#include "util/u_math.h"
#include "util/u_memory.h"
+#include "util/u_rect.h"
+#include "util/u_sse.h"
#include "lp_perf.h"
#include "lp_setup_context.h"
#include "lp_rast.h"
+#include "lp_state_fs.h"
+#include "lp_state_setup.h"
#define NUM_CHANNELS 4
-
-/**
- * Compute a0 for a constant-valued coefficient (GL_FLAT shading).
- */
-static void constant_coef( struct lp_setup_context *setup,
- struct lp_rast_triangle *tri,
- unsigned slot,
- const float value,
- unsigned i )
+#if defined(PIPE_ARCH_SSE)
+#include <emmintrin.h>
+#endif
+
+static INLINE int
+subpixel_snap(float a)
{
- tri->inputs.a0[slot][i] = value;
- tri->inputs.dadx[slot][i] = 0.0f;
- tri->inputs.dady[slot][i] = 0.0f;
+ return util_iround(FIXED_ONE * a);
}
-
-/**
- * Compute a0, dadx and dady for a linearly interpolated coefficient,
- * for a triangle.
- */
-static void linear_coef( struct lp_setup_context *setup,
- struct lp_rast_triangle *tri,
- float oneoverarea,
- unsigned slot,
- const float (*v1)[4],
- const float (*v2)[4],
- const float (*v3)[4],
- unsigned vert_attr,
- unsigned i)
+static INLINE float
+fixed_to_float(int a)
{
- float a1 = v1[vert_attr][i];
- float a2 = v2[vert_attr][i];
- float a3 = v3[vert_attr][i];
+ return a * (1.0 / FIXED_ONE);
+}
- float da12 = a1 - a2;
- float da31 = a3 - a1;
- float dadx = (da12 * tri->dy31 - tri->dy12 * da31) * oneoverarea;
- float dady = (da31 * tri->dx12 - tri->dx31 * da12) * oneoverarea;
- tri->inputs.dadx[slot][i] = dadx;
- tri->inputs.dady[slot][i] = dady;
- /* calculate a0 as the value which would be sampled for the
- * fragment at (0,0), taking into account that we want to sample at
- * pixel centers, in other words (0.5, 0.5).
- *
- * this is neat but unfortunately not a good way to do things for
- * triangles with very large values of dadx or dady as it will
- * result in the subtraction and re-addition from a0 of a very
- * large number, which means we'll end up loosing a lot of the
- * fractional bits and precision from a0. the way to fix this is
- * to define a0 as the sample at a pixel center somewhere near vmin
- * instead - i'll switch to this later.
- */
- tri->inputs.a0[slot][i] = (a1 -
- (dadx * (v1[0][0] - setup->pixel_offset) +
- dady * (v1[0][1] - setup->pixel_offset)));
-}
-/**
- * Compute a0, dadx and dady for a perspective-corrected interpolant,
- * for a triangle.
- * We basically multiply the vertex value by 1/w before computing
- * the plane coefficients (a0, dadx, dady).
- * Later, when we compute the value at a particular fragment position we'll
- * divide the interpolated value by the interpolated W at that fragment.
- */
-static void perspective_coef( struct lp_setup_context *setup,
- struct lp_rast_triangle *tri,
- float oneoverarea,
- unsigned slot,
- const float (*v1)[4],
- const float (*v2)[4],
- const float (*v3)[4],
- unsigned vert_attr,
- unsigned i)
-{
- /* premultiply by 1/w (v[0][3] is always 1/w):
- */
- float a1 = v1[vert_attr][i] * v1[0][3];
- float a2 = v2[vert_attr][i] * v2[0][3];
- float a3 = v3[vert_attr][i] * v3[0][3];
- float da12 = a1 - a2;
- float da31 = a3 - a1;
- float dadx = (da12 * tri->dy31 - tri->dy12 * da31) * oneoverarea;
- float dady = (da31 * tri->dx12 - tri->dx31 * da12) * oneoverarea;
-
- tri->inputs.dadx[slot][i] = dadx;
- tri->inputs.dady[slot][i] = dady;
- tri->inputs.a0[slot][i] = (a1 -
- (dadx * (v1[0][0] - setup->pixel_offset) +
- dady * (v1[0][1] - setup->pixel_offset)));
-}
/**
- * Special coefficient setup for gl_FragCoord.
- * X and Y are trivial
- * Z and W are copied from position_coef which should have already been computed.
- * We could do a bit less work if we'd examine gl_FragCoord's swizzle mask.
+ * Alloc space for a new triangle plus the input.a0/dadx/dady arrays
+ * immediately after it.
+ * The memory is allocated from the per-scene pool, not per-tile.
+ * \param tri_size returns number of bytes allocated
+ * \param num_inputs number of fragment shader inputs
+ * \return pointer to triangle space
*/
-static void
-setup_fragcoord_coef(struct lp_setup_context *setup,
- struct lp_rast_triangle *tri,
- float oneoverarea,
- unsigned slot,
- const float (*v1)[4],
- const float (*v2)[4],
- const float (*v3)[4])
+struct lp_rast_triangle *
+lp_setup_alloc_triangle(struct lp_scene *scene,
+ unsigned nr_inputs,
+ unsigned nr_planes,
+ unsigned *tri_size)
{
- /*X*/
- tri->inputs.a0[slot][0] = 0.0;
- tri->inputs.dadx[slot][0] = 1.0;
- tri->inputs.dady[slot][0] = 0.0;
- /*Y*/
- tri->inputs.a0[slot][1] = 0.0;
- tri->inputs.dadx[slot][1] = 0.0;
- tri->inputs.dady[slot][1] = 1.0;
- /*Z*/
- linear_coef(setup, tri, oneoverarea, slot, v1, v2, v3, 0, 2);
- /*W*/
- linear_coef(setup, tri, oneoverarea, slot, v1, v2, v3, 0, 3);
-}
+ unsigned input_array_sz = NUM_CHANNELS * (nr_inputs + 1) * sizeof(float);
+ unsigned plane_sz = nr_planes * sizeof(struct lp_rast_plane);
+ struct lp_rast_triangle *tri;
+ *tri_size = (sizeof(struct lp_rast_triangle) +
+ 3 * input_array_sz +
+ plane_sz);
-/**
- * Setup the fragment input attribute with the front-facing value.
