- Fix voxel speed regression in the palette software renderer

src/swrenderer/drawers/r_draw.h

@@ -28,6 +28,7 @@ namespace swrenderer
 	class WallDrawerArgs;
 	class SpanDrawerArgs;
 	class SpriteDrawerArgs;
+	class VoxelBlock;
 
 	extern uint8_t shadetables[/*NUMCOLORMAPS*16*256*/];
 	extern FDynamicColormap ShadeFakeColormap[16];
@@ -79,6 +80,7 @@ namespace swrenderer
 		virtual void DrawSubClampTranslatedColumn(const SpriteDrawerArgs &args) = 0;
 		virtual void DrawRevSubClampColumn(const SpriteDrawerArgs &args) = 0;
 		virtual void DrawRevSubClampTranslatedColumn(const SpriteDrawerArgs &args) = 0;
+		virtual void DrawVoxelBlocks(const SpriteDrawerArgs &args, const VoxelBlock *blocks, int blockcount) = 0;
 		virtual void DrawSpan(const SpanDrawerArgs &args) = 0;
 		virtual void DrawSpanMasked(const SpanDrawerArgs &args) = 0;
 		virtual void DrawSpanTranslucent(const SpanDrawerArgs &args) = 0;

src/swrenderer/drawers/r_draw_pal.cpp

@@ -3011,4 +3011,107 @@ namespace swrenderer
 	{
 		return "DrawParticle";
 	}
+
+	/////////////////////////////////////////////////////////////////////////////
+
+	DrawVoxelBlocksPalCommand::DrawVoxelBlocksPalCommand(const SpriteDrawerArgs &args, const VoxelBlock *blocks, int blockcount) : args(args), blocks(blocks), blockcount(blockcount)
+	{
+	}
+
+	void DrawVoxelBlocksPalCommand::Execute(DrawerThread *thread)
+	{
+		int destpitch = args.Viewport()->RenderTarget->GetPitch();
+		uint8_t *destorig = args.Viewport()->RenderTarget->GetBuffer();
+		const uint8_t *colormap = args.Colormap(args.Viewport());
+
+		for (int i = 0; i < blockcount; i++)
+		{
+			const VoxelBlock &block = blocks[i];
+
+			const uint8_t *source = block.voxels;
+
+			fixed_t fracpos = block.vPos;
+			fixed_t iscale = block.vStep;
+			int count = block.height;
+			int width = block.width;
+			int pitch = destpitch;
+			uint8_t *dest = destorig + (block.x + block.y * pitch);
+
+			count = thread->count_for_thread(block.y, count);
+			dest = thread->dest_for_thread(block.y, pitch, dest);
+			fracpos += iscale * thread->skipped_by_thread(block.y);
+			iscale *= thread->num_cores;
+			pitch *= thread->num_cores;
+
+			if (width == 1)
+			{
+				while (count > 0)
+				{
+					uint8_t color = colormap[source[fracpos >> FRACBITS]];
+					*dest = color;
+					dest += pitch;
+					fracpos += iscale;
+					count--;
+				}
+			}
+			else if (width == 2)
+			{
+				while (count > 0)
+				{
+					uint8_t color = colormap[source[fracpos >> FRACBITS]];
+					dest[0] = color;
+					dest[1] = color;
+					dest += pitch;
+					fracpos += iscale;
+					count--;
+				}
+			}
+			else if (width == 3)
+			{
+				while (count > 0)
+				{
+					uint8_t color = colormap[source[fracpos >> FRACBITS]];
+					dest[0] = color;
+					dest[1] = color;
+					dest[2] = color;
+					dest += pitch;
+					fracpos += iscale;
+					count--;
+				}
+			}
+			else if (width == 4)
+			{
+				while (count > 0)
+				{
+					uint8_t color = colormap[source[fracpos >> FRACBITS]];
+					dest[0] = color;
+					dest[1] = color;
+					dest[2] = color;
+					dest[3] = color;
+					dest += pitch;
+					fracpos += iscale;
+					count--;
+				}
+			}
+			else
+			{
+				while (count > 0)
+				{
+					uint8_t color = colormap[source[fracpos >> FRACBITS]];
+
+					for (int x = 0; x < width; x++)
+						dest[x] = color;
+
+					dest += pitch;
+					fracpos += iscale;
+					count--;
+				}
+			}
+		}
+	}
+
+	FString DrawVoxelBlocksPalCommand::DebugInfo()
+	{
+		return "DrawVoxelBlocks";
+	}
 }

src/swrenderer/drawers/r_draw_pal.h

@@ -49,10 +49,10 @@ namespace swrenderer
 		PalColumnCommand(const SpriteDrawerArgs &args);
 		FString DebugInfo() override { return "PalColumnCommand"; }
 
