tr_surface: parallel MD5 model CPU runtime computation

Author: illwieckz

src/engine/renderer/tr_surface.cpp

@@ -1058,9 +1058,14 @@ static void Tess_SurfaceMD5( md5Surface_t *srf )
 	shaderVertex_t *tessVertex = tess.verts + tess.numVertexes;
 	shaderVertex_t *lastVertex = tessVertex + srf->numVerts;
 
+	size_t num_chunks = Omp::GetThreads();
+
+	size_t chunk_size = srf->numVerts / num_chunks;
+
 	// Deform the vertices by the lerped bones.
 	if ( tess.skipTangents )
 	{
+if ( !r_parallelNotbntoq.Get() ) {
 		for ( ; tessVertex < lastVertex; tessVertex++,
 			surfaceVertex++ )
 		{
@@ -1084,49 +1089,108 @@ static void Tess_SurfaceMD5( md5Surface_t *srf )
 
 			Vector2Copy( surfaceVertex->texCoords, tessVertex->texCoords );
 		}
+} else {
+		auto task = [&]( size_t chunk_index ) -> void
+		{
+			size_t chunk_offset = chunk_index * chunk_size;
+
+			shaderVertex_t *chunk_tessVertex = tessVertex + chunk_offset;
+			md5Vertex_t *chunk_surfaceVertex = surfaceVertex + chunk_offset;
+
+			shaderVertex_t *chunk_lastVertex =
+				chunk_index == num_chunks - 1
+				? lastVertex
+				: chunk_tessVertex + chunk_size;
+
+			for ( ; chunk_tessVertex < chunk_lastVertex;
+				chunk_tessVertex++,
+				chunk_surfaceVertex++ )
+			{
+				vec3_t position = {};
+
+				float *chunk_boneWeight = chunk_surfaceVertex->boneWeights;
+				float *chunk_lastWeight = chunk_boneWeight + chunk_surfaceVertex->numWeights;
+				uint32_t *chunk_boneIndex = chunk_surfaceVertex->boneIndexes;
+				vec4_t *chunk_surfacePosition = &chunk_surfaceVertex->position;
+
+				for ( ; chunk_boneWeight < chunk_lastWeight;
+					chunk_boneWeight++,
+					chunk_boneIndex++ )
+				{
+					vec3_t tmp;
+
+					TransformPoint( &bones[ *chunk_boneIndex ], *chunk_surfacePosition, tmp );
+					VectorMA( position, *chunk_boneWeight, tmp, position );
+				}
+
+				VectorCopy( position, chunk_tessVertex->xyz );
+
+				Vector2Copy( chunk_surfaceVertex->texCoords, chunk_tessVertex->texCoords );
+			}
+		};
+
+		Omp::Tasker( task, num_chunks );
+}
 	}
 	else
 	{
-		for ( ; tessVertex < lastVertex; tessVertex++,
-			surfaceVertex++ )
+		auto task = [&]( size_t chunk_index ) -> void
 		{
-			vec3_t tangent = {}, binormal = {}, normal = {}, position = {};
+			size_t chunk_offset = chunk_index * chunk_size;
 
-			float *boneWeight = surfaceVertex->boneWeights;
-			float *lastWeight = boneWeight + surfaceVertex->numWeights;
-			uint32_t *boneIndex = surfaceVertex->boneIndexes;
-			vec4_t *surfacePosition = &surfaceVertex->position;
-			vec4_t *surfaceNormal = &surfaceVertex->normal;
-			vec4_t *surfaceTangent = &surfaceVertex->tangent;
-			vec4_t *surfaceBinormal = &surfaceVertex->binormal;
+			shaderVertex_t *chunk_tessVertex = tessVertex + chunk_offset;
+			md5Vertex_t *chunk_surfaceVertex = surfaceVertex + chunk_offset;
 
-			for ( ; boneWeight < lastWeight; boneWeight++,
-				boneIndex++ )
+			shaderVertex_t *chunk_lastVertex =
+				chunk_index == num_chunks - 1
+				? lastVertex
+				: chunk_tessVertex + chunk_size;
+
+			for ( ; chunk_tessVertex < chunk_lastVertex;
+				chunk_tessVertex++,
+				chunk_surfaceVertex++ )
 			{
-				vec3_t tmp;
+				vec3_t tangent = {}, binormal = {}, normal = {}, position = {};
+
+				float *chunk_boneWeight = chunk_surfaceVertex->boneWeights;
+				float *chunk_lastWeight = chunk_boneWeight + chunk_surfaceVertex->numWeights;
+				uint32_t *chunk_boneIndex = chunk_surfaceVertex->boneIndexes;
+				vec4_t *chunk_surfacePosition = &chunk_surfaceVertex->position;
+				vec4_t *chunk_surfaceNormal = &chunk_surfaceVertex->normal;
+				vec4_t *chunk_surfaceTangent = &chunk_surfaceVertex->tangent;
+				vec4_t *chunk_surfaceBinormal = &chunk_surfaceVertex->binormal;
+
+				for ( ; chunk_boneWeight < chunk_lastWeight;
+					chunk_boneWeight++,
+					chunk_boneIndex++ )
+				{
+					vec3_t tmp;
 
-				TransformPoint( &bones[ *boneIndex ], *surfacePosition, tmp );
-				VectorMA( position, *boneWeight, tmp, position );
+					TransformPoint( &bones[ *chunk_boneIndex ], *chunk_surfacePosition, tmp );
+					VectorMA( position, *chunk_boneWeight, tmp, position );
 
-				TransformNormalVector( &bones[ *boneIndex ], *surfaceNormal, tmp );
-				VectorMA( normal, *boneWeight, tmp, normal );
+					TransformNormalVector( &bones[ *chunk_boneIndex ], *chunk_surfaceNormal, tmp );
+					VectorMA( normal, *chunk_boneWeight, tmp, normal );
 
-				TransformNormalVector( &bones[ *boneIndex ], *surfaceTangent, tmp );
-				VectorMA( tangent, *boneWeight, tmp, tangent );
+					TransformNormalVector( &bones[ *chunk_boneIndex ], *chunk_surfaceTangent, tmp );
+					VectorMA( tangent, *chunk_boneWeight, tmp, tangent );
 
-				TransformNormalVector( &bones[ *boneIndex ], *surfaceBinormal, tmp );
-				VectorMA( binormal, *boneWeight, tmp, binormal );
-			}
+					TransformNormalVector( &bones[ *chunk_boneIndex ], *chunk_surfaceBinormal, tmp );
+					VectorMA( binormal, *chunk_boneWeight, tmp, binormal );
+				}
 
-			VectorNormalizeFast( normal );
-			VectorNormalizeFast( tangent );
-			VectorNormalizeFast( binormal );
-			VectorCopy( position, tessVertex->xyz );
+				VectorNormalizeFast( normal );
+				VectorNormalizeFast( tangent );
+				VectorNormalizeFast( binormal );
+				VectorCopy( position, chunk_tessVertex->xyz );
 
-			R_TBNtoQtangentsFast( tangent, binormal, normal, tessVertex->qtangents );
+				R_TBNtoQtangentsFast( tangent, binormal, normal, chunk_tessVertex->qtangents );
 
-			Vector2Copy( surfaceVertex->texCoords, tessVertex->texCoords );
-		}
+				Vector2Copy( chunk_surfaceVertex->texCoords, chunk_tessVertex->texCoords );
+			}
+		};
+
+		Omp::Tasker( task, num_chunks );
 	}
 
 	tess.numIndexes += numIndexes;