From 6b3f9bff2af04b42540fd224e1a7bc5c0cba0f76 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Emmanuel=20Beno=C3=AEt?= <tseeker@nocternity.net>
Date: Mon, 25 Dec 2017 20:39:56 +0100
Subject: [PATCH] Scripting - Handle compute shader groups automatically

---
 test/demo.srd                   | 27 +++++----------------------
 test/shaders/chunks/dof-cs.glsl |  5 -----
 test/shaders/dof-pass1.c.glsl   |  6 ++++++
 test/shaders/dof-pass2.c.glsl   |  6 ++++++
 ui-opemu.cc                     | 16 ++++++++++++++--
 ui-shaders.cc                   |  1 +
 ui-shaders.hh                   |  5 +++++
 7 files changed, 37 insertions(+), 29 deletions(-)

diff --git a/test/demo.srd b/test/demo.srd
index f16a03a..2b2ed6c 100644
--- a/test/demo.srd
+++ b/test/demo.srd
@@ -19,13 +19,6 @@
 
 	(set use-compute 1)
 
-	(set render-compute-size-x 8)
-	(set render-compute-size-y 8)
-	(set combine-compute-size-x 8)
-	(set combine-compute-size-y 32)
-	(set dof-compute-size-x 64)
-	(set dof-compute-size-y 16)
-
 	(call scene-init)
 	(call dof-init)
 	(call bloom-init)
@@ -150,10 +143,7 @@
 		(use-pipeline pl-scene-p1)
 		(if $use-compute (
 			(image 0 tx-scene-output 0)
-			(compute
-				(add 1 (div $vp-width $render-compute-size-x))
-				(add 1 (div $vp-height $render-compute-size-y))
-				1)
+			(compute $vp-width $vp-height 1)
 		)(
 			(uniforms prg-scene-p1 1 $vp-width $vp-height)
 			(use-framebuffer rt-scene)
@@ -252,11 +242,7 @@
 		(uniforms-i prg-dof-pass1 0 0)
 		(uniforms-i prg-dof-pass2 0 0)
 
-		(if $use-compute (
-			(locals dof-cx dof-cy)
-			(set dof-cx (add 1 (div $vp-width $dof-compute-size-x)))
-			(set dof-cy (add 1 (div $vp-height $dof-compute-size-y)))
-		)(
+		(if (not $use-compute) (
 			(use-texture 1 in-depth smp-dof)
 			(uniforms-i prg-dof-pass1 1 1)
 			(uniforms-i prg-dof-pass2 1 1)
@@ -268,7 +254,7 @@
 		(use-pipeline pl-dof-pass1)
 		(if $use-compute (
 			(image 0 tx-dof-pass1 0)
-			(compute $dof-cx $dof-cy 1)
+			(compute $vp-width $vp-height 1)
 		)(
 			(use-framebuffer rt-dof-pass1)
 			(viewport 0 0 $vp-width $vp-height)
@@ -281,7 +267,7 @@
 		(use-pipeline pl-dof-pass2)
 		(if $use-compute (
 			(image 0 tx-dof-pass2 0)
-			(compute $dof-cx $dof-cy 1)
+			(compute $vp-width $vp-height 1)
 		)(
 			(use-framebuffer rt-dof-pass2)
 			(viewport 0 0 $vp-width $vp-height)
@@ -567,10 +553,7 @@
 
 		(if $use-compute (
 			(image 0 tx-combined 0)
-			(compute
-				(add 1 (div $vp-width $combine-compute-size-x))
-				(add 1 (div $vp-height $combine-compute-size-y))
-				1)
+			(compute $vp-width $vp-height 1)
 		)(
 			(viewport 0 0 $vp-width $vp-height)
 			(fullscreen)
diff --git a/test/shaders/chunks/dof-cs.glsl b/test/shaders/chunks/dof-cs.glsl
index 7cf03d0..7130d21 100644
--- a/test/shaders/chunks/dof-cs.glsl
+++ b/test/shaders/chunks/dof-cs.glsl
@@ -1,10 +1,5 @@
 //! type chunk
 
