adding primitive restart, integer texture sampling and mip/lod management in sampling

src/soft/device/Renderer.zig

@@ -53,6 +53,7 @@ pub const DynamicState = struct {
 };
 
 pub const Vertex = struct {
+    primitive_restart: bool,
     position: F32x4,
     outputs: [spv.SPIRV_MAX_OUTPUT_LOCATIONS][4]?struct {
         interpolation_type: enum { smooth, flat, noperspective },
@@ -100,6 +101,7 @@ pub const DrawCall = struct {
         };
 
         for (self.vertices) |*vertex| {
+            vertex.primitive_restart = false;
             for (&vertex.outputs) |*location| {
                 @memset(location, null);
             }
@@ -157,19 +159,19 @@ pub fn init(device: *SoftDevice, state: *PipelineState, active_occlusion_queries
 
 pub fn draw(self: *Self, vertex_count: usize, instance_count: usize, first_vertex: usize, first_instance: usize) VkError!void {
     var bounded_allocator: BoundedAllocator = .init(self.device.device_allocator.allocator(), @"1GiB");
-    try self.drawCall(&bounded_allocator, vertex_count, instance_count, first_vertex, first_instance, null);
+    try self.drawCall(&bounded_allocator, vertex_count, instance_count, first_vertex, first_instance, null, null);
 }
 
 pub fn drawIndexed(self: *Self, index_count: usize, instance_count: usize, first_index: usize, first_instance: usize, vertex_offset: i32) VkError!void {
     var bounded_allocator: BoundedAllocator = .init(self.device.device_allocator.allocator(), @"1GiB");
     const allocator = bounded_allocator.allocator();
 
-    const indices = try self.readIndexBuffer(allocator, index_count, first_index, vertex_offset);
+    const indexed_draw = try self.readIndexBuffer(allocator, index_count, first_index, vertex_offset);
 
-    try self.drawCall(&bounded_allocator, index_count, instance_count, 0, first_instance, indices);
+    try self.drawCall(&bounded_allocator, index_count, instance_count, 0, first_instance, indexed_draw.indices, indexed_draw.primitive_restart);
 }
 
-fn drawCall(self: *Self, bounded_allocator: *BoundedAllocator, vertex_count: usize, instance_count: usize, first_vertex: usize, first_instance: usize, indices: ?[]const i32) VkError!void {
+fn drawCall(self: *Self, bounded_allocator: *BoundedAllocator, vertex_count: usize, instance_count: usize, first_vertex: usize, first_instance: usize, indices: ?[]const u32, primitive_restart: ?[]const bool) VkError!void {
     const io = self.device.interface.io();
     const allocator = bounded_allocator.allocator();
 
@@ -215,7 +217,7 @@ fn drawCall(self: *Self, bounded_allocator: *BoundedAllocator, vertex_count: usi
         }
     }
 
-    self.vertexShaderStage(allocator, &draw_call, vertex_count, instance_count, first_vertex, first_instance, indices) catch |err| {
+    self.vertexShaderStage(allocator, &draw_call, vertex_count, instance_count, first_vertex, first_instance, indices, primitive_restart) catch |err| {
         std.log.scoped(.@"Vertex stage").err("catched a '{s}'", .{@errorName(err)});
         if (comptime base.config.logs == .verbose) {
             if (@errorReturnTrace()) |trace| {
@@ -232,7 +234,7 @@ fn drawCall(self: *Self, bounded_allocator: *BoundedAllocator, vertex_count: usi
     try rasterizer.processThenFragmentStage(self, allocator, &draw_call);
 }
 
-fn vertexShaderStage(self: *Self, allocator: std.mem.Allocator, draw_call: *DrawCall, vertex_count: usize, instance_count: usize, first_vertex: usize, first_instance: usize, indices: ?[]const i32) !void {
+fn vertexShaderStage(self: *Self, allocator: std.mem.Allocator, draw_call: *DrawCall, vertex_count: usize, instance_count: usize, first_vertex: usize, first_instance: usize, indices: ?[]const u32, primitive_restart: ?[]const bool) !void {
     const pipeline = self.state.pipeline orelse return;
     const batch_size = (pipeline.stages.getPtr(.vertex) orelse return).runtimes.len;
 