- * \param frontface is the triangle front facing?
- */
-static void setup_facing_coef( struct lp_setup_context *setup,
- struct lp_rast_triangle *tri,
- unsigned slot,
- boolean frontface )
-{
- /* convert TRUE to 1.0 and FALSE to -1.0 */
- constant_coef( setup, tri, slot, 2.0f * frontface - 1.0f, 0 );
- constant_coef( setup, tri, slot, 0.0f, 1 ); /* wasted */
- constant_coef( setup, tri, slot, 0.0f, 2 ); /* wasted */
- constant_coef( setup, tri, slot, 0.0f, 3 ); /* wasted */
-}
+ tri = lp_scene_alloc_aligned( scene, *tri_size, 16 );
+ if (tri == NULL)
+ return NULL;
+ tri->inputs.stride = input_array_sz;
-/**
- * Compute the tri->coef[] array dadx, dady, a0 values.
- */
-static void setup_tri_coefficients( struct lp_setup_context *setup,
- struct lp_rast_triangle *tri,
- float oneoverarea,
- const float (*v1)[4],
- const float (*v2)[4],
- const float (*v3)[4],
- boolean frontface)
-{
- unsigned slot;
+ {
+ char *a = (char *)tri;
+ char *b = (char *)&GET_PLANES(tri)[nr_planes];
+ assert(b - a == *tri_size);
+ }
- /* The internal position input is in slot zero:
- */
- setup_fragcoord_coef(setup, tri, oneoverarea, 0, v1, v2, v3);
+ return tri;
+}
- /* setup interpolation for all the remaining attributes:
- */
- for (slot = 0; slot < setup->fs.nr_inputs; slot++) {
- unsigned vert_attr = setup->fs.input[slot].src_index;
- unsigned i;
-
- switch (setup->fs.input[slot].interp) {
- case LP_INTERP_CONSTANT:
- for (i = 0; i < NUM_CHANNELS; i++)
- constant_coef(setup, tri, slot+1, v3[vert_attr][i], i);
- break;
-
- case LP_INTERP_LINEAR:
- for (i = 0; i < NUM_CHANNELS; i++)
- linear_coef(setup, tri, oneoverarea, slot+1, v1, v2, v3, vert_attr, i);
- break;
-
- case LP_INTERP_PERSPECTIVE:
- for (i = 0; i < NUM_CHANNELS; i++)
- perspective_coef(setup, tri, oneoverarea, slot+1, v1, v2, v3, vert_attr, i);
- break;
-
- case LP_INTERP_POSITION:
- /* XXX: fix me - duplicates the values in slot zero.
- */
- setup_fragcoord_coef(setup, tri, oneoverarea, slot+1, v1, v2, v3);
- break;
+void
+lp_setup_print_vertex(struct lp_setup_context *setup,
+ const char *name,
+ const float (*v)[4])
+{
+ const struct lp_setup_variant_key *key = &setup->setup.variant->key;
+ int i, j;
- case LP_INTERP_FACING:
- setup_facing_coef(setup, tri, slot+1, frontface);
- break;
+ debug_printf(" wpos (%s[0]) xyzw %f %f %f %f\n",
+ name,
+ v[0][0], v[0][1], v[0][2], v[0][3]);
- default:
- assert(0);
- }
- }
-}
+ for (i = 0; i < key->num_inputs; i++) {
+ const float *in = v[key->inputs[i].src_index];
+ debug_printf(" in[%d] (%s[%d]) %s%s%s%s ",
+ i,
+ name, key->inputs[i].src_index,
+ (key->inputs[i].usage_mask & 0x1) ? "x" : " ",
+ (key->inputs[i].usage_mask & 0x2) ? "y" : " ",
+ (key->inputs[i].usage_mask & 0x4) ? "z" : " ",
+ (key->inputs[i].usage_mask & 0x8) ? "w" : " ");
+ for (j = 0; j < 4; j++)
+ if (key->inputs[i].usage_mask & (1<<j))
+ debug_printf("%.5f ", in[j]);
-static INLINE int subpixel_snap( float a )
-{
- return util_iround(FIXED_ONE * a - (FIXED_ONE / 2));
+ debug_printf("\n");
+ }
}
-
/**
- * Alloc space for a new triangle plus the input.a0/dadx/dady arrays
- * immediately after it.
- * The memory is allocated from the per-scene pool, not per-tile.
- * \param tri_size returns number of bytes allocated
- * \param nr_inputs number of fragment shader inputs
- * \return pointer to triangle space
+ * Print triangle vertex attribs (for debug).
*/
-static INLINE struct lp_rast_triangle *
-alloc_triangle(struct lp_scene *scene, unsigned nr_inputs, unsigned *tri_size)
+void
+lp_setup_print_triangle(struct lp_setup_context *setup,
+ const float (*v0)[4],
+ const float (*v1)[4],
+ const float (*v2)[4])
{
- unsigned input_array_sz = NUM_CHANNELS * (nr_inputs + 1) * sizeof(float);
- struct lp_rast_triangle *tri;
- unsigned bytes;
- char *inputs;
+ debug_printf("triangle\n");
- assert(sizeof(*tri) % 16 == 0);
+ {
+ const float ex = v0[0][0] - v2[0][0];
+ const float ey = v0[0][1] - v2[0][1];
+ const float fx = v1[0][0] - v2[0][0];
+ const float fy = v1[0][1] - v2[0][1];
+
+ /* det = cross(e,f).z */
+ const float det = ex * fy - ey * fx;
+ if (det < 0.0f)
+ debug_printf(" - ccw\n");
+ else if (det > 0.0f)
+ debug_printf(" - cw\n");
+ else
+ debug_printf(" - zero area\n");
+ }
- bytes = sizeof(*tri) + (3 * input_array_sz);
+ lp_setup_print_vertex(setup, "v0", v0);
+ lp_setup_print_vertex(setup, "v1", v1);
+ lp_setup_print_vertex(setup, "v2", v2);
+}
- tri = lp_scene_alloc_aligned( scene, bytes, 16 );
- inputs = (char *) (tri + 1);
- tri->inputs.a0 = (float (*)[4]) inputs;
- tri->inputs.dadx = (float (*)[4]) (inputs + input_array_sz);
- tri->inputs.dady = (float (*)[4]) (inputs + 2 * input_array_sz);
+#define MAX_PLANES 8
+static unsigned
+lp_rast_tri_tab[MAX_PLANES+1] = {
+ 0, /* should be impossible */
+ LP_RAST_OP_TRIANGLE_1,
+ LP_RAST_OP_TRIANGLE_2,
+ LP_RAST_OP_TRIANGLE_3,
+ LP_RAST_OP_TRIANGLE_4,
+ LP_RAST_OP_TRIANGLE_5,
+ LP_RAST_OP_TRIANGLE_6,
+ LP_RAST_OP_TRIANGLE_7,
+ LP_RAST_OP_TRIANGLE_8
+};
- *tri_size = bytes;
-
- return tri;
-}
/**
- * Print triangle vertex attribs (for debug).
