Optional antialiasing

src/assets/texture.zig

@@ -29,13 +29,14 @@ pub fn load(path: []const u8, alloc: std.mem.Allocator) Assets.LoadError!@This()
     const bytes_per_pixel = 4;
     const mip_level = if (std.math.isPowerOfTwo(width) and width == height) @as(u32, Graphics.MIP_LEVEL) else @as(u32, 1);
 
-    const texture = Graphics.createTexture(
-        width,
-        height,
-        target_format,
-        sdl.GPU_TEXTUREUSAGE_SAMPLER | sdl.GPU_TEXTUREUSAGE_COLOR_TARGET,
-        mip_level,
-    );
+    const texture = sdl.CreateGPUTexture(Graphics.device, &.{
+        .width = width,
+        .height = height,
+        .layer_count_or_depth = 1,
+        .format = target_format,
+        .usage = sdl.GPU_TEXTUREUSAGE_SAMPLER | sdl.GPU_TEXTUREUSAGE_COLOR_TARGET,
+        .num_levels = mip_level,
+    }) orelse err.sdl();
     errdefer Graphics.freeTexture(texture);
 
     const transfer_buffer_capacity = Graphics.TRANSFER_BUFFER_DEFAULT_CAPACITY;

src/graphics.zig

@@ -2,6 +2,7 @@ const std = @import("std");
 const sdl = @import("sdl");
 const err = @import("error.zig");
 const presets = @import("graphics/presets.zig");
+const Game = @import("game.zig");
 const Assets = @import("assets.zig");
 
 pub const Transform = @import("graphics/transform.zig");
@@ -13,21 +14,20 @@ pub var device: *sdl.GPUDevice = undefined;
 var command_buffer: ?*sdl.GPUCommandBuffer = null;
 var render_pass: ?*sdl.GPURenderPass = null;
 var render_target: ?*sdl.GPUTexture = null;
-var render_fsaa: bool = undefined;
+var batches: Batches = undefined;
 
 var shader_vert: *sdl.GPUShader = undefined;
 var shader_frag: *sdl.GPUShader = undefined;
 
 var depth_texture: *sdl.GPUTexture = undefined;
-var fsaa_target: *sdl.GPUTexture = undefined;
+var antialias: Antialias = undefined;
+var aa_target: *sdl.GPUTexture = undefined;
 var pipeline: *sdl.GPUGraphicsPipeline = undefined;
 
 pub var window_width: u32 = undefined;
 pub var window_height: u32 = undefined;
 pub var pixel_width: u32 = undefined;
 pub var pixel_height: u32 = undefined;
-var fsaa_scale: u32 = 4;
-var fsaa_level: u32 = 3;
 var render_width: u32 = undefined;
 var render_height: u32 = undefined;
 
@@ -38,6 +38,53 @@ const DEPTH_FORMAT = sdl.GPU_TEXTUREFORMAT_D32_FLOAT;
 pub const TRANSFER_BUFFER_DEFAULT_CAPACITY = 512 * 1024;
 pub const MIP_LEVEL = 4;
 
+const Antialias = union(enum) {
+    none,
+    msaa: enum(sdl.GPUSampleCount) {
+        @"2" = sdl.GPU_SAMPLECOUNT_2,
+        @"4" = sdl.GPU_SAMPLECOUNT_4,
+        @"8" = sdl.GPU_SAMPLECOUNT_8,
+    },
+    fsaa: enum {
+        @"2",
+        @"4",
+        @"8",
+    },
+
+    pub fn getMsaaSamples(self: @This()) sdl.GPUSampleCount {
+        if (self != .msaa) return sdl.GPU_SAMPLECOUNT_1;
+        return @intFromEnum(self.msaa);
+    }
+    pub fn getFsaaScale(self: @This()) u8 {
+        if (self != .fsaa) return 1;
+        return switch (self.fsaa) {
+            .@"2" => 2,
+            .@"4" => 4,
+            .@"8" => 8,
+        };
+    }
+    pub fn getFsaaLevel(self: @This()) u8 {
+        if (self != .fsaa) return 1;
+        return switch (self.fsaa) {
+            .@"2" => 2,
+            .@"4" => 3,
+            .@"8" => 4,
+        };
+    }
+};
+
+const Batch = struct {
+    object: *Assets.Object,
+    transform: Transform,
+    z: f32,
+
+    fn orderLessThan(ctx: void, lhs: Batch, rhs: Batch) bool {
+        _ = ctx;
+        return lhs.z > rhs.z;
+    }
+};
+const Batches = std.ArrayListUnmanaged(Batch);
+
 const Graphics = @This();
 pub fn create() void {
     // Init
@@ -62,6 +109,8 @@ pub fn create() void {
     Graphics.render_width = pixel_width;
     Graphics.render_height = pixel_height;
 
