mirror/86Box
1
0
Fork 0
mirror of https://github.com/86Box/86Box.git synced 2026-02-22 09:35:32 -07:00
Code Issues Packages Projects Releases Activity

Voodoo: smooth V1/V2 sync and Windows waits

Browse Source 
Make LFB relax/front-sync the default for V1/V2, keep wait stats opt-in, document sync behavior, and reduce Windows swap/event stalls.
This commit is contained in:
AITUS95
2026-01-18 03:13:34 +01:00
parent b3b6943055
commit e414d9ab55
3 changed files with 524 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
src/include/86box/vid_voodoo_common.h
View File

@@ -74,6 +74,13 @@ typedef union rgba_u {
#define FIFO_FULL ((voodoo->fifo_write_idx - voodoo->fifo_read_idx) >= FIFO_SIZE - 4)
#define FIFO_EMPTY (voodoo->fifo_read_idx == voodoo->fifo_write_idx)
#define VOODOO_BUF_FRONT 0
#define VOODOO_BUF_BACK 1
#define VOODOO_BUF_AUX 2
#define VOODOO_BUF_UNKNOWN 3
#define VOODOO_BUF_COUNT 4
#define VOODOO_BUF_NONE 0xff
#define FIFO_TYPE 0xff000000
#define FIFO_ADDR 0x00ffffff
@@ -98,6 +105,8 @@ typedef struct
{
uint32_t addr_type;
uint32_t val;
uint8_t target_buf;
uint8_t pad[3];
} fifo_entry_t;
typedef struct voodoo_params_t {
@@ -356,6 +365,8 @@ typedef struct voodoo_t {
event_t *wake_fifo_thread;
event_t *wake_main_thread;
event_t *fifo_not_full_event;
event_t *fifo_empty_event;
ATOMIC_INT fifo_empty_signaled;
event_t *render_not_full_event[4];
event_t *wake_render_thread[4];
@@ -399,6 +410,8 @@ typedef struct voodoo_t {
ATOMIC_INT cmd_written;
ATOMIC_INT cmd_written_fifo;
ATOMIC_INT cmd_written_fifo_2;
ATOMIC_INT pending_fb_writes_buf[VOODOO_BUF_COUNT];
ATOMIC_INT pending_draw_cmds_buf[VOODOO_BUF_COUNT];
voodoo_params_t params_buffer[PARAM_SIZE];
ATOMIC_INT params_read_idx[4];
@@ -627,6 +640,12 @@ typedef struct voodoo_t {
int fb_write_buffer;
int fb_draw_buffer;
int buffer_cutoff;
int queued_disp_buffer;
int queued_draw_buffer;
int queued_fb_write_buffer;
int queued_fb_draw_buffer;
uint32_t queued_lfbMode;
uint32_t queued_fbzMode;
uint32_t tile_base;
uint32_t tile_stride;
@@ -657,6 +676,32 @@ typedef struct voodoo_t {
uint64_t time;
int render_time[4];
uint64_t fifo_full_waits;
uint64_t fifo_full_wait_ticks;
uint64_t fifo_full_spin_checks;
uint64_t fifo_empty_waits;
uint64_t fifo_empty_wait_ticks;
uint64_t fifo_empty_spin_checks;
uint64_t render_waits;
uint64_t render_wait_ticks;
uint64_t render_wait_spin_checks;
uint64_t readl_fb_count;
uint64_t readl_fb_sync_count;
uint64_t readl_fb_nosync_count;
uint64_t readl_fb_relaxed_count;
uint64_t readl_fb_sync_buf[3];
uint64_t readl_fb_nosync_buf[3];
uint64_t readl_fb_relaxed_buf[3];
uint64_t readl_reg_count;
uint64_t readl_tex_count;
int wait_stats_enabled;
int wait_stats_explicit;
int lfb_relax_enabled;
int lfb_relax_full;
int lfb_relax_ignore_cmdfifo;
int lfb_relax_ignore_draw;
int lfb_relax_ignore_fb_writes;
int lfb_relax_front_sync;
int force_blit_count;
int can_blit;

321
src/video/vid_voodoo.c
View File

@@ -16,6 +16,7 @@
#include <stdarg.h>
#include <stdio.h>
#include <stdint.h>
#include <inttypes.h>
#include <string.h>
#include <stdlib.h>
#include <stddef.h>
@@ -73,6 +74,66 @@ voodoo_log(const char *fmt, ...)