+		SpriteDrawerArgs args;
+
 	protected:
 		uint8_t AddLights(uint8_t fg, uint8_t material, uint32_t lit_r, uint32_t lit_g, uint32_t lit_b);
-		
-		SpriteDrawerArgs args;
 	};
 
 	class DrawColumnPalCommand : public PalColumnCommand { public: using PalColumnCommand::PalColumnCommand; void Execute(DrawerThread *thread) override; };
@@ -205,6 +205,19 @@ namespace swrenderer
 		uint32_t _fracposx;
 	};
 
+	class DrawVoxelBlocksPalCommand : public DrawerCommand
+	{
+	public:
+		DrawVoxelBlocksPalCommand(const SpriteDrawerArgs &args, const VoxelBlock *blocks, int blockcount);
+		void Execute(DrawerThread *thread) override;
+		FString DebugInfo() override;
+
+	private:
+		SpriteDrawerArgs args;
+		const VoxelBlock *blocks;
+		int blockcount;
+	};
+
 	class SWPalDrawers : public SWPixelFormatDrawers
 	{
 	public:
@@ -244,6 +257,7 @@ namespace swrenderer
 		void DrawSubClampTranslatedColumn(const SpriteDrawerArgs &args) override { Queue->Push<DrawColumnSubClampTranslatedPalCommand>(args); }
 		void DrawRevSubClampColumn(const SpriteDrawerArgs &args) override { Queue->Push<DrawColumnRevSubClampPalCommand>(args); }
 		void DrawRevSubClampTranslatedColumn(const SpriteDrawerArgs &args) override { Queue->Push<DrawColumnRevSubClampTranslatedPalCommand>(args); }
+		void DrawVoxelBlocks(const SpriteDrawerArgs &args, const VoxelBlock *blocks, int blockcount) override { Queue->Push<DrawVoxelBlocksPalCommand>(args, blocks, blockcount); }
 		void DrawSpan(const SpanDrawerArgs &args) override { Queue->Push<DrawSpanPalCommand>(args); }
 		void DrawSpanMasked(const SpanDrawerArgs &args) override { Queue->Push<DrawSpanMaskedPalCommand>(args); }
 		void DrawSpanTranslucent(const SpanDrawerArgs &args) override { Queue->Push<DrawSpanTranslucentPalCommand>(args); }

src/swrenderer/drawers/r_draw_rgba.cpp

@@ -201,6 +201,11 @@ namespace swrenderer
 		Queue->Push<DrawSpriteTranslatedRevSubClamp32Command>(args);
 	}
 
+	void SWTruecolorDrawers::DrawVoxelBlocks(const SpriteDrawerArgs &args, const VoxelBlock *blocks, int blockcount)
+	{
+		Queue->Push<DrawVoxelBlocksRGBACommand>(args, blocks, blockcount);
+	}
+
 	void SWTruecolorDrawers::DrawSpan(const SpanDrawerArgs &args)
 	{
 		Queue->Push<DrawSpan32Command>(args);
@@ -858,4 +863,44 @@ namespace swrenderer
 		return "DrawParticle";
 	}
 
