* Binning code for triangles
*/
-#include "lp_setup_context.h"
-#include "lp_rast.h"
#include "util/u_math.h"
#include "util/u_memory.h"
+#include "util/u_rect.h"
+#include "lp_perf.h"
+#include "lp_setup_context.h"
+#include "lp_setup_coef.h"
+#include "lp_rast.h"
+#include "lp_state_fs.h"
#define NUM_CHANNELS 4
-/**
- * Compute a0 for a constant-valued coefficient (GL_FLAT shading).
- */
-static void constant_coef( struct lp_rast_triangle *tri,
- unsigned slot,
- const float value,
- unsigned i )
+
+
+static INLINE int
+subpixel_snap(float a)
{
- tri->inputs.a0[slot][i] = value;
- tri->inputs.dadx[slot][i] = 0;
- tri->inputs.dady[slot][i] = 0;
+ 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_rast_triangle *tri,
- 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) * tri->oneoverarea;
- float dady = (da31 * tri->dx12 - tri->dx31 * da12) * tri->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] = (v1[vert_attr][i] -
- (dadx * (v1[0][0] - 0.5f) +
- dady * (v1[0][1] - 0.5f)));
-}
-/**
- * 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_rast_triangle *tri,
- 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) * tri->oneoverarea;
- float dady = (da31 * tri->dx12 - tri->dx31 * da12) * tri->oneoverarea;
-
-
- tri->inputs.dadx[slot][i] = dadx;
- tri->inputs.dady[slot][i] = dady;
- tri->inputs.a0[slot][i] = (a1 -
- (dadx * (v1[0][0] - 0.5f) +
- dady * (v1[0][1] - 0.5f)));
-}
/**
- * Special coefficient setup for gl_FragCoord.
- * X and Y are trivial, though Y has to be inverted for OpenGL.
- * 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 nr_inputs number of fragment shader inputs
+ * \return pointer to triangle space
*/
-static void
-setup_fragcoord_coef(struct lp_rast_triangle *tri,
- 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(tri, slot, v1, v2, v3, 0, 2);
- /*W*/
- linear_coef(tri, slot, v1, v2, v3, 0, 3);
-}
+ unsigned input_array_sz = NUM_CHANNELS * (nr_inputs + 1) * sizeof(float);
+ struct lp_rast_triangle *tri;
+ unsigned tri_bytes, bytes;
+ char *inputs;
+ tri_bytes = align(Offset(struct lp_rast_triangle, plane[nr_planes]), 16);
+ bytes = tri_bytes + (3 * input_array_sz);
-static void setup_facing_coef( struct lp_rast_triangle *tri,
- unsigned slot,
- boolean frontface )
-{
- constant_coef( tri, slot, 1.0f - frontface, 0 );
- constant_coef( tri, slot, 0.0f, 1 ); /* wasted */
- constant_coef( tri, slot, 0.0f, 2 ); /* wasted */
- constant_coef( tri, slot, 0.0f, 3 ); /* wasted */
-}
+ tri = lp_scene_alloc_aligned( scene, bytes, 16 );
+ if (tri) {
+ inputs = ((char *)tri) + tri_bytes;
+ 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);
-/**
- * Compute the tri->coef[] array dadx, dady, a0 values.