-layout(
-	local_size_x = 64 ,
-	local_size_y = 16
-) in;
-
 //#define DOF_USE_RANDOM
 
 layout( location = 0 ) uniform sampler2D u_Input;
diff --git a/test/shaders/dof-pass1.c.glsl b/test/shaders/dof-pass1.c.glsl
index fd4142c..0e9315a 100644
--- a/test/shaders/dof-pass1.c.glsl
+++ b/test/shaders/dof-pass1.c.glsl
@@ -1,6 +1,12 @@
 #version 450 core
 
 //! type compute
+
+layout(
+	local_size_x = 8 ,
+	local_size_y = 64
+) in;
+
 //! include chunks/dof-cs.glsl
 
 void main()
diff --git a/test/shaders/dof-pass2.c.glsl b/test/shaders/dof-pass2.c.glsl
index d1238b0..d322ce1 100644
--- a/test/shaders/dof-pass2.c.glsl
+++ b/test/shaders/dof-pass2.c.glsl
@@ -1,6 +1,12 @@
 #version 450 core
 
 //! type compute
+
+layout(
+	local_size_x = 8 ,
+	local_size_y = 20
+) in;
+
 //! include chunks/dof-cs.glsl
 
 void main()
diff --git a/ui-opemu.cc b/ui-opemu.cc
index 45d2597..6b99cef 100644
--- a/ui-opemu.cc
+++ b/ui-opemu.cc
@@ -714,8 +714,20 @@ void T_OpContext::run(
 		    {
 			ensureStack( instr , 3 );
 			const auto ss( stack.size( ) );
-			glDispatchCompute( stack[ ss - 1 ].f , stack[ ss - 2 ].f ,
-					stack[ ss - 3 ].f );
+			const auto cpl{ UI::Shaders( ).currentPipeline( ) };
+			const auto pid{ cpl.valid( ) ? cpl.program( E_ShaderType::COMPUTE ) : 0 };
+			if ( pid ) {
+				int32_t wg[ 3 ];
+				glGetProgramiv( pid , GL_COMPUTE_WORK_GROUP_SIZE , wg );
+
+				int32_t ds[ 3 ];
+				for ( auto i = 0u ; i < 3u ; i ++ ) {
+					const auto sv{ uint32_t( stack[ ss - 1 - i ].f ) };
+					const auto m{ sv % wg[ i ] };
+					ds[ i ] = ( sv / wg[ i ] ) + ( m ? 1 : 0 );
+				}
+				glDispatchCompute( ds[ 0 ] , ds[ 1 ] , ds[ 2 ] );
+			}
 			stack.resize( ss - 3 );
 			break;
 		    }
diff --git a/ui-shaders.cc b/ui-shaders.cc
index e5aabe7..c400d02 100644
--- a/ui-shaders.cc
+++ b/ui-shaders.cc
@@ -220,6 +220,7 @@ void T_ShaderPipeline::enable( ) const
 	if ( pl && pl->id ) {
 		glUseProgram( 0 );
 		glBindProgramPipeline( pl->id );
+		UI::Shaders( ).cPipeline_ = id_;
 	}
 }
 
diff --git a/ui-shaders.hh b/ui-shaders.hh
index 21bbf22..5700046 100644
--- a/ui-shaders.hh
+++ b/ui-shaders.hh
@@ -83,6 +83,9 @@ struct T_ShaderManager
 	bool& uiEnabled( )
 		{ return uiEnabled_; }
 
+	T_ShaderPipeline currentPipeline( ) const noexcept
+		{ return T_ShaderPipeline{ cPipeline_ }; }
+
     private:
 	struct T_Pipeline_
 	{
@@ -129,6 +132,8 @@ struct T_ShaderManager
 	T_KeyValueTable< T_String , T_Array< T_String > > missing_;
 	T_Array< T_String > updates_;
 
+	T_String cPipeline_{ };
+
 	// Load/use existing program for use with pipelines
 	void loadProgram(
 			T_String const& pipeline ,