hal/gles: stencil and vertex state

wgpu-hal/src/gles/adapter.rs

@@ -232,7 +232,8 @@ impl super::Adapter {
                 0
             },
             max_vertex_buffers: gl.get_parameter_i32(glow::MAX_VERTEX_ATTRIB_BINDINGS) as u32,
-            max_vertex_attributes: gl.get_parameter_i32(glow::MAX_VERTEX_ATTRIBS) as u32,
+            max_vertex_attributes: (gl.get_parameter_i32(glow::MAX_VERTEX_ATTRIBS) as u32)
+                .min(super::MAX_VERTEX_ATTRIBUTES as u32),
             max_vertex_buffer_array_stride: 2048,
             max_push_constant_size: 0,
         };

wgpu-hal/src/gles/command.rs

@@ -1,6 +1,26 @@
 use super::{conv, Command as C};
+use arrayvec::ArrayVec;
 use std::{mem, ops::Range};
 
+bitflags::bitflags! {
+    #[derive(Default)]
+    struct Dirty: u32 {
+        const VERTEX_BUFFERS = 0x0001;
+    }
+}
+
+#[derive(Default)]
+pub(super) struct State {
+    topology: u32,
+    index_format: wgt::IndexFormat,
+    index_offset: wgt::BufferAddress,
+    vertex_buffers: [super::VertexBufferDesc; crate::MAX_VERTEX_BUFFERS],
+    vertex_attributes: ArrayVec<[super::AttributeDesc; super::MAX_VERTEX_ATTRIBUTES]>,
+    stencil: super::StencilState,
+    has_pass_label: bool,
+    dirty: Dirty,
+}
+
 impl super::CommandBuffer {
     fn clear(&mut self) {
         self.label = None;
@@ -16,9 +36,62 @@ impl super::CommandBuffer {
     }
 }
 
+impl super::CommandEncoder {
+    fn rebind_stencil_func(&mut self) {
+        fn make(s: &super::StencilSide, face: u32) -> C {
+            C::SetStencilFunc {
+                face,
+                function: s.function,
+                reference: s.reference,
+                read_mask: s.mask_read,
+            }
+        }
+
+        let s = &self.state.stencil;
+        if s.front.function == s.back.function
+            && s.front.mask_read == s.back.mask_read
+            && s.front.reference == s.back.reference
+        {
+            self.cmd_buffer
+                .commands
+                .push(make(&s.front, glow::FRONT_AND_BACK));
+        } else {
+            self.cmd_buffer.commands.push(make(&s.front, glow::FRONT));
+            self.cmd_buffer.commands.push(make(&s.back, glow::BACK));
+        }
+    }
+
+    fn rebind_vertex_attributes(&mut self, first_instance: u32) {
+        for attribute in self.state.vertex_attributes.iter() {
+            let vb = self.state.vertex_buffers[attribute.buffer_index as usize].clone();
+
+            let mut vat = attribute.clone();
+            vat.offset += vb.offset as u32;
+
+            if vb.step == wgt::InputStepMode::Instance {
+                vat.offset += vb.stride * first_instance;
+            }
+
+            self.cmd_buffer
+                .commands
+                .push(C::SetVertexAttribute(vat, vb));
+        }
+    }
+
+    fn prepare_draw(&mut self, first_instance: u32) {
+        if first_instance != 0 {
+            self.rebind_vertex_attributes(first_instance);
+            self.state.dirty.set(Dirty::VERTEX_BUFFERS, true);
+        } else if self.state.dirty.contains(Dirty::VERTEX_BUFFERS) {
+            self.rebind_vertex_attributes(0);
+            self.state.dirty.set(Dirty::VERTEX_BUFFERS, false);
+        }
+    }
+}
+
 impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
     unsafe fn begin_encoding(&mut self, label: crate::Label) -> Result<(), crate::DeviceError> {
-        self.state = super::CommandState::default();
+        self.state = State::default();
         self.cmd_buffer.label = label.map(str::to_string);
         Ok(())
     }
@@ -63,18 +136,22 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
         {
             return;
         }
+
+        let mut combined_usage = crate::TextureUse::empty();
         for bar in barriers {
             // GLES only synchronizes storage -> anything explicitly
             if !bar.usage.start.contains(crate::TextureUse::STORAGE_STORE) {
                 continue;
             }
-            let raw = match bar.texture.inner {
+            // unlike buffers, there is no need for a concrete texture
-                super::TextureInner::Texture { raw, .. } => raw,
+            // object to be bound anywhere for a barrier
-                super::TextureInner::Renderbuffer { .. } => continue,
+            combined_usage |= bar.usage.end;
-            };
+        }
+
+        if !combined_usage.is_empty() {
             self.cmd_buffer
                 .commands
-                .push(C::TextureBarrier(raw, bar.usage.end));
+                .push(C::TextureBarrier(combined_usage));
         }
     }
 