# define voodoo_log(fmt, ...)
#endif
static int
voodoo_env_is_disabled(const char *value)
{
/* Accept common "off" values for env overrides. */
return !strcmp(value, "0") || !strcmp(value, "off") || !strcmp(value, "false") || !strcmp(value, "disabled");
}
static void
voodoo_init_relax_settings(voodoo_t *voodoo)
{
const char *relax_env = getenv("VOODOO_LFB_RELAX");
const char *wait_env = getenv("VOODOO_WAIT_STATS");
int relax_enabled = 1;
/* Default to front-sync relax mode; wait stats are opt-in. */
if (!relax_env || !*relax_env) {
relax_env = "4";
} else if (voodoo_env_is_disabled(relax_env)) {
relax_enabled = 0;
}
voodoo->wait_stats_explicit = (wait_env && *wait_env);
voodoo->wait_stats_enabled = voodoo->wait_stats_explicit && !voodoo_env_is_disabled(wait_env);
voodoo->lfb_relax_enabled = relax_enabled;
voodoo->lfb_relax_full = relax_enabled && (strcmp(relax_env, "full") == 0);
voodoo->lfb_relax_ignore_cmdfifo = relax_enabled && (!strcmp(relax_env, "nocmdfifo") || !strcmp(relax_env, "2") || !strcmp(relax_env, "3") || !strcmp(relax_env, "4") || !strcmp(relax_env, "frontsync"));
voodoo->lfb_relax_ignore_draw = relax_enabled && (!strcmp(relax_env, "nodraw") || !strcmp(relax_env, "2") || !strcmp(relax_env, "3") || !strcmp(relax_env, "4") || !strcmp(relax_env, "frontsync"));
voodoo->lfb_relax_ignore_fb_writes = relax_enabled && (!strcmp(relax_env, "nowrites") || !strcmp(relax_env, "3") || !strcmp(relax_env, "4") || !strcmp(relax_env, "frontsync"));
voodoo->lfb_relax_front_sync = relax_enabled && (!strcmp(relax_env, "4") || !strcmp(relax_env, "frontsync"));
}
static void
voodoo_update_queued_buffers(voodoo_t *voodoo)
{
switch (voodoo->queued_lfbMode & LFB_WRITE_MASK) {
case LFB_WRITE_FRONT:
voodoo->queued_fb_write_buffer = voodoo->queued_disp_buffer;
break;
case LFB_WRITE_BACK:
voodoo->queued_fb_write_buffer = voodoo->queued_draw_buffer;
break;
default:
voodoo->queued_fb_write_buffer = voodoo->queued_disp_buffer;
break;
}
switch (voodoo->queued_fbzMode & FBZ_DRAW_MASK) {
case FBZ_DRAW_FRONT:
voodoo->queued_fb_draw_buffer = voodoo->queued_disp_buffer;
break;
case FBZ_DRAW_BACK:
voodoo->queued_fb_draw_buffer = voodoo->queued_draw_buffer;
break;
default:
voodoo->queued_fb_draw_buffer = voodoo->queued_draw_buffer;
break;
}
}
void
voodoo_recalc(voodoo_t *voodoo)
{
@@ -167,11 +228,69 @@ voodoo_readw(uint32_t addr, void *priv)
voodoo = set->voodoos[0];
}
voodoo->flush = 1;
while (!FIFO_EMPTY)
voodoo_wake_fifo_thread_now(voodoo);
voodoo_wait_for_render_thread_idle(voodoo);
voodoo->flush = 0;
/* Reads from aux/draw/write regions must see completed rendering. */
int need_sync = (voodoo->fb_read_offset == voodoo->params.aux_offset) ||
(voodoo->fb_read_offset == voodoo->params.draw_offset) ||
(voodoo->fb_read_offset == voodoo->fb_write_offset);
int do_sync = 0;
int read_buf = -1;
if (voodoo->fb_read_offset == voodoo->params.front_offset)
read_buf = VOODOO_BUF_FRONT;
else if (voodoo->fb_read_offset == voodoo->back_offset)
read_buf = VOODOO_BUF_BACK;
else if (voodoo->fb_read_offset == voodoo->params.aux_offset)
read_buf = VOODOO_BUF_AUX;
if (!need_sync && voodoo->lfb_relax_front_sync && read_buf >= 0 && read_buf != VOODOO_BUF_BACK)
need_sync = 1;
if (need_sync) {
if (!voodoo->lfb_relax_enabled)
do_sync = 1;
else if (voodoo->lfb_relax_full)
do_sync = 0;
else {
/* In relax mode, only back-buffer reads can skip the full FIFO flush. */
int pending_buf = 0;
int pending_unknown = 0;
if (read_buf >= 0 && read_buf < VOODOO_BUF_COUNT) {
if (!voodoo->lfb_relax_ignore_fb_writes)
pending_buf += voodoo->pending_fb_writes_buf[read_buf];
if (!voodoo->lfb_relax_ignore_draw)
pending_buf += voodoo->pending_draw_cmds_buf[read_buf];
}
if (!voodoo->lfb_relax_ignore_fb_writes)
pending_unknown += voodoo->pending_fb_writes_buf[VOODOO_BUF_UNKNOWN];
if (!voodoo->lfb_relax_ignore_draw)
