From e11e4ba7b569bfa2f23e5adf7a471ab7c829f63c Mon Sep 17 00:00:00 2001
From: Pavel <68122101+red-prig@users.noreply.github.com>
Date: Thu, 8 Aug 2024 11:10:43 +0300
Subject: [PATCH] +

---
 chip/pm4_me.pas    | 15 +++++--------
 sys/dev/dev_gc.pas | 56 ++++++++++++++++++++++++++++++++++++++++++++--
 2 files changed, 59 insertions(+), 12 deletions(-)

diff --git a/chip/pm4_me.pas b/chip/pm4_me.pas
index de4d26e2..632d0d65 100644
--- a/chip/pm4_me.pas
+++ b/chip/pm4_me.pas
@@ -75,7 +75,7 @@ type
   end;
   //
   event:PRTLEvent;
-  //on_idle:TProcedure;
+  on_idle:TProcedure;
   on_submit_flip_eop:t_on_submit_flip_eop;
   //
   started:Pointer;
@@ -162,7 +162,7 @@ type
   //
   procedure switch_task;
   procedure next_task;
-  //procedure on_idle;
+  procedure on_idle;
  end;
 
 var
@@ -673,7 +673,6 @@ begin
  end;
 end;
 
-{
 procedure t_me_render_context.on_idle;
 begin
  if (me^.on_idle<>nil) then
@@ -681,7 +680,6 @@ begin
   me^.on_idle();
  end;
 end;
-}
 
 //
 
@@ -1892,7 +1890,6 @@ begin
   ctx.me^.knote_eventid($40,0,curr*NSEC_PER_UNIT,0); //(absolute time) (freq???)
  end;
 
- //ctx.on_idle;
 end;
 
 procedure pm4_SubmitFlipEop(var ctx:t_me_render_context;node:p_pm4_node_SubmitFlipEop);
@@ -1922,7 +1919,6 @@ begin
   ctx.me^.knote_eventid($40,0,curr*NSEC_PER_UNIT,0); //(absolute time) (freq???)
  end;
 
- //ctx.on_idle;
 end;
 
 function get_compute_pipe_id(buft:t_pm4_stream_type):Byte; inline;
@@ -2016,7 +2012,6 @@ begin
   ctx.me^.knote_eventid(get_compute_pipe_id(ctx.stream^.buft),0,curr*NSEC_PER_UNIT,0); //(absolute time) (freq???)
  end;
 
- //ctx.on_idle;
 end;
 
 procedure pm4_EventWrite(var ctx:t_me_render_context;node:p_pm4_node_EventWrite);
@@ -2104,7 +2099,6 @@ begin
    Assert(false,'pm4_EventWriteEos');
  end;
 
- //ctx.on_idle;
 end;
 
 procedure pm4_WriteData(var ctx:t_me_render_context;node:p_pm4_node_WriteData);
@@ -2299,7 +2293,7 @@ begin
  begin
   Inc(ctx.stream^.hint_loop);
   //
-  if (ctx.stream^.hint_loop>1000) then
+  if (ctx.stream^.hint_loop>10000) then
   begin
    //loop detection
    Writeln(stderr,'WaitReg loop detected -> skip');
@@ -2472,8 +2466,9 @@ begin
   me^.reset_sheduler;
 
   ctx.rel_time:=0; //reset time
+
   //
-  //ctx.on_idle;
+  ctx.on_idle;
   //
 
   RTLEventWaitFor(me^.event);
diff --git a/sys/dev/dev_gc.pas b/sys/dev/dev_gc.pas
index ea0bc9fc..fa475d45 100644
--- a/sys/dev/dev_gc.pas
+++ b/sys/dev/dev_gc.pas
@@ -10,6 +10,9 @@ uses
 
 procedure gc_initialize();
 
+var
+ sync_me_submit:Boolean=False; //forced wait for all tasks to complete on the GPU side after submit
+
 implementation
 
 uses
@@ -158,6 +161,9 @@ var
  GC_SRI_event:PRTLEvent=nil;
  GC_SRI_label:QWORD=0;
 
+ me_idle_event:PRTLEvent=nil;
+ me_idle_label:QWORD=0;
+
  parse_gfx_started:Pointer=nil;
  parse_gfx_td:p_kthread;
 
@@ -431,7 +437,7 @@ begin
  end;
 end;
 
-procedure gc_idle; {register;}
+procedure gc_idle;
 begin
  if (GC_SRI_event<>nil) then
  begin
@@ -445,6 +451,38 @@ begin
  end;
 end;
 
+procedure me_idle; register;
+begin
+ if (me_idle_event<>nil) then
+ begin
+  RTLEventSetEvent(me_idle_event);
+ end;
+ me_idle_label:=0;
+end;
+
+procedure wait_me_idle;
+begin
+ if (me_idle_label=0) then Exit;
+
+ if (me_idle_event<>nil) then
+ begin
+  RTLEventWaitFor(me_idle_event);
+ end;
+end;
+
+procedure reset_me_idle;
+begin
+ me_idle_label:=1;
+
+ if (me_idle_event=nil) then
+ begin
+  me_idle_event:=RTLEventCreate;
+ end else
+ begin
+  RTLEventResetEvent(me_idle_event);
+ end;
+end;
+
 procedure gc_imdone;
 var
  prev:QWORD;
@@ -798,6 +836,12 @@ begin
 
   $C0108102: //submit
             begin
+             if sync_me_submit then
+             begin
+              reset_me_idle;
+             end;
+             //
+
              start_gfx_ring;
 
              rw_wlock(ring_gfx_lock);
@@ -811,7 +855,15 @@ begin
              if (Result=0) then
              begin
               retrigger_watchdog;
+
+              if sync_me_submit then
+              begin
+               //force wait GPU idle
+               wait_me_idle;
+              end;
+              //msleep_td(hz);
              end;
+
             end;
 
   $C020810C: //submit eop
@@ -1142,7 +1194,7 @@ begin
  gc_ring_create(@ring_gfx,GC_RING_SIZE);
 
  pm4_me_gfx.Init(@gc_knlist);
- //pm4_me_gfx.on_idle:=@gc_idle;
+ pm4_me_gfx.on_idle:=@me_idle;
  pm4_me_gfx.on_submit_flip_eop:=@dev_dce.TriggerFlipEop;
 end;