+	/////////////////////////////////////////////////////////////////////////////
+
+	DrawVoxelBlocksRGBACommand::DrawVoxelBlocksRGBACommand(const SpriteDrawerArgs &args, const VoxelBlock *blocks, int blockcount) : args(args), blocks(blocks), blockcount(blockcount)
+	{
+	}
+
+	void DrawVoxelBlocksRGBACommand::Execute(DrawerThread *thread)
+	{
+		int pitch = args.Viewport()->RenderTarget->GetPitch();
+		uint8_t *destorig = args.Viewport()->RenderTarget->GetBuffer();
+
+		DrawSprite32Command drawer(args);
+		drawer.args.dc_texturefracx = 0;
+		drawer.args.dc_source2 = 0;
+		for (int i = 0; i < blockcount; i++)
+		{
+			const VoxelBlock &block = blocks[i];
+
+			double v = block.vPos / (double)block.voxelsCount / FRACUNIT;
+			double vstep = block.vStep / (double)block.voxelsCount / FRACUNIT;
+			drawer.args.dc_texturefrac = (int)(v * (1 << 30));
+			drawer.args.dc_iscale = (int)(vstep * (1 << 30));
+			drawer.args.dc_source = block.voxels;
+			drawer.args.dc_textureheight = block.voxelsCount;
+			drawer.args.dc_count = block.height;
+			drawer.args.dc_dest_y = block.y;
+			drawer.args.dc_dest = destorig + (block.x + block.y * pitch) * 4;
+
+			for (int j = 0; j < block.width; j++)
+			{
+				drawer.Execute(thread);
+				drawer.args.dc_dest += 4;
+			}
+		}
+	}
+
+	FString DrawVoxelBlocksRGBACommand::DebugInfo()
+	{
+		return "DrawVoxelBlocks";
+	}
 }

src/swrenderer/drawers/r_draw_rgba.h

@@ -215,6 +215,21 @@ namespace swrenderer
 
 	/////////////////////////////////////////////////////////////////////////////
 
+	class DrawVoxelBlocksRGBACommand : public DrawerCommand
+	{
+	public:
+		DrawVoxelBlocksRGBACommand(const SpriteDrawerArgs &args, const VoxelBlock *blocks, int blockcount);
+		void Execute(DrawerThread *thread) override;
+		FString DebugInfo() override;
+
+	private:
+		SpriteDrawerArgs args;
+		const VoxelBlock *blocks;
+		int blockcount;
+	};
+
+	/////////////////////////////////////////////////////////////////////////////
+
 	class SWTruecolorDrawers : public SWPixelFormatDrawers
 	{
 	public:
@@ -246,6 +261,7 @@ namespace swrenderer
 		void DrawSubClampTranslatedColumn(const SpriteDrawerArgs &args) override;
 		void DrawRevSubClampColumn(const SpriteDrawerArgs &args) override;
 		void DrawRevSubClampTranslatedColumn(const SpriteDrawerArgs &args) override;
+		void DrawVoxelBlocks(const SpriteDrawerArgs &args, const VoxelBlock *blocks, int blockcount) override;
 		void DrawSpan(const SpanDrawerArgs &args) override;
 		void DrawSpanMasked(const SpanDrawerArgs &args) override;
 		void DrawSpanTranslucent(const SpanDrawerArgs &args) override;

src/swrenderer/drawers/r_draw_sprite32.h

@@ -56,10 +56,9 @@ namespace swrenderer
 	template<typename BlendT, typename SamplerT>
 	class DrawSprite32T : public DrawerCommand
 	{
-	protected:
+	public:
 		SpriteDrawerArgs args;
 
-	public:
 		DrawSprite32T(const SpriteDrawerArgs &drawerargs) : args(drawerargs) { }
 
 		void Execute(DrawerThread *thread) override

src/swrenderer/drawers/r_draw_sprite32_sse2.h

@@ -56,10 +56,9 @@ namespace swrenderer
 	template<typename BlendT, typename SamplerT>
 	class DrawSprite32T : public DrawerCommand
 	{
-	protected:
+	public:
 		SpriteDrawerArgs args;
 
-	public:
 		DrawSprite32T(const SpriteDrawerArgs &drawerargs) : args(drawerargs) { }
 
 		void Execute(DrawerThread *thread) override

src/swrenderer/things/r_voxel.cpp

@@ -220,11 +220,12 @@ namespace swrenderer
 		{
 			CheckOffscreenBuffer(viewport->RenderTarget->GetWidth(), viewport->RenderTarget->GetHeight(), !!(flags & DVF_SPANSONLY));
 		}
+		*/
+
 		if (spr->bInMirror)
 		{
 			flags |= DVF_MIRRORED;
 		}
-		*/
 