+ * The primitive covers the whole tile- shade whole tile.
+ *
+ * \param tx, ty the tile position in tiles, not pixels
*/
-static void
-print_triangle(struct lp_setup_context *setup,
- const float (*v1)[4],
- const float (*v2)[4],
- const float (*v3)[4])
+static boolean
+lp_setup_whole_tile(struct lp_setup_context *setup,
+ const struct lp_rast_shader_inputs *inputs,
+ int tx, int ty)
{
- uint i;
+ struct lp_scene *scene = setup->scene;
- debug_printf("llvmpipe triangle\n");
- for (i = 0; i < setup->fs.nr_inputs; i++) {
- debug_printf(" v1[%d]: %f %f %f %f\n", i,
- v1[i][0], v1[i][1], v1[i][2], v1[i][3]);
- }
- for (i = 0; i < setup->fs.nr_inputs; i++) {
- debug_printf(" v2[%d]: %f %f %f %f\n", i,
- v2[i][0], v2[i][1], v2[i][2], v2[i][3]);
- }
- for (i = 0; i < setup->fs.nr_inputs; i++) {
- debug_printf(" v3[%d]: %f %f %f %f\n", i,
- v3[i][0], v3[i][1], v3[i][2], v3[i][3]);
+ LP_COUNT(nr_fully_covered_64);
+
+ /* if variant is opaque and scissor doesn't effect the tile */
+ if (inputs->opaque) {
+ if (!scene->fb.zsbuf) {
+ /*
+ * All previous rendering will be overwritten so reset the bin.
+ */
+ lp_scene_bin_reset( scene, tx, ty );
+ }
+
+ LP_COUNT(nr_shade_opaque_64);
+ return lp_scene_bin_cmd_with_state( scene, tx, ty,
+ setup->fs.stored,
+ LP_RAST_OP_SHADE_TILE_OPAQUE,
+ lp_rast_arg_inputs(inputs) );
+ } else {
+ LP_COUNT(nr_shade_64);
+ return lp_scene_bin_cmd_with_state( scene, tx, ty,
+ setup->fs.stored,
+ LP_RAST_OP_SHADE_TILE,
+ lp_rast_arg_inputs(inputs) );
}
}
* framebuffer tiles are touched. Put the triangle in the scene's
* bins for the tiles which we overlap.
*/
-static void
+static boolean
do_triangle_ccw(struct lp_setup_context *setup,
+ const float (*v0)[4],
const float (*v1)[4],
const float (*v2)[4],
- const float (*v3)[4],
boolean frontfacing )
{
- /* x/y positions in fixed point */
- const int x1 = subpixel_snap(v1[0][0] + 0.5 - setup->pixel_offset);
- const int x2 = subpixel_snap(v2[0][0] + 0.5 - setup->pixel_offset);
- const int x3 = subpixel_snap(v3[0][0] + 0.5 - setup->pixel_offset);
- const int y1 = subpixel_snap(v1[0][1] + 0.5 - setup->pixel_offset);
- const int y2 = subpixel_snap(v2[0][1] + 0.5 - setup->pixel_offset);
- const int y3 = subpixel_snap(v3[0][1] + 0.5 - setup->pixel_offset);
-
- struct lp_scene *scene = lp_setup_get_current_scene(setup);
+ struct lp_scene *scene = setup->scene;
+ const struct lp_setup_variant_key *key = &setup->setup.variant->key;
struct lp_rast_triangle *tri;
- int area;
- float oneoverarea;
- int minx, maxx, miny, maxy;
+ struct lp_rast_plane *plane;
+ int x[4];
+ int y[4];
+ struct u_rect bbox;
unsigned tri_bytes;
+ int nr_planes = 3;
if (0)
- print_triangle(setup, v1, v2, v3);
+ lp_setup_print_triangle(setup, v0, v1, v2);
- tri = alloc_triangle(scene, setup->fs.nr_inputs, &tri_bytes);
-
-#ifdef DEBUG
- tri->v[0][0] = v1[0][0];
- tri->v[1][0] = v2[0][0];
- tri->v[2][0] = v3[0][0];
- tri->v[0][1] = v1[0][1];
- tri->v[1][1] = v2[0][1];
- tri->v[2][1] = v3[0][1];
-#endif
-
- tri->dx12 = x1 - x2;
- tri->dx23 = x2 - x3;
- tri->dx31 = x3 - x1;
+ if (setup->scissor_test) {
+ nr_planes = 7;
+ }
+ else {
+ nr_planes = 3;
+ }
- tri->dy12 = y1 - y2;
- tri->dy23 = y2 - y3;
- tri->dy31 = y3 - y1;
+ /* x/y positions in fixed point */
+ x[0] = subpixel_snap(v0[0][0] - setup->pixel_offset);
+ x[1] = subpixel_snap(v1[0][0] - setup->pixel_offset);
+ x[2] = subpixel_snap(v2[0][0] - setup->pixel_offset);
+ x[3] = 0;
+ y[0] = subpixel_snap(v0[0][1] - setup->pixel_offset);
+ y[1] = subpixel_snap(v1[0][1] - setup->pixel_offset);
+ y[2] = subpixel_snap(v2[0][1] - setup->pixel_offset);
+ y[3] = 0;
+
- area = (tri->dx12 * tri->dy31 - tri->dx31 * tri->dy12);
+ /* Bounding rectangle (in pixels) */
+ {
+ /* Yes this is necessary to accurately calculate bounding boxes
+ * with the two fill-conventions we support. GL (normally) ends
+ * up needing a bottom-left fill convention, which requires
+ * slightly different rounding.
