From fcb5f6d1d3675b1b0bd50a186918a78b46975744 Mon Sep 17 00:00:00 2001
From: Volker Ruppert <Volker.Ruppert@t-online.de>
Date: Tue, 15 Aug 2017 21:30:45 +0000
Subject: [PATCH] Voodoo2: Now only start processing of the CMDFIFO at the
 vertical retrace. I noticed a speed improvement of around 10 % (tested with
 donut demo only).

---
 bochs/iodev/display/voodoo.cc     | 17 ++++++++++++++++-
 bochs/iodev/display/voodoo_data.h |  3 +++
 bochs/iodev/display/voodoo_func.h | 10 ++--------
 3 files changed, 21 insertions(+), 9 deletions(-)

diff --git a/bochs/iodev/display/voodoo.cc b/bochs/iodev/display/voodoo.cc
index 3fd2dc836..ab217e12d 100644
--- a/bochs/iodev/display/voodoo.cc
+++ b/bochs/iodev/display/voodoo.cc
@@ -166,15 +166,22 @@ BX_THREAD_FUNC(cmdfifo_thread, indata)
   while (1) {
 #ifdef WIN32
     if (WaitForSingleObject(v->fbi.cmdfifo[0].event, 1) == WAIT_OBJECT_0) {
+      BX_LOCK(cmdfifo_mutex);
       while (v->fbi.cmdfifo[0].enable && (v->fbi.cmdfifo[0].depth >= v->fbi.cmdfifo[0].depth_needed)) {
         cmdfifo_process();
       }
+      v->fbi.cmdfifo[0].cmd_ready = 0;
+      BX_UNLOCK(cmdfifo_mutex);
     }
 #else
+    while (!v->fbi.cmdfifo[0].event) BX_MSLEEP(1);
+    BX_LOCK(cmdfifo_mutex);
     while (v->fbi.cmdfifo[0].enable && (v->fbi.cmdfifo[0].depth >= v->fbi.cmdfifo[0].depth_needed)) {
       cmdfifo_process();
     }
-    BX_MSLEEP(1);
+    v->fbi.cmdfifo[0].cmd_ready = 0;
+    v->fbi.cmdfifo[0].event = 0;
+    BX_UNLOCK(cmdfifo_mutex);
 #endif
   }
   BX_THREAD_EXIT;
@@ -603,6 +610,14 @@ void bx_voodoo_c::vertical_timer_handler(void *this_ptr)
 
   BX_VOODOO_THIS s.vdraw.frame_start = bx_virt_timer.time_usec(BX_VOODOO_THIS s.vdraw.realtime);
 
+  if (v->fbi.cmdfifo[0].cmd_ready) {
+#ifdef WIN32
+    SetEvent(v->fbi.cmdfifo[0].event);
+#else
+    v->fbi.cmdfifo[0].event = 1;
+#endif
+  }
+
   if (v->fbi.vblank_swap_pending) {
     swap_buffers(v);
   }
diff --git a/bochs/iodev/display/voodoo_data.h b/bochs/iodev/display/voodoo_data.h
index 81bbf084c..3659a7ffa 100644
--- a/bochs/iodev/display/voodoo_data.h
+++ b/bochs/iodev/display/voodoo_data.h
@@ -1439,8 +1439,11 @@ struct _cmdfifo_info
   Bit32u  depth;        /* current depth */
   Bit32u  depth_needed; /* depth needed for command */
   Bit32u  holes;        /* number of holes */
+  bx_bool cmd_ready;
 #ifdef WIN32
   HANDLE  event;
+#else
+  bx_bool event;
 #endif
 };
 
diff --git a/bochs/iodev/display/voodoo_func.h b/bochs/iodev/display/voodoo_func.h
index aeb612283..cdafd36ed 100644
--- a/bochs/iodev/display/voodoo_func.h
+++ b/bochs/iodev/display/voodoo_func.h
@@ -2406,9 +2406,7 @@ Bit32u cmdfifo_calc_depth_needed(void)
 
   if (v->fbi.cmdfifo[0].depth == 0)
     return needed;
-  BX_LOCK(cmdfifo_mutex);
   command = *(Bit32u*)(&v->fbi.ram[v->fbi.cmdfifo[0].rdptr & v->fbi.mask]);
-  BX_UNLOCK(cmdfifo_mutex);
   type = (Bit8u)(command & 0x07);
   switch (type) {
     case 0:
@@ -2462,26 +2460,22 @@ void cmdfifo_w(Bit32u fbi_offset, Bit32u data)
   BX_LOCK(cmdfifo_mutex);
   *(Bit32u*)(&v->fbi.ram[fbi_offset]) = data;
   v->fbi.cmdfifo[0].depth++;
-  BX_UNLOCK(cmdfifo_mutex);
   if (v->fbi.cmdfifo[0].depth_needed == BX_MAX_BIT32U) {
     v->fbi.cmdfifo[0].depth_needed = cmdfifo_calc_depth_needed();
   }
-#ifdef WIN32
   if (v->fbi.cmdfifo[0].depth >= v->fbi.cmdfifo[0].depth_needed) {
-    SetEvent(v->fbi.cmdfifo[0].event);
+    v->fbi.cmdfifo[0].cmd_ready = 1;
   }
-#endif
+  BX_UNLOCK(cmdfifo_mutex);
 }
 
 Bit32u cmdfifo_r(void)
 {
   Bit32u data;
 
-  BX_LOCK(cmdfifo_mutex);
   data = *(Bit32u*)(&v->fbi.ram[v->fbi.cmdfifo[0].rdptr & v->fbi.mask]);
   v->fbi.cmdfifo[0].rdptr += 4;
   v->fbi.cmdfifo[0].depth--;
-  BX_UNLOCK(cmdfifo_mutex);
   return data;
 }