pending_unknown += voodoo->pending_draw_cmds_buf[VOODOO_BUF_UNKNOWN];
if (!voodoo->lfb_relax_ignore_cmdfifo) {
if ((voodoo->cmdfifo_depth_rd != voodoo->cmdfifo_depth_wr) || voodoo->cmdfifo_in_sub)
pending_unknown++;
if ((voodoo->cmdfifo_depth_rd_2 != voodoo->cmdfifo_depth_wr_2) || voodoo->cmdfifo_in_sub_2)
pending_unknown++;
}
if (read_buf != VOODOO_BUF_BACK)
do_sync = 1;
else
do_sync = (pending_buf || pending_unknown);
}
if (do_sync) {
voodoo->flush = 1;
while (!FIFO_EMPTY) {
voodoo_wake_fifo_thread_now(voodoo);
thread_wait_event(voodoo->fifo_empty_event, -1);
}
voodoo_wait_for_render_thread_idle(voodoo);
voodoo->flush = 0;
} else if (voodoo->lfb_relax_enabled && !voodoo->lfb_relax_full) {
voodoo_wait_for_render_thread_idle(voodoo);
}
}
return voodoo_fb_readw(addr, voodoo);
}
@@ -191,8 +310,17 @@ voodoo_readl(uint32_t addr, void *priv)
cycles -= voodoo->read_time;
if (addr & 0x800000) { /*Texture*/
if (voodoo->wait_stats_enabled)
voodoo->readl_tex_count++;
} else if (addr & 0x400000) /*Framebuffer*/
{
uint64_t fifo_wait_start = 0;
uint64_t fifo_wait_spins = 0;
int fifo_wait_active = 0;
int need_sync = 0;
int do_sync = 0;
int read_buf = -1;
if (SLI_ENABLED) {
const voodoo_set_t *set = voodoo->set;
int y = (addr >> 11) & 0x3ff;
@@ -203,23 +331,116 @@ voodoo_readl(uint32_t addr, void *priv)
voodoo = set->voodoos[0];
}
voodoo->flush = 1;
while (!FIFO_EMPTY) {
voodoo_wake_fifo_thread_now(voodoo);
thread_wait_event(voodoo->fifo_not_full_event, 1);
if (voodoo->wait_stats_enabled)
voodoo->readl_fb_count++;
if (voodoo->fb_read_offset == voodoo->params.front_offset)
read_buf = VOODOO_BUF_FRONT;
else if (voodoo->fb_read_offset == voodoo->back_offset)
read_buf = VOODOO_BUF_BACK;
else if (voodoo->fb_read_offset == voodoo->params.aux_offset)
read_buf = VOODOO_BUF_AUX;
/* Reads from aux/draw/write regions must see completed rendering. */
need_sync = (voodoo->fb_read_offset == voodoo->params.aux_offset) ||
(voodoo->fb_read_offset == voodoo->params.draw_offset) ||
(voodoo->fb_read_offset == voodoo->fb_write_offset);
if (!need_sync && voodoo->lfb_relax_front_sync && read_buf >= 0 && read_buf != VOODOO_BUF_BACK)
need_sync = 1;
if (need_sync) {
if (!voodoo->lfb_relax_enabled)
do_sync = 1;
else if (voodoo->lfb_relax_full)
do_sync = 0;
else {
/* In relax mode, only back-buffer reads can skip the full FIFO flush. */
int pending_buf = 0;
int pending_unknown = 0;
if (read_buf >= 0 && read_buf < VOODOO_BUF_COUNT) {
if (!voodoo->lfb_relax_ignore_fb_writes)
pending_buf += voodoo->pending_fb_writes_buf[read_buf];
if (!voodoo->lfb_relax_ignore_draw)
pending_buf += voodoo->pending_draw_cmds_buf[read_buf];
}
if (!voodoo->lfb_relax_ignore_fb_writes)
pending_unknown += voodoo->pending_fb_writes_buf[VOODOO_BUF_UNKNOWN];
if (!voodoo->lfb_relax_ignore_draw)
pending_unknown += voodoo->pending_draw_cmds_buf[VOODOO_BUF_UNKNOWN];
if (!voodoo->lfb_relax_ignore_cmdfifo) {
if ((voodoo->cmdfifo_depth_rd != voodoo->cmdfifo_depth_wr) || voodoo->cmdfifo_in_sub)
pending_unknown++;
if ((voodoo->cmdfifo_depth_rd_2 != voodoo->cmdfifo_depth_wr_2) || voodoo->cmdfifo_in_sub_2)
pending_unknown++;
}
if (read_buf != VOODOO_BUF_BACK)
do_sync = 1;
else
do_sync = (pending_buf || pending_unknown);
}
}
if (voodoo->wait_stats_enabled) {
if (do_sync)
voodoo->readl_fb_sync_count++;
else
voodoo->readl_fb_nosync_count++;
if (read_buf >= 0) {
if (do_sync)
voodoo->readl_fb_sync_buf[read_buf]++;
else
voodoo->readl_fb_nosync_buf[read_buf]++;
}
if (need_sync && voodoo->lfb_relax_enabled && !do_sync) {
voodoo->readl_fb_relaxed_count++;
if (read_buf >= 0)
voodoo->readl_fb_relaxed_buf[read_buf]++;
}
}
if (do_sync) {
voodoo->flush = 1;
while (!FIFO_EMPTY) {
if (voodoo->wait_stats_enabled) {
if (!fifo_wait_active) {
fifo_wait_active = 1;
fifo_wait_start = plat_timer_read();
voodoo->fifo_empty_waits++;
}
fifo_wait_spins++;