+ */
+ int adj = (setup->pixel_offset != 0) ? 1 : 0;
- LP_COUNT(nr_tris);
+ bbox.x0 = (MIN3(x[0], x[1], x[2]) + (FIXED_ONE-1)) >> FIXED_ORDER;
+ bbox.x1 = (MAX3(x[0], x[1], x[2]) + (FIXED_ONE-1)) >> FIXED_ORDER;
+ bbox.y0 = (MIN3(y[0], y[1], y[2]) + (FIXED_ONE-1) + adj) >> FIXED_ORDER;
+ bbox.y1 = (MAX3(y[0], y[1], y[2]) + (FIXED_ONE-1) + adj) >> FIXED_ORDER;
- /* Cull non-ccw and zero-sized triangles.
- *
- * XXX: subject to overflow??
- */
- if (area <= 0) {
- lp_scene_putback_data( scene, tri_bytes );
- LP_COUNT(nr_culled_tris);
- return;
+ /* Inclusive coordinates:
+ */
+ bbox.x1--;
+ bbox.y1--;
}
- /* Bounding rectangle (in pixels) */
- minx = (MIN3(x1, x2, x3) + (FIXED_ONE-1)) >> FIXED_ORDER;
- maxx = (MAX3(x1, x2, x3) + (FIXED_ONE-1)) >> FIXED_ORDER;
- miny = (MIN3(y1, y2, y3) + (FIXED_ONE-1)) >> FIXED_ORDER;
- maxy = (MAX3(y1, y2, y3) + (FIXED_ONE-1)) >> FIXED_ORDER;
-
- if (setup->scissor_test) {
- minx = MAX2(minx, setup->scissor.current.minx);
- maxx = MIN2(maxx, setup->scissor.current.maxx);
- miny = MAX2(miny, setup->scissor.current.miny);
- maxy = MIN2(maxy, setup->scissor.current.maxy);
+ if (bbox.x1 < bbox.x0 ||
+ bbox.y1 < bbox.y0) {
+ if (0) debug_printf("empty bounding box\n");
+ LP_COUNT(nr_culled_tris);
+ return TRUE;
}
- if (miny == maxy ||
- minx == maxx) {
- lp_scene_putback_data( scene, tri_bytes );
+ if (!u_rect_test_intersection(&setup->draw_region, &bbox)) {
+ if (0) debug_printf("offscreen\n");
LP_COUNT(nr_culled_tris);
- return;
+ return TRUE;
}
- /*
+ /* Can safely discard negative regions, but need to keep hold of
+ * information about when the triangle extends past screen
+ * boundaries. See trimmed_box in lp_setup_bin_triangle().
*/
- oneoverarea = ((float)FIXED_ONE) / (float)area;
+ bbox.x0 = MAX2(bbox.x0, 0);
+ bbox.y0 = MAX2(bbox.y0, 0);
+
+ tri = lp_setup_alloc_triangle(scene,
+ key->num_inputs,
+ nr_planes,
+ &tri_bytes);
+ if (!tri)
+ return FALSE;
+
+#if 0
+ tri->v[0][0] = v0[0][0];
+ tri->v[1][0] = v1[0][0];
+ tri->v[2][0] = v2[0][0];
+ tri->v[0][1] = v0[0][1];
+ tri->v[1][1] = v1[0][1];
+ tri->v[2][1] = v2[0][1];
+#endif
+
+ LP_COUNT(nr_tris);
/* Setup parameter interpolants:
*/
- setup_tri_coefficients( setup, tri, oneoverarea, v1, v2, v3, frontfacing );
+ setup->setup.variant->jit_function( v0,
+ v1,
+ v2,
+ frontfacing,
+ GET_A0(&tri->inputs),
+ GET_DADX(&tri->inputs),
+ GET_DADY(&tri->inputs) );
+
+ tri->inputs.frontfacing = frontfacing;
+ tri->inputs.disable = FALSE;
+ tri->inputs.opaque = setup->fs.current.variant->opaque;
- tri->inputs.facing = frontfacing ? 1.0F : -1.0F;
+ if (0)
+ lp_dump_setup_coef(&setup->setup.variant->key,
+ (const float (*)[4])GET_A0(&tri->inputs),
+ (const float (*)[4])GET_DADX(&tri->inputs),
+ (const float (*)[4])GET_DADY(&tri->inputs));
- /* half-edge constants, will be interated over the whole render target.
- */
- tri->c1 = tri->dy12 * x1 - tri->dx12 * y1;
- tri->c2 = tri->dy23 * x2 - tri->dx23 * y2;
- tri->c3 = tri->dy31 * x3 - tri->dx31 * y3;
+ plane = GET_PLANES(tri);
- /* correct for top-left fill convention:
- */
- if (tri->dy12 < 0 || (tri->dy12 == 0 && tri->dx12 > 0)) tri->c1++;
- if (tri->dy23 < 0 || (tri->dy23 == 0 && tri->dx23 > 0)) tri->c2++;
- if (tri->dy31 < 0 || (tri->dy31 == 0 && tri->dx31 > 0)) tri->c3++;
-
- tri->dy12 *= FIXED_ONE;
- tri->dy23 *= FIXED_ONE;
- tri->dy31 *= FIXED_ONE;
-
- tri->dx12 *= FIXED_ONE;
- tri->dx23 *= FIXED_ONE;
- tri->dx31 *= FIXED_ONE;
-
- /* find trivial reject offsets for each edge for a single-pixel
- * sized block. These will be scaled up at each recursive level to
- * match the active blocksize. Scaling in this way works best if
- * the blocks are square.
