radeonsi: always prefer SWITCH_ON_EOP(0) on CIK

The code is rewritten to take known constraints into account, while always
using 0 by default.

This should improve performance for multi-SE parts in theory.

A debug option is also added for easier debugging. (If there are hangs,
use the option. If the hangs go away, you have found the problem.)

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

v2: fix a typo, set max_se for evergreen GPUs according to the kernel driver
This commit is contained in:
Marek Olšák 2014-08-06 03:18:06 +02:00
parent 515269b3a7
commit 0e229b8c5a
4 changed files with 46 additions and 10 deletions

View file

@ -239,7 +239,6 @@ static const struct debug_named_value common_debug_options[] = {
{ "vm", DBG_VM, "Print virtual addresses when creating resources" },
{ "trace_cs", DBG_TRACE_CS, "Trace cs and write rlockup_<csid>.c file with faulty cs" },
/* shaders */
{ "fs", DBG_FS, "Print fetch shaders" },
{ "vs", DBG_VS, "Print vertex shaders" },
@ -254,6 +253,7 @@ static const struct debug_named_value common_debug_options[] = {
{ "noinvalrange", DBG_NO_DISCARD_RANGE, "Disable handling of INVALIDATE_RANGE map flags" },
{ "no2d", DBG_NO_2D_TILING, "Disable 2D tiling" },
{ "notiling", DBG_NO_TILING, "Disable tiling" },
{ "switch_on_eop", DBG_SWITCH_ON_EOP, "Program WD/IA to switch on end-of-packet." },
DEBUG_NAMED_VALUE_END /* must be last */
};

View file

@ -93,6 +93,7 @@
#define DBG_NO_DISCARD_RANGE (1 << 12)
#define DBG_NO_2D_TILING (1 << 13)
#define DBG_NO_TILING (1 << 14)
#define DBG_SWITCH_ON_EOP (1 << 15)
/* The maximum allowed bit is 15. */
#define R600_MAP_BUFFER_ALIGNMENT 64

View file

@ -401,25 +401,40 @@ static bool si_update_draw_info_state(struct si_context *sctx,
if (sctx->b.chip_class >= CIK) {
struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
bool wd_switch_on_eop = prim == V_008958_DI_PT_POLYGON ||
prim == V_008958_DI_PT_LINELOOP ||
prim == V_008958_DI_PT_TRIFAN ||
prim == V_008958_DI_PT_TRISTRIP_ADJ ||
info->primitive_restart ||
(rs ? rs->line_stipple_enable : false);
/* If the WD switch is false, the IA switch must be false too. */
bool ia_switch_on_eop = wd_switch_on_eop;
unsigned primgroup_size = 64;
/* SWITCH_ON_EOP(0) is always preferable. */
bool wd_switch_on_eop = false;
bool ia_switch_on_eop = false;
/* WD_SWITCH_ON_EOP has no effect on GPUs with less than
* 4 shader engines. Set 1 to pass the assertion below.
* The other cases are hardware requirements. */
if (sctx->b.screen->info.max_se < 4 ||
prim == V_008958_DI_PT_POLYGON ||
prim == V_008958_DI_PT_LINELOOP ||
prim == V_008958_DI_PT_TRIFAN ||
prim == V_008958_DI_PT_TRISTRIP_ADJ ||
info->primitive_restart)
wd_switch_on_eop = true;
/* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
* We don't know that for indirect drawing, so treat it as
* always problematic. */
if (sctx->b.family == CHIP_HAWAII &&
(info->indirect || info->instance_count > 1)) {
(info->indirect || info->instance_count > 1))
wd_switch_on_eop = true;
/* This is a hardware requirement. */
if ((rs && rs->line_stipple_enable) ||
(sctx->b.screen->debug_flags & DBG_SWITCH_ON_EOP)) {
ia_switch_on_eop = true;
wd_switch_on_eop = true;
}
/* If the WD switch is false, the IA switch must be false too. */
assert(wd_switch_on_eop || !ia_switch_on_eop);
si_pm4_set_reg(pm4, R_028B74_VGT_DISPATCH_DRAW_INDEX,
ib->index_size == 4 ? 0xFC000000 : 0xFC00);

View file

@ -392,6 +392,26 @@ static boolean do_winsys_init(struct radeon_drm_winsys *ws)
radeon_get_drm_value(ws->fd, RADEON_INFO_MAX_SE, NULL,
&ws->info.max_se);
if (!ws->info.max_se) {
switch (ws->info.family) {
default:
ws->info.max_se = 1;
break;
case CHIP_CYPRESS:
case CHIP_HEMLOCK:
case CHIP_BARTS:
case CHIP_CAYMAN:
case CHIP_TAHITI:
case CHIP_PITCAIRN:
case CHIP_BONAIRE:
ws->info.max_se = 2;
break;
case CHIP_HAWAII:
ws->info.max_se = 4;
break;
}
}
radeon_get_drm_value(ws->fd, RADEON_INFO_MAX_SH_PER_SE, NULL,
&ws->info.max_sh_per_se);