}
voodoo_wake_fifo_thread_now(voodoo);
thread_wait_event(voodoo->fifo_empty_event, -1);
}
if (fifo_wait_active) {
voodoo->fifo_empty_wait_ticks += plat_timer_read() - fifo_wait_start;
voodoo->fifo_empty_spin_checks += fifo_wait_spins;
}
voodoo_wait_for_render_thread_idle(voodoo);
voodoo->flush = 0;
} else if (need_sync && voodoo->lfb_relax_enabled && !voodoo->lfb_relax_full) {
voodoo_wait_for_render_thread_idle(voodoo);
}
voodoo_wait_for_render_thread_idle(voodoo);
voodoo->flush = 0;
temp = voodoo_fb_readl(addr, voodoo);
} else
} else {
if (voodoo->wait_stats_enabled)
voodoo->readl_reg_count++;
switch (addr & 0x3fc) {
case SST_status:
{
int fifo_entries = FIFO_ENTRIES;
int swap_count = voodoo->swap_count;
int written = voodoo->cmd_written + voodoo->cmd_written_fifo + voodoo->cmd_written_fifo_2;
int busy = (written - voodoo->cmd_read) || (voodoo->cmdfifo_depth_rd != voodoo->cmdfifo_depth_wr);
int busy = (written - voodoo->cmd_read) ||
(voodoo->cmdfifo_depth_rd != voodoo->cmdfifo_depth_wr) ||
voodoo->voodoo_busy ||
voodoo->render_voodoo_busy[0] ||
(voodoo->render_threads >= 2 && voodoo->render_voodoo_busy[1]) ||
(voodoo->render_threads == 4 && (voodoo->render_voodoo_busy[2] || voodoo->render_voodoo_busy[3]));
if (SLI_ENABLED && voodoo->type != VOODOO_2) {
voodoo_t *voodoo_other = (voodoo == voodoo->set->voodoos[0]) ? voodoo->set->voodoos[1] : voodoo->set->voodoos[0];
@@ -229,7 +450,12 @@ voodoo_readl(uint32_t addr, void *priv)
swap_count = voodoo_other->swap_count;
if ((voodoo_other->fifo_write_idx - voodoo_other->fifo_read_idx) > fifo_entries)
fifo_entries = voodoo_other->fifo_write_idx - voodoo_other->fifo_read_idx;
if ((other_written - voodoo_other->cmd_read) || (voodoo_other->cmdfifo_depth_rd != voodoo_other->cmdfifo_depth_wr))
if ((other_written - voodoo_other->cmd_read) ||
(voodoo_other->cmdfifo_depth_rd != voodoo_other->cmdfifo_depth_wr) ||
voodoo_other->voodoo_busy ||
voodoo_other->render_voodoo_busy[0] ||
(voodoo_other->render_threads >= 2 && voodoo_other->render_voodoo_busy[1]) ||
(voodoo_other->render_threads == 4 && (voodoo_other->render_voodoo_busy[2] || voodoo_other->render_voodoo_busy[3])))
busy = 1;
if (!voodoo_other->voodoo_busy)
voodoo_wake_fifo_thread(voodoo_other);
@@ -384,6 +610,7 @@ voodoo_readl(uint32_t addr, void *priv)
voodoo_log("voodoo_readl : bad addr %08X\n", addr);
temp = 0xffffffff;
}
}
return temp;
}
@@ -537,6 +764,11 @@ voodoo_writel(uint32_t addr, uint32_t val, void *priv)
happen here on a real Voodoo*/
voodoo->disp_buffer = 0;
voodoo->draw_buffer = 1;
voodoo->queued_disp_buffer = voodoo->disp_buffer;
voodoo->queued_draw_buffer = voodoo->draw_buffer;
voodoo->queued_lfbMode = voodoo->lfbMode;
voodoo->queued_fbzMode = voodoo->params.fbzMode;
voodoo_update_queued_buffers(voodoo);
voodoo_recalc(voodoo);
voodoo->front_offset = voodoo->params.front_offset;
}
@@ -925,6 +1157,7 @@ voodoo_card_init(void)
voodoo_t *voodoo = malloc(sizeof(voodoo_t));
memset(voodoo, 0, sizeof(voodoo_t));
voodoo_init_relax_settings(voodoo);
voodoo->bilinear_enabled = device_get_config_int("bilinear");
voodoo->dithersub_enabled = device_get_config_int("dithersub");
voodoo->scrfilter = device_get_config_int("dacfilter");
@@ -992,6 +1225,9 @@ voodoo_card_init(void)
voodoo->wake_render_thread[3] = thread_create_event();
voodoo->wake_main_thread = thread_create_event();
voodoo->fifo_not_full_event = thread_create_event();
voodoo->fifo_empty_event = thread_create_event();
thread_set_event(voodoo->fifo_empty_event);
ATOMIC_STORE(voodoo->fifo_empty_signaled, 1);
voodoo->render_not_full_event[0] = thread_create_event();
voodoo->render_not_full_event[1] = thread_create_event();
voodoo->render_not_full_event[2] = thread_create_event();
@@ -1065,6 +1301,11 @@ voodoo_card_init(void)
voodoo->disp_buffer = 0;
voodoo->draw_buffer = 1;