@@ -235,9 +312,7 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
         }
 
         // set the framebuffer
-        self.cmd_buffer
+        self.cmd_buffer.commands.push(C::ResetFramebuffer);
-            .commands
-            .push(C::ResetFramebuffer(desc.extent));
         for (i, cat) in desc.color_attachments.iter().enumerate() {
             let attachment = glow::COLOR_ATTACHMENT0 + i as u32;
             self.cmd_buffer.commands.push(C::SetFramebufferAttachment {
@@ -257,9 +332,22 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
             });
         }
 
+        // set the draw buffers and states
         self.cmd_buffer
             .commands
             .push(C::SetDrawColorBuffers(desc.color_attachments.len() as u8));
+        let rect = crate::Rect {
+            x: 0,
+            y: 0,
+            w: desc.extent.width as i32,
+            h: desc.extent.height as i32,
+        };
+        self.cmd_buffer.commands.push(C::SetScissor(rect.clone()));
+        self.cmd_buffer.commands.push(C::SetViewport {
+            rect,
+            depth: 0.0..1.0,
+        });
+
         // issue the clears
         for (i, cat) in desc.color_attachments.iter().enumerate() {
             if !cat.ops.contains(crate::AttachmentOp::LOAD) {
@@ -302,6 +390,7 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
             self.cmd_buffer.commands.push(C::PopDebugGroup);
             self.state.has_pass_label = false;
         }
+        self.state.dirty = Dirty::empty();
     }
 
     unsafe fn set_bind_group(
@@ -335,6 +424,46 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
 
     unsafe fn set_render_pipeline(&mut self, pipeline: &super::RenderPipeline) {
         self.state.topology = conv::map_primitive_topology(pipeline.primitive.topology);
+        self.state.dirty |= Dirty::VERTEX_BUFFERS;
+
+        self.state.vertex_attributes.clear();
+        for vat in pipeline.vertex_attributes.iter() {
+            self.state.vertex_attributes.push(vat.clone());
+        }
+        for (state_desc, pipe_desc) in self
+            .state
+            .vertex_buffers
+            .iter_mut()
+            .zip(pipeline.vertex_buffers.iter())
+        {
+            state_desc.step = pipe_desc.step;
+            state_desc.stride = pipe_desc.stride;
+        }
+
+        if let Some(ref stencil) = pipeline.stencil {
+            self.state.stencil = stencil.clone();
+            self.rebind_stencil_func();
+            if stencil.front.ops == stencil.back.ops
+                && stencil.front.mask_write == stencil.back.mask_write
+            {
+                self.cmd_buffer.commands.push(C::SetStencilOps {
+                    face: glow::FRONT_AND_BACK,
+                    write_mask: stencil.front.mask_write,
+                    ops: stencil.front.ops.clone(),
+                });
+            } else {
+                self.cmd_buffer.commands.push(C::SetStencilOps {
+                    face: glow::FRONT,
+                    write_mask: stencil.front.mask_write,
+                    ops: stencil.front.ops.clone(),
+                });
+                self.cmd_buffer.commands.push(C::SetStencilOps {
+                    face: glow::BACK,
+                    write_mask: stencil.back.mask_write,
+                    ops: stencil.back.ops.clone(),
+                });
+            }
+        }
     }
 