- */
- tri->eo1 = 0;
- if (tri->dy12 < 0) tri->eo1 -= tri->dy12;
- if (tri->dx12 > 0) tri->eo1 += tri->dx12;
+#if defined(PIPE_ARCH_SSE)
+ {
+ __m128i vertx, verty;
+ __m128i shufx, shufy;
+ __m128i dcdx, dcdy, c;
+ __m128i unused;
+ __m128i dcdx_neg_mask;
+ __m128i dcdy_neg_mask;
+ __m128i dcdx_zero_mask;
+ __m128i top_left_flag;
+ __m128i c_inc_mask, c_inc;
+ __m128i eo, p0, p1, p2;
+ __m128i zero = _mm_setzero_si128();
- tri->eo2 = 0;
- if (tri->dy23 < 0) tri->eo2 -= tri->dy23;
- if (tri->dx23 > 0) tri->eo2 += tri->dx23;
+ vertx = _mm_loadu_si128((__m128i *)x); /* vertex x coords */
+ verty = _mm_loadu_si128((__m128i *)y); /* vertex y coords */
- tri->eo3 = 0;
- if (tri->dy31 < 0) tri->eo3 -= tri->dy31;
- if (tri->dx31 > 0) tri->eo3 += tri->dx31;
+ shufx = _mm_shuffle_epi32(vertx, _MM_SHUFFLE(3,0,2,1));
+ shufy = _mm_shuffle_epi32(verty, _MM_SHUFFLE(3,0,2,1));
- /* Calculate trivial accept offsets from the above.
- */
- tri->ei1 = tri->dx12 - tri->dy12 - tri->eo1;
- tri->ei2 = tri->dx23 - tri->dy23 - tri->eo2;
- tri->ei3 = tri->dx31 - tri->dy31 - tri->eo3;
+ dcdx = _mm_sub_epi32(verty, shufy);
+ dcdy = _mm_sub_epi32(vertx, shufx);
- /* Fill in the inputs.step[][] arrays.
- * We've manually unrolled some loops here.
- */
+ dcdx_neg_mask = _mm_srai_epi32(dcdx, 31);
+ dcdx_zero_mask = _mm_cmpeq_epi32(dcdx, zero);
+ dcdy_neg_mask = _mm_srai_epi32(dcdy, 31);
+
+ top_left_flag = _mm_set1_epi32((setup->pixel_offset == 0) ? ~0 : 0);
+
+ c_inc_mask = _mm_or_si128(dcdx_neg_mask,
+ _mm_and_si128(dcdx_zero_mask,
+ _mm_xor_si128(dcdy_neg_mask,
+ top_left_flag)));
+
+ c_inc = _mm_srli_epi32(c_inc_mask, 31);
+
+ c = _mm_sub_epi32(mm_mullo_epi32(dcdx, vertx),
+ mm_mullo_epi32(dcdy, verty));
+
+ c = _mm_add_epi32(c, c_inc);
+
+ /* Scale up to match c:
+ */
+ dcdx = _mm_slli_epi32(dcdx, FIXED_ORDER);
+ dcdy = _mm_slli_epi32(dcdy, FIXED_ORDER);
+
+ /* Calculate trivial reject values:
+ */
+ eo = _mm_sub_epi32(_mm_andnot_si128(dcdy_neg_mask, dcdy),
+ _mm_and_si128(dcdx_neg_mask, dcdx));
+
+ /* ei = _mm_sub_epi32(_mm_sub_epi32(dcdy, dcdx), eo); */
+
+ /* Pointless transpose which gets undone immediately in
+ * rasterization:
+ */
+ transpose4_epi32(&c, &dcdx, &dcdy, &eo,
+ &p0, &p1, &p2, &unused);
+
+ _mm_store_si128((__m128i *)&plane[0], p0);
+ _mm_store_si128((__m128i *)&plane[1], p1);
+ _mm_store_si128((__m128i *)&plane[2], p2);
+ }
+#else
{
- const int xstep1 = -tri->dy12;
- const int xstep2 = -tri->dy23;
- const int xstep3 = -tri->dy31;
- const int ystep1 = tri->dx12;
- const int ystep2 = tri->dx23;
- const int ystep3 = tri->dx31;
-
-#define SETUP_STEP(i, x, y) \
- do { \
- tri->inputs.step[0][i] = x * xstep1 + y * ystep1; \
- tri->inputs.step[1][i] = x * xstep2 + y * ystep2; \
- tri->inputs.step[2][i] = x * xstep3 + y * ystep3; \
- } while (0)
-
- SETUP_STEP(0, 0, 0);
- SETUP_STEP(1, 1, 0);
- SETUP_STEP(2, 0, 1);
- SETUP_STEP(3, 1, 1);
-
- SETUP_STEP(4, 2, 0);
- SETUP_STEP(5, 3, 0);
- SETUP_STEP(6, 2, 1);
- SETUP_STEP(7, 3, 1);
-
- SETUP_STEP(8, 0, 2);
- SETUP_STEP(9, 1, 2);
- SETUP_STEP(10, 0, 3);
- SETUP_STEP(11, 1, 3);
-
- SETUP_STEP(12, 2, 2);
- SETUP_STEP(13, 3, 2);
- SETUP_STEP(14, 2, 3);
- SETUP_STEP(15, 3, 3);
-#undef STEP
+ int i;
+ plane[0].dcdy = x[0] - x[1];
+ plane[1].dcdy = x[1] - x[2];
+ plane[2].dcdy = x[2] - x[0];
+ plane[0].dcdx = y[0] - y[1];
+ plane[1].dcdx = y[1] - y[2];
+ plane[2].dcdx = y[2] - y[0];
+
+ for (i = 0; i < 3; i++) {
+ /* half-edge constants, will be interated over the whole render
+ * target.
+ */
+ plane[i].c = plane[i].dcdx * x[i] - plane[i].dcdy * y[i];
+
+ /* correct for top-left vs. bottom-left fill convention.
+ *
+ * note that we're overloading gl_rasterization_rules to mean
+ * both (0.5,0.5) pixel centers *and* bottom-left filling
+ * convention.
+ *
+ * GL actually has a top-left filling convention, but GL's
+ * notion of "top" differs from gallium's...
