Add SHADER_FLOAT16 feature (#2646)

* Add SHADER_FLOAT16 feature

* vk: check SHADER_FLOAT16 feature

deno_webgpu/02_idl_types.js

@@ -120,6 +120,7 @@
       "texture-compression-astc",
       "timestamp-query",
       "indirect-first-instance",
+      "shader-f16",
       // extended from spec
       "mappable-primary-buffers",
       "texture-binding-array",

deno_webgpu/src/lib.rs

@@ -146,6 +146,9 @@ fn deserialize_features(features: &wgpu_types::Features) -> Vec<&'static str> {
     if features.contains(wgpu_types::Features::INDIRECT_FIRST_INSTANCE) {
         return_features.push("indirect-first-instance");
     }
+    if features.contains(wgpu_types::Features::SHADER_FLOAT16) {
+        return_features.push("shader-f16")
+    }
 
     // extended from spec
     if features.contains(wgpu_types::Features::MAPPABLE_PRIMARY_BUFFERS) {
@@ -306,6 +309,10 @@ impl From<GpuRequiredFeatures> for wgpu_types::Features {
             wgpu_types::Features::INDIRECT_FIRST_INSTANCE,
             required_features.0.contains("indirect-first-instance"),
         );
+        features.set(
+            wgpu_types::Features::SHADER_FLOAT16,
+            required_features.0.contains("shader-f16"),
+        );
 
         // extended from spec
         features.set(

deno_webgpu/webgpu.idl

@@ -90,6 +90,7 @@ enum GPUFeatureName {
     "texture-compression-astc",
     "timestamp-query",
     "indirect-first-instance",
+    "shader-f16",
 };
 
 [Exposed=(Window, DedicatedWorker), SecureContext]

wgpu-hal/src/metal/adapter.rs

@@ -751,7 +751,8 @@ impl super::PrivateCapabilities {
             | F::PUSH_CONSTANTS
             | F::POLYGON_MODE_LINE
             | F::CLEAR_TEXTURE
-            | F::TEXTURE_FORMAT_16BIT_NORM;
+            | F::TEXTURE_FORMAT_16BIT_NORM
+            | F::SHADER_FLOAT16;
 
         features.set(F::TEXTURE_COMPRESSION_ASTC_LDR, self.format_astc);
         features.set(F::TEXTURE_COMPRESSION_ASTC_HDR, self.format_astc_hdr);

wgpu-hal/src/vulkan/adapter.rs

@@ -25,6 +25,10 @@ pub struct PhysicalDeviceFeatures {
     depth_clip_enable: Option<vk::PhysicalDeviceDepthClipEnableFeaturesEXT>,
     multiview: Option<vk::PhysicalDeviceMultiviewFeaturesKHR>,
     astc_hdr: Option<vk::PhysicalDeviceTextureCompressionASTCHDRFeaturesEXT>,
+    shader_float16: Option<(
+        vk::PhysicalDeviceShaderFloat16Int8Features,
+        vk::PhysicalDevice16BitStorageFeatures,
+    )>,
 }
 
 // This is safe because the structs have `p_next: *mut c_void`, which we null out/never read.
@@ -62,6 +66,10 @@ impl PhysicalDeviceFeatures {
         if let Some(ref mut feature) = self.astc_hdr {
             info = info.push_next(feature);
         }
+        if let Some((ref mut f16_i8_feature, ref mut _16bit_feature)) = self.shader_float16 {
+            info = info.push_next(f16_i8_feature);
+            info = info.push_next(_16bit_feature);
+        }
         info
     }
 
@@ -326,6 +334,19 @@ impl PhysicalDeviceFeatures {
             } else {
                 None
             },
+            shader_float16: if requested_features.contains(wgt::Features::SHADER_FLOAT16) {
+                Some((
+                    vk::PhysicalDeviceShaderFloat16Int8Features::builder()
+                        .shader_float16(true)
+                        .build(),
+                    vk::PhysicalDevice16BitStorageFeatures::builder()
+                        .storage_buffer16_bit_access(true)
+                        .uniform_and_storage_buffer16_bit_access(true)
+                        .build(),
+                ))
+            } else {
+                None
+            },
         }
     }
 
@@ -531,6 +552,15 @@ impl PhysicalDeviceFeatures {
             );
         }
 
+        if let Some((ref f16_i8, ref bit16)) = self.shader_float16 {
+            features.set(
+                F::SHADER_FLOAT16,
+                f16_i8.shader_float16 != 0
+                    && bit16.storage_buffer16_bit_access != 0
+                    && bit16.uniform_and_storage_buffer16_bit_access != 0,
+            );
+        }
+
         (features, dl_flags)
     }
 
@@ -642,7 +672,12 @@ impl PhysicalDeviceCapabilities {
         extensions.push(vk::KhrPortabilitySubsetFn::name());
 
         if requested_features.contains(wgt::Features::TEXTURE_COMPRESSION_ASTC_HDR) {
-            extensions.push(vk::ExtTextureCompressionAstcHdrFn::name())
+            extensions.push(vk::ExtTextureCompressionAstcHdrFn::name());
+        }
+
+        if requested_features.contains(wgt::Features::SHADER_FLOAT16) {
+            extensions.push(vk::KhrShaderFloat16Int8Fn::name());
+            extensions.push(vk::Khr16bitStorageFn::name());
         }
 
         extensions
@@ -883,6 +918,16 @@ impl super::InstanceShared {
                     .insert(vk::PhysicalDeviceTextureCompressionASTCHDRFeaturesEXT::default());
                 builder = builder.push_next(next);
             }
+            if capabilities.supports_extension(vk::KhrShaderFloat16Int8Fn::name())
+                && capabilities.supports_extension(vk::Khr16bitStorageFn::name())
+            {
+                let next = features.shader_float16.insert((
+                    vk::PhysicalDeviceShaderFloat16Int8FeaturesKHR::default(),
+                    vk::PhysicalDevice16BitStorageFeaturesKHR::default(),
+                ));
+                builder = builder.push_next(&mut next.0);
+                builder = builder.push_next(&mut next.1);
+            }
 
             let mut features2 = builder.build();
             unsafe {

wgpu-types/src/lib.rs

@@ -241,6 +241,16 @@ bitflags::bitflags! {
         ///
         /// This is a web and native feature.
         const PIPELINE_STATISTICS_QUERY = 1 << 4;
+        /// Allows shaders to acquire the FP16 ability
+        ///
+        /// Note: this is not supported in naga yet，only through spir-v passthrough right now.
+        ///
+        /// Supported Platforms:
+        /// - Vulkan
+        /// - Metal
+        ///
+        /// This is a web and native feature.
+        const SHADER_FLOAT16 = 1 << 5;
         /// Webgpu only allows the MAP_READ and MAP_WRITE buffer usage to be matched with
         /// COPY_DST and COPY_SRC respectively. This removes this requirement.
         ///