     unsafe fn set_index_buffer<'a>(
@@ -353,10 +482,35 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
         index: u32,
         binding: crate::BufferBinding<'a, super::Api>,
     ) {
+        self.state.dirty |= Dirty::VERTEX_BUFFERS;
+        let vb = &mut self.state.vertex_buffers[index as usize];
+        vb.raw = binding.buffer.raw;
+        vb.offset = binding.offset;
+    }
+    unsafe fn set_viewport(&mut self, rect: &crate::Rect<f32>, depth: Range<f32>) {
+        self.cmd_buffer.commands.push(C::SetViewport {
+            rect: crate::Rect {
+                x: rect.x as i32,
+                y: rect.y as i32,
+                w: rect.w as i32,
+                h: rect.h as i32,
+            },
+            depth,
+        });
+    }
+    unsafe fn set_scissor_rect(&mut self, rect: &crate::Rect<u32>) {
+        self.cmd_buffer.commands.push(C::SetScissor(crate::Rect {
+            x: rect.x as i32,
+            y: rect.y as i32,
+            w: rect.w as i32,
+            h: rect.h as i32,
+        }));
+    }
+    unsafe fn set_stencil_reference(&mut self, value: u32) {
+        self.state.stencil.front.reference = value;
+        self.state.stencil.back.reference = value;
+        self.rebind_stencil_func();
     }
-    unsafe fn set_viewport(&mut self, rect: &crate::Rect<f32>, depth_range: Range<f32>) {}
-    unsafe fn set_scissor_rect(&mut self, rect: &crate::Rect<u32>) {}
-    unsafe fn set_stencil_reference(&mut self, value: u32) {}
     unsafe fn set_blend_constants(&mut self, color: &wgt::Color) {}
 
     unsafe fn draw(
@@ -366,7 +520,7 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
         start_instance: u32,
         instance_count: u32,
     ) {
-        debug_assert_eq!(start_instance, 0);
+        self.prepare_draw(start_instance);
         self.cmd_buffer.commands.push(C::Draw {
             topology: self.state.topology,
             start_vertex,
@@ -382,7 +536,7 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
         start_instance: u32,
         instance_count: u32,
     ) {
-        debug_assert_eq!(start_instance, 0);
+        self.prepare_draw(start_instance);
         let (index_size, index_type) = match self.state.index_format {
             wgt::IndexFormat::Uint16 => (2, glow::UNSIGNED_SHORT),
             wgt::IndexFormat::Uint32 => (4, glow::UNSIGNED_INT),
@@ -403,6 +557,7 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
         offset: wgt::BufferAddress,
         draw_count: u32,
     ) {
+        self.prepare_draw(0);
         for draw in 0..draw_count as wgt::BufferAddress {
             let indirect_offset =
                 offset + draw * mem::size_of::<wgt::DrawIndirectArgs>() as wgt::BufferAddress;
@@ -419,6 +574,7 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
         offset: wgt::BufferAddress,
         draw_count: u32,
     ) {
+        self.prepare_draw(0);
         let index_type = match self.state.index_format {
             wgt::IndexFormat::Uint16 => glow::UNSIGNED_SHORT,
             wgt::IndexFormat::Uint32 => glow::UNSIGNED_INT,
@@ -469,6 +625,7 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
             self.cmd_buffer.commands.push(C::PopDebugGroup);
             self.state.has_pass_label = false;
         }
+        self.state.dirty = Dirty::empty();
     }
 
