- fix sky drawers not staying within their numa node

src/swrenderer/drawers/r_draw_pal.cpp

@@ -551,7 +551,6 @@ namespace swrenderer
 	void DrawSingleSky1PalCommand::Execute(DrawerThread *thread)
 	{
 		uint8_t *dest = args.Dest();
-		int count = args.Count();
 		int pitch = args.Viewport()->RenderTarget->GetPitch();
 		const uint8_t *source0 = args.FrontTexturePixels();
 		int textureheight0 = args.FrontTextureHeight();
@@ -559,6 +558,25 @@ namespace swrenderer
 		int32_t frac = args.TextureVPos();
 		int32_t fracstep = args.TextureVStep();
 
+		if (!args.FadeSky())
+		{
+			int count = thread->count_for_thread(args.DestY(), args.Count());
+
+			for (int index = 0; index < count; index++)
+			{
+				uint32_t sample_index = (((((uint32_t)frac) << 8) >> FRACBITS) * textureheight0) >> FRACBITS;
+				*dest = source0[sample_index];
+				dest += pitch;
+				frac += fracstep;
+			}
+
+			return;
+		}
+
+		int num_cores = thread->num_cores;
+		int skipped = thread->skipped_by_thread(args.DestY());
+		int count = skipped + thread->count_for_thread(args.DestY(), args.Count()) * num_cores;
+
 		// Find bands for top solid color, top fade, center textured, bottom fade, bottom solid color:
 		int start_fade = 2; // How fast it should fade out
 		int fade_length = (1 << (24 - start_fade));
@@ -571,28 +589,11 @@ namespace swrenderer
 		start_fadebottom_y = clamp(start_fadebottom_y, 0, count);
 		end_fadebottom_y = clamp(end_fadebottom_y, 0, count);
 
-		int num_cores = thread->num_cores;
-		int skipped = thread->skipped_by_thread(args.DestY());
 		dest = thread->dest_for_thread(args.DestY(), pitch, dest);
 		frac += fracstep * skipped;
 		fracstep *= num_cores;
 		pitch *= num_cores;
 
-		if (!args.FadeSky())
-		{
-			count = thread->count_for_thread(args.DestY(), count);
-
-			for (int index = 0; index < count; index++)
-			{
-				uint32_t sample_index = (((((uint32_t)frac) << 8) >> FRACBITS) * textureheight0) >> FRACBITS;
-				*dest = source0[sample_index];
-				dest += pitch;
-				frac += fracstep;
-			}
-
-			return;
-		}
-
 		uint32_t solid_top = args.SolidTopColor();
 		uint32_t solid_bottom = args.SolidBottomColor();
 

src/swrenderer/drawers/r_draw_sky32.h

@@ -39,7 +39,6 @@ namespace swrenderer
 		void Execute(DrawerThread *thread) override
 		{
 			uint32_t *dest = (uint32_t *)args.Dest();
-			int count = args.Count();
 			int pitch = args.Viewport()->RenderTarget->GetPitch();
 			const uint32_t *source0 = (const uint32_t *)args.FrontTexturePixels();
 			int textureheight0 = args.FrontTextureHeight();
@@ -51,6 +50,25 @@ namespace swrenderer
 			uint32_t solid_bottom = args.SolidBottomColor();
 			bool fadeSky = args.FadeSky();
 
+			if (!fadeSky)
+			{
+				int count = thread->count_for_thread(args.DestY(), args.Count());
+
+				for (int index = 0; index < count; index++)
+				{
+					uint32_t sample_index = (((((uint32_t)frac) << 8) >> FRACBITS) * textureheight0) >> FRACBITS;
+					*dest = source0[sample_index];
+					dest += pitch;
+					frac += fracstep;
+				}
+
+				return;
+			}
+
+			int num_cores = thread->num_cores;
+			int skipped = thread->skipped_by_thread(args.DestY());
+			int count = skipped + thread->count_for_thread(args.DestY(), args.Count()) * num_cores;
+
 			// Find bands for top solid color, top fade, center textured, bottom fade, bottom solid color:
 			int start_fade = 2; // How fast it should fade out
 			int fade_length = (1 << (24 - start_fade));
@@ -63,28 +81,11 @@ namespace swrenderer
 			start_fadebottom_y = clamp(start_fadebottom_y, 0, count);
 			end_fadebottom_y = clamp(end_fadebottom_y, 0, count);
 
-			int num_cores = thread->num_cores;
-			int skipped = thread->skipped_by_thread(args.DestY());
 			dest = thread->dest_for_thread(args.DestY(), pitch, dest);
 			frac += fracstep * skipped;
 			fracstep *= num_cores;
 			pitch *= num_cores;
 