@@ -248,6 +250,7 @@ fn vertexShaderStage(self: *Self, allocator: std.mem.Allocator, draw_call: *Draw
                 .first_vertex = first_vertex,
                 .first_instance = first_instance,
                 .indices = indices,
+                .primitive_restart = primitive_restart,
                 .instance_index = instance_index,
                 .draw_call = draw_call,
             };
@@ -258,7 +261,12 @@ fn vertexShaderStage(self: *Self, allocator: std.mem.Allocator, draw_call: *Draw
     wg.await(self.device.interface.io()) catch return VkError.DeviceLost;
 }
 
-fn readIndexBuffer(self: *Self, allocator: std.mem.Allocator, index_count: usize, first_index: usize, vertex_offset: i32) VkError![]i32 {
+const IndexedDrawData = struct {
+    indices: []u32,
+    primitive_restart: ?[]bool,
+};
+
+fn readIndexBuffer(self: *Self, allocator: std.mem.Allocator, index_count: usize, first_index: usize, vertex_offset: i32) VkError!IndexedDrawData {
     const index_buffer = self.state.data.graphics.index_buffer;
     const buffer = index_buffer.buffer;
     const buffer_memory = if (buffer.interface.memory) |memory| memory else return VkError.InvalidDeviceMemoryDrv;
@@ -271,7 +279,11 @@ fn readIndexBuffer(self: *Self, allocator: std.mem.Allocator, index_count: usize
     const byte_size = index_count * index_size;
     const index_memory: []const u8 = try buffer_memory.map(byte_offset, byte_size);
 
-    const indices = allocator.alloc(i32, index_count) catch return VkError.OutOfDeviceMemory;
+    const indices = allocator.alloc(u32, index_count) catch return VkError.OutOfDeviceMemory;
+    const restart_enabled = (self.state.pipeline orelse return VkError.InvalidPipelineDrv).interface.mode.graphics.input_assembly.primitive_restart_enable == .true;
+    const restart_index = primitiveRestartIndex(index_buffer.index_type);
+    const primitive_restart = if (restart_enabled) allocator.alloc(bool, index_count) catch return VkError.OutOfDeviceMemory else null;
+
     for (indices, 0..) |*index, i| {
         const offset = i * index_size;
         const raw_index: u32 = switch (index_size) {
@@ -280,10 +292,21 @@ fn readIndexBuffer(self: *Self, allocator: std.mem.Allocator, index_count: usize
             4 => @intCast(std.mem.readInt(u32, index_memory[offset..][0..4], .little)),
             else => unreachable,
         };
-        index.* = vertex_offset + @as(i32, @intCast(raw_index));
+        if (primitive_restart) |restart| {
+            restart[i] = raw_index == restart_index;
+            if (restart[i]) {
+                index.* = 0;
+                continue;
+            }
+        }
+        const shifted = @as(i64, raw_index) + @as(i64, vertex_offset);
+        index.* = @as(u32, @truncate(@as(u64, @bitCast(shifted))));
     }
 
-    return indices;
+    return .{
+        .indices = indices,
+        .primitive_restart = primitive_restart,
+    };
 }
 
 fn indexTypeSize(index_type: vk.IndexType) ?usize {
@@ -295,6 +318,15 @@ fn indexTypeSize(index_type: vk.IndexType) ?usize {
     };
 }
 
