hal/vulkan: Fix multi draw indirect emulating and respect maxDrawIndirectCount (#8665)

Author: beicause

wgpu-hal/src/vulkan/adapter.rs

@@ -1912,6 +1912,7 @@ impl super::Instance {
                 depth_stencil_required_flags(),
             ),
             multi_draw_indirect: phd_features.core.multi_draw_indirect != 0,
+            max_draw_indirect_count: phd_capabilities.properties.limits.max_draw_indirect_count,
             non_coherent_map_mask: phd_capabilities.properties.limits.non_coherent_atom_size - 1,
             can_present: true,
             //TODO: make configurable

wgpu-hal/src/vulkan/command.rs

@@ -1129,23 +1129,46 @@ impl crate::CommandEncoder for super::CommandEncoder {
         offset: wgt::BufferAddress,
         draw_count: u32,
     ) {
-        unsafe {
-            self.device.raw.cmd_draw_indirect(
-                self.active,
-                buffer.raw,
-                offset,
-                draw_count,
-                size_of::<wgt::DrawIndirectArgs>() as u32,
-            )
-        };
+        if draw_count >= 1
+            && self.device.private_caps.multi_draw_indirect
+            && draw_count <= self.device.private_caps.max_draw_indirect_count
+        {
+            unsafe {
+                self.device.raw.cmd_draw_indirect(
+                    self.active,
+                    buffer.raw,
+                    offset,
+                    draw_count,
+                    size_of::<wgt::DrawIndirectArgs>() as u32,
+                )
+            };
+        } else {
+            for i in 0..draw_count {
+                let indirect_offset = offset
+                    + i as wgt::BufferAddress
+                        * size_of::<wgt::DrawIndirectArgs>() as wgt::BufferAddress;
+                unsafe {
+                    self.device.raw.cmd_draw_indirect(
+                        self.active,
+                        buffer.raw,
+                        indirect_offset,
+                        1,
+                        size_of::<wgt::DrawIndirectArgs>() as u32,
+                    )
+                };
+            }
+        }
     }
     unsafe fn draw_indexed_indirect(
         &mut self,
         buffer: &super::Buffer,
         offset: wgt::BufferAddress,
         draw_count: u32,
     ) {
-        if draw_count >= 1 && self.device.private_caps.multi_draw_indirect {
+        if draw_count >= 1
+            && self.device.private_caps.multi_draw_indirect
+            && draw_count <= self.device.private_caps.max_draw_indirect_count
+        {
             unsafe {
                 self.device.raw.cmd_draw_indexed_indirect(
                     self.active,
@@ -1156,12 +1179,15 @@ impl crate::CommandEncoder for super::CommandEncoder {
                 )
             };
         } else {
-            for _ in 0..draw_count {
+            for i in 0..draw_count {
+                let indirect_offset = offset
+                    + i as wgt::BufferAddress
+                        * size_of::<wgt::DrawIndexedIndirectArgs>() as wgt::BufferAddress;
                 unsafe {
                     self.device.raw.cmd_draw_indexed_indirect(
                         self.active,
                         buffer.raw,
-                        offset,
+                        indirect_offset,
                         1,
                         size_of::<wgt::DrawIndexedIndirectArgs>() as u32,
                     )

wgpu-hal/src/vulkan/mod.rs

@@ -319,6 +319,7 @@ struct PrivateCapabilities {
     can_present: bool,
     non_coherent_map_mask: wgt::BufferAddress,
     multi_draw_indirect: bool,
+    max_draw_indirect_count: u32,
 
     /// True if this adapter advertises the [`robustBufferAccess`][vrba] feature.
     ///