+    Graphics.antialias = .none;
+
     // Device
     Graphics.device = sdl.CreateGPUDevice(
         sdl.GPU_SHADERFORMAT_SPIRV,
@@ -95,20 +144,24 @@ pub fn create() void {
     const target_format = sdl.GetGPUSwapchainTextureFormat(Graphics.device, Graphics.window);
     if (target_format == sdl.GPU_TEXTUREFORMAT_INVALID) err.sdl();
 
-    Graphics.depth_texture = createTexture(
-        @as(u32, @intCast(Graphics.render_width)) * Graphics.fsaa_scale,
-        @as(u32, @intCast(Graphics.render_height)) * Graphics.fsaa_scale,
-        DEPTH_FORMAT,
-        sdl.GPU_TEXTUREUSAGE_DEPTH_STENCIL_TARGET,
-        1,
-    );
-    Graphics.fsaa_target = createTexture(
-        @as(u32, @intCast(Graphics.render_width)) * Graphics.fsaa_scale,
-        @as(u32, @intCast(Graphics.render_height)) * Graphics.fsaa_scale,
-        target_format,
-        sdl.GPU_TEXTUREUSAGE_COLOR_TARGET | sdl.GPU_TEXTUREUSAGE_SAMPLER,
-        fsaa_level,
-    );
+    Graphics.depth_texture = sdl.CreateGPUTexture(device, &.{
+        .width = Graphics.render_width * Graphics.antialias.getFsaaScale(),
+        .height = Graphics.render_height * Graphics.antialias.getFsaaScale(),
+        .layer_count_or_depth = 1,
+        .format = DEPTH_FORMAT,
+        .usage = sdl.GPU_TEXTUREUSAGE_DEPTH_STENCIL_TARGET,
+        .num_levels = 1,
+        .sample_count = Graphics.antialias.getMsaaSamples(),
+    }) orelse err.sdl();
+    Graphics.aa_target = sdl.CreateGPUTexture(device, &.{
+        .width = Graphics.render_width * Graphics.antialias.getFsaaScale(),
+        .height = Graphics.render_height * Graphics.antialias.getFsaaScale(),
+        .layer_count_or_depth = 1,
+        .format = target_format,
+        .usage = if (Graphics.antialias == .fsaa) sdl.GPU_TEXTUREUSAGE_COLOR_TARGET | sdl.GPU_TEXTUREUSAGE_SAMPLER else sdl.GPU_TEXTUREUSAGE_COLOR_TARGET,
+        .num_levels = Graphics.antialias.getFsaaLevel(),
+        .sample_count = Graphics.antialias.getMsaaSamples(),
+    }) orelse err.sdl();
 
     Graphics.pipeline = sdl.CreateGPUGraphicsPipeline(Graphics.device, &.{
         .vertex_shader = Graphics.shader_vert,
@@ -136,6 +189,9 @@ pub fn create() void {
             },
             .num_vertex_attributes = 2,
         },
+        .multisample_state = .{
+            .sample_count = Graphics.antialias.getMsaaSamples(),
+        },
         .primitive_type = sdl.GPU_PRIMITIVETYPE_TRIANGLELIST,
         .rasterizer_state = presets.RASTERIZER_CULL,
         .depth_stencil_state = presets.DEPTH_ENABLED,
@@ -150,6 +206,8 @@ pub fn create() void {
         },
     }) orelse err.sdl();
 
