#include "brw_state.h"
#include "brw_defines.h"
+/**
+ * When the GS is not in use, we assign the entire URB space to the VS. When
+ * the GS is in use, we split the URB space evenly between the VS and the GS.
+ * This is not ideal, but it's simple.
+ *
+ * URB size / 2 URB size / 2
+ * _____________-______________ _____________-______________
+ * / \ / \
+ * +-------------------------------------------------------------+
+ * | Vertex Shader Entries | Geometry Shader Entries |
+ * +-------------------------------------------------------------+
+ *
+ * Sandybridge GT1 has 32kB of URB space, while GT2 has 64kB.
+ * (See the Sandybridge PRM, Volume 2, Part 1, Section 1.4.7: 3DSTATE_URB.)
+ */
static void
-prepare_urb( struct brw_context *brw )
+gen6_upload_urb( struct brw_context *brw )
{
- brw->urb.nr_vs_entries = 24;
- if (brw->gs.prog_bo)
- brw->urb.nr_gs_entries = 4;
- else
- brw->urb.nr_gs_entries = 0;
+ int nr_vs_entries, nr_gs_entries;
+ int total_urb_size = brw->urb.size * 1024; /* in bytes */
+
/* CACHE_NEW_VS_PROG */
- brw->urb.vs_size = MIN2(brw->vs.prog_data->urb_entry_size, 1);
+ unsigned vs_size = MAX2(brw->vs.prog_data->base.urb_entry_size, 1);
- /* Check that the number of URB rows (8 floats each) allocated is less
- * than the URB space.
+ /* We use the same VUE layout for VS outputs and GS outputs (as it's what
+ * the SF and Clipper expect), so we can simply make the GS URB entry size
+ * the same as for the VS. This may technically be too large in cases
+ * where we have few vertex attributes and a lot of varyings, since the VS
+ * size is determined by the larger of the two. For now, it's safe.
*/
- assert((brw->urb.nr_vs_entries +
- brw->urb.nr_gs_entries) * brw->urb.vs_size * 8 < 64 * 1024);
-}
+ unsigned gs_size = vs_size;
-static void
-upload_urb(struct brw_context *brw)
-{
- struct intel_context *intel = &brw->intel;
+ /* Calculate how many entries fit in each stage's section of the URB */
+ if (brw->ff_gs.prog_active) {
+ nr_vs_entries = (total_urb_size/2) / (vs_size * 128);
+ nr_gs_entries = (total_urb_size/2) / (gs_size * 128);
+ } else {
+ nr_vs_entries = total_urb_size / (vs_size * 128);
+ nr_gs_entries = 0;
+ }
+
+ /* Then clamp to the maximum allowed by the hardware */
+ if (nr_vs_entries > brw->urb.max_vs_entries)
+ nr_vs_entries = brw->urb.max_vs_entries;
+
+ if (nr_gs_entries > brw->urb.max_gs_entries)
+ nr_gs_entries = brw->urb.max_gs_entries;
+ /* Finally, both must be a multiple of 4 (see 3DSTATE_URB in the PRM). */
+ brw->urb.nr_vs_entries = ROUND_DOWN_TO(nr_vs_entries, 4);
+ brw->urb.nr_gs_entries = ROUND_DOWN_TO(nr_gs_entries, 4);
+
+ assert(brw->urb.nr_vs_entries >= brw->urb.min_vs_entries);
assert(brw->urb.nr_vs_entries % 4 == 0);
assert(brw->urb.nr_gs_entries % 4 == 0);
- /* GS requirement */
- assert(!brw->gs.prog_bo || brw->urb.vs_size < 5);
+ assert(vs_size < 5);
+ assert(gs_size < 5);
BEGIN_BATCH(3);
- OUT_BATCH(CMD_URB << 16 | (3 - 2));
- OUT_BATCH(((brw->urb.vs_size - 1) << GEN6_URB_VS_SIZE_SHIFT) |
+ OUT_BATCH(_3DSTATE_URB << 16 | (3 - 2));
+ OUT_BATCH(((vs_size - 1) << GEN6_URB_VS_SIZE_SHIFT) |
((brw->urb.nr_vs_entries) << GEN6_URB_VS_ENTRIES_SHIFT));
- OUT_BATCH(((brw->urb.vs_size - 1) << GEN6_URB_GS_SIZE_SHIFT) |
+ OUT_BATCH(((gs_size - 1) << GEN6_URB_GS_SIZE_SHIFT) |
((brw->urb.nr_gs_entries) << GEN6_URB_GS_ENTRIES_SHIFT));
ADVANCE_BATCH();
+
+ /* From the PRM Volume 2 part 1, section 1.4.7:
+ *
+ * Because of a urb corruption caused by allocating a previous gsunit’s
+ * urb entry to vsunit software is required to send a "GS NULL
+ * Fence"(Send URB fence with VS URB size == 1 and GS URB size == 0) plus
+ * a dummy DRAW call before any case where VS will be taking over GS URB
+ * space.
+ *
+ * It is not clear exactly what this means ("URB fence" is a command that
+ * doesn't exist on Gen6). So for now we just do a full pipeline flush as
+ * a workaround.
+ */
+ if (brw->urb.gen6_gs_previously_active && !brw->ff_gs.prog_active)
+ intel_batchbuffer_emit_mi_flush(brw);
+ brw->urb.gen6_gs_previously_active = brw->ff_gs.prog_active;
}
const struct brw_tracked_state gen6_urb = {
.dirty = {
.mesa = 0,
.brw = BRW_NEW_CONTEXT,
- .cache = CACHE_NEW_VS_PROG,
+ .cache = (CACHE_NEW_VS_PROG | CACHE_NEW_FF_GS_PROG),
},
- .prepare = prepare_urb,
- .emit = upload_urb,
+ .emit = gen6_upload_urb,
};