 		// Render the voxel, either directly to the screen or offscreen.
 		DrawVoxel(thread, drawerargs, spr->pa.vpos, spr->pa.vang, spr->gpos, spr->Angle,
@@ -276,18 +277,18 @@ namespace swrenderer
 		fixed_t cosang, sinang, sprcosang, sprsinang;
 		int backx, backy, gxinc, gyinc;
 		int daxscalerecip, dayscalerecip, cnt, gxstart, gystart, dazscale;
-		int lx, rx, nx, ny, x1=0, y1=0, x2=0, y2=0, yinc=0;
-		int yoff, xs=0, ys=0, xe, ye, xi=0, yi=0, cbackx, cbacky, dagxinc, dagyinc;
+		int lx, rx, nx, ny, x1 = 0, y1 = 0, x2 = 0, y2 = 0, yinc = 0;
+		int yoff, xs = 0, ys = 0, xe, ye, xi = 0, yi = 0, cbackx, cbacky, dagxinc, dagyinc;
 		kvxslab_t *voxptr, *voxend;
 		FVoxelMipLevel *mip;
 		int z1a[64], z2a[64], yplc[64];
 
 		auto viewport = thread->Viewport.get();
 
-		const int nytooclose = viewport->viewwindow.centerxwide * 2100, nytoofar = 32768*32768 - 1048576;
+		const int nytooclose = viewport->viewwindow.centerxwide * 2100, nytoofar = 32768 * 32768 - 1048576;
 		const int xdimenscale = FLOAT2FIXED(viewport->viewwindow.centerxwide * viewport->YaspectMul / 160);
 		const double centerxwide_f = viewport->viewwindow.centerxwide;
-		const double centerxwidebig_f = centerxwide_f * 65536*65536*8;
+		const double centerxwidebig_f = centerxwide_f * 65536 * 65536 * 8;
 
 		// Convert to Build's coordinate system.
 		fixed_t globalposx = xs_Fix<4>::ToFix(globalpos.X);
@@ -329,14 +330,14 @@ namespace swrenderer
 		minslabz >>= k;
 		maxslabz >>= k;
 
-		daxscale <<= (k+8); dayscale <<= (k+8);
+		daxscale <<= (k + 8); dayscale <<= (k + 8);
 		dazscale = FixedDiv(dayscale, FLOAT2FIXED(viewport->BaseYaspectMul));
 		daxscale = fixed_t(daxscale / viewport->YaspectMul);
 		daxscale = Scale(daxscale, xdimenscale, viewport->viewwindow.centerxwide << 9);
 		dayscale = Scale(dayscale, FixedMul(xdimenscale, viewport->viewingrangerecip), viewport->viewwindow.centerxwide << 9);
 
-		daxscalerecip = (1<<30) / daxscale;
-		dayscalerecip = (1<<30) / dayscale;
+		daxscalerecip = (1 << 30) / daxscale;
+		dayscalerecip = (1 << 30) / dayscale;
 
 		fixed_t piv_x = fixed_t(mip->Pivot.X*256.);
 		fixed_t piv_y = fixed_t(mip->Pivot.Y*256.);
@@ -376,18 +377,24 @@ namespace swrenderer
 		syoff = DivScale21(globalposz - dasprz, FixedMul(dazscale, 0xE800)) + (piv_z << 7);
 		yoff = (abs(gxinc) + abs(gyinc)) >> 1;
 
