Support of vk-graph performance work

CHANGELOG.md

@@ -1,26 +1,23 @@
-# Changelog
+# nvidia-sdk Changelog
 
-## nvidia-nrd 0.1.1 (2026-09-10)
+This changelog covers SDK acquisition and staging. Integration release notes:
 
-- Fix repeated native compilation in downstream projects by watching SDK source
-  files individually instead of recursively watching the entire SDK directory.
+- [nvidia-dlss](https://github.com/attackgoat/nvidia/blob/main/crates/nvidia-dlss/CHANGELOG.md)
+- [nvidia-nrd](https://github.com/attackgoat/nvidia/blob/main/crates/nvidia-nrd/CHANGELOG.md)
+- [nvidia-omm](https://github.com/attackgoat/nvidia/blob/main/crates/nvidia-omm/CHANGELOG.md)
+- [nvidia-streamline](https://github.com/attackgoat/nvidia/blob/main/crates/nvidia-streamline/CHANGELOG.md)
 
 ## 0.1.0 (2026-09-05)
 
+Published `nvidia-sdk` to crates.io.
+
 Initial extraction of the NVIDIA integrations from private codebase.
 
-- `nvidia-sdk`: pinned, SHA-256-verified SDK acquisition, cache/offline support,
+- Pinned, SHA-256-verified SDK acquisition, cache/offline support,
   explicit SDK overrides, source dependency assembly, and runtime staging.
-- `nvidia-dlss`: Vulkan NGX DLSS Ray Reconstruction on x86-64 GNU Linux.
-- `nvidia-streamline`: Vulkan interposer and DLSS Ray Reconstruction on
-  x86-64 MSVC Windows.
-- `nvidia-nrd`: NRD RELAX SH on x86-64 Linux/Windows and Apple silicon macOS.
-- `nvidia-omm`: native CPU baking on x86-64 Linux/Windows and Intel/Apple
-  silicon macOS.
-- SHARC 1.8.3 shader-header acquisition on all targets through `nvidia-sdk`.
-- Default-enabled optional `vk-graph` adapters and graph-independent Vulkan APIs.
-- SDK-free documentation builds; no NVIDIA SDK payloads vendored in these crates.
+- SHARC 1.8.3 shader-header acquisition on all targets.
+- SDK-free documentation builds; no NVIDIA SDK payloads vendored in the crate.
 
 The declared minimum Rust version is 1.95, including default-feature dependencies.
-Native SDK components retain their upstream licenses separately from these
-MIT OR Apache-2.0 bindings.
+Native SDK components retain their upstream licenses separately from this
+MIT OR Apache-2.0 crate.

Cargo.toml

@@ -27,7 +27,7 @@ serde_json = "1"
 sha2 = "0.10"
 toml = "0.9"
 ureq = { version = "2.12", default-features = false, features = ["tls"] }
-vk-graph = "0.14.7"
+vk-graph = { version = "0.15", git = "https://github.com/attackgoat/vk-graph", branch = "performance" }
 zip = { version = "4.6.1", default-features = false, features = ["deflate"] }
 
 [workspace.lints.rust]

README.md

@@ -15,7 +15,7 @@ Rust integrations for selected NVIDIA graphics SDKs.
 | OMM CPU baker | [![nvidia-omm](https://img.shields.io/crates/v/nvidia-omm.svg)](https://crates.io/crates/nvidia-omm) | [![docs.rs](https://docs.rs/nvidia-omm/badge.svg)](https://docs.rs/nvidia-omm) |
 | Streamline | [![nvidia-streamline](https://img.shields.io/crates/v/nvidia-streamline.svg)](https://crates.io/crates/nvidia-streamline) | [![docs.rs](https://docs.rs/nvidia-streamline/badge.svg)](https://docs.rs/nvidia-streamline) |
 
-See the [changelog](https://github.com/attackgoat/nvidia/blob/main/CHANGELOG.md)
+See the [SDK changelog and integration changelog links](https://github.com/attackgoat/nvidia/blob/main/CHANGELOG.md)
 and [release checklist](https://github.com/attackgoat/nvidia/blob/main/RELEASE.md).
 
