* Keith Whitwell <keith@tungstengraphics.com>
*/
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
+#include "util/u_prim.h"
#include "draw/draw_context.h"
#include "draw/draw_private.h"
-//#include "draw/draw_vbuf.h"
-//#include "draw/draw_vertex.h"
#include "draw/draw_pt.h"
-#define CACHE_MAX 32
-#define FETCH_MAX 128
+#define CACHE_MAX 256
+#define FETCH_MAX 256
#define DRAW_MAX (16*1024)
struct vcache_frontend {
unsigned draw_count;
unsigned fetch_count;
-
+ unsigned fetch_max;
+
struct draw_pt_middle_end *middle;
unsigned input_prim;
unsigned output_prim;
+
+ unsigned middle_prim;
+ unsigned opt;
};
-static void vcache_flush( struct vcache_frontend *vcache )
+static INLINE void
+vcache_flush( struct vcache_frontend *vcache )
{
+ if (vcache->middle_prim != vcache->output_prim) {
+ vcache->middle_prim = vcache->output_prim;
+ vcache->middle->prepare( vcache->middle,
+ vcache->input_prim,
+ vcache->middle_prim,
+ vcache->opt,
+ &vcache->fetch_max );
+ }
+
if (vcache->draw_count) {
vcache->middle->run( vcache->middle,
vcache->fetch_elts,
vcache->draw_count = 0;
}
-static void vcache_check_flush( struct vcache_frontend *vcache )
+static INLINE void
+vcache_check_flush( struct vcache_frontend *vcache )
{
if ( vcache->draw_count + 6 >= DRAW_MAX ||
vcache->fetch_count + 4 >= FETCH_MAX )
}
-static void vcache_elt( struct vcache_frontend *vcache,
- unsigned felt )
+static INLINE void
+vcache_elt( struct vcache_frontend *vcache,
+ unsigned felt,
+ ushort flags )
{
unsigned idx = felt % CACHE_MAX;
vcache->fetch_elts[vcache->fetch_count++] = felt;
}
- vcache->draw_elts[vcache->draw_count++] = vcache->out[idx];
+ vcache->draw_elts[vcache->draw_count++] = vcache->out[idx] | flags;
}
-static unsigned add_edgeflag( struct vcache_frontend *vcache,
- unsigned idx,
- unsigned mask )
+
+
+static INLINE void
+vcache_triangle( struct vcache_frontend *vcache,
+ unsigned i0,
+ unsigned i1,
+ unsigned i2 )
{
- if (mask && draw_get_edgeflag(vcache->draw, idx))
- return idx | DRAW_PT_EDGEFLAG;
- else
- return idx;
+ vcache_elt(vcache, i0, 0);
+ vcache_elt(vcache, i1, 0);
+ vcache_elt(vcache, i2, 0);
+ vcache_check_flush(vcache);
}
-
-static unsigned add_reset_stipple( unsigned idx,
- unsigned reset )
+
+static INLINE void
+vcache_triangle_flags( struct vcache_frontend *vcache,
+ ushort flags,
+ unsigned i0,
+ unsigned i1,
+ unsigned i2 )
{
- if (reset)
- return idx | DRAW_PT_RESET_STIPPLE;
- else
- return idx;
+ vcache_elt(vcache, i0, flags);
+ vcache_elt(vcache, i1, 0);
+ vcache_elt(vcache, i2, 0);
+ vcache_check_flush(vcache);
}
-
-static void vcache_triangle( struct vcache_frontend *vcache,
- unsigned i0,
- unsigned i1,
- unsigned i2 )
+static INLINE void
+vcache_line( struct vcache_frontend *vcache,
+ unsigned i0,