+    Graphics.batches = Batches.empty;
+
     Graphics.camera = Camera{
         .transform = .{},
         .near = 1.0 / 16.0,
@@ -164,7 +222,7 @@ pub fn destroy() void {
     sdl.DestroyWindow(Graphics.window);
 
     sdl.ReleaseGPUGraphicsPipeline(Graphics.device, Graphics.pipeline);
-    sdl.ReleaseGPUTexture(Graphics.device, Graphics.fsaa_target);
+    sdl.ReleaseGPUTexture(Graphics.device, Graphics.aa_target);
     sdl.ReleaseGPUTexture(Graphics.device, Graphics.depth_texture);
 
     sdl.ReleaseGPUShader(Graphics.device, Graphics.shader_vert);
@@ -175,6 +233,7 @@ pub fn destroy() void {
         Graphics.command_buffer = null;
     }
     sdl.DestroyGPUDevice(Graphics.device);
+    Graphics.batches.clearAndFree(Game.alloc);
 }
 
 /// If window is minimized returns `false`, `render_pass` remains null
@@ -195,14 +254,14 @@ pub fn beginDraw() bool {
         Graphics.camera.aspect = @as(f32, @floatFromInt(Graphics.render_width)) / @as(f32, @floatFromInt(Graphics.render_height));
     }
 
-    Graphics.render_fsaa = Graphics.fsaa_level > 1;
     Graphics.render_pass = sdl.BeginGPURenderPass(Graphics.command_buffer.?, &.{
         .clear_color = .{ .r = 0.0, .g = 0.0, .b = 0.0, .a = 1.0 },
         .cycle = false,
-        .load_op = sdl.GPU_LOADOP_CLEAR,
-        .store_op = sdl.GPU_STOREOP_STORE,
+        .load_op = sdl.GPU_LOADOP_DONT_CARE,
+        .store_op = if (Graphics.antialias == .msaa) sdl.GPU_STOREOP_RESOLVE else sdl.GPU_STOREOP_STORE,
         .mip_level = 0,
-        .texture = if (Graphics.render_fsaa) Graphics.fsaa_target else Graphics.render_target,
+        .texture = if (Graphics.antialias == .fsaa or Graphics.antialias == .msaa) Graphics.aa_target else Graphics.render_target,
+        .resolve_texture = if (Graphics.antialias == .msaa) Graphics.render_target else null,
     }, 1, &.{
         .clear_depth = 0.0,
         .load_op = sdl.GPU_LOADOP_CLEAR,
@@ -219,34 +278,13 @@ pub fn beginDraw() bool {
     return true;
 }
 
-pub fn clearDepth() void {
-    sdl.EndGPURenderPass(Graphics.render_pass.?);
+fn finishPass() void {
+    std.sort.block(Batch, Graphics.batches.items, {}, Batch.orderLessThan);
 
-    Graphics.render_pass = sdl.BeginGPURenderPass(Graphics.command_buffer.?, &.{
-        .clear_color = .{ .r = 0.0, .g = 0.0, .b = 0.0, .a = 1.0 },
-        .cycle = false,
-        .load_op = sdl.GPU_LOADOP_LOAD,
-        .store_op = sdl.GPU_STOREOP_STORE,
-        .mip_level = 0,
-        .texture = if (Graphics.render_fsaa) Graphics.fsaa_target else Graphics.render_target,
-    }, 1, &.{
-        .clear_depth = 0.0,
-        .load_op = sdl.GPU_LOADOP_CLEAR,
-        .store_op = sdl.GPU_STOREOP_DONT_CARE,
-        .stencil_load_op = sdl.GPU_STOREOP_DONT_CARE,
-        .stencil_store_op = sdl.GPU_STOREOP_DONT_CARE,
-        .texture = Graphics.depth_texture,
-    }) orelse err.sdl();
+    for (Graphics.batches.items) |*batch| {
+        const asset_object = batch.object.get() orelse continue;
 