voodoo->queued_disp_buffer = voodoo->disp_buffer;
voodoo->queued_draw_buffer = voodoo->draw_buffer;
voodoo->queued_lfbMode = voodoo->lfbMode;
voodoo->queued_fbzMode = voodoo->params.fbzMode;
voodoo_update_queued_buffers(voodoo);
voodoo->force_blit_count = 0;
voodoo->can_blit = 0;
@@ -1080,6 +1321,7 @@ voodoo_2d3d_card_init(int type)
voodoo_t *voodoo = malloc(sizeof(voodoo_t));
memset(voodoo, 0, sizeof(voodoo_t));
voodoo_init_relax_settings(voodoo);
voodoo->bilinear_enabled = device_get_config_int("bilinear");
voodoo->dithersub_enabled = device_get_config_int("dithersub");
voodoo->scrfilter = device_get_config_int("dacfilter");
@@ -1116,6 +1358,9 @@ voodoo_2d3d_card_init(int type)
voodoo->wake_render_thread[3] = thread_create_event();
voodoo->wake_main_thread = thread_create_event();
voodoo->fifo_not_full_event = thread_create_event();
voodoo->fifo_empty_event = thread_create_event();
thread_set_event(voodoo->fifo_empty_event);
ATOMIC_STORE(voodoo->fifo_empty_signaled, 1);
voodoo->render_not_full_event[0] = thread_create_event();
voodoo->render_not_full_event[1] = thread_create_event();
voodoo->render_not_full_event[2] = thread_create_event();
@@ -1189,6 +1434,11 @@ voodoo_2d3d_card_init(int type)
voodoo->disp_buffer = 0;
voodoo->draw_buffer = 1;
voodoo->queued_disp_buffer = voodoo->disp_buffer;
voodoo->queued_draw_buffer = voodoo->draw_buffer;
voodoo->queued_lfbMode = voodoo->lfbMode;
voodoo->queued_fbzMode = voodoo->params.fbzMode;
voodoo_update_queued_buffers(voodoo);
voodoo->force_blit_count = 0;
voodoo->can_blit = 0;
@@ -1277,6 +1527,7 @@ voodoo_card_close(voodoo_t *voodoo)
thread_wait(voodoo->render_thread[3]);
}
thread_destroy_event(voodoo->fifo_not_full_event);
thread_destroy_event(voodoo->fifo_empty_event);
thread_destroy_event(voodoo->wake_main_thread);
thread_destroy_event(voodoo->wake_fifo_thread);
thread_destroy_event(voodoo->wake_render_thread[0]);
@@ -1284,6 +1535,48 @@ voodoo_card_close(voodoo_t *voodoo)
thread_destroy_event(voodoo->render_not_full_event[0]);
thread_destroy_event(voodoo->render_not_full_event[1]);
if (voodoo->wait_stats_enabled && voodoo->wait_stats_explicit) {
pclog("Voodoo wait stats (type=%d): fifo_full waits=%" PRIu64 " ticks=%" PRIu64 " spins=%" PRIu64
", fifo_empty waits=%" PRIu64 " ticks=%" PRIu64 " spins=%" PRIu64
", render_wait waits=%" PRIu64 " ticks=%" PRIu64 " spins=%" PRIu64
", readl fb=%" PRIu64 " sync=%" PRIu64 " nosync=%" PRIu64 " relaxed=%" PRIu64 " relax=%d full=%d nocmdfifo=%d nodraw=%d nowrites=%d frontsync=%d"
" sync_buf f=%" PRIu64 " b=%" PRIu64 " a=%" PRIu64
" nosync_buf f=%" PRIu64 " b=%" PRIu64 " a=%" PRIu64
" relaxed_buf f=%" PRIu64 " b=%" PRIu64 " a=%" PRIu64
" reg=%" PRIu64 " tex=%" PRIu64 "\n",
voodoo->type,
voodoo->fifo_full_waits,
voodoo->fifo_full_wait_ticks,
voodoo->fifo_full_spin_checks,
voodoo->fifo_empty_waits,
voodoo->fifo_empty_wait_ticks,
voodoo->fifo_empty_spin_checks,
voodoo->render_waits,
voodoo->render_wait_ticks,
voodoo->render_wait_spin_checks,
voodoo->readl_fb_count,
voodoo->readl_fb_sync_count,
voodoo->readl_fb_nosync_count,
voodoo->readl_fb_relaxed_count,
voodoo->lfb_relax_enabled,
voodoo->lfb_relax_full,
voodoo->lfb_relax_ignore_cmdfifo,
voodoo->lfb_relax_ignore_draw,
voodoo->lfb_relax_ignore_fb_writes,
voodoo->lfb_relax_front_sync,
voodoo->readl_fb_sync_buf[0],
voodoo->readl_fb_sync_buf[1],
voodoo->readl_fb_sync_buf[2],
voodoo->readl_fb_nosync_buf[0],
voodoo->readl_fb_nosync_buf[1],
voodoo->readl_fb_nosync_buf[2],
voodoo->readl_fb_relaxed_buf[0],
voodoo->readl_fb_relaxed_buf[1],
voodoo->readl_fb_relaxed_buf[2],
voodoo->readl_reg_count,
voodoo->readl_tex_count);
}
for (uint8_t c = 0; c < TEX_CACHE_MAX; c++) {
if (voodoo->dual_tmus)
free(voodoo->texture_cache[1][c].data);

191
src/video/vid_voodoo_fifo.c
View File

@@ -62,8 +62,97 @@ voodoo_fifo_log(const char *fmt, ...)