+ unsigned i1 )
{
- vcache_elt(vcache, i0 | DRAW_PT_EDGEFLAG | DRAW_PT_RESET_STIPPLE);
- vcache_elt(vcache, i1 | DRAW_PT_EDGEFLAG);
- vcache_elt(vcache, i2 | DRAW_PT_EDGEFLAG);
+ vcache_elt(vcache, i0, 0);
+ vcache_elt(vcache, i1, 0);
vcache_check_flush(vcache);
}
-
-static void vcache_ef_triangle( struct vcache_frontend *vcache,
- boolean reset_stipple,
- unsigned ef_mask,
- unsigned i0,
- unsigned i1,
- unsigned i2 )
+
+static INLINE void
+vcache_line_flags( struct vcache_frontend *vcache,
+ ushort flags,
+ unsigned i0,
+ unsigned i1 )
{
- i0 = add_edgeflag( vcache, i0, (ef_mask >> 0) & 1 );
- i1 = add_edgeflag( vcache, i1, (ef_mask >> 1) & 1 );
- i2 = add_edgeflag( vcache, i2, (ef_mask >> 2) & 1 );
+ vcache_elt(vcache, i0, flags);
+ vcache_elt(vcache, i1, 0);
+ vcache_check_flush(vcache);
+}
- i0 = add_reset_stipple( i0, reset_stipple );
- vcache_elt(vcache, i0);
- vcache_elt(vcache, i1);
- vcache_elt(vcache, i2);
+static INLINE void
+vcache_point( struct vcache_frontend *vcache,
+ unsigned i0 )
+{
+ vcache_elt(vcache, i0, 0);
vcache_check_flush(vcache);
+}
- if (0) debug_printf("emit tri ef: %d %d %d\n",
- !!(i0 & DRAW_PT_EDGEFLAG),
- !!(i1 & DRAW_PT_EDGEFLAG),
- !!(i2 & DRAW_PT_EDGEFLAG));
-
+static INLINE void
+vcache_quad( struct vcache_frontend *vcache,
+ unsigned i0,
+ unsigned i1,
+ unsigned i2,
+ unsigned i3 )
+{
+ if (vcache->draw->rasterizer->flatshade_first) {
+ /* pass last quad vertex as first triangle vertex */
+ vcache_triangle( vcache, i3, i0, i1 );
+ vcache_triangle( vcache, i3, i1, i2 );
+ }
+ else {
+ /* pass last quad vertex as last triangle vertex */
+ vcache_triangle( vcache, i0, i1, i3 );
+ vcache_triangle( vcache, i1, i2, i3 );
+ }
}
+static INLINE void
+vcache_ef_quad( struct vcache_frontend *vcache,
+ unsigned i0,
+ unsigned i1,
+ unsigned i2,
+ unsigned i3 )
+{
+ if (vcache->draw->rasterizer->flatshade_first) {
+ /* pass last quad vertex as first triangle vertex */
+ vcache_triangle_flags( vcache,
+ ( DRAW_PIPE_RESET_STIPPLE |
+ DRAW_PIPE_EDGE_FLAG_0 |
+ DRAW_PIPE_EDGE_FLAG_1 ),
+ i3, i0, i1 );
+
+ vcache_triangle_flags( vcache,
+ ( DRAW_PIPE_EDGE_FLAG_1 |
+ DRAW_PIPE_EDGE_FLAG_2 ),
+ i3, i1, i2 );
+ }
+ else {
+ /* pass last quad vertex as last triangle vertex */
+ vcache_triangle_flags( vcache,
+ ( DRAW_PIPE_RESET_STIPPLE |
+ DRAW_PIPE_EDGE_FLAG_0 |
+ DRAW_PIPE_EDGE_FLAG_2 ),
+ i0, i1, i3 );
+
+ vcache_triangle_flags( vcache,
+ ( DRAW_PIPE_EDGE_FLAG_0 |
+ DRAW_PIPE_EDGE_FLAG_1 ),
+ i1, i2, i3 );
+ }
+}
-static void vcache_line( struct vcache_frontend *vcache,
- boolean reset_stipple,
- unsigned i0,
- unsigned i1 )
+/* At least for now, we're back to using a template include file for
+ * this. The two paths aren't too different though - it may be
+ * possible to reunify them.
