Implement pre-vertex-shader backface culling on the GPU.

Decreases frametime by ~16% on my hardware.
Not as much as I hoped, but still worth a lot.

assets/cubyz/shaders/chunks/chunk_vertex.vs

@@ -42,7 +42,7 @@ struct ChunkData {
 	int visibilityMask;
 	int voxelSize;
 	uint vertexStartOpaque;
-	uint vertexCountOpaque;
+	uint faceCountsByNormalOpaque[7];
 	uint vertexStartTransparent;
 	uint vertexCountTransparent;
 };

assets/cubyz/shaders/chunks/fillIndirectBuffer.glsl

@@ -12,7 +12,7 @@ struct ChunkData {
 	int visibilityMask;
 	int voxelSize;
 	uint vertexStartOpaque;
-	uint vertexCountOpaque;
+	uint faceCountsByNormalOpaque[7];
 	uint vertexStartTransparent;
 	uint vertexCountTransparent;
 };
@@ -40,26 +40,62 @@ uniform uint chunkIDIndex;
 uniform uint commandIndexStart;
 uniform uint size;
 uniform bool isTransparent;
+uniform ivec3 playerPositionInteger;
+
+bool isVisible(int dir, ivec3 relativePlayerPos, int voxelSize) {
+	switch(dir) {
+	case 0: // dirUp
+		return relativePlayerPos.z >= 0;
+	case 1: // dirDown
+		return relativePlayerPos.z < 32*voxelSize;
+	case 2: // dirPosX
+		return relativePlayerPos.x >= 0;
+	case 3: // dirNegX
+		return relativePlayerPos.x < 32*voxelSize;
+	case 4: // dirPosY
+		return relativePlayerPos.y >= 0;
+	case 5: // dirNegY
+		return relativePlayerPos.y < 32*voxelSize;
+	}
+	return true;
+}
+
+DrawElementsIndirectCommand addCommand(uint indices, uint vertexOffset, uint chunkID) {
+	return DrawElementsIndirectCommand(indices, 1, 0, int(vertexOffset), chunkID);
+}
 
 void main() {
 	uint chunkID = chunkIDs[chunkIDIndex + gl_GlobalInvocationID.x];
-	uint commandIndex = commandIndexStart + gl_GlobalInvocationID.x;
 	if(gl_GlobalInvocationID.x >= size) return;
 	if(isTransparent) {
-		commands[commandIndex] = DrawElementsIndirectCommand(
+		uint commandIndex = commandIndexStart + gl_GlobalInvocationID.x;
-			chunks[chunkID].vertexCountTransparent,
+		commands[commandIndex] = addCommand(chunks[chunkID].vertexCountTransparent, chunks[chunkID].vertexStartTransparent, chunkID);
-			1,
-			0,
-			int(chunks[chunkID].vertexStartTransparent),
-			chunkID
-		);
 	} else {
-		commands[commandIndex] = DrawElementsIndirectCommand(
+		uint commandIndex = commandIndexStart + gl_GlobalInvocationID.x*4;
-			chunks[chunkID].vertexCountOpaque,
+		uint commandIndexEnd = commandIndex + 4;
-			1,
+		uint groupFaceOffset = 0;
-			0,
+		uint groupFaceCount = 0;
-			int(chunks[chunkID].vertexStartOpaque),
+		for(int i = 0; i < 7; i++) {
-			chunkID
+			uint faceCount = chunks[chunkID].faceCountsByNormalOpaque[i];
-		);
+			if(isVisible(i, playerPositionInteger - chunks[chunkID].position.xyz, chunks[chunkID].voxelSize) || faceCount == 0) {
+				groupFaceCount += faceCount;
+			} else {
+				if(groupFaceCount != 0) {
+					commands[commandIndex] = addCommand(6*groupFaceCount, chunks[chunkID].vertexStartOpaque + 4*groupFaceOffset, chunkID);
+					commandIndex += 1;
+					groupFaceOffset += groupFaceCount;
+					groupFaceCount = 0;
+				}
+				groupFaceOffset += faceCount;
+			}
+		}
+		if(groupFaceCount != 0) {
+			commands[commandIndex] = addCommand(6*groupFaceCount, chunks[chunkID].vertexStartOpaque + 4*groupFaceOffset, chunkID);
+			commandIndex += 1;
+		}
+
+		for(; commandIndex < commandIndexEnd; commandIndex++) {
+			commands[commandIndex] = DrawElementsIndirectCommand(0, 0, 0, 0, 0);
+		}
 	}
 }

src/graphics.zig

@@ -2015,7 +2015,7 @@ pub fn generateBlockTexture(blockType: u16) Texture {
 			.visibilityMask = 255,
 			.voxelSize = 1,
 			.vertexStartOpaque = undefined,
-			.vertexCountOpaque = undefined,
+			.faceCountsByNormalOpaque = undefined,
 			.vertexStartTransparent = undefined,
 			.vertexCountTransparent = undefined,
 		}}, &chunkAllocation);

src/renderer/chunk_meshing.zig

@@ -51,6 +51,7 @@ pub var commandUniforms: struct {
 	commandIndexStart: c_int,
 	size: c_int,
 	isTransparent: c_int,
+	playerPositionInteger: c_int,
 } = undefined;
 var vao: c_uint = undefined;
 var vbo: c_uint = undefined;
@@ -158,16 +159,18 @@ pub fn bindTransparentShaderAndUniforms(projMatrix: Mat4f, ambient: Vec3f, playe
 }
 