-    sdl.BindGPUGraphicsPipeline(Graphics.render_pass, Graphics.pipeline);
-    sdl.PushGPUVertexUniformData(Graphics.command_buffer, 0, &Graphics.camera.matrix, 16 * 4);
-}
-
-pub fn drawObject(object: *Assets.Object, transform: Transform) void {
-    if (Graphics.render_pass == null) return;
-    const asset_object = object.get() orelse return;
-
-    sdl.PushGPUVertexUniformData(Graphics.command_buffer, 1, &transform.matrix(), 16 * 4);
+        sdl.PushGPUVertexUniformData(Graphics.command_buffer, 1, &batch.transform.matrix(), 16 * 4);
         for (asset_object.nodes) |node| {
             const mesh = &asset_object.meshes[node.mesh];
 
@@ -261,22 +299,72 @@ pub fn drawObject(object: *Assets.Object, transform: Transform) void {
                 sdl.DrawGPUIndexedPrimitives(Graphics.render_pass, primitive.indices, 1, 0, 0, 0);
             }
         }
+    }
+    Graphics.batches.clearRetainingCapacity();
+
+    sdl.EndGPURenderPass(Graphics.render_pass.?);
+}
+
+pub fn clearDepth() void {
+    Graphics.finishPass();
+
+    Graphics.render_pass = sdl.BeginGPURenderPass(Graphics.command_buffer.?, &.{
+        .clear_color = .{ .r = 0.0, .g = 0.0, .b = 0.0, .a = 1.0 },
+        .cycle = false,
+        .load_op = sdl.GPU_LOADOP_LOAD,
+        .store_op = if (Graphics.antialias == .msaa) sdl.GPU_STOREOP_RESOLVE else sdl.GPU_STOREOP_STORE,
+        .mip_level = 0,
+        .texture = if (Graphics.antialias == .fsaa or Graphics.antialias == .msaa) Graphics.aa_target else Graphics.render_target,
+        .resolve_texture = if (Graphics.antialias == .msaa) Graphics.render_target else null,
+    }, 1, &.{
+        .clear_depth = 0.0,
+        .load_op = sdl.GPU_LOADOP_CLEAR,
+        .store_op = sdl.GPU_STOREOP_DONT_CARE,
+        .stencil_load_op = sdl.GPU_STOREOP_DONT_CARE,
+        .stencil_store_op = sdl.GPU_STOREOP_DONT_CARE,
+        .texture = Graphics.depth_texture,
+    }) orelse err.sdl();
+
+    sdl.BindGPUGraphicsPipeline(Graphics.render_pass, Graphics.pipeline);
+    sdl.PushGPUVertexUniformData(Graphics.command_buffer, 0, &Graphics.camera.matrix, 16 * 4);
+}
+
+// `object`: pointer MUST be vaild until current render pass ends
+pub fn drawObject(object: *Assets.Object, transform: Transform) void {
+    if (Graphics.render_pass == null) return;
+
+    @setFloatMode(.optimized);
+    const z = Graphics.camera.matrix[8] * transform.position[0] +
+        Graphics.camera.matrix[9] * transform.position[1] +
+        Graphics.camera.matrix[10] * transform.position[2] +
+        Graphics.camera.matrix[11];
+    var w = Graphics.camera.matrix[12] * transform.position[0] +
+        Graphics.camera.matrix[13] * transform.position[1] +
+        Graphics.camera.matrix[14] * transform.position[2] +
+        Graphics.camera.matrix[15];
+    if (w == 0) w = 1;
+
+    Graphics.batches.append(Game.alloc, .{
+        .object = object,
+        .transform = transform,
+        .z = z / w,
+    }) catch err.oom();
 }
 
 pub fn endDraw() void {
     defer Graphics.command_buffer = null;
     defer Graphics.render_pass = null;
-    if (Graphics.render_pass) |pass| {
-        sdl.EndGPURenderPass(pass);
 