+ */
+#define TRIANGLE(vc,flags,i0,i1,i2) vcache_triangle_flags(vc,flags,i0,i1,i2)
+#define QUAD(vc,i0,i1,i2,i3) vcache_ef_quad(vc,i0,i1,i2,i3)
+#define LINE(vc,flags,i0,i1) vcache_line_flags(vc,flags,i0,i1)
+#define POINT(vc,i0) vcache_point(vc,i0)
+#define FUNC vcache_run_extras
+#include "draw_pt_vcache_tmp.h"
+
+#define TRIANGLE(vc,flags,i0,i1,i2) vcache_triangle(vc,i0,i1,i2)
+#define QUAD(vc,i0,i1,i2,i3) vcache_quad(vc,i0,i1,i2,i3)
+#define LINE(vc,flags,i0,i1) vcache_line(vc,i0,i1)
+#define POINT(vc,i0) vcache_point(vc,i0)
+#define FUNC vcache_run
+#include "draw_pt_vcache_tmp.h"
+
+static INLINE void
+rebase_uint_elts( const unsigned *src,
+ unsigned count,
+ int delta,
+ ushort *dest )
{
- i0 = add_reset_stipple( i0, reset_stipple );
+ unsigned i;
- vcache_elt(vcache, i0);
- vcache_elt(vcache, i1);
- vcache_check_flush(vcache);
+ for (i = 0; i < count; i++)
+ dest[i] = (ushort)(src[i] + delta);
+}
+
+static INLINE void
+rebase_ushort_elts( const ushort *src,
+ unsigned count,
+ int delta,
+ ushort *dest )
+{
+ unsigned i;
+
+ for (i = 0; i < count; i++)
+ dest[i] = (ushort)(src[i] + delta);
}
+static INLINE void
+rebase_ubyte_elts( const ubyte *src,
+ unsigned count,
+ int delta,
+ ushort *dest )
+{
+ unsigned i;
+
+ for (i = 0; i < count; i++)
+ dest[i] = (ushort)(src[i] + delta);
+}
-static void vcache_point( struct vcache_frontend *vcache,
- unsigned i0 )
+
+
+static INLINE void
+translate_uint_elts( const unsigned *src,
+ unsigned count,
+ ushort *dest )
{
- vcache_elt(vcache, i0);
- vcache_check_flush(vcache);
+ unsigned i;
+
+ for (i = 0; i < count; i++)
+ dest[i] = (ushort)(src[i]);
}
-static void vcache_quad( struct vcache_frontend *vcache,
- unsigned i0,
- unsigned i1,
- unsigned i2,
- unsigned i3 )
+static INLINE void
+translate_ushort_elts( const ushort *src,
+ unsigned count,
+ ushort *dest )
{
- vcache_triangle( vcache, i0, i1, i3 );
- vcache_triangle( vcache, i1, i2, i3 );
+ unsigned i;
+
+ for (i = 0; i < count; i++)
+ dest[i] = (ushort)(src[i]);
}
-static void vcache_ef_quad( struct vcache_frontend *vcache,
- unsigned i0,
- unsigned i1,
- unsigned i2,
- unsigned i3 )
+static INLINE void
+translate_ubyte_elts( const ubyte *src,
+ unsigned count,
+ ushort *dest )
{
- const unsigned omitEdge2 = ~(1 << 1);
- const unsigned omitEdge3 = ~(1 << 2);
- vcache_ef_triangle( vcache, 1, omitEdge2, i0, i1, i3 );
- vcache_ef_triangle( vcache, 0, omitEdge3, i1, i2, i3 );
+ unsigned i;
+
+ for (i = 0; i < count; i++)
+ dest[i] = (ushort)(src[i]);
}
-static void vcache_run( struct draw_pt_front_end *frontend,
- pt_elt_func get_elt,
- const void *elts,
- unsigned count )
+#if 0
+static INLINE enum pipe_format
+format_from_get_elt( pt_elt_func get_elt )
{
- struct vcache_frontend *vcache = (struct vcache_frontend *)frontend;
+ switch (draw->pt.user.eltSize) {
+ case 1: return PIPE_FORMAT_R8_UNORM;