+fn primitiveRestartIndex(index_type: vk.IndexType) u32 {
+    return switch (index_type) {
+        .uint8 => std.math.maxInt(u8),
+        .uint16 => std.math.maxInt(u16),
+        .uint32 => std.math.maxInt(u32),
+        else => unreachable,
+    };
+}
+
 fn resolveViewport(self: *Self, viewport_index: usize) VkError!vk.Viewport {
     const pipeline_data =
         &(self.state.pipeline orelse return VkError.InvalidPipelineDrv).interface.mode.graphics;

src/soft/device/clip.zig

@@ -169,6 +169,7 @@ fn interpolateBlob(allocator: std.mem.Allocator, a: []const u8, b: []const u8, s
 
 fn interpolateVertexForClipping(allocator: std.mem.Allocator, a: *const Vertex, b: *const Vertex, t: f32) VkError!Vertex {
     var result: Vertex = .{
+        .primitive_restart = false,
         .position = a.position + ((b.position - a.position) * zm.f32x4s(t)),
         .outputs = undefined,
     };

src/soft/device/rasterizer.zig

@@ -121,6 +121,9 @@ pub fn processThenFragmentStage(renderer: *Renderer, allocator: std.mem.Allocato
 
     switch (topology) {
         .point_list => for (draw_call.vertices) |*vertex| {
+            if (vertex.primitive_restart)
+                continue;
+
             try clipTransformAndRasterizePoint(
                 allocator,
                 draw_call,
@@ -148,56 +151,71 @@ pub fn processThenFragmentStage(renderer: *Renderer, allocator: std.mem.Allocato
                 if (stencil_attachment_access) |*access| access else null,
             );
         },
-        .triangle_fan => if (draw_call.vertices.len >= 3) {
-            const v0 = &draw_call.vertices[0];
-            for (1..(draw_call.vertices.len - 1)) |vertex_index| {
-                const v1 = &draw_call.vertices[vertex_index];
-                const v2 = &draw_call.vertices[vertex_index + 1];
+        .triangle_fan => {
+            var segment_start = firstNonRestart(draw_call, 0);
+            while (segment_start < draw_call.vertices.len) {
+                const segment_end = nextRestart(draw_call, segment_start);
+                if (segment_end - segment_start >= 3) {
+                    const v0 = &draw_call.vertices[segment_start];
+                    for ((segment_start + 1)..(segment_end - 1)) |vertex_index| {
+                        const v1 = &draw_call.vertices[vertex_index];
+                        const v2 = &draw_call.vertices[vertex_index + 1];
 
-                try clipTransformAndRasterizeTriangle(
-                    renderer,
-                    allocator,
-                    draw_call,
-                    v0,
-                    v1,
-                    v2,
-                    color_attachment_access,
-                    if (depth_attachment_access) |*access| access else null,
-                    if (stencil_attachment_access) |*access| access else null,
-                );
+                        try clipTransformAndRasterizeTriangle(
+                            renderer,
+                            allocator,
+                            draw_call,
+                            v0,
+                            v1,
+                            v2,
+                            color_attachment_access,
+                            if (depth_attachment_access) |*access| access else null,
+                            if (stencil_attachment_access) |*access| access else null,
+                        );
+                    }
+                }
+                segment_start = firstNonRestart(draw_call, segment_end + 1);
             }
         },
-        .triangle_strip => if (draw_call.vertices.len >= 3) {
-            for (0..(draw_call.vertices.len - 2)) |vertex_index| {
-                const v0 = &draw_call.vertices[vertex_index + 0];
-                const v1 = &draw_call.vertices[vertex_index + 1];
-                const v2 = &draw_call.vertices[vertex_index + 2];
+        .triangle_strip => {
+            var segment_start = firstNonRestart(draw_call, 0);
+            while (segment_start < draw_call.vertices.len) {
+                const segment_end = nextRestart(draw_call, segment_start);
+                if (segment_end - segment_start >= 3) {
+                    for (segment_start..(segment_end - 2)) |vertex_index| {
+                        const local_index = vertex_index - segment_start;
+                        const v0 = &draw_call.vertices[vertex_index + 0];
+                        const v1 = &draw_call.vertices[vertex_index + 1];
+                        const v2 = &draw_call.vertices[vertex_index + 2];
 