-			if (!fadeSky)
-			{
-				count = thread->count_for_thread(args.DestY(), count);
-
-				for (int index = 0; index < count; index++)
-				{
-					uint32_t sample_index = (((((uint32_t)frac) << 8) >> FRACBITS) * textureheight0) >> FRACBITS;
-					*dest = source0[sample_index];
-					dest += pitch;
-					frac += fracstep;
-				}
-
-				return;
-			}
-
 			BgraColor solid_top_fill = solid_top;
 			BgraColor solid_bottom_fill = solid_bottom;
 

src/swrenderer/drawers/r_draw_sky32_sse2.h

@@ -38,7 +38,6 @@ namespace swrenderer
 		void Execute(DrawerThread *thread) override
 		{
 			uint32_t *dest = (uint32_t *)args.Dest();
-			int count = args.Count();
 			int pitch = args.Viewport()->RenderTarget->GetPitch();
 			const uint32_t *source0 = (const uint32_t *)args.FrontTexturePixels();
 			int textureheight0 = args.FrontTextureHeight();
@@ -50,6 +49,25 @@ namespace swrenderer
 			uint32_t solid_bottom = args.SolidBottomColor();
 			bool fadeSky = args.FadeSky();
 
+			if (!fadeSky)
+			{
+				int count = thread->count_for_thread(args.DestY(), args.Count());
+
+				for (int index = 0; index < count; index++)
+				{
+					uint32_t sample_index = (((((uint32_t)frac) << 8) >> FRACBITS) * textureheight0) >> FRACBITS;
+					*dest = source0[sample_index];
+					dest += pitch;
+					frac += fracstep;
+				}
+
+				return;
+			}
+
+			int num_cores = thread->num_cores;
+			int skipped = thread->skipped_by_thread(args.DestY());
+			int count = skipped + thread->count_for_thread(args.DestY(), args.Count()) * num_cores;
+
 			// Find bands for top solid color, top fade, center textured, bottom fade, bottom solid color:
 			int start_fade = 2; // How fast it should fade out
 			int fade_length = (1 << (24 - start_fade));
@@ -62,28 +80,11 @@ namespace swrenderer
 			start_fadebottom_y = clamp(start_fadebottom_y, 0, count);
 			end_fadebottom_y = clamp(end_fadebottom_y, 0, count);
 
-			int num_cores = thread->num_cores;
-			int skipped = thread->skipped_by_thread(args.DestY());
 			dest = thread->dest_for_thread(args.DestY(), pitch, dest);
 			frac += fracstep * skipped;
 			fracstep *= num_cores;
 			pitch *= num_cores;
 
-			if (!fadeSky)
-			{
-				count = thread->count_for_thread(args.DestY(), count);
-
-				for (int index = 0; index < count; index++)
-				{
-					uint32_t sample_index = (((((uint32_t)frac) << 8) >> FRACBITS) * textureheight0) >> FRACBITS;
-					*dest = source0[sample_index];
-					dest += pitch;
-					frac += fracstep;
-				}
-
-				return;
-			}
-
 			__m128i solid_top_fill = _mm_unpacklo_epi8(_mm_cvtsi32_si128(solid_top), _mm_setzero_si128());
 			__m128i solid_bottom_fill = _mm_unpacklo_epi8(_mm_cvtsi32_si128(solid_bottom), _mm_setzero_si128());
 

src/swrenderer/drawers/r_thread.cpp

@@ -140,6 +140,8 @@ void DrawerThreads::WorkerMain(DrawerThread *thread)
 		// Grab the commands
 		DrawerCommandQueuePtr list = active_commands[thread->current_queue];
 		thread->current_queue++;
+		thread->numa_start_y = thread->numa_node * viewheight / thread->num_numa_nodes;
+		thread->numa_end_y = (thread->numa_node + 1) * viewheight / thread->num_numa_nodes;
 		start_lock.unlock();
 
 		// Do the work:
@@ -206,8 +208,6 @@ void DrawerThreads::StartThreads()
 					thread->num_cores = I_GetNumaNodeThreadCount(numaNode);
 					thread->numa_node = numaNode;
 					thread->num_numa_nodes = I_GetNumaNodeCount();
-					thread->numa_start_y = numaNode * viewheight / I_GetNumaNodeCount();
-					thread->numa_end_y = (numaNode + 1) * viewheight / I_GetNumaNodeCount();
 					thread->thread = std::thread([=]() { queue->WorkerMain(thread); });
 					I_SetThreadNumaNode(thread->thread, numaNode);
 				}
@@ -223,8 +223,6 @@ void DrawerThreads::StartThreads()
 				thread->num_cores = num_threads;
 				thread->numa_node = 0;
 				thread->num_numa_nodes = 1;
-				thread->numa_start_y = 0;
-				thread->numa_end_y = viewheight;
 				thread->thread = std::thread([=]() { queue->WorkerMain(thread); });
 				I_SetThreadNumaNode(thread->thread, 0);
 			}
@@ -288,7 +286,4 @@ void MemcpyCommand::Execute(DrawerThread *thread)
 		d += dstep;
 		s += sstep;
 	}
-
-	thread->numa_start_y = thread->numa_node * viewheight / thread->num_numa_nodes;
-	thread->numa_end_y = (thread->numa_node + 1) * viewheight / thread->num_numa_nodes;
 }