- */
-static void setup_tri_coefficients( struct setup_context *setup,
- struct lp_rast_triangle *tri,
- const float (*v1)[4],
- const float (*v2)[4],
- const float (*v3)[4],
- boolean frontface )
+ *tri_size = bytes;
+ }
+
+ return tri;
+}
+
+void
+lp_setup_print_vertex(struct lp_setup_context *setup,
+ const char *name,
+ const float (*v)[4])
{
- unsigned slot;
+ int i, j;
- /* The internal position input is in slot zero:
- */
- setup_fragcoord_coef(tri, 0, v1, v2, v3);
+ debug_printf(" wpos (%s[0]) xyzw %f %f %f %f\n",
+ name,
+ v[0][0], v[0][1], v[0][2], v[0][3]);
- /* setup interpolation for all the remaining attrbutes:
- */
- 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(tri, slot+1, v3[vert_attr][i], i);
- break;
-
- case LP_INTERP_LINEAR:
- for (i = 0; i < NUM_CHANNELS; i++)
- linear_coef(tri, slot+1, v1, v2, v3, vert_attr, i);
- break;
-
- case LP_INTERP_PERSPECTIVE:
- for (i = 0; i < NUM_CHANNELS; i++)
- perspective_coef(tri, 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(tri, slot+1, v1, v2, v3);
- break;
+ for (i = 0; i < setup->fs.nr_inputs; i++) {
+ const float *in = v[setup->fs.input[i].src_index];
- case LP_INTERP_FACING:
- setup_facing_coef(tri, slot+1, frontface);
- break;
+ debug_printf(" in[%d] (%s[%d]) %s%s%s%s ",
+ i,
+ name, setup->fs.input[i].src_index,
+ (setup->fs.input[i].usage_mask & 0x1) ? "x" : " ",
+ (setup->fs.input[i].usage_mask & 0x2) ? "y" : " ",
+ (setup->fs.input[i].usage_mask & 0x4) ? "z" : " ",
+ (setup->fs.input[i].usage_mask & 0x8) ? "w" : " ");
- default:
- assert(0);
- }
+ for (j = 0; j < 4; j++)
+ if (setup->fs.input[i].usage_mask & (1<<j))
+ debug_printf("%.5f ", in[j]);
+
+ debug_printf("\n");
}
}
-
-/* XXX: do this by add/subtracting a large floating point number:
+/**
+ * Print triangle vertex attribs (for debug).
*/
-static inline float subpixel_snap( float a )
+void
+lp_setup_print_triangle(struct lp_setup_context *setup,
+ const float (*v0)[4],
+ const float (*v1)[4],
+ const float (*v2)[4])
{
- int i = a * 16;
- return (float)i * (1.0/16);
-}
+ debug_printf("triangle\n");
+ {
+ 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");
+ }
-static INLINE void bin_triangle( struct cmd_block_list *list,
- const struct lp_rast_triangle arg )
-{
+ lp_setup_print_vertex(setup, "v0", v0);
+ lp_setup_print_vertex(setup, "v1", v1);
+ lp_setup_print_vertex(setup, "v2", v2);
}
-/* to avoid having to allocate power-of-four, square render targets,
- * end up having a specialized version of the above that runs only at
- * the topmost level.
+#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
+};
+
+
+
+/**
+ * The primitive covers the whole tile- shade whole tile.
*
- * at the topmost level there may be an arbitary number of steps on
- * either dimension, so this loop needs to be either separately
- * code-generated and unrolled for each render target size, or kept as
- * generic looping code:
+ * \param tx, ty the tile position in tiles, not pixels
*/
+static boolean
+lp_setup_whole_tile(struct lp_setup_context *setup,
+ const struct lp_rast_shader_inputs *inputs,
+ int tx, int ty)
+{
+ struct lp_scene *scene = setup->scene;
+
+ 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) );
+ }
+}
-#define MIN3(a,b,c) MIN2(MIN2(a,b),c)
-#define MAX3(a,b,c) MAX2(MAX2(a,b),c)
-static void
-do_triangle_ccw(struct setup_context *setup,
+/**
+ * Do basic setup for triangle rasterization and determine which
+ * framebuffer tiles are touched. Put the triangle in the scene's
+ * bins for the tiles which we overlap.