+ *
+ * Also, sometimes (in FBO cases) GL will render upside down
+ * to its usual method, in which case it will probably want
+ * to use the opposite, top-left convention.
+ */
+ if (plane[i].dcdx < 0) {
+ /* both fill conventions want this - adjust for left edges */
+ plane[i].c++;
+ }
+ else if (plane[i].dcdx == 0) {
+ if (setup->pixel_offset == 0) {
+ /* correct for top-left fill convention:
+ */
+ if (plane[i].dcdy > 0) plane[i].c++;
+ }
+ else {
+ /* correct for bottom-left fill convention:
+ */
+ if (plane[i].dcdy < 0) plane[i].c++;
+ }
+ }
+
+ plane[i].dcdx *= FIXED_ONE;
+ plane[i].dcdy *= FIXED_ONE;
+
+ /* find trivial reject offsets for each edge for a single-pixel
+ * sized block. These will be scaled up at each recursive level to
+ * match the active blocksize. Scaling in this way works best if
+ * the blocks are square.
+ */
+ plane[i].eo = 0;
+ if (plane[i].dcdx < 0) plane[i].eo -= plane[i].dcdx;
+ if (plane[i].dcdy > 0) plane[i].eo += plane[i].dcdy;
+ }
+ }
+#endif
+
+ if (0) {
+ debug_printf("p0: %08x/%08x/%08x/%08x\n",
+ plane[0].c,
+ plane[0].dcdx,
+ plane[0].dcdy,
+ plane[0].eo);
+
+ debug_printf("p1: %08x/%08x/%08x/%08x\n",
+ plane[1].c,
+ plane[1].dcdx,
+ plane[1].dcdy,
+ plane[1].eo);
+
+ debug_printf("p0: %08x/%08x/%08x/%08x\n",
+ plane[2].c,
+ plane[2].dcdx,
+ plane[2].dcdy,
+ plane[2].eo);
+ }
+
+
+ /*
+ * When rasterizing scissored tris, use the intersection of the
+ * triangle bounding box and the scissor rect to generate the
+ * scissor planes.
+ *
+ * This permits us to cut off the triangle "tails" that are present
+ * in the intermediate recursive levels caused when two of the
+ * triangles edges don't diverge quickly enough to trivially reject
+ * exterior blocks from the triangle.
+ *
+ * It's not really clear if it's worth worrying about these tails,
+ * but since we generate the planes for each scissored tri, it's
+ * free to trim them in this case.
+ *
+ * Note that otherwise, the scissor planes only vary in 'C' value,
+ * and even then only on state-changes. Could alternatively store
+ * these planes elsewhere.
+ */
+ if (nr_planes == 7) {
+ const struct u_rect *scissor = &setup->scissor;
+
+ plane[3].dcdx = -1;
+ plane[3].dcdy = 0;
+ plane[3].c = 1-scissor->x0;
+ plane[3].eo = 1;
+
+ plane[4].dcdx = 1;
+ plane[4].dcdy = 0;
+ plane[4].c = scissor->x1+1;
+ plane[4].eo = 0;
+
+ plane[5].dcdx = 0;
+ plane[5].dcdy = 1;
+ plane[5].c = 1-scissor->y0;
+ plane[5].eo = 1;
+
+ plane[6].dcdx = 0;
+ plane[6].dcdy = -1;
+ plane[6].c = scissor->y1+1;
+ plane[6].eo = 0;
}
- /*
- * All fields of 'tri' are now set. The remaining code here is
- * concerned with binning.
+ return lp_setup_bin_triangle( setup, tri, &bbox, nr_planes );
+}
+
+/*
+ * Round to nearest less or equal power of two of the input.
+ *
+ * Undefined if no bit set exists, so code should check against 0 first.
+ */
+static INLINE uint32_t
+floor_pot(uint32_t n)
+{
+#if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86)
+ if (n == 0)
+ return 0;
+
+ __asm__("bsr %1,%0"
+ : "=r" (n)
+ : "rm" (n));
+ return 1 << n;
+#else
+ n |= (n >> 1);
+ n |= (n >> 2);
+ n |= (n >> 4);
+ n |= (n >> 8);
+ n |= (n >> 16);
+ return n - (n >> 1);
+#endif
+}
+
+
+boolean
+lp_setup_bin_triangle( struct lp_setup_context *setup,
+ struct lp_rast_triangle *tri,
+ const struct u_rect *bbox,
+ int nr_planes )
+{
+ struct lp_scene *scene = setup->scene;
+ struct u_rect trimmed_box = *bbox;
+ int i;
+
+ /* What is the largest power-of-two boundary this triangle crosses:
*/
+ int dx = floor_pot((bbox->x0 ^ bbox->x1) |
+ (bbox->y0 ^ bbox->y1));
- /* Convert to tile coordinates:
+ /* The largest dimension of the rasterized area of the triangle
+ * (aligned to a 4x4 grid), rounded down to the nearest power of two:
*/
- minx = minx / TILE_SIZE;
- miny = miny / TILE_SIZE;
- maxx = maxx / TILE_SIZE;
- maxy = maxy / TILE_SIZE;
+ int sz = floor_pot((bbox->x1 - (bbox->x0 & ~3)) |
+ (bbox->y1 - (bbox->y0 & ~3)));
- /*
- * Clamp to framebuffer size
+ /* Now apply scissor, etc to the bounding box. Could do this
+ * earlier, but it confuses the logic for tri-16 and would force
+ * the rasterizer to also respect scissor, etc, just for the rare
+ * cases where a small triangle extends beyond the scissor.