-                if ((vertex_index & 1) == 0) {
-                    try clipTransformAndRasterizeTriangle(
-                        renderer,
-                        allocator,
-                        draw_call,
-                        v0,
-                        v1,
-                        v2,
-                        color_attachment_access,
-                        if (depth_attachment_access) |*access| access else null,
-                        if (stencil_attachment_access) |*access| access else null,
-                    );
-                } else {
-                    try clipTransformAndRasterizeTriangle(
-                        renderer,
-                        allocator,
-                        draw_call,
-                        v1,
-                        v0,
-                        v2,
-                        color_attachment_access,
-                        if (depth_attachment_access) |*access| access else null,
-                        if (stencil_attachment_access) |*access| access else null,
-                    );
+                        if ((local_index & 1) == 0) {
+                            try clipTransformAndRasterizeTriangle(
+                                renderer,
+                                allocator,
+                                draw_call,
+                                v0,
+                                v1,
+                                v2,
+                                color_attachment_access,
+                                if (depth_attachment_access) |*access| access else null,
+                                if (stencil_attachment_access) |*access| access else null,
+                            );
+                        } else {
+                            try clipTransformAndRasterizeTriangle(
+                                renderer,
+                                allocator,
+                                draw_call,
+                                v1,
+                                v0,
+                                v2,
+                                color_attachment_access,
+                                if (depth_attachment_access) |*access| access else null,
+                                if (stencil_attachment_access) |*access| access else null,
+                            );
+                        }
+                    }
                 }
+                segment_start = firstNonRestart(draw_call, segment_end + 1);
             }
         },
         .line_list => for (0..@divTrunc(draw_call.vertices.len, 2)) |line_index| {
@@ -215,20 +233,27 @@ pub fn processThenFragmentStage(renderer: *Renderer, allocator: std.mem.Allocato
                 if (stencil_attachment_access) |*access| access else null,
             );
         },
-        .line_strip => if (draw_call.vertices.len >= 2) {
-            for (0..(draw_call.vertices.len - 1)) |vertex_index| {
-                const v0 = &draw_call.vertices[vertex_index + 0];
-                const v1 = &draw_call.vertices[vertex_index + 1];
+        .line_strip => {
+            var segment_start = firstNonRestart(draw_call, 0);
+            while (segment_start < draw_call.vertices.len) {
+                const segment_end = nextRestart(draw_call, segment_start);
+                if (segment_end - segment_start >= 2) {
+                    for (segment_start..(segment_end - 1)) |vertex_index| {
+                        const v0 = &draw_call.vertices[vertex_index + 0];
+                        const v1 = &draw_call.vertices[vertex_index + 1];
 
-                try clipTransformAndRasterizeLine(
-                    allocator,
-                    draw_call,
-                    v0,
-                    v1,
-                    color_attachment_access,
-                    if (depth_attachment_access) |*access| access else null,
-                    if (stencil_attachment_access) |*access| access else null,
-                );
+                        try clipTransformAndRasterizeLine(
+                            allocator,
+                            draw_call,
+                            v0,
+                            v1,
+                            color_attachment_access,
+                            if (depth_attachment_access) |*access| access else null,
+                            if (stencil_attachment_access) |*access| access else null,
+                        );
+                    }
+                }
+                segment_start = firstNonRestart(draw_call, segment_end + 1);
             }
         },
         else => base.unsupported("primitive topology {any}", .{topology}),
@@ -237,6 +262,18 @@ pub fn processThenFragmentStage(renderer: *Renderer, allocator: std.mem.Allocato
     draw_call.rasterizer_wait_group.await(io) catch return VkError.DeviceLost;
 }
 