+ */
+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 )
{
- const int rt_width = setup->fb.width;
- const int rt_height = setup->fb.height;
+ struct lp_scene *scene = setup->scene;
+ struct lp_rast_triangle *tri;
+ int x[3];
+ int y[3];
+ struct u_rect bbox;
+ unsigned tri_bytes;
+ int i;
+ int nr_planes = 3;
+
+ if (0)
+ lp_setup_print_triangle(setup, v0, v1, v2);
+
+ if (setup->scissor_test) {
+ nr_planes = 7;
+ }
+ else {
+ nr_planes = 3;
+ }
- const float y1 = subpixel_snap(v1[0][1]);
- const float y2 = subpixel_snap(v2[0][1]);
- const float y3 = subpixel_snap(v3[0][1]);
+ /* 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);
+ 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);
- const float x1 = subpixel_snap(v1[0][0]);
- const float x2 = subpixel_snap(v2[0][0]);
- const float x3 = subpixel_snap(v3[0][0]);
-
- struct lp_rast_triangle *tri = get_data( &setup->data, sizeof *tri );
- float area;
- float c1, c2, c3;
- int minx, maxx, miny, maxy;
- tri->inputs.state = setup->fs.stored;
+ /* 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;
- tri->dx12 = x1 - x2;
- tri->dx23 = x2 - x3;
- tri->dx31 = x3 - x1;
+ 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;
- tri->dy12 = y1 - y2;
- tri->dy23 = y2 - y3;
- tri->dy31 = y3 - y1;
+ /* Inclusive coordinates:
+ */
+ bbox.x1--;
+ bbox.y1--;
+ }
- area = (tri->dx12 * tri->dy31 -
- tri->dx31 * tri->dy12);
+ if (bbox.x1 < bbox.x0 ||
+ bbox.y1 < bbox.y0) {
+ if (0) debug_printf("empty bounding box\n");
+ LP_COUNT(nr_culled_tris);
+ return TRUE;
+ }
- /* Cull non-ccw and zero-sized triangles.
- */
- if (area <= 0 || util_is_inf_or_nan(area))
- return;
-
- // Bounding rectangle
- minx = util_iround(MIN3(x1, x2, x3) - .5);
- maxx = util_iround(MAX3(x1, x2, x3) + .5);
- miny = util_iround(MIN3(y1, y2, y3) - .5);
- maxy = util_iround(MAX3(y1, y2, y3) + .5);
-
- /* Clamp to framebuffer (or tile) dimensions:
- */
- miny = MAX2(0, miny);
- minx = MAX2(0, minx);
- maxy = MIN2(rt_height, maxy);
- maxx = MIN2(rt_width, maxx);
+ if (!u_rect_test_intersection(&setup->draw_region, &bbox)) {
+ if (0) debug_printf("offscreen\n");
+ LP_COUNT(nr_culled_tris);
+ return TRUE;
+ }
- if (miny == maxy || minx == maxx)
- return;
+ u_rect_find_intersection(&setup->draw_region, &bbox);
- /* The only divide in this code. Is it really needed?
- */
- tri->oneoverarea = 1.0f / area;
+ tri = lp_setup_alloc_triangle(scene,
+ setup->fs.nr_inputs,
+ nr_planes,
+ &tri_bytes);
+ if (!tri)
+ return FALSE;
+
+#ifdef DEBUG
+ 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
+
+ tri->plane[0].dcdy = x[0] - x[1];
+ tri->plane[1].dcdy = x[1] - x[2];
+ tri->plane[2].dcdy = x[2] - x[0];
+
+ tri->plane[0].dcdx = y[0] - y[1];
+ tri->plane[1].dcdx = y[1] - y[2];
+ tri->plane[2].dcdx = y[2] - y[0];
+
+ LP_COUNT(nr_tris);
/* Setup parameter interpolants:
*/
- setup_tri_coefficients( setup, tri, v1, v2, v3, frontfacing );
+ lp_setup_tri_coef( setup, &tri->inputs, v0, v1, v2, frontfacing );
- /* half-edge constants, will be interated over the whole
- * rendertarget.
