From 52892cb7ef21e9bd745b5c9ecf63992245966a08 Mon Sep 17 00:00:00 2001
From: Magnus Norddahl <dpjudas@users.noreply.github.com>
Date: Sat, 24 Dec 2016 01:50:54 +0100
Subject: [PATCH] Move particle drawing to a command to prevent pipeline stalls

---
 src/r_draw_rgba.h    | 17 +++++++++++++++
 src/r_drawt_rgba.cpp | 52 ++++++++++++++++++++++++++++++++++++++++++++
 src/r_things.cpp     | 24 +-------------------
 3 files changed, 70 insertions(+), 23 deletions(-)

diff --git a/src/r_draw_rgba.h b/src/r_draw_rgba.h
index afb3bc22d6..6d16558c06 100644
--- a/src/r_draw_rgba.h
+++ b/src/r_draw_rgba.h
@@ -440,6 +440,23 @@ namespace swrenderer
 		FString DebugInfo() override;
 	};
 
+	/////////////////////////////////////////////////////////////////////////////
+
+	class DrawParticleColumnRGBACommand : public DrawerCommand
+	{
+	public:
+		DrawParticleColumnRGBACommand(uint32_t *dest, int dest_y, int pitch, int count, uint32_t fg, uint32_t alpha);
+		void Execute(DrawerThread *thread) override;
+		FString DebugInfo() override;
+
+	private:
+		uint32_t *_dest;
+		int _pitch;
+		int _count;
+		uint32_t _fg;
+		uint32_t _alpha;
+	};
+
 	/////////////////////////////////////////////////////////////////////////////
 	// Pixel shading inline functions:
 
diff --git a/src/r_drawt_rgba.cpp b/src/r_drawt_rgba.cpp
index 5d3064e5d7..3609956fe4 100644
--- a/src/r_drawt_rgba.cpp
+++ b/src/r_drawt_rgba.cpp
@@ -231,4 +231,56 @@ namespace swrenderer
 	{
 		return "FillColumnHoriz";
 	}
+
+	/////////////////////////////////////////////////////////////////////////////
+
+	DrawParticleColumnRGBACommand::DrawParticleColumnRGBACommand(uint32_t *dest, int dest_y, int pitch, int count, uint32_t fg, uint32_t alpha)
+	{
+		_dest = dest;
+		_pitch = pitch;
+		_count = count;
+		_fg = fg;
+		_alpha = alpha;
+		_dest_y = dest_y;
+	}
+
+	void DrawParticleColumnRGBACommand::Execute(DrawerThread *thread)
+	{
+		int count = thread->count_for_thread(_dest_y, _count);
+		if (count <= 0)
+			return;
+
+		uint32_t *dest = thread->dest_for_thread(_dest_y, _pitch, _dest);
+		int pitch = _pitch * thread->num_cores;
+
+		uint32_t alpha = _alpha;
+		uint32_t inv_alpha = 256 - alpha;
+
+		uint32_t fg_red = (_fg >> 16) & 0xff;
+		uint32_t fg_green = (_fg >> 8) & 0xff;
+		uint32_t fg_blue = _fg & 0xff;
+
+		fg_red *= alpha;
+		fg_green *= alpha;
+		fg_blue *= alpha;
+
+		for (int y = 0; y < count; y++)
+		{
+			uint32_t bg_red = (*dest >> 16) & 0xff;
+			uint32_t bg_green = (*dest >> 8) & 0xff;
+			uint32_t bg_blue = (*dest) & 0xff;
+
+			uint32_t red = (fg_red + bg_red * inv_alpha) / 256;
+			uint32_t green = (fg_green + bg_green * inv_alpha) / 256;
+			uint32_t blue = (fg_blue + bg_blue * inv_alpha) / 256;
+
+			*dest = 0xff000000 | (red << 16) | (green << 8) | blue;
+			dest += pitch;
+		}
+	}
+
+	FString DrawParticleColumnRGBACommand::DebugInfo()
+	{
+		return "DrawParticle";
+	}
 }
diff --git a/src/r_things.cpp b/src/r_things.cpp
index 6e1261c1bb..bbd9eb9cf8 100644
--- a/src/r_things.cpp
+++ b/src/r_things.cpp
@@ -2865,21 +2865,11 @@ void R_DrawParticle_rgba(vissprite_t *vis)
 
 	R_DrawMaskedSegsBehindParticle(vis);
 	
-	DrawerCommandQueue::WaitForWorkers();
-
 	uint32_t fg = LightBgra::shade_pal_index_simple(color, LightBgra::calc_light_multiplier(LIGHTSCALE(0, vis->Style.ColormapNum << FRACBITS)));
-	uint32_t fg_red = (fg >> 16) & 0xff;
-	uint32_t fg_green = (fg >> 8) & 0xff;
-	uint32_t fg_blue = fg & 0xff;
 
 	// vis->renderflags holds translucency level (0-255)
 	fixed_t fglevel = ((vis->renderflags + 1) << 8) & ~0x3ff;
 	uint32_t alpha = fglevel * 256 / FRACUNIT;
-	uint32_t inv_alpha = 256 - alpha;
-
-	fg_red *= alpha;
-	fg_green *= alpha;
-	fg_blue *= alpha;
 
 	spacing = RenderTarget->GetPitch();
 
@@ -2889,19 +2879,7 @@ void R_DrawParticle_rgba(vissprite_t *vis)
 		if (R_ClipSpriteColumnWithPortals(vis))
 			continue;
 		dest = ylookup[yl] + x + (uint32_t*)dc_destorg;
-		for (int y = 0; y < ycount; y++)
-		{
-			uint32_t bg_red = (*dest >> 16) & 0xff;
-			uint32_t bg_green = (*dest >> 8) & 0xff;
-			uint32_t bg_blue = (*dest) & 0xff;
-
-			uint32_t red = (fg_red + bg_red * inv_alpha) / 256;
-			uint32_t green = (fg_green + bg_green * inv_alpha) / 256;
-			uint32_t blue = (fg_blue + bg_blue * inv_alpha) / 256;
-
-			*dest = 0xff000000 | (red << 16) | (green << 8) | blue;
-			dest += spacing;
-		}
+		DrawerCommandQueue::QueueCommand<DrawParticleColumnRGBACommand>(dest, yl, spacing, ycount, fg, alpha);
 	}
 }