     unsafe fn set_compute_pipeline(&mut self, pipeline: &super::ComputePipeline) {}

wgpu-hal/src/gles/conv.rs

@@ -232,3 +232,44 @@ pub fn map_view_dimension(dim: wgt::TextureViewDimension) -> u32 {
         Tvd::D3 => glow::TEXTURE_3D,
     }
 }
+
+fn map_stencil_op(operation: wgt::StencilOperation) -> u32 {
+    use wgt::StencilOperation as So;
+    match operation {
+        So::Keep => glow::KEEP,
+        So::Zero => glow::ZERO,
+        So::Replace => glow::REPLACE,
+        So::Invert => glow::INVERT,
+        So::IncrementClamp => glow::INCR,
+        So::DecrementClamp => glow::DECR,
+        So::IncrementWrap => glow::INCR_WRAP,
+        So::DecrementWrap => glow::DECR_WRAP,
+    }
+}
+
+fn map_stencil_ops(face: &wgt::StencilFaceState) -> super::StencilOps {
+    super::StencilOps {
+        pass: map_stencil_op(face.pass_op),
+        fail: map_stencil_op(face.fail_op),
+        depth_fail: map_stencil_op(face.depth_fail_op),
+    }
+}
+
+pub(super) fn map_stencil(state: &wgt::StencilState) -> super::StencilState {
+    super::StencilState {
+        front: super::StencilSide {
+            function: map_compare_func(state.front.compare),
+            mask_read: state.read_mask,
+            mask_write: state.write_mask,
+            reference: 0,
+            ops: map_stencil_ops(&state.front),
+        },
+        back: super::StencilSide {
+            function: map_compare_func(state.back.compare),
+            mask_read: state.read_mask,
+            mask_write: state.write_mask,
+            reference: 0,
+            ops: map_stencil_ops(&state.back),
+        },
+    }
+}

wgpu-hal/src/gles/device.rs

@@ -518,7 +518,7 @@ impl crate::Device<super::Api> for super::Device {
     ) -> Result<super::CommandEncoder, crate::DeviceError> {
         Ok(super::CommandEncoder {
             cmd_buffer: super::CommandBuffer::default(),
-            state: super::CommandState::default(),
+            state: Default::default(),
             private_caps: self.shared.private_caps,
         })
     }
@@ -665,11 +665,16 @@ impl crate::Device<super::Api> for super::Device {
         );
         let inner = self.create_pipeline(shaders, desc.layout)?;
 
-        let attributes = {
+        let (vertex_buffers, vertex_attributes) = {
-            let gl = &self.shared.context;
+            let mut buffers = Vec::new();
             let mut attributes = Vec::new();
-
             for (index, vb_layout) in desc.vertex_buffers.iter().enumerate() {
+                buffers.push(super::VertexBufferDesc {
+                    raw: 0,
+                    offset: 0,
+                    step: vb_layout.step_mode,
+                    stride: vb_layout.array_stride as u32,
+                });
                 for vat in vb_layout.attributes.iter() {
                     let format_desc = conv::describe_vertex_format(vat.format);
                     attributes.push(super::AttributeDesc {
@@ -680,15 +685,23 @@ impl crate::Device<super::Api> for super::Device {
                     });
                 }
             }
-
+            (buffers.into_boxed_slice(), attributes.into_boxed_slice())
-            attributes.into_boxed_slice()
         };
 
         Ok(super::RenderPipeline {
             inner,
             primitive: desc.primitive,
-            attributes,
+            vertex_buffers,
-            depth: desc.depth_stencil.clone(),
+            vertex_attributes,
+            depth: desc.depth_stencil.as_ref().map(|ds| super::DepthState {
+                function: conv::map_compare_func(ds.depth_compare),
+                mask: ds.depth_write_enabled,
+            }),
+            depth_bias: desc.depth_stencil.as_ref().map(|ds| ds.bias),
+            stencil: desc
+                .depth_stencil
+                .as_ref()
+                .map(|ds| conv::map_stencil(&ds.stencil)),
         })
     }
     unsafe fn destroy_render_pipeline(&self, pipeline: super::RenderPipeline) {

wgpu-hal/src/gles/mod.rs

@@ -22,6 +22,7 @@ pub struct Api;
 //Note: we can support more samplers if not every one of them is used at a time,
 // but it probably doesn't worth it.
 const MAX_TEXTURE_SLOTS: usize = 16;
+const MAX_VERTEX_ATTRIBUTES: usize = 16;
 
 impl crate::Api for Api {
     type Instance = Instance;
@@ -159,6 +160,12 @@ enum VertexAttribKind {
            //Double,  // glVertexAttribLPointer
 }
 
+impl Default for VertexAttribKind {
+    fn default() -> Self {
+        Self::Float
+    }
+}
+
 #[derive(Debug)]
 struct TextureFormatDesc {
     internal: u32,
@@ -296,12 +303,14 @@ pub struct ShaderModule {
     naga: crate::NagaShader,
 }
 