+fn firstNonRestart(draw_call: *const DrawCall, start: usize) usize {
+    var index = start;
+    while (index < draw_call.vertices.len and draw_call.vertices[index].primitive_restart) : (index += 1) {}
+    return index;
+}
+
+fn nextRestart(draw_call: *const DrawCall, start: usize) usize {
+    var index = start;
+    while (index < draw_call.vertices.len and !draw_call.vertices[index].primitive_restart) : (index += 1) {}
+    return index;
+}
+
 fn clipTransformAndRasterizePoint(
     allocator: std.mem.Allocator,
     draw_call: *DrawCall,
@@ -253,10 +290,10 @@ fn clipTransformAndRasterizePoint(
     clip.viewportTransformVertex(draw_call.viewport, &transformed);
 
     const point_size = 1.0;
-    const min_x: i32 = @intFromFloat(@floor(transformed.position[0] - (point_size / 2.0)));
-    const max_x: i32 = @intFromFloat(@ceil(transformed.position[0] + (point_size / 2.0)) - 1.0);
-    const min_y: i32 = @intFromFloat(@floor(transformed.position[1] - (point_size / 2.0)));
-    const max_y: i32 = @intFromFloat(@ceil(transformed.position[1] + (point_size / 2.0)) - 1.0);
+    const min_x: i32 = @intFromFloat(@ceil(transformed.position[0] - (point_size / 2.0) - 0.5));
+    const max_x: i32 = @intFromFloat(@ceil(transformed.position[0] + (point_size / 2.0) - 0.5) - 1.0);
+    const min_y: i32 = @intFromFloat(@ceil(transformed.position[1] - (point_size / 2.0) - 0.5));
+    const max_y: i32 = @intFromFloat(@ceil(transformed.position[1] + (point_size / 2.0) - 0.5) - 1.0);
     const pipeline = draw_call.renderer.state.pipeline orelse return;
     const has_fragment_shader = pipeline.stages.getPtr(.fragment) != null;
 

src/soft/device/rasterizer/common.zig

@@ -268,27 +268,34 @@ inline fn blendOp(op: vk.BlendOp, src: F32x4, dst: F32x4) F32x4 {
     };
 }
 
-inline fn blendColor(src: F32x4, dst: F32x4, state: vk.PipelineColorBlendAttachmentState, constants: [4]f32) F32x4 {
+inline fn blendColor(src: F32x4, dst: F32x4, state: vk.PipelineColorBlendAttachmentState, constants: [4]f32, format: vk.Format) F32x4 {
     if (state.blend_enable == .false)
         return src;
 
-    const constant = F32x4{ constants[0], constants[1], constants[2], constants[3] };
-    const color_src = if (state.color_blend_op == .min or state.color_blend_op == .max)
-        src
+    const min_value = zm.f32x4s(base.format.minElementValue(format));
+    const max_value = zm.f32x4s(base.format.maxElementValue(format));
+    const clamped_src = if (base.format.isFloat(format)) src else std.math.clamp(src, min_value, max_value);
+    const constant = if (base.format.isFloat(format))
+        F32x4{ constants[0], constants[1], constants[2], constants[3] }
     else
-        src * blendFactor(state.src_color_blend_factor, src, dst, constant);
+        std.math.clamp(F32x4{ constants[0], constants[1], constants[2], constants[3] }, min_value, max_value);
+
+    const color_src = if (state.color_blend_op == .min or state.color_blend_op == .max)
+        clamped_src
+    else
+        clamped_src * blendFactor(state.src_color_blend_factor, clamped_src, dst, constant);
     const color_dst = if (state.color_blend_op == .min or state.color_blend_op == .max)
         dst
     else
-        dst * blendFactor(state.dst_color_blend_factor, src, dst, constant);
+        dst * blendFactor(state.dst_color_blend_factor, clamped_src, dst, constant);
     const alpha_src = if (state.alpha_blend_op == .min or state.alpha_blend_op == .max)
-        src
+        clamped_src
     else
-        src * blendFactor(state.src_alpha_blend_factor, src, dst, constant);
+        clamped_src * blendFactor(state.src_alpha_blend_factor, clamped_src, dst, constant);
     const alpha_dst = if (state.alpha_blend_op == .min or state.alpha_blend_op == .max)
         dst
     else
-        dst * blendFactor(state.dst_alpha_blend_factor, src, dst, constant);
+        dst * blendFactor(state.dst_alpha_blend_factor, clamped_src, dst, constant);
 