 pub fn drawChunksIndirect(chunkIDs: []const u32, projMatrix: Mat4f, ambient: Vec3f, playerPos: Vec3d, transparent: bool) void {
+	const drawCallsEstimate: u31 = @intCast(if(transparent) chunkIDs.len else chunkIDs.len*4);
 	var chunkIDAllocation: main.graphics.SubAllocation = .{.start = 0, .len = 0};
 	chunkIDBuffer.uploadData(chunkIDs, &chunkIDAllocation);
 	defer chunkIDBuffer.free(chunkIDAllocation);
-	const allocation = commandBuffer.rawAlloc(@intCast(chunkIDs.len));
+	const allocation = commandBuffer.rawAlloc(drawCallsEstimate);
 	defer commandBuffer.free(allocation);
 	commandShader.bind();
 	c.glUniform1ui(commandUniforms.chunkIDIndex, chunkIDAllocation.start);
 	c.glUniform1ui(commandUniforms.commandIndexStart, allocation.start);
 	c.glUniform1ui(commandUniforms.size, @intCast(chunkIDs.len));
 	c.glUniform1i(commandUniforms.isTransparent, @intFromBool(transparent));
+	c.glUniform3i(commandUniforms.playerPositionInteger, @intFromFloat(playerPos[0]), @intFromFloat(playerPos[1]), @intFromFloat(playerPos[2]));
 	c.glDispatchCompute(@intCast(@divFloor(chunkIDs.len + 63, 64)), 1, 1); // TODO: Replace with @divCeil once available
 	c.glMemoryBarrier(c.GL_SHADER_STORAGE_BARRIER_BIT);
 
@@ -177,7 +180,7 @@ pub fn drawChunksIndirect(chunkIDs: []const u32, projMatrix: Mat4f, ambient: Vec
 		bindShaderAndUniforms(projMatrix, ambient, playerPos);
 	}
 	c.glBindBuffer(c.GL_DRAW_INDIRECT_BUFFER, commandBuffer.ssbo.bufferID);
-	c.glMultiDrawElementsIndirect(c.GL_TRIANGLES, c.GL_UNSIGNED_INT, @ptrFromInt(allocation.start*@sizeOf(IndirectData)), @intCast(chunkIDs.len), 0);
+	c.glMultiDrawElementsIndirect(c.GL_TRIANGLES, c.GL_UNSIGNED_INT, @ptrFromInt(allocation.start*@sizeOf(IndirectData)), drawCallsEstimate, 0);
 }
 
 pub const FaceData = extern struct {
@@ -208,7 +211,7 @@ pub const ChunkData = extern struct {
 	visibilityMask: i32,
 	voxelSize: i32,
 	vertexStartOpaque: u32,
-	vertexCountOpaque: u32,
+	faceCountsByNormalOpaque: [7]u32,
 	vertexStartTransparent: u32,
 	vertexCountTransparent: u32,
 };
@@ -232,6 +235,7 @@ const PrimitiveMesh = struct {
 	mutex: std.Thread.Mutex = .{},
 	bufferAllocation: graphics.SubAllocation = .{.start = 0, .len = 0},
 	vertexCount: u31 = 0,
+	byNormalCount: [7]u32 = .{0} ** 7,
 	wasChanged: bool = false,
 
 	fn deinit(self: *PrimitiveMesh) void {
@@ -497,12 +501,30 @@ const PrimitiveMesh = struct {
 		}
 		const fullBuffer = faceBuffer.allocateAndMapRange(len, &self.bufferAllocation);
 		defer faceBuffer.unmapRange(fullBuffer);
-		@memcpy(fullBuffer[0..self.coreLen], self.completeList[0..self.coreLen]);
+		// Sort the faces by normal to allow for backface culling on the GPU:
-		var i: usize = self.coreLen;
+		var i: u32 = 0;
-		for(0..6) |n| {
+		var iStart = i;
-			@memcpy(fullBuffer[i..][0..list[n].len], list[n]);
+		const coreList = self.completeList[0..self.coreLen];
-			i += list[n].len;
+		for(0..7) |normal| {
+			for(coreList) |face| {
+				if(main.models.extraQuadInfos.items[face.blockAndQuad.quadIndex].alignedNormalDirection) |normalDir| {
+					if(normalDir == normal) {
+						fullBuffer[i] = face;
+						i += 1;
 					}
+				} else if(normal == 6) {
+					fullBuffer[i] = face;
+					i += 1;
+				}
+			}
+			if(normal < 6) {
+				@memcpy(fullBuffer[i..][0..list[normal ^ 1].len], list[normal ^ 1]);
+				i += @intCast(list[normal ^ 1].len);
+			}
+			self.byNormalCount[normal] = i - iStart;
+			iStart = i;
+		}
+		std.debug.assert(i == fullBuffer.len);
 		self.vertexCount = @intCast(6*fullBuffer.len);
 		self.wasChanged = true;
 	}
@@ -1102,7 +1124,7 @@ pub const ChunkMesh = struct {
 			.voxelSize = self.pos.voxelSize,
 			.visibilityMask = self.visibilityMask,
 			.vertexStartOpaque = self.opaqueMesh.bufferAllocation.start*4,
-			.vertexCountOpaque = self.opaqueMesh.vertexCount,
+			.faceCountsByNormalOpaque = self.opaqueMesh.byNormalCount,
 			.vertexStartTransparent = self.transparentMesh.bufferAllocation.start*4,
 			.vertexCountTransparent = self.transparentMesh.bufferAllocation.len*6,
 		}}, &self.chunkAllocation);