- */
- tri->c1 = tri->dy12 * x1 - tri->dx12 * y1;
- tri->c2 = tri->dy23 * x2 - tri->dx23 * y2;
- tri->c3 = tri->dy31 * x3 - tri->dx31 * y3;
+ tri->inputs.facing = frontfacing ? 1.0F : -1.0F;
+ tri->inputs.disable = FALSE;
+ tri->inputs.opaque = setup->fs.current.variant->opaque;
- /* correct for top-left fill convention:
- */
- if (tri->dy12 < 0 || (tri->dy12 == 0 && tri->dx12 > 0)) c1++;
- if (tri->dy23 < 0 || (tri->dy23 == 0 && tri->dx23 > 0)) c2++;
- if (tri->dy31 < 0 || (tri->dy31 == 0 && tri->dx31 > 0)) c3++;
-
- /* 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;
+
+ for (i = 0; i < 3; i++) {
+ struct lp_rast_plane *plane = &tri->plane[i];
- tri->eo2 = 0;
- if (tri->dy23 < 0) tri->eo2 -= tri->dy23;
- if (tri->dx23 > 0) tri->eo2 += tri->dx23;
+ /* half-edge constants, will be interated over the whole render
+ * target.
+ */
+ plane->c = plane->dcdx * x[i] - plane->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->dcdx < 0) {
+ /* both fill conventions want this - adjust for left edges */
+ plane->c++;
+ }
+ else if (plane->dcdx == 0) {
+ if (setup->pixel_offset == 0) {
+ /* correct for top-left fill convention:
+ */
+ if (plane->dcdy > 0) plane->c++;
+ }
+ else {
+ /* correct for bottom-left fill convention:
+ */
+ if (plane->dcdy < 0) plane->c++;
+ }
+ }
- tri->eo3 = 0;
- if (tri->dy31 < 0) tri->eo3 -= tri->dy31;
- if (tri->dx31 > 0) tri->eo3 += tri->dx31;
+ plane->dcdx *= FIXED_ONE;
+ plane->dcdy *= FIXED_ONE;
- /* Calculate trivial accept offsets from the above.
+ /* 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->eo = 0;
+ if (plane->dcdx < 0) plane->eo -= plane->dcdx;
+ if (plane->dcdy > 0) plane->eo += plane->dcdy;
+
+ /* Calculate trivial accept offsets from the above.
+ */
+ plane->ei = plane->dcdy - plane->dcdx - plane->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.
*/
- tri->ei1 = tri->dx12 - tri->dy12 - tri->eo1;
- tri->ei2 = tri->dx23 - tri->dy23 - tri->eo2;
- tri->ei3 = tri->dx31 - tri->dy31 - tri->eo3;
+ if (nr_planes == 7) {
+ tri->plane[3].dcdx = -1;
+ tri->plane[3].dcdy = 0;
+ tri->plane[3].c = 1-bbox.x0;
+ tri->plane[3].ei = 0;
+ tri->plane[3].eo = 1;
+
+ tri->plane[4].dcdx = 1;
+ tri->plane[4].dcdy = 0;
+ tri->plane[4].c = bbox.x1+1;
+ tri->plane[4].ei = -1;
+ tri->plane[4].eo = 0;
+
+ tri->plane[5].dcdx = 0;
+ tri->plane[5].dcdy = 1;
+ tri->plane[5].c = 1-bbox.y0;
+ tri->plane[5].ei = 0;
+ tri->plane[5].eo = 1;
+
+ tri->plane[6].dcdx = 0;
+ tri->plane[6].dcdy = -1;
+ tri->plane[6].c = bbox.y1+1;
+ tri->plane[6].ei = -1;
+ tri->plane[6].eo = 0;
+ }
- minx &= ~(TILESIZE-1); /* aligned blocks */
- miny &= ~(TILESIZE-1); /* aligned blocks */
+ return lp_setup_bin_triangle( setup, tri, &bbox, nr_planes );
+}
- c1 = tri->c1 + tri->dx12 * miny - tri->dy12 * minx;
- c2 = tri->c2 + tri->dx23 * miny - tri->dy23 * minx;
- c3 = tri->c3 + tri->dx31 * miny - tri->dy31 * minx;
+/*
+ * 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
+}
- minx /= TILESIZE;
- miny /= TILESIZE;
- maxx /= TILESIZE;
- maxy /= TILESIZE;
- /* Convert to tile coordinates:
+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;
+ 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));
+
