 = x + 0;
- tri->quad[1].input.x0 = x + 2;
+static inline unsigned
+build_mask_linear(int32_t c, int32_t dcdx, int32_t dcdy)
+{
+ unsigned mask = 0;
+
+ 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;
+}
- for (iy = 0; iy < 4; iy += 2) {
- tri->quad[0].inout.mask = 0xf;
- tri->quad[1].inout.mask = 0xf;
- tri->quad[0].input.y0 = y + iy;
- tri->quad[1].input.y0 = y + iy;
+static inline void
+build_masks(int32_t c,
+ int32_t cdiff,
+ int32_t dcdx,
+ int32_t dcdy,
+ unsigned *outmask,
+ unsigned *partmask)
+{
+ *outmask |= build_mask_linear(c, dcdx, dcdy);
+ *partmask |= build_mask_linear(c + cdiff, dcdx, dcdy);
+}
- /* XXX: don't bother with this ptrs business */
- ptrs[0] = &tri->quad[0];
- ptrs[1] = &tri->quad[1];
+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);
+}
- tri->llvmpipe->quad.first->run( tri->llvmpipe->quad.first, ptrs, 2 );
- }
+void
+lp_rast_triangle_3_4(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
+{
+ lp_rast_triangle_3_16(task, arg);
}
-#endif
-static void
-do_quad( struct lp_rasterizer *rast,
- int x, int y,
- float c1, float c2, float c3 )
+void
+lp_rast_triangle_4_16(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
{
- struct triangle *tri = rast->tri;
- struct quad_header *quad = &rast->quad[0];
-
- float xstep1 = -tri->dy12;
- float xstep2 = -tri->dy23;
- float xstep3 = -tri->dy31;
-
- float ystep1 = tri->dx12;
- float ystep2 = tri->dx23;
- float ystep3 = tri->dx31;
-
- quad->input.x0 = x;
- quad->input.y0 = y;
- quad->inout.mask = 0;
-
- if (c1 > 0 &&
- c2 > 0 &&
- c3 > 0)
- quad->inout.mask |= 1;
-
- if (c1 + xstep1 > 0 &&
- c2 + xstep2 > 0 &&
- c3 + xstep3 > 0)
- quad->inout.mask |= 2;
-
- if (c1 + ystep1 > 0 &&
- c2 + ystep2 > 0 &&
- c3 + ystep3 > 0)
- quad->inout.mask |= 4;
-
- if (c1 + ystep1 + xstep1 > 0 &&
- c2 + ystep2 + xstep2 > 0 &&
- c3 + ystep3 + xstep3 > 0)
- quad->inout.mask |= 8;
-
- if (quad->inout.mask)
- rast->state->run( rast->state->state, &quad, 1 );
+ 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);
}
-/* Evaluate each pixel in a block, generate a mask and possibly render
- * the quad:
- */
-static void
-do_block( struct triangle *tri,
- int x, int y,
- float c1,
- float c2,
- float c3 )
+#if defined(PIPE_ARCH_SSE)
+
+#include <emmintrin.h>
+#include "util/u_sse.h"
+
+
+static inline void
+build_masks_sse(int c,
+ int cdiff,
+ int dcdx,
+ int dcdy,
+ unsigned *outmask,
+ unsigned *partmask)
{
- const int step = 2;
+ __m128i cstep0 = _mm_setr_epi32(c, c+dcdx, c+dcdx*2, c+dcdx*3);
+ __m128i xdcdy = _mm_set1_epi32(dcdy);
- float xstep1 = -step * tri->dy12;
- float xstep2 = -step * tri->dy23;
- float xstep3 = -step * tri->dy31;
+ /* 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);
- float ystep1 = step * tri->dx12;
- float ystep2 = step * tri->dx23;
- float ystep3 = step * tri->dx31;
+ {
+ __m128i cstep01, cstep23, result;
+
+ cstep01 = _mm_packs_epi32(cstep0, cstep1);
+ cstep23 = _mm_packs_epi32(cstep2, cstep3);
+ result = _mm_packs_epi16(cstep01, cstep23);
- int ix, iy;
+ *outmask |= _mm_movemask_epi8(result);
+ }
- for (iy = 0; iy < BLOCKSIZE; iy += 2) {
- float cx1 = c1;
- float cx2 = c2;
- float cx3 = c3;
- for (ix = 0; ix < BLOCKSIZE; ix += 2) {
+ {
+ __m128i cio4 = _mm_set1_epi32(cdiff);
+ __m128i cstep01, cstep23, result;
- do_quad(tri, x+ix, y+iy, cx1, cx2, cx3);
+ cstep0 = _mm_add_epi32(cstep0, cio4);
+ cstep1 = _mm_add_epi32(cstep1, cio4);
+ cstep2 = _mm_add_epi32(cstep2, cio4);
+ cstep3 = _mm_add_epi32(cstep3, cio4);
- cx1 += xstep1;
- cx2 += xstep2;
- cx3 += xstep3;
- }
+ cstep01 = _mm_packs_epi32(cstep0, cstep1);
+ cstep23 = _mm_packs_epi32(cstep2, cstep3);
+ result = _mm_packs_epi16(cstep01, cstep23);
- c1 += ystep1;
- c2 += ystep2;
- c3 += ystep3;
+ *partmask |= _mm_movemask_epi8(result);
}
}
+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);
-/* Scan the tile in chunks and figure out which pixels to rasterize
- * for this triangle:
- */
-void lp_rast_triangle( struct lp_rasterizer *rast,
- const struct lp_rast_triangle *tri )
+ /* 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)
{
- int minx, maxx, miny, maxy;
- /* Clamp to tile dimensions:
+ /* Adjust the step values
*/
- minx = MAX2(tri->maxx, rast->x);
- miny = MAX2(tri->miny, rast->y);
- maxx = MIN2(tri->maxx, rast->x + TILESIZE);
- maxy = MIN2(tri->maxy, rast->y + TILESIZE);
-
- if (miny == maxy ||
- minx == maxx) {
- debug_printf("%s: non-intersecting triangle in bin\n", __FUNCTION__);
- //assert(0);
- return;
- }
+ __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);
+}
+
- /* Bind parameter interpolants:
+#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 = _mm_setzero_si128();
+
+ __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;
+
+ transpose4_epi32(&p0, &p1, &p2, &zero,
+ &c, &dcdx, &dcdy, &rej4);
+
+ /* Adjust dcdx;
*/
- for (i = 0; i < Elements(rast->quad); i++) {
- rast->quad[i].coef = tri->coef;
- rast->quad[i].posCoef = &tri->position_coef;
+ 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));
}
- /* Small area?