+		bool useSlabDataBgra = !drawerargs.DrawerNeedsPalInput() && viewport->RenderTarget->IsBgra();
+
+		const int maxoutblocks = 100;
+		VoxelBlock *outblocks = thread->FrameMemory->AllocMemory<VoxelBlock>(maxoutblocks);
+		int nextoutblock = 0;
+
 		for (cnt = 0; cnt < 8; cnt++)
 		{
 			switch (cnt)
 			{
 			case 0: xs = 0;				ys = 0;				xi = 1; yi = 1; break;
-				case 1: xs = mip->SizeX-1;	ys = 0;				xi = -1; yi =  1; break;
-				case 2: xs = 0;				ys = mip->SizeY-1;	xi =  1; yi = -1; break;
-				case 3: xs = mip->SizeX-1;	ys = mip->SizeY-1;	xi = -1; yi = -1; break;
+			case 1: xs = mip->SizeX - 1;	ys = 0;				xi = -1; yi = 1; break;
+			case 2: xs = 0;				ys = mip->SizeY - 1;	xi = 1; yi = -1; break;
+			case 3: xs = mip->SizeX - 1;	ys = mip->SizeY - 1;	xi = -1; yi = -1; break;
 			case 4: xs = 0;				ys = cbacky;		xi = 1; yi = 2; break;
-				case 5: xs = mip->SizeX-1;	ys = cbacky;		xi = -1; yi =  2; break;
+			case 5: xs = mip->SizeX - 1;	ys = cbacky;		xi = -1; yi = 2; break;
 			case 6: xs = cbackx;		ys = 0;				xi = 2; yi = 1; break;
-				case 7: xs = cbackx;		ys = mip->SizeY-1;	xi =  2; yi = -1; break;
+			case 7: xs = cbackx;		ys = mip->SizeY - 1;	xi = 2; yi = -1; break;
 			}
 			xe = cbackx; ye = cbacky;
 			if (cnt < 4)
@@ -407,21 +414,21 @@ namespace swrenderer
 			}
 
 			i = sgn(ys - backy) + sgn(xs - backx) * 3 + 4;
-			switch(i)
+			switch (i)
 			{
 			case 6: case 7: x1 = 0;				y1 = 0;				break;
 			case 8: case 5: x1 = gxinc;			y1 = gyinc;			break;
 			case 0: case 3: x1 = gyinc;			y1 = -gxinc;		break;
-				case 2: case 1: x1 = gxinc+gyinc;	y1 = gyinc-gxinc;	break;
+			case 2: case 1: x1 = gxinc + gyinc;	y1 = gyinc - gxinc;	break;
 			}
-			switch(i)
+			switch (i)
 			{
 			case 2: case 5: x2 = 0;				y2 = 0;				break;
 			case 0: case 1: x2 = gxinc;			y2 = gyinc;			break;
 			case 8: case 7: x2 = gyinc;			y2 = -gxinc;		break;
-				case 6: case 3: x2 = gxinc+gyinc;	y2 = gyinc-gxinc;	break;
+			case 6: case 3: x2 = gxinc + gyinc;	y2 = gyinc - gxinc;	break;
 			}
-			uint8_t oand = (1 << int(xs<backx)) + (1 << (int(ys<backy)+2));
+			uint8_t oand = (1 << int(xs < backx)) + (1 << (int(ys < backy) + 2));
 			uint8_t oand16 = oand + 16;
 			uint8_t oand32 = oand + 32;
 
@@ -446,7 +453,7 @@ namespace swrenderer
 				{
 					if ((ny <= nytooclose) || (ny >= nytoofar)) continue;
 					voxptr = (kvxslab_t *)(slabxoffs + xyoffs[y]);
-					voxend = (kvxslab_t *)(slabxoffs + xyoffs[y+1]);
+					voxend = (kvxslab_t *)(slabxoffs + xyoffs[y + 1]);
 					if (voxptr >= voxend) continue;
 
 					lx = xs_RoundToInt(nx * centerxwide_f / (ny + y1)) + viewport->viewwindow.centerx;
@@ -516,7 +523,7 @@ namespace swrenderer
 						}
 						else
 						{
-							if (z2-z1 >= 1024) yinc = FixedDiv(zleng, z2 - z1);
+							if (z2 - z1 >= 1024) yinc = FixedDiv(zleng, z2 - z1);
 							else yinc = (((1 << 24) - 1) / (z2 - z1)) * zleng >> 8;
 						}
 						// [RH] Clip each column separately, not just by the first one.
@@ -549,7 +556,7 @@ namespace swrenderer
 							}
 
 							const uint8_t *columnColors = col;
-							if (!drawerargs.DrawerNeedsPalInput() && viewport->RenderTarget->IsBgra())
+							if (useSlabDataBgra)
 							{
 								// The true color slab data array is identical, except its using uint32_t instead of uint8.
 								//
@@ -575,12 +582,31 @@ namespace swrenderer
 										break;
 								}
 
