radeonsi: don't count IB space for states, just use an upper bound

Since we don't put any resource descriptors in IBs, the space used by draw
calls is quite small.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
This commit is contained in:
Marek Olšák
2015-08-30 03:49:15 +02:00
parent fc95058add
commit ec9d5e181e
2 changed files with 5 additions and 55 deletions
+5 -48
View File
@@ -35,8 +35,9 @@ void si_need_cs_space(struct si_context *ctx, unsigned num_dw,
/* There are two memory usage counters in the winsys for all buffers
* that have been added (cs_add_reloc) and two counters in the pipe
* driver for those that haven't been added yet.
* */
if (!ctx->b.ws->cs_memory_below_limit(ctx->b.rings.gfx.cs, ctx->b.vram, ctx->b.gtt)) {
*/
if (unlikely(!ctx->b.ws->cs_memory_below_limit(ctx->b.rings.gfx.cs,
ctx->b.vram, ctx->b.gtt))) {
ctx->b.gtt = 0;
ctx->b.vram = 0;
ctx->b.rings.gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
@@ -46,54 +47,10 @@ void si_need_cs_space(struct si_context *ctx, unsigned num_dw,
ctx->b.vram = 0;
/* If the CS is sufficiently large, don't count the space needed
* and just flush if there is less than 8096 dwords left.
* and just flush if there is not enough space left.
*/
if (cs->max_dw >= 24 * 1024) {
if (cs->cdw > cs->max_dw - 8 * 1024)
ctx->b.rings.gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
return;
}
/* The number of dwords we already used in the CS so far. */
num_dw += cs->cdw;
if (count_draw_in) {
unsigned mask = ctx->dirty_atoms;
while (mask)
num_dw += ctx->atoms.array[u_bit_scan(&mask)]->num_dw;
/* The number of dwords all the dirty states would take. */
num_dw += si_pm4_dirty_dw(ctx);
/* The upper-bound of how much a draw command would take. */
num_dw += SI_MAX_DRAW_CS_DWORDS;
}
/* Count in queries_suspend. */
num_dw += ctx->b.num_cs_dw_nontimer_queries_suspend +
ctx->b.num_cs_dw_timer_queries_suspend;
/* Count in streamout_end at the end of CS. */
if (ctx->b.streamout.begin_emitted) {
num_dw += ctx->b.streamout.num_dw_for_end;
}
/* Count in render_condition(NULL) at the end of CS. */
if (ctx->b.predicate_drawing) {
num_dw += 3;
}
/* Count in framebuffer cache flushes at the end of CS. */
num_dw += ctx->atoms.s.cache_flush->num_dw;
if (ctx->screen->b.trace_bo)
num_dw += SI_TRACE_CS_DWORDS * 2;
/* Flush if there's not enough space. */
if (num_dw > cs->max_dw) {
if (unlikely(cs->cdw > cs->max_dw - 2048))
ctx->b.rings.gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
}
}
void si_context_gfx_flush(void *context, unsigned flags,
-7
View File
@@ -43,13 +43,6 @@
#define SI_RESTART_INDEX_UNKNOWN INT_MIN
#define SI_NUM_SMOOTH_AA_SAMPLES 8
#define SI_TRACE_CS_DWORDS 7
#define SI_MAX_DRAW_CS_DWORDS \
(/*scratch:*/ 3 + /*derived prim state:*/ 3 + \
/*draw regs:*/ 18 + /*draw packets:*/ 31 +\
/*derived tess state:*/ 19)
/* Instruction cache. */
#define SI_CONTEXT_INV_ICACHE (R600_CONTEXT_PRIVATE_FLAG << 0)
/* Cache used by scalar memory (SMEM) instructions. They also use TC