+ 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;
+
+ __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 = _mm_setzero_si128();
+
+ __m128i c;
+ __m128i dcdx;
+ __m128i dcdy;
+
+ __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;
+
+ transpose4_epi32(&p0, &p1, &p2, &zero,
+ &c, &dcdx, &dcdy, &unused);
+
+ /* Adjust dcdx;
*/
- if (miny + 16 > maxy &&
- minx + 16 > maxx)
+ 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);
+
+ transpose4_epi32(&zero, &dcdx, &dcdx2, &dcdx3,
+ &span_0, &span_1, &span_2, &unused);
+
+
{
- const int step = 2;
-
- float xstep1 = -step * tri->dy12;
- float xstep2 = -step * tri->dy23;
- float xstep3 = -step * tri->dy31;
-
- float ystep1 = step * tri->dx12;
- float ystep2 = step * tri->dx23;
- float ystep3 = step * tri->dx31;
-
- float eo1 = tri->eo1 * step;
- float eo2 = tri->eo2 * step;
- float eo3 = tri->eo3 * step;
-
- int x, y;
-
- minx &= ~(step-1);
- maxx &= ~(step-1);
-
- /* Subdivide space into NxM blocks, where each block is square and
- * power-of-four in dimension.
- *
- * Trivially accept or reject blocks, else jump to per-pixel
- * examination above.
- */
- for (y = miny; y < maxy; y += step)
- {
- float cx1 = c1;
- float cx2 = c2;
- float cx3 = c3;
-
- for (x = minx; x < maxx; x += step)
- {
- if (cx1 + eo1 < 0 ||
- cx2 + eo2 < 0 ||
- cx3 + eo3 < 0)
- {
- }
- else
- {
- do_quad(&tri, x, y, cx1, cx2, cx3);
- }
-
- /* Iterate cx values across the region:
- */
- cx1 += xstep1;
- cx2 += xstep2;
- cx3 += xstep3;
- }
+ __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);
- /* Iterate c values down the region:
- */
- c1 += ystep1;
- c2 += ystep2;
- c3 += ystep3;
- }
+ __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);
}
- else
+}
+
+#undef NR_PLANES
+
+#else
+
+#if defined(_ARCH_PWR8) && defined(PIPE_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
+ */
+ __m128i cstep1 = vec_add_epi32(cstep0, xdcdy);
+ __m128i cstep2 = vec_add_epi32(cstep1, xdcdy);
+ __m128i cstep3 = vec_add_epi32(cstep2, xdcdy);
+
{
- const int step = BLOCKSIZE;
-
- float ei1 = tri->ei1 * step;
- float ei2 = tri->ei2 * step;
- float ei3 = tri->ei3 * step;
-
- float eo1 = tri->eo1 * step;
- float eo2 = tri->eo2 * step;
- float eo3 = tri->eo3 * step;
-
- float xstep1 = -step * tri->dy12;
- float xstep2 = -step * tri->dy23;
- float xstep3 = -step * tri->dy31;
-
- float ystep1 = step * tri->dx12;
- float ystep2 = step * tri->dx23;
- float ystep3 = step * tri->dx31;
- int x, y;
-
- minx &= ~(step-1);
- miny &= ~(step-1);
-
- for (y = miny; y < maxy; y += step)
- {
- float cx1 = c1;
- float cx2 = c2;
- float cx3 = c3;
-
- 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(&tri, x, y); /* trivial accept */
- }
- else
- {
- do_block(&tri, x, y, cx1, cx2, cx3);
- }
-
- /* Iterate cx values across the region:
- */
- cx1 += xstep1;
- cx2 += xstep2;
- cx3 += xstep3;
- }
-
- /* Iterate c values down the region:
- */
- c1 += ystep1;
- c2 += ystep2;
- c3 += ystep3;
+ __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);
+ }
+
+
+ {
+ __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
+ */
+ __m128i result = vec_packs_epi16(cstep01, cstep23);
+
+ /* 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;
+
+#ifdef PIPE_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));
+
+ __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 && PIPE_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) && defined(PIPE_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"
+
projects / mesa.git / blobdiff