+ /* The largest dimension of the rasterized area of the triangle
+ * (aligned to a 4x4 grid), rounded down to the nearest power of two:
+ */
+ int sz = floor_pot((bbox->x1 - (bbox->x0 & ~3)) |
+ (bbox->y1 - (bbox->y0 & ~3)));
+
+ /* 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:
*/
- bin_command( &setup->tile[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
{
- const int step = TILESIZE;
-
- 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;
+ int c[MAX_PLANES];
+ int ei[MAX_PLANES];
+ int eo[MAX_PLANES];
+ int xstep[MAX_PLANES];
+ int ystep[MAX_PLANES];
+ int x, y;
- float xstep1 = -step * tri->dy12;
- float xstep2 = -step * tri->dy23;
- float xstep3 = -step * tri->dy31;
+ int ix0 = bbox->x0 / TILE_SIZE;
+ int iy0 = bbox->y0 / TILE_SIZE;
+ int ix1 = bbox->x1 / TILE_SIZE;
+ int iy1 = bbox->y1 / TILE_SIZE;
+
+ for (i = 0; i < nr_planes; i++) {
+ c[i] = (tri->plane[i].c +
+ tri->plane[i].dcdy * iy0 * TILE_SIZE -
+ tri->plane[i].dcdx * ix0 * TILE_SIZE);
+
+ ei[i] = tri->plane[i].ei << TILE_ORDER;
+ eo[i] = tri->plane[i].eo << TILE_ORDER;
+ xstep[i] = -(tri->plane[i].dcdx << TILE_ORDER);
+ ystep[i] = tri->plane[i].dcdy << TILE_ORDER;
+ }
- float ystep1 = step * tri->dx12;
- float ystep2 = step * tri->dx23;
- float ystep3 = step * tri->dx31;
- int x, y;
- /* 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.
+ /* 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++)
{
- float cx1 = c1;
- float cx2 = c2;
- float 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 */
- }
- else if (cx1 + ei1 > 0 &&
- cx2 + ei2 > 0 &&
- cx3 + ei3 > 0)
- {
- /* shade whole tile */
- bin_command( &setup->tile[x][y], lp_rast_shade_tile,
- lp_rast_arg_inputs(&tri->inputs) );
- }
- else
- {
- /* shade partial tile */
- bin_command( &setup->tile[x][y],
- lp_rast_triangle,
- lp_rast_arg_triangle(tri) );
- }
+ 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);
+ }
+ 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 (!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;
}
-static void triangle_cw( struct setup_context *setup,
- const float (*v0)[4],
- const float (*v1)[4],
- const float (*v2)[4] )
+
+/**
+ * Try to draw the triangle, restart the scene on failure.
+ */
+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;
}
-static void triangle_ccw( struct setup_context *setup,
+
+/**
+ * Draw triangle if it's CW, cull otherwise.
+ */
+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);
}
-static void triangle_both( struct setup_context *setup,
+/**
+ * Draw triangle whether it's CW or CCW.
+ */
+static void triangle_both( struct lp_setup_context *setup,
const float (*v0)[4],
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)
- triangle_ccw( setup, v0, v1, v2 );
- else
- triangle_cw( setup, v0, v1, v2 );
+ float area = calc_area(v0, v1, v2);
+
+ 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 );
}
-static void triangle_nop( struct setup_context *setup,
+
+static void triangle_nop( struct lp_setup_context *setup,
const float (*v0)[4],
const float (*v1)[4],
const float (*v2)[4] )
void
-lp_setup_choose_triangle( struct setup_context *setup )
+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;
break;
}
}
-
-
projects / mesa.git / blobdiff