+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);
+ }
+
+
+ {
+ __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);
+ }
+}
+
+
+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)
+{
+
+ /* 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;
+
+ 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++;