+ case 2: return PIPE_FORMAT_R16_UNORM;
+ case 4: return PIPE_FORMAT_R32_UNORM;
+ default: return PIPE_FORMAT_NONE;
+ }
+}
+#endif
+
+static INLINE void
+vcache_check_run( struct draw_pt_front_end *frontend,
+ pt_elt_func get_elt,
+ const void *elts,
+ int elt_bias,
+ unsigned draw_count )
+{
+ struct vcache_frontend *vcache = (struct vcache_frontend *)frontend;
struct draw_context *draw = vcache->draw;
+ unsigned min_index = draw->pt.user.min_index;
+ unsigned max_index = draw->pt.user.max_index;
+ unsigned index_size = draw->pt.user.eltSize;
+ unsigned fetch_count = max_index + 1 - min_index;
+ const ushort *transformed_elts;
+ ushort *storage = NULL;
+ boolean ok = FALSE;
+
+
+ if (0) debug_printf("fetch_count %d fetch_max %d draw_count %d\n", fetch_count,
+ vcache->fetch_max,
+ draw_count);
+
+ if (elt_bias + max_index >= DRAW_PIPE_MAX_VERTICES ||
+ fetch_count >= UNDEFINED_VERTEX_ID ||
+ fetch_count > draw_count) {
+ if (0) debug_printf("fail\n");
+ goto fail;
+ }
- boolean unfilled = (draw->rasterizer->fill_cw != PIPE_POLYGON_MODE_FILL ||
- draw->rasterizer->fill_ccw != PIPE_POLYGON_MODE_FILL);
+ if (vcache->middle_prim != vcache->input_prim) {
+ vcache->middle_prim = vcache->input_prim;
+ vcache->middle->prepare( vcache->middle,
+ vcache->input_prim,
+ vcache->middle_prim,
+ vcache->opt,
+ &vcache->fetch_max );
+ }
- boolean flatfirst = (draw->rasterizer->flatshade &&
- draw->rasterizer->flatshade_first);
- unsigned i;
-// debug_printf("%s (%d) %d/%d\n", __FUNCTION__, draw->prim, start, count );
+ assert((elt_bias >= 0 && min_index + elt_bias >= min_index) ||
+ (elt_bias < 0 && min_index + elt_bias < min_index));
- switch (vcache->input_prim) {
- case PIPE_PRIM_POINTS:
- for (i = 0; i < count; i ++) {
- vcache_point( vcache,
- get_elt(elts, i + 0) );
- }
- break;
-
- case PIPE_PRIM_LINES:
- for (i = 0; i+1 < count; i += 2) {
- vcache_line( vcache,
- TRUE,
- get_elt(elts, i + 0),
- get_elt(elts, i + 1));
- }
- break;
-
- case PIPE_PRIM_LINE_LOOP:
- if (count >= 2) {
- for (i = 1; i < count; i++) {
- vcache_line( vcache,
- i == 1, /* XXX: only if vb not split */
- get_elt(elts, i - 1),
- get_elt(elts, i ));
- }
-
- vcache_line( vcache,
- 0,
- get_elt(elts, count - 1),
- get_elt(elts, 0 ));
- }
- break;
-
- case PIPE_PRIM_LINE_STRIP:
- for (i = 1; i < count; i++) {
- vcache_line( vcache,
- i == 1,
- get_elt(elts, i - 1),
- get_elt(elts, i ));
- }
- break;
-
- case PIPE_PRIM_TRIANGLES:
- if (unfilled) {
- for (i = 0; i+2 < count; i += 3) {
- vcache_ef_triangle( vcache,
- 1,
- ~0,
- get_elt(elts, i + 0),
- get_elt(elts, i + 1),
- get_elt(elts, i + 2 ));
- }
- }
- else {
- for (i = 0; i+2 < count; i += 3) {
- vcache_triangle( vcache,
- get_elt(elts, i + 0),
- get_elt(elts, i + 1),
- get_elt(elts, i + 2 ));
- }
- }
- break;
-
- case PIPE_PRIM_TRIANGLE_STRIP:
- if (flatfirst) {
- for (i = 0; i+2 < count; i++) {