*/
- minx = MAX2(minx, 0);
- miny = MAX2(miny, 0);
- maxx = MIN2(maxx, scene->tiles_x - 1);
- maxy = MIN2(maxy, scene->tiles_y - 1);
+ u_rect_find_intersection(&setup->draw_region, &trimmed_box);
/* Determine which tile(s) intersect the triangle's bounding box
*/
- if (miny == maxy && minx == maxx)
+ if (dx < TILE_SIZE)
{
+ int ix0 = bbox->x0 / TILE_SIZE;
+ int iy0 = bbox->y0 / TILE_SIZE;
+ int px = bbox->x0 & 63 & ~3;
+ int py = bbox->y0 & 63 & ~3;
+ int mask = px | (py << 8);
+
+ assert(iy0 == bbox->y1 / TILE_SIZE &&
+ ix0 == bbox->x1 / TILE_SIZE);
+
+ if (nr_planes == 3) {
+ if (sz < 4)
+ {
+ /* Triangle is contained in a single 4x4 stamp:
+ */
+ return lp_scene_bin_cmd_with_state( scene, ix0, iy0,
+ setup->fs.stored,
+ LP_RAST_OP_TRIANGLE_3_4,
+ lp_rast_arg_triangle(tri, mask) );
+ }
+
+ if (sz < 16)
+ {
+ /* Triangle is contained in a single 16x16 block:
+ */
+ return lp_scene_bin_cmd_with_state( scene, ix0, iy0,
+ setup->fs.stored,
+ LP_RAST_OP_TRIANGLE_3_16,
+ lp_rast_arg_triangle(tri, mask) );
+ }
+ }
+ else if (nr_planes == 4 && sz < 16)
+ {
+ return lp_scene_bin_cmd_with_state(scene, ix0, iy0,
+ setup->fs.stored,
+ LP_RAST_OP_TRIANGLE_4_16,
+ lp_rast_arg_triangle(tri, mask) );
+ }
+
+
/* Triangle is contained in a single tile:
*/
- lp_scene_bin_command( scene, minx, miny, lp_rast_triangle,
- lp_rast_arg_triangle(tri) );
+ return lp_scene_bin_cmd_with_state( scene, ix0, iy0, setup->fs.stored,
+ lp_rast_tri_tab[nr_planes],
+ lp_rast_arg_triangle(tri, (1<<nr_planes)-1) );
}
- else
+ else
{
- int c1 = (tri->c1 +
- tri->dx12 * miny * TILE_SIZE -
- tri->dy12 * minx * TILE_SIZE);
- int c2 = (tri->c2 +
- tri->dx23 * miny * TILE_SIZE -
- tri->dy23 * minx * TILE_SIZE);
- int c3 = (tri->c3 +
- tri->dx31 * miny * TILE_SIZE -
- tri->dy31 * minx * TILE_SIZE);
-
- int ei1 = tri->ei1 << TILE_ORDER;
- int ei2 = tri->ei2 << TILE_ORDER;
- int ei3 = tri->ei3 << TILE_ORDER;
-
- int eo1 = tri->eo1 << TILE_ORDER;
- int eo2 = tri->eo2 << TILE_ORDER;
- int eo3 = tri->eo3 << TILE_ORDER;
-
- int xstep1 = -(tri->dy12 << TILE_ORDER);
- int xstep2 = -(tri->dy23 << TILE_ORDER);
- int xstep3 = -(tri->dy31 << TILE_ORDER);
-
- int ystep1 = tri->dx12 << TILE_ORDER;
- int ystep2 = tri->dx23 << TILE_ORDER;
- int ystep3 = tri->dx31 << TILE_ORDER;
+ struct lp_rast_plane *plane = GET_PLANES(tri);
+ int c[MAX_PLANES];
+ int ei[MAX_PLANES];
+
+ int eo[MAX_PLANES];
+ int xstep[MAX_PLANES];
+ int ystep[MAX_PLANES];
int x, y;
+ int ix0 = trimmed_box.x0 / TILE_SIZE;
+ int iy0 = trimmed_box.y0 / TILE_SIZE;
+ int ix1 = trimmed_box.x1 / TILE_SIZE;
+ int iy1 = trimmed_box.y1 / TILE_SIZE;
+
+ for (i = 0; i < nr_planes; i++) {
+ c[i] = (plane[i].c +
+ plane[i].dcdy * iy0 * TILE_SIZE -
+ plane[i].dcdx * ix0 * TILE_SIZE);
+
+ ei[i] = (plane[i].dcdy -
+ plane[i].dcdx -
+ plane[i].eo) << TILE_ORDER;
+
+ eo[i] = plane[i].eo << TILE_ORDER;
+ xstep[i] = -(plane[i].dcdx << TILE_ORDER);
+ ystep[i] = plane[i].dcdy << TILE_ORDER;
+ }
+
+
/* Test tile-sized blocks against the triangle.
* Discard blocks fully outside the tri. If the block is fully
* contained inside the tri, bin an lp_rast_shade_tile command.
* Else, bin a lp_rast_triangle command.
*/
- for (y = miny; y <= maxy; y++)
+ for (y = iy0; y <= iy1; y++)
{
- int cx1 = c1;
- int cx2 = c2;
- int cx3 = c3;
boolean in = FALSE; /* are we inside the triangle? */
+ int cx[MAX_PLANES];
+
+ for (i = 0; i < nr_planes; i++)
+ cx[i] = c[i];
- for (x = minx; x <= maxx; x++)
+ for (x = ix0; x <= ix1; x++)
{
- if (cx1 + eo1 < 0 ||
- cx2 + eo2 < 0 ||
- cx3 + eo3 < 0)
- {
- /* do nothing */
+ int out = 0;
+ int partial = 0;
+
+ for (i = 0; i < nr_planes; i++) {
+ int planeout = cx[i] + eo[i];
+ int planepartial = cx[i] + ei[i] - 1;
+ out |= (planeout >> 31);
+ partial |= (planepartial >> 31) & (1<<i);
+ }
+
+ if (out) {
+ /* do nothing */
+ if (in)
+ break; /* exiting triangle, all done with this row */
LP_COUNT(nr_empty_64);
- if (in)
- break; /* exiting triangle, all done with this row */
- }
- else if (cx1 + ei1 > 0 &&
- cx2 + ei2 > 0 &&
- cx3 + ei3 > 0)
- {
+ }
+ else if (partial) {
+ /* Not trivially accepted by at least one plane -
+ * rasterize/shade partial tile
+ */
+ int count = util_bitcount(partial);
+ in = TRUE;
+
+ if (!lp_scene_bin_cmd_with_state( scene, x, y,
+ setup->fs.stored,
+ lp_rast_tri_tab[count],
+ lp_rast_arg_triangle(tri, partial) ))
+ goto fail;
+
+ LP_COUNT(nr_partially_covered_64);
+ }
+ else {
/* triangle covers the whole tile- shade whole tile */
LP_COUNT(nr_fully_covered_64);
- in = TRUE;
- if(setup->fs.current.opaque) {
- lp_scene_bin_reset( scene, x, y );
- lp_scene_bin_command( scene, x, y,
- lp_rast_set_state,
- lp_rast_arg_state(setup->fs.stored) );
- }
- lp_scene_bin_command( scene, x, y,
- lp_rast_shade_tile,
- lp_rast_arg_inputs(&tri->inputs) );
- }
- else
- {
- /* rasterizer/shade partial tile */
- LP_COUNT(nr_partially_covered_64);
- in = TRUE;
- lp_scene_bin_command( scene, x, y,
- lp_rast_triangle,
- lp_rast_arg_triangle(tri) );
- }
+ in = TRUE;
+ if (!lp_setup_whole_tile(setup, &tri->inputs, x, y))
+ goto fail;
+ }
/* Iterate cx values across the region:
*/
- cx1 += xstep1;
- cx2 += xstep2;
- cx3 += xstep3;
+ for (i = 0; i < nr_planes; i++)
+ cx[i] += xstep[i];
}
/* Iterate c values down the region:
*/
- c1 += ystep1;
- c2 += ystep2;
- c3 += ystep3;
+ for (i = 0; i < nr_planes; i++)
+ c[i] += ystep[i];
}
}
+
+ return TRUE;
+
+fail:
+ /* Need to disable any partially binned triangle. This is easier
+ * than trying to locate all the triangle, shade-tile, etc,
+ * commands which may have been binned.