+#[derive(Clone, Debug, Default)]
 struct VertexFormatDesc {
     element_count: i32,
     element_format: u32,
     attrib_kind: VertexAttribKind,
 }
 
+#[derive(Clone, Debug, Default)]
 struct AttributeDesc {
     location: u32,
     offset: u32,
@@ -309,6 +318,14 @@ struct AttributeDesc {
     format_desc: VertexFormatDesc,
 }
 
+#[derive(Clone, Debug, Default)]
+struct VertexBufferDesc {
+    raw: glow::Buffer,
+    offset: wgt::BufferAddress,
+    step: wgt::InputStepMode,
+    stride: u32,
+}
+
 #[derive(Clone)]
 struct UniformDesc {
     location: glow::UniformLocation,
@@ -326,13 +343,20 @@ struct PipelineInner {
     uniforms: Box<[UniformDesc]>,
 }
 
+struct DepthState {
+    function: u32,
+    mask: bool,
+}
+
 pub struct RenderPipeline {
     inner: PipelineInner,
     //blend_targets: Vec<pso::ColorBlendDesc>,
-    attributes: Box<[AttributeDesc]>,
+    vertex_buffers: Box<[VertexBufferDesc]>,
-    //vertex_buffers: Box<[wgt::VertexBufferLayout]>,
+    vertex_attributes: Box<[AttributeDesc]>,
     primitive: wgt::PrimitiveState,
-    depth: Option<wgt::DepthStencilState>,
+    depth: Option<DepthState>,
+    depth_bias: Option<wgt::DepthBiasState>,
+    stencil: Option<StencilState>,
 }
 
 pub struct ComputePipeline {
@@ -387,6 +411,50 @@ struct TextureCopyInfo {
     texel_size: u8,
 }
 
+#[derive(Clone, Debug, PartialEq)]
+struct StencilOps {
+    pass: u32,
+    fail: u32,
+    depth_fail: u32,
+}
+
+impl Default for StencilOps {
+    fn default() -> Self {
+        Self {
+            pass: glow::KEEP,
+            fail: glow::KEEP,
+            depth_fail: glow::KEEP,
+        }
+    }
+}
+
+#[derive(Clone, Debug, PartialEq)]
+struct StencilSide {
+    function: u32,
+    mask_read: u32,
+    mask_write: u32,
+    reference: u32,
+    ops: StencilOps,
+}
+
+impl Default for StencilSide {
+    fn default() -> Self {
+        Self {
+            function: glow::ALWAYS,
+            mask_read: 0xFF,
+            mask_write: 0xFF,
+            reference: 0,
+            ops: StencilOps::default(),
+        }
+    }
+}
+
+#[derive(Clone, Default)]
+struct StencilState {
+    front: StencilSide,
+    back: StencilSide,
+}
+
 #[derive(Debug)]
 enum Command {
     Draw {
@@ -463,7 +531,7 @@ enum Command {
         dst_target: BindTarget,
         dst_offset: wgt::BufferAddress,
     },
-    ResetFramebuffer(wgt::Extent3d),
+    ResetFramebuffer,
     SetFramebufferAttachment {
         attachment: u32,
         view: TextureView,
@@ -475,7 +543,24 @@ enum Command {
     ClearDepth(f32),
     ClearStencil(u32),
     BufferBarrier(glow::Buffer, crate::BufferUse),
-    TextureBarrier(glow::Texture, crate::TextureUse),
+    TextureBarrier(crate::TextureUse),
+    SetViewport {
+        rect: crate::Rect<i32>,
+        depth: Range<f32>,
+    },
+    SetScissor(crate::Rect<i32>),
+    SetStencilFunc {
+        face: u32,
+        function: u32,
+        reference: u32,
+        read_mask: u32,
+    },
+    SetStencilOps {
+        face: u32,
+        write_mask: u32,
+        ops: StencilOps,
+    },
+    SetVertexAttribute(AttributeDesc, VertexBufferDesc),
     InsertDebugMarker(Range<u32>),
     PushDebugGroup(Range<u32>),
     PopDebugGroup,
@@ -489,20 +574,12 @@ pub struct CommandBuffer {
     data_words: Vec<u32>,
 }
 