#define WAKE_DELAY_DEFAULT (TIMER_USEC * 100)
/* Per-card wake delay: Voodoo1 uses a larger delay to reduce FIFO wake frequency */
#define WAKE_DELAY_OF(v) ((v)->type == VOODOO_1 ? (TIMER_USEC * 500) : WAKE_DELAY_DEFAULT)
/* Per-card wake delay: keep all Voodoo cards at the default */
#define WAKE_DELAY_OF(v) (WAKE_DELAY_DEFAULT)
static __inline uint8_t
voodoo_queue_color_buf_tag(const voodoo_t *voodoo, int buf)
{
if (buf == voodoo->queued_disp_buffer)
return VOODOO_BUF_FRONT;
if (buf == voodoo->queued_draw_buffer)
return VOODOO_BUF_BACK;
return VOODOO_BUF_UNKNOWN;
}
static __inline void
voodoo_queue_recalc_buffers(voodoo_t *voodoo)
{
switch (voodoo->queued_lfbMode & LFB_WRITE_MASK) {
case LFB_WRITE_FRONT:
voodoo->queued_fb_write_buffer = voodoo->queued_disp_buffer;
break;
case LFB_WRITE_BACK:
voodoo->queued_fb_write_buffer = voodoo->queued_draw_buffer;
break;
default:
voodoo->queued_fb_write_buffer = voodoo->queued_disp_buffer;
break;
}
switch (voodoo->queued_fbzMode & FBZ_DRAW_MASK) {
case FBZ_DRAW_FRONT:
voodoo->queued_fb_draw_buffer = voodoo->queued_disp_buffer;
break;
case FBZ_DRAW_BACK:
voodoo->queued_fb_draw_buffer = voodoo->queued_draw_buffer;
break;
default:
voodoo->queued_fb_draw_buffer = voodoo->queued_draw_buffer;
break;
}
}
static __inline void
voodoo_queue_apply_reg(voodoo_t *voodoo, uint32_t addr, uint32_t val)
{
switch (addr & 0x3fc) {
case SST_lfbMode:
voodoo->queued_lfbMode = val;
voodoo_queue_recalc_buffers(voodoo);
break;
case SST_fbzMode:
voodoo->queued_fbzMode = val;
voodoo_queue_recalc_buffers(voodoo);
break;
case SST_swapbufferCMD:
if (TRIPLE_BUFFER) {
voodoo->queued_disp_buffer = (voodoo->queued_disp_buffer + 1) % 3;
voodoo->queued_draw_buffer = (voodoo->queued_draw_buffer + 1) % 3;
} else {
voodoo->queued_disp_buffer = !voodoo->queued_disp_buffer;
voodoo->queued_draw_buffer = !voodoo->queued_draw_buffer;
}
voodoo_queue_recalc_buffers(voodoo);
break;
default:
break;
}
}
static __inline uint8_t
voodoo_queue_reg_target_buf(voodoo_t *voodoo, uint32_t addr)
{
switch (addr & 0x3fc) {
case SST_triangleCMD:
case SST_ftriangleCMD:
case SST_fastfillCMD:
return voodoo_queue_color_buf_tag(voodoo, voodoo->queued_fb_draw_buffer);
case SST_lfbMode:
case SST_fbzMode:
case SST_swapbufferCMD:
return VOODOO_BUF_UNKNOWN;
default:
return VOODOO_BUF_NONE;
}
}
static __inline void
voodoo_cmdfifo_reg_writel(voodoo_t *voodoo, uint32_t addr, uint32_t val)
{
voodoo_reg_writel(addr, val, voodoo);
voodoo_queue_apply_reg(voodoo, addr, val);
}
void
voodoo_wake_fifo_thread(voodoo_t *voodoo)
{
@@ -94,8 +183,19 @@ void
voodoo_queue_command(voodoo_t *voodoo, uint32_t addr_type, uint32_t val)
{
fifo_entry_t *fifo = &voodoo->fifo[voodoo->fifo_write_idx & FIFO_MASK];
uint64_t fifo_wait_start = 0;
uint64_t fifo_wait_spins = 0;
int fifo_wait_active = 0;
while (FIFO_FULL) {
if (voodoo->wait_stats_enabled) {
if (!fifo_wait_active) {
fifo_wait_active = 1;
fifo_wait_start = plat_timer_read();
voodoo->fifo_full_waits++;
}
fifo_wait_spins++;
}
thread_reset_event(voodoo->fifo_not_full_event);
if (FIFO_FULL) {
thread_wait_event(voodoo->fifo_not_full_event, 1); /*Wait for room in ringbuffer*/
@@ -104,8 +204,38 @@ voodoo_queue_command(voodoo_t *voodoo, uint32_t addr_type, uint32_t val)
}
}
fifo->val = val;
fifo->addr_type = addr_type;
if (fifo_wait_active) {
voodoo->fifo_full_wait_ticks += plat_timer_read() - fifo_wait_start;