- if (i & 1) {
- vcache_triangle( vcache,
- get_elt(elts, i + 0),
- get_elt(elts, i + 2),
- get_elt(elts, i + 1 ));
- }
- else {
- vcache_triangle( vcache,
- get_elt(elts, i + 0),
- get_elt(elts, i + 1),
- get_elt(elts, i + 2 ));
- }
+ if (min_index == 0 &&
+ index_size == 2)
+ {
+ transformed_elts = (const ushort *)elts;
+ }
+ else
+ {
+ storage = MALLOC( draw_count * sizeof(ushort) );
+ if (!storage)
+ goto fail;
+
+ if (min_index == 0) {
+ switch(index_size) {
+ case 1:
+ translate_ubyte_elts( (const ubyte *)elts,
+ draw_count,
+ storage );
+ break;
+
+ case 2:
+ translate_ushort_elts( (const ushort *)elts,
+ draw_count,
+ storage );
+ break;
+
+ case 4:
+ translate_uint_elts( (const uint *)elts,
+ draw_count,
+ storage );
+ break;
+
+ default:
+ assert(0);
+ FREE(storage);
+ return;
}
}
else {
- for (i = 0; i+2 < count; i++) {
- if (i & 1) {
- vcache_triangle( vcache,
- get_elt(elts, i + 1),
- get_elt(elts, i + 0),
- get_elt(elts, i + 2 ));
- }
- else {
- vcache_triangle( vcache,
- get_elt(elts, i + 0),
- get_elt(elts, i + 1),
- get_elt(elts, i + 2 ));
- }
+ switch(index_size) {
+ case 1:
+ rebase_ubyte_elts( (const ubyte *)elts,
+ draw_count,
+ 0 - (int)min_index,
+ storage );
+ break;
+
+ case 2:
+ rebase_ushort_elts( (const ushort *)elts,
+ draw_count,
+ 0 - (int)min_index,
+ storage );
+ break;
+
+ case 4:
+ rebase_uint_elts( (const uint *)elts,
+ draw_count,
+ 0 - (int)min_index,
+ storage );
+ break;
+
+ default:
+ assert(0);
+ FREE(storage);
+ return;
}
}
- break;
-
- case PIPE_PRIM_TRIANGLE_FAN:
- if (count >= 3) {
- if (flatfirst) {
- for (i = 0; i+2 < count; i++) {
- vcache_triangle( vcache,
- get_elt(elts, i + 1),
- get_elt(elts, i + 2),
- get_elt(elts, 0 ));
- }
- }
- else {
- for (i = 0; i+2 < count; i++) {
- vcache_triangle( vcache,
- get_elt(elts, 0),
- get_elt(elts, i + 1),
- get_elt(elts, i + 2 ));
- }
- }
- }
- break;
-
-
- case PIPE_PRIM_QUADS:
- if (unfilled) {
- for (i = 0; i+3 < count; i += 4) {
- vcache_ef_quad( vcache,
- get_elt(elts, i + 0),
- get_elt(elts, i + 1),
- get_elt(elts, i + 2),
- get_elt(elts, i + 3));
- }
- }
- else {
- for (i = 0; i+3 < count; i += 4) {
- vcache_quad( vcache,
- get_elt(elts, i + 0),
- get_elt(elts, i + 1),
- get_elt(elts, i + 2),
- get_elt(elts, i + 3));
- }
- }
- break;
-
- case PIPE_PRIM_QUAD_STRIP:
- if (unfilled) {
- for (i = 0; i+3 < count; i += 2) {
- vcache_ef_quad( vcache,
- get_elt(elts, i + 2),
- get_elt(elts, i + 0),
- get_elt(elts, i + 1),
- get_elt(elts, i + 3));
- }
- }
- else {
- for (i = 0; i+3 < count; i += 2) {
- vcache_quad( vcache,
- get_elt(elts, i + 2),
- get_elt(elts, i + 0),
- get_elt(elts, i + 1),
- get_elt(elts, i + 3));
- }
- }
- break;
-
- case PIPE_PRIM_POLYGON:
- if (unfilled) {
- /* These bitflags look a little odd because we submit the
- * vertices as (1,2,0) to satisfy flatshade requirements.