-#[derive(Default)]
-struct CommandState {
-    topology: u32,
-    index_format: wgt::IndexFormat,
-    index_offset: wgt::BufferAddress,
-    has_pass_label: bool,
-}
-
 //TODO: we would have something like `Arc<typed_arena::Arena>`
 // here and in the command buffers. So that everything grows
 // inside the encoder and stays there until `reset_all`.
 
 pub struct CommandEncoder {
     cmd_buffer: CommandBuffer,
-    state: CommandState,
+    state: command::State,
     private_caps: PrivateCapability,
 }

wgpu-hal/src/gles/queue.rs

@@ -292,7 +292,7 @@ impl super::Queue {
                 gl.bind_buffer(dst_target, Some(dst));
                 gl.buffer_sub_data_u8_slice(dst_target, dst_offset as i32, query_data);
             }
-            C::ResetFramebuffer(extent) => {
+            C::ResetFramebuffer => {
                 gl.bind_framebuffer(glow::DRAW_FRAMEBUFFER, Some(self.draw_fbo));
                 gl.framebuffer_texture_2d(
                     glow::DRAW_FRAMEBUFFER,
@@ -317,8 +317,6 @@ impl super::Queue {
                 gl.disable(glow::DEPTH_TEST);
                 gl.disable(glow::STENCIL_TEST);
                 gl.disable(glow::SCISSOR_TEST);
-                gl.scissor(0, 0, extent.width as i32, extent.height as i32);
-                gl.viewport(0, 0, extent.width as i32, extent.height as i32);
             }
             C::SetFramebufferAttachment {
                 attachment,
@@ -409,7 +407,7 @@ impl super::Queue {
                 }
                 gl.memory_barrier(flags);
             }
-            C::TextureBarrier(_raw, usage) => {
+            C::TextureBarrier(usage) => {
                 let mut flags = 0;
                 if usage.contains(crate::TextureUse::SAMPLED) {
                     flags |= glow::TEXTURE_FETCH_BARRIER_BIT;
@@ -431,6 +429,55 @@ impl super::Queue {
                 }
                 gl.memory_barrier(flags);
             }
+            C::SetViewport {
+                ref rect,
+                ref depth,
+            } => {
+                gl.viewport(rect.x, rect.y, rect.w, rect.h);
+                gl.depth_range_f32(depth.start, depth.end);
+            }
+            C::SetScissor(ref rect) => {
+                gl.scissor(rect.x, rect.y, rect.w, rect.h);
+            }
+            C::SetStencilFunc {
+                face,
+                function,
+                reference,
+                read_mask,
+            } => {
+                gl.stencil_func_separate(face, function, reference as i32, read_mask);
+            }
+            C::SetStencilOps {
+                face,
+                write_mask,
+                ref ops,
+            } => {
+                gl.stencil_mask_separate(face, write_mask);
+                gl.stencil_op_separate(face, ops.fail, ops.depth_fail, ops.pass);
+            }
+            C::SetVertexAttribute(ref vat, ref vb) => {
+                gl.bind_buffer(glow::ARRAY_BUFFER, Some(vb.raw));
+                let offset = vat.offset as i32 + vb.offset as i32;
+                match vat.format_desc.attrib_kind {
+                    super::VertexAttribKind::Float => gl.vertex_attrib_pointer_f32(
+                        vat.location,
+                        vat.format_desc.element_count,
+                        vat.format_desc.element_format,
+                        true, // always normalized
+                        vb.stride as i32,
+                        offset,
+                    ),
+                    super::VertexAttribKind::Integer => gl.vertex_attrib_pointer_i32(
+                        vat.location,
+                        vat.format_desc.element_count,
+                        vat.format_desc.element_format,
+                        vb.stride as i32,
+                        offset,
+                    ),
+                }
+                gl.vertex_attrib_divisor(vat.location, vb.step as u32);
+                gl.enable_vertex_attrib_array(vat.location);
+            }
             C::InsertDebugMarker(ref range) => {
                 let marker = extract_marker(data_bytes, range);
                 gl.debug_message_insert(