-        if (Graphics.render_fsaa) {
-            sdl.GenerateMipmapsForGPUTexture(Graphics.command_buffer, Graphics.fsaa_target);
+    Graphics.finishPass();
+
+    if (Graphics.antialias == .fsaa) {
+        sdl.GenerateMipmapsForGPUTexture(Graphics.command_buffer, Graphics.aa_target);
         sdl.BlitGPUTexture(Graphics.command_buffer, &.{
             .source = .{
-                    .texture = Graphics.fsaa_target,
+                .texture = Graphics.aa_target,
                 .w = Graphics.render_width,
                 .h = Graphics.render_height,
-                    .mip_level = fsaa_level - 1,
+                .mip_level = Graphics.antialias.getFsaaLevel() - 1,
             },
             .destination = .{
                 .texture = Graphics.render_target,
@@ -287,7 +375,6 @@ pub fn endDraw() void {
             .filter = sdl.GPU_FILTER_NEAREST,
         });
     }
-    }
     if (!sdl.SubmitGPUCommandBuffer(Graphics.command_buffer)) err.sdl();
 }
 
@@ -304,18 +391,6 @@ fn loadShader(path: []const u8, info: sdl.GPUShaderCreateInfo) *sdl.GPUShader {
     return sdl.CreateGPUShader(device, &updated_info) orelse err.sdl();
 }
 
-pub fn createTexture(width: u32, height: u32, format: c_uint, usage: c_uint, mip_level: u32) *sdl.GPUTexture {
-    return sdl.CreateGPUTexture(device, &.{
-        .format = format,
-        .layer_count_or_depth = 1,
-        .width = width,
-        .height = height,
-        .num_levels = mip_level,
-        .sample_count = sdl.GPU_SAMPLECOUNT_1,
-        .usage = usage,
-    }) orelse err.sdl();
-}
-
 pub fn freeTexture(texture: *sdl.GPUTexture) void {
     sdl.ReleaseGPUTexture(Graphics.device, texture);
 }
@@ -340,24 +415,32 @@ pub fn freeSampler(sampler: *sdl.GPUSampler) void {
 
 fn resetTextures(width: u32, height: u32) void {
     sdl.ReleaseGPUTexture(Graphics.device, Graphics.depth_texture);
-    Graphics.depth_texture = createTexture(
-        width * Graphics.fsaa_scale,
-        height * Graphics.fsaa_scale,
-        DEPTH_FORMAT,
-        sdl.GPU_TEXTUREUSAGE_DEPTH_STENCIL_TARGET,
-        1,
-    );
+    Graphics.depth_texture = sdl.CreateGPUTexture(device, &.{
+        .width = width * Graphics.antialias.getFsaaScale(),
+        .height = height * Graphics.antialias.getFsaaScale(),
+        .layer_count_or_depth = 1,
+        .format = DEPTH_FORMAT,
+        .usage = sdl.GPU_TEXTUREUSAGE_DEPTH_STENCIL_TARGET,
+        .num_levels = 1,
+        .sample_count = Graphics.antialias.getMsaaSamples(),
+    }) orelse err.sdl();
 
     const target_format = sdl.GetGPUSwapchainTextureFormat(Graphics.device, Graphics.window);
 
-    sdl.ReleaseGPUTexture(Graphics.device, Graphics.fsaa_target);
-    Graphics.fsaa_target = createTexture(
-        width * Graphics.fsaa_scale,
-        height * Graphics.fsaa_scale,
-        target_format,
-        sdl.GPU_TEXTUREUSAGE_COLOR_TARGET | sdl.GPU_TEXTUREUSAGE_SAMPLER,
-        fsaa_level,
-    );
+    sdl.ReleaseGPUTexture(Graphics.device, Graphics.aa_target);
+    Graphics.aa_target = sdl.CreateGPUTexture(device, &.{
+        .width = width * Graphics.antialias.getFsaaScale(),
+        .height = height * Graphics.antialias.getFsaaScale(),
+        .layer_count_or_depth = 1,
+        .format = target_format,
+        .usage = switch (Graphics.antialias) {
+            .none => sdl.GPU_TEXTUREUSAGE_COLOR_TARGET,
+            .fsaa => sdl.GPU_TEXTUREUSAGE_COLOR_TARGET | sdl.GPU_TEXTUREUSAGE_SAMPLER,
+            .msaa => sdl.GPU_TEXTUREUSAGE_COLOR_TARGET,
+        },
+        .num_levels = Graphics.antialias.getFsaaLevel(),
+        .sample_count = Graphics.antialias.getMsaaSamples(),
+    }) orelse err.sdl();
 }
 
 pub fn windowId() sdl.WindowID {