+								outblocks[nextoutblock].x = lxt + xxl;
+								outblocks[nextoutblock].y = z1;
+								outblocks[nextoutblock].width = xxr - xxl;
+								outblocks[nextoutblock].height = z2 - z1;
+								outblocks[nextoutblock].vPos = yplc[xxl];
+								outblocks[nextoutblock].vStep = yinc;
+								outblocks[nextoutblock].voxels = columnColors;
+								outblocks[nextoutblock].voxelsCount = zleng;
+								nextoutblock++;
+
+								if (nextoutblock == maxoutblocks)
+								{
+									drawerargs.DrawVoxelBlocks(thread, outblocks, nextoutblock);
+									outblocks = thread->FrameMemory->AllocMemory<VoxelBlock>(maxoutblocks);
+									nextoutblock = 0;
+								}
+
+								/*
 								for (int x = xxl; x < xxr; ++x)
 								{
 									drawerargs.SetDest(viewport, lxt + x, z1);
 									drawerargs.SetCount(z2 - z1);
 									drawerargs.DrawVoxelColumn(thread, yplc[xxl], yinc, columnColors, zleng);
 								}
+								*/
 
 								/*
 								if (!(flags & DVF_OFFSCREEN))
@@ -618,6 +644,11 @@ namespace swrenderer
 				}
 			}
 		}
+
+		if (nextoutblock != 0)
+		{
+			drawerargs.DrawVoxelBlocks(thread, outblocks, nextoutblock);
+		}
 	}
 
 	kvxslab_t *RenderVoxel::GetSlabStart(const FVoxelMipLevel &mip, int x, int y)

src/swrenderer/viewport/r_spritedrawer.cpp

@@ -494,8 +494,11 @@ namespace swrenderer
 		thread->Drawers(dc_viewport)->FillColumn(*this);
 	}
 
-	void SpriteDrawerArgs::DrawVoxelColumn(RenderThread *thread, fixed_t vPos, fixed_t vStep, const uint8_t *voxels, int voxelsCount)
+	void SpriteDrawerArgs::DrawVoxelBlocks(RenderThread *thread, const VoxelBlock *blocks, int blockcount)
 	{
+		SetDest(thread->Viewport.get(), 0, 0);
+		thread->Drawers(dc_viewport)->DrawVoxelBlocks(*this, blocks, blockcount);
+#if 0
 		if (dc_viewport->RenderTarget->IsBgra())
 		{
 			double v = vPos / (double)voxelsCount / FRACUNIT;
@@ -514,6 +517,7 @@ namespace swrenderer
 		dc_source2 = 0;
 		dc_textureheight = voxelsCount;
 		(thread->Drawers(dc_viewport)->*colfunc)(*this);
+#endif
 	}
 
 	void SpriteDrawerArgs::SetDest(RenderViewport *viewport, int x, int y)

src/swrenderer/viewport/r_spritedrawer.h

@@ -10,6 +10,17 @@ namespace swrenderer
 {
 	class RenderThread;
 	
+	class VoxelBlock
+	{
+	public:
+		int x, y;
+		uint16_t width, height;
+		fixed_t vPos;
+		fixed_t vStep;
+		const uint8_t *voxels;
+		int voxelsCount;
+	};
+
 	class SpriteDrawerArgs : public DrawerArgs
 	{
 	public:
@@ -24,7 +35,7 @@ namespace swrenderer
 
 		void DrawMaskedColumn(RenderThread *thread, int x, fixed_t iscale, FTexture *texture, fixed_t column, double spryscale, double sprtopscreen, bool sprflipvert, const short *mfloorclip, const short *mceilingclip, bool unmasked = false);
 		void FillColumn(RenderThread *thread);
-		void DrawVoxelColumn(RenderThread *thread, fixed_t vPos, fixed_t vStep, const uint8_t *voxels, int voxelsCount);
+		void DrawVoxelBlocks(RenderThread *thread, const VoxelBlock *blocks, int blockcount);
 
 		uint8_t *Dest() const { return dc_dest; }
 		int DestY() const { return dc_dest_y; }
@@ -95,5 +106,8 @@ namespace swrenderer
 		SpriteDrawerFunc colfunc;
 
 		RenderViewport *dc_viewport = nullptr;
+
+		friend class DrawVoxelBlocksRGBACommand;
+		friend class DrawVoxelBlocksPalCommand;
 	};
 }