- */
- const unsigned edge_first = (1<<2);
- const unsigned edge_middle = (1<<0);
- const unsigned edge_last = (1<<1);
-
- for (i = 0; i+2 < count; i++) {
- unsigned ef_mask = edge_middle;
-
- if (i == 0)
- ef_mask |= edge_first;
-
- if (i + 3 == count)
- ef_mask |= edge_last;
-
- vcache_ef_triangle( vcache,
- i == 0,
- ef_mask,
- get_elt(elts, i + 1),
- get_elt(elts, i + 2),
- get_elt(elts, 0));
- }
- }
- else {
- for (i = 0; i+2 < count; i++) {
- vcache_triangle( vcache,
- get_elt(elts, i + 1),
- get_elt(elts, i + 2),
- get_elt(elts, 0));
- }
- }
- break;
-
- default:
- assert(0);
- break;
+ transformed_elts = storage;
}
+
+ if (fetch_count < UNDEFINED_VERTEX_ID)
+ ok = vcache->middle->run_linear_elts( vcache->middle,
+ min_index + elt_bias, /* start */
+ fetch_count,
+ transformed_elts,
+ draw_count );
- vcache_flush( vcache );
-}
+ FREE(storage);
+ if (ok)
+ return;
+ debug_printf("failed to execute atomic draw elts for %d/%d, splitting up\n",
+ fetch_count, draw_count);
-static unsigned reduced_prim[PIPE_PRIM_POLYGON + 1] = {
- PIPE_PRIM_POINTS,
- PIPE_PRIM_LINES,
- PIPE_PRIM_LINES,
- PIPE_PRIM_LINES,
- PIPE_PRIM_TRIANGLES,
- PIPE_PRIM_TRIANGLES,
- PIPE_PRIM_TRIANGLES,
- PIPE_PRIM_TRIANGLES,
- PIPE_PRIM_TRIANGLES,
- PIPE_PRIM_TRIANGLES
-};
+ fail:
+ vcache_run( frontend, get_elt, elts, elt_bias, draw_count );
+}
-static void vcache_prepare( struct draw_pt_front_end *frontend,
- unsigned prim,
- struct draw_pt_middle_end *middle,
- unsigned opt )
+
+static void
+vcache_prepare( struct draw_pt_front_end *frontend,
+ unsigned in_prim,
+ unsigned out_prim,
+ struct draw_pt_middle_end *middle,
+ unsigned opt )
{
struct vcache_frontend *vcache = (struct vcache_frontend *)frontend;
-/*
- if (vcache->draw->rasterizer->flatshade_first)
- vcache->base.run = vcache_run_pv0;
+ if (opt & PT_PIPELINE)
+ {
+ vcache->base.run = vcache_run_extras;
+ }
else
- vcache->base.run = vcache_run_pv2;
-*/
+ {
+ vcache->base.run = vcache_check_run;
+ }
- vcache->base.run = vcache_run;
- vcache->input_prim = prim;
- vcache->output_prim = reduced_prim[prim];
+ vcache->input_prim = in_prim;
+ vcache->output_prim = u_reduced_prim(out_prim);
vcache->middle = middle;
- middle->prepare( middle, vcache->output_prim, opt );
+ vcache->opt = opt;
+
+ /* Have to run prepare here, but try and guess a good prim for
+ * doing so:
+ */
+ vcache->middle_prim = (opt & PT_PIPELINE) ? vcache->output_prim : vcache->input_prim;
+ middle->prepare( middle, vcache->input_prim,
+ vcache->middle_prim, opt, &vcache->fetch_max );
}
-static void vcache_finish( struct draw_pt_front_end *frontend )
+static void
+vcache_finish( struct draw_pt_front_end *frontend )
{
struct vcache_frontend *vcache = (struct vcache_frontend *)frontend;
vcache->middle->finish( vcache->middle );
vcache->middle = NULL;
}
-static void vcache_destroy( struct draw_pt_front_end *frontend )
+static void
+vcache_destroy( struct draw_pt_front_end *frontend )
{
FREE(frontend);
}
struct draw_pt_front_end *draw_pt_vcache( struct draw_context *draw )
{
struct vcache_frontend *vcache = CALLOC_STRUCT( vcache_frontend );
+ if (vcache == NULL)
+ return NULL;
vcache->base.prepare = vcache_prepare;
vcache->base.run = NULL;