voodoo->fifo_full_spin_checks += fifo_wait_spins;
}
#ifdef _WIN32
/* Reset only after an empty signal to avoid heavy ResetEvent churn on Windows. */
if (ATOMIC_LOAD(voodoo->fifo_empty_signaled)) {
ATOMIC_STORE(voodoo->fifo_empty_signaled, 0);
thread_reset_event(voodoo->fifo_empty_event);
}
#else
thread_reset_event(voodoo->fifo_empty_event);
#endif
fifo->val = val;
fifo->addr_type = addr_type;
fifo->target_buf = VOODOO_BUF_NONE;
if (((addr_type & FIFO_TYPE) == FIFO_WRITEW_FB) ||
((addr_type & FIFO_TYPE) == FIFO_WRITEL_FB)) {
fifo->target_buf = voodoo_queue_color_buf_tag(voodoo, voodoo->queued_fb_write_buffer);
ATOMIC_INC(voodoo->pending_fb_writes_buf[fifo->target_buf]);
} else if ((addr_type & FIFO_TYPE) == FIFO_WRITEL_REG) {
uint8_t reg_buf = voodoo_queue_reg_target_buf(voodoo, addr_type & FIFO_ADDR);
if (reg_buf != VOODOO_BUF_NONE) {
fifo->target_buf = reg_buf;
ATOMIC_INC(voodoo->pending_draw_cmds_buf[fifo->target_buf]);
}
voodoo_queue_apply_reg(voodoo, addr_type & FIFO_ADDR, val);
}
voodoo->fifo_write_idx++;
voodoo->cmd_status &= ~(1 << 24);
@@ -120,7 +250,7 @@ voodoo_flush(voodoo_t *voodoo)
voodoo->flush = 1;
while (!FIFO_EMPTY) {
voodoo_wake_fifo_thread_now(voodoo);
thread_wait_event(voodoo->fifo_not_full_event, 1);
thread_wait_event(voodoo->fifo_empty_event, -1);
}
voodoo_wait_for_render_thread_idle(voodoo);
voodoo->flush = 0;
@@ -138,9 +268,7 @@ void
voodoo_wait_for_swap_complete(voodoo_t *voodoo)
{
while (voodoo->swap_pending) {
thread_wait_event(voodoo->wake_fifo_thread, -1);
thread_reset_event(voodoo->wake_fifo_thread);
/* Avoid waiting on wake_fifo_thread here; main thread may be draining the FIFO. */
thread_wait_mutex(voodoo->swap_mutex);
if ((voodoo->swap_pending && voodoo->flush) || FIFO_FULL) {
/*Main thread is waiting for FIFO to empty, so skip vsync wait and just swap*/
@@ -153,6 +281,13 @@ voodoo_wait_for_swap_complete(voodoo_t *voodoo)
break;
} else
thread_release_mutex(voodoo->swap_mutex);
/* Yield briefly while waiting for the swap to complete. */
#ifdef _WIN32
/* Sleep(1) can add measurable stalls on Windows. */
plat_delay_ms(0);
#else
plat_delay_ms(1);
#endif
}
}
@@ -261,9 +396,15 @@ voodoo_fifo_thread(void *param)
switch (fifo->addr_type & FIFO_TYPE) {
case FIFO_WRITEL_REG:
while ((fifo->addr_type & FIFO_TYPE) == FIFO_WRITEL_REG) {
voodoo_reg_writel(fifo->addr_type & FIFO_ADDR, fifo->val, voodoo);
uint32_t reg_addr = fifo->addr_type & FIFO_ADDR;
uint8_t target_buf = fifo->target_buf;
voodoo_reg_writel(reg_addr, fifo->val, voodoo);
fifo->addr_type = FIFO_INVALID;
voodoo->fifo_read_idx++;
if (target_buf < VOODOO_BUF_COUNT) {
ATOMIC_DEC(voodoo->pending_draw_cmds_buf[target_buf]);
}
if (FIFO_EMPTY)
break;
fifo = &voodoo->fifo[voodoo->fifo_read_idx & FIFO_MASK];
@@ -272,9 +413,14 @@ voodoo_fifo_thread(void *param)
case FIFO_WRITEW_FB:
voodoo_wait_for_render_thread_idle(voodoo);
while ((fifo->addr_type & FIFO_TYPE) == FIFO_WRITEW_FB) {
uint8_t target_buf = fifo->target_buf;
voodoo_fb_writew(fifo->addr_type & FIFO_ADDR, fifo->val, voodoo);
fifo->addr_type = FIFO_INVALID;
voodoo->fifo_read_idx++;
if (target_buf < VOODOO_BUF_COUNT) {
ATOMIC_DEC(voodoo->pending_fb_writes_buf[target_buf]);
}
if (FIFO_EMPTY)
break;
fifo = &voodoo->fifo[voodoo->fifo_read_idx & FIFO_MASK];
@@ -283,9 +429,14 @@ voodoo_fifo_thread(void *param)
case FIFO_WRITEL_FB:
voodoo_wait_for_render_thread_idle(voodoo);