+ */
+ tri->inputs.disable = TRUE;
+ return FALSE;
}
/**
- * Draw triangle if it's CW, cull otherwise.
+ * Try to draw the triangle, restart the scene on failure.
*/
-static void triangle_cw( struct lp_setup_context *setup,
- const float (*v0)[4],
- const float (*v1)[4],
- const float (*v2)[4] )
+static void retry_triangle_ccw( struct lp_setup_context *setup,
+ const float (*v0)[4],
+ const float (*v1)[4],
+ const float (*v2)[4],
+ boolean front)
+{
+ if (!do_triangle_ccw( setup, v0, v1, v2, front ))
+ {
+ if (!lp_setup_flush_and_restart(setup))
+ return;
+
+ if (!do_triangle_ccw( setup, v0, v1, v2, front ))
+ return;
+ }
+}
+
+static INLINE float
+calc_area(const float (*v0)[4],
+ const float (*v1)[4],
+ const float (*v2)[4])
{
- do_triangle_ccw( setup, v1, v0, v2, !setup->ccw_is_frontface );
+ float dx01 = v0[0][0] - v1[0][0];
+ float dy01 = v0[0][1] - v1[0][1];
+ float dx20 = v2[0][0] - v0[0][0];
+ float dy20 = v2[0][1] - v0[0][1];
+ return dx01 * dy20 - dx20 * dy01;
}
/**
- * Draw triangle if it's CCW, cull otherwise.
+ * Draw triangle if it's CW, cull otherwise.
*/
-static void triangle_ccw( struct lp_setup_context *setup,
+static void triangle_cw( struct lp_setup_context *setup,
const float (*v0)[4],
const float (*v1)[4],
const float (*v2)[4] )
{
- do_triangle_ccw( setup, v0, v1, v2, setup->ccw_is_frontface );
+ float area = calc_area(v0, v1, v2);
+
+ if (area < 0.0f)
+ retry_triangle_ccw(setup, v0, v2, v1, !setup->ccw_is_frontface);
}
+static void triangle_ccw( struct lp_setup_context *setup,
+ const float (*v0)[4],
+ const float (*v1)[4],
+ const float (*v2)[4])
+{
+ float area = calc_area(v0, v1, v2);
+
+ if (area > 0.0f)
+ retry_triangle_ccw(setup, v0, v1, v2, setup->ccw_is_frontface);
+}
/**
* Draw triangle whether it's CW or CCW.
const float (*v1)[4],
const float (*v2)[4] )
{
- /* edge vectors e = v0 - v2, f = v1 - v2 */
- const float ex = v0[0][0] - v2[0][0];
- const float ey = v0[0][1] - v2[0][1];
- const float fx = v1[0][0] - v2[0][0];
- const float fy = v1[0][1] - v2[0][1];
-
- /* det = cross(e,f).z */
- if (ex * fy - ey * fx < 0.0f)
- triangle_ccw( setup, v0, v1, v2 );
- else
- triangle_cw( setup, v0, v1, v2 );
+ float area = calc_area(v0, v1, v2);
+
+ if (0) {
+ assert(!util_is_inf_or_nan(v0[0][0]));
+ assert(!util_is_inf_or_nan(v0[0][1]));
+ assert(!util_is_inf_or_nan(v1[0][0]));
+ assert(!util_is_inf_or_nan(v1[0][1]));
+ assert(!util_is_inf_or_nan(v2[0][0]));
+ assert(!util_is_inf_or_nan(v2[0][1]));
+ assert(!util_is_inf_or_nan(area));
+ }
+
+ if (area > 0.0f)
+ retry_triangle_ccw( setup, v0, v1, v2, setup->ccw_is_frontface );
+ else if (area < 0.0f)
+ retry_triangle_ccw( setup, v0, v2, v1, !setup->ccw_is_frontface );
}
lp_setup_choose_triangle( struct lp_setup_context *setup )
{
switch (setup->cullmode) {
- case PIPE_WINDING_NONE:
+ case PIPE_FACE_NONE:
setup->triangle = triangle_both;
break;
- case PIPE_WINDING_CCW:
- setup->triangle = triangle_cw;
+ case PIPE_FACE_BACK:
+ setup->triangle = setup->ccw_is_frontface ? triangle_ccw : triangle_cw;
break;
- case PIPE_WINDING_CW:
- setup->triangle = triangle_ccw;
+ case PIPE_FACE_FRONT:
+ setup->triangle = setup->ccw_is_frontface ? triangle_cw : triangle_ccw;
break;
default:
setup->triangle = triangle_nop;