     var blended = blendOp(state.color_blend_op, color_src, color_dst);
     blended[3] = blendOp(state.alpha_blend_op, alpha_src, alpha_dst)[3];
@@ -389,7 +396,7 @@ pub fn writeToTargets(
                 if (location >= attachments.len)
                     break :blk src;
                 const constants = draw_call.renderer.dynamic_state.blend_constants orelse pipeline_data.color_blend.constants;
-                const blended = blendColor(src, dst, attachments[location], constants);
+                const blended = blendColor(src, dst, attachments[location], constants, color.format);
                 break :blk applyColorWriteMask(blended, dst, attachments[location].color_write_mask);
             } else src;
             const encoded_color = if (base.format.isSrgb(color.format)) zm.rgbToSrgb(final_color) else final_color;

src/soft/device/vertex_dispatcher.zig

@@ -21,7 +21,8 @@ pub const RunData = struct {
     vertex_count: usize,
     first_vertex: usize,
     first_instance: usize,
-    indices: ?[]const i32,
+    indices: ?[]const u32,
+    primitive_restart: ?[]const bool,
     instance_index: usize,
     draw_call: *Renderer.DrawCall,
 };
@@ -51,13 +52,22 @@ inline fn run(data: RunData) !void {
 
     var invocation_index: usize = data.batch_id;
     while (invocation_index < data.vertex_count) : (invocation_index += data.batch_size) {
+        const output: *Renderer.Vertex = &data.draw_call.vertices[(data.instance_index * data.vertex_count) + invocation_index];
+        if (data.primitive_restart) |primitive_restart| {
+            if (primitive_restart[invocation_index]) {
+                output.primitive_restart = true;
+                continue;
+            }
+        }
+
         rt.resetInvocation(data.allocator);
         try rt.populatePushConstants(data.draw_call.renderer.state.push_constant_blob[0..]);
 
-        const vertex_index: usize = if (data.indices) |indices| @intCast(indices[invocation_index]) else data.first_vertex + invocation_index;
+        const vertex_index_u32: u32 = if (data.indices) |indices| indices[invocation_index] else @intCast(data.first_vertex + invocation_index);
+        const vertex_index: usize = vertex_index_u32;
         const instance_index = data.first_instance + data.instance_index;
 
-        setupBuiltins(rt, vertex_index, instance_index) catch |err| switch (err) {
+        setupBuiltins(rt, vertex_index_u32, instance_index) catch |err| switch (err) {
             SpvRuntimeError.NotFound => {},
             else => return err,
         };
@@ -88,7 +98,6 @@ inline fn run(data: RunData) !void {
             else => return err,
         };
 
-        const output: *Renderer.Vertex = &data.draw_call.vertices[(data.instance_index * data.vertex_count) + invocation_index];
         try rt.readBuiltIn(std.mem.asBytes(&output.position), .Position);
 
         for (0..spv.SPIRV_MAX_OUTPUT_LOCATIONS) |location| {
@@ -112,8 +121,7 @@ inline fn run(data: RunData) !void {
     }
 }
 
-fn setupBuiltins(rt: *spv.Runtime, vertex_index: usize, instance_index: usize) !void {
-    const vertex_index_u32: u32 = @intCast(vertex_index);
+fn setupBuiltins(rt: *spv.Runtime, vertex_index_u32: u32, instance_index: usize) !void {
     const instance_index_u32: u32 = @intCast(instance_index);
 
     try rt.writeBuiltIn(std.mem.asBytes(&vertex_index_u32), .VertexIndex);