while ((fifo->addr_type & FIFO_TYPE) == FIFO_WRITEL_FB) {
uint8_t target_buf = fifo->target_buf;
voodoo_fb_writel(fifo->addr_type & FIFO_ADDR, fifo->val, voodoo);
fifo->addr_type = FIFO_INVALID;
voodoo->fifo_read_idx++;
if (target_buf < VOODOO_BUF_COUNT) {
ATOMIC_DEC(voodoo->pending_fb_writes_buf[target_buf]);
}
if (FIFO_EMPTY)
break;
fifo = &voodoo->fifo[voodoo->fifo_read_idx & FIFO_MASK];
@@ -326,6 +477,8 @@ voodoo_fifo_thread(void *param)
voodoo->cmd_status |= (1 << 24);
voodoo->cmd_status_2 |= (1 << 24);
thread_set_event(voodoo->fifo_empty_event);
ATOMIC_STORE(voodoo->fifo_empty_signaled, 1);
while (voodoo->cmdfifo_enabled && (voodoo->cmdfifo_depth_rd != voodoo->cmdfifo_depth_wr || voodoo->cmdfifo_in_sub)) {
uint64_t start_time = plat_timer_read();
@@ -418,7 +571,7 @@ voodoo_fifo_thread(void *param)
if (voodoo->type >= VOODOO_BANSHEE && (addr & 0x3ff) == SST_swapbufferCMD)
voodoo->cmd_written_fifo++;
voodoo_reg_writel(addr, val, voodoo);
voodoo_cmdfifo_reg_writel(voodoo, addr, val);
}
if (header & (1 << 15))
@@ -447,7 +600,7 @@ voodoo_fifo_thread(void *param)
num = (header >> 29) & 7;
mask = header; //(header >> 10) & 0xff;
smode = (header >> 22) & 0xf;
voodoo_reg_writel(SST_sSetupMode, ((header >> 10) & 0xff) | (smode << 16), voodoo);
voodoo_cmdfifo_reg_writel(voodoo, SST_sSetupMode, ((header >> 10) & 0xff) | (smode << 16));
num_verticies = (header >> 6) & 0xf;
v_num = 0;
if (((header >> 3) & 7) == 2)
@@ -492,9 +645,9 @@ voodoo_fifo_thread(void *param)
voodoo->verts[3].sT1 = cmdfifo_get_f(voodoo);
}
if (v_num)
voodoo_reg_writel(SST_sDrawTriCMD, 0, voodoo);
voodoo_cmdfifo_reg_writel(voodoo, SST_sDrawTriCMD, 0);
else
voodoo_reg_writel(SST_sBeginTriCMD, 0, voodoo);
voodoo_cmdfifo_reg_writel(voodoo, SST_sBeginTriCMD, 0);
v_num++;
if (v_num == 3 && ((header >> 3) & 7) == 0)
v_num = 0;
@@ -528,7 +681,7 @@ voodoo_fifo_thread(void *param)
if (voodoo->type >= VOODOO_BANSHEE && (addr & 0x3ff) == SST_swapbufferCMD)
voodoo->cmd_written_fifo++;
voodoo_reg_writel(addr, val, voodoo);
voodoo_cmdfifo_reg_writel(voodoo, addr, val);
}
}
@@ -709,7 +862,7 @@ voodoo_fifo_thread(void *param)
if (voodoo->type >= VOODOO_BANSHEE && (addr & 0x3ff) == SST_swapbufferCMD)
voodoo->cmd_written_fifo_2++;
voodoo_reg_writel(addr, val, voodoo);
voodoo_cmdfifo_reg_writel(voodoo, addr, val);
}
if (header & (1 << 15))
@@ -738,7 +891,7 @@ voodoo_fifo_thread(void *param)
num = (header >> 29) & 7;
mask = header; //(header >> 10) & 0xff;
smode = (header >> 22) & 0xf;
voodoo_reg_writel(SST_sSetupMode, ((header >> 10) & 0xff) | (smode << 16), voodoo);
voodoo_cmdfifo_reg_writel(voodoo, SST_sSetupMode, ((header >> 10) & 0xff) | (smode << 16));
num_verticies = (header >> 6) & 0xf;
v_num = 0;
if (((header >> 3) & 7) == 2)
@@ -783,9 +936,9 @@ voodoo_fifo_thread(void *param)
voodoo->verts[3].sT1 = cmdfifo_get_f_2(voodoo);
}
if (v_num)
voodoo_reg_writel(SST_sDrawTriCMD, 0, voodoo);
voodoo_cmdfifo_reg_writel(voodoo, SST_sDrawTriCMD, 0);
else
voodoo_reg_writel(SST_sBeginTriCMD, 0, voodoo);
voodoo_cmdfifo_reg_writel(voodoo, SST_sBeginTriCMD, 0);
v_num++;
if (v_num == 3 && ((header >> 3) & 7) == 0)
v_num = 0;
@@ -819,7 +972,7 @@ voodoo_fifo_thread(void *param)
if (voodoo->type >= VOODOO_BANSHEE && (addr & 0x3ff) == SST_swapbufferCMD)
voodoo->cmd_written_fifo_2++;
voodoo_reg_writel(addr, val, voodoo);
voodoo_cmdfifo_reg_writel(voodoo, addr, val);
}
}