 > [!NOTE]

crates/nvidia-dlss/CHANGELOG.md

@@ -0,0 +1,20 @@
+# nvidia-dlss Changelog
+
+## 0.2.0 (Unreleased)
+
+- **Breaking:** Update the optional `vk-graph` adapter from 0.14.7 to 0.15.
+
+Publication is pending the crates.io release of `vk-graph` 0.15 and completion
+of the release gates.
+
+## 0.1.0 (2026-09-05)
+
+Published `nvidia-dlss` to crates.io. Initial extraction from private codebase.
+
+- Vulkan NGX DLSS Ray Reconstruction on x86-64 GNU Linux.
+- Default-enabled optional `vk-graph` 0.14.7 adapter and graph-independent Vulkan APIs.
+- SDK-free documentation builds; no NVIDIA SDK payloads vendored in the crate.
+
+The declared minimum Rust version is 1.95, including default-feature dependencies.
+Native SDK components retain their upstream licenses separately from these
+MIT OR Apache-2.0 bindings.

crates/nvidia-dlss/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "nvidia-dlss"
-version.workspace = true
+version = "0.2.0"
 edition.workspace = true
 rust-version.workspace = true
 license.workspace = true

crates/nvidia-dlss/native/ngx.cpp

@@ -26,6 +26,8 @@ constexpr uint32_t kSpecularHitDistance = 1;
 constexpr uint32_t kDlssRrQualityBalanced = 0;
 constexpr uint32_t kDlssRrQualityQuality = 1;
 constexpr uint32_t kDlssRrQualityDlaa = 2;
+constexpr uint32_t kDlssRrQualityPerformance = 3;
+constexpr uint32_t kDlssRrQualityUltraPerformance = 4;
 
 struct NgxIdentity
 {
@@ -273,7 +275,7 @@ bool validReflectionGuide(const NgxDlssRrResources& resources)
 
 bool validDlssRrQuality(uint32_t quality)
 {
-    return quality <= kDlssRrQualityDlaa;
+    return quality <= kDlssRrQualityUltraPerformance;
 }
 
 NVSDK_NGX_PerfQuality_Value dlssRrQuality(uint32_t quality)
@@ -284,6 +286,10 @@ NVSDK_NGX_PerfQuality_Value dlssRrQuality(uint32_t quality)
         return NVSDK_NGX_PerfQuality_Value_MaxQuality;
     case kDlssRrQualityDlaa:
         return NVSDK_NGX_PerfQuality_Value_DLAA;
+    case kDlssRrQualityPerformance:
+        return NVSDK_NGX_PerfQuality_Value_MaxPerf;
+    case kDlssRrQualityUltraPerformance:
+        return NVSDK_NGX_PerfQuality_Value_UltraPerformance;
     default:
         return NVSDK_NGX_PerfQuality_Value_Balanced;
     }
@@ -351,6 +357,14 @@ NVSDK_NGX_Result ensureFeature(
         NVSDK_NGX_RayReconstruction_Hint_Render_Preset_E);
 
     NVSDK_NGX_DLSSD_Create_Params create{};
+    NVSDK_NGX_Parameter_SetUI(
+        context->parameters,
+        NVSDK_NGX_Parameter_RayReconstruction_Hint_Render_Preset_Performance,
+        NVSDK_NGX_RayReconstruction_Hint_Render_Preset_E);
+    NVSDK_NGX_Parameter_SetUI(
+        context->parameters,
+        NVSDK_NGX_Parameter_RayReconstruction_Hint_Render_Preset_UltraPerformance,
+        NVSDK_NGX_RayReconstruction_Hint_Render_Preset_E);
     create.InDenoiseMode = NVSDK_NGX_DLSS_Denoise_Mode_DLUnified;
     create.InRoughnessMode = NVSDK_NGX_DLSS_Roughness_Mode_Packed;
     create.InUseHWDepth = context->config.depth == 2

crates/nvidia-dlss/native/ngx_test.cpp

@@ -35,6 +35,7 @@ static float expectedScale[2];
 static int expectedFlags;
 static unsigned expectedDepth;
 static unsigned expectedGuide;
+static int expectedQuality;
 static unsigned evaluations;
 static unsigned creations;
 
@@ -57,6 +58,7 @@ static NVSDK_NGX_Result captureCreation(
     assert(parameter<unsigned>(params, NVSDK_NGX_Parameter_Height) == 32);
     assert(parameter<unsigned>(params, NVSDK_NGX_Parameter_OutWidth) == 128);
     assert(parameter<unsigned>(params, NVSDK_NGX_Parameter_OutHeight) == 64);
+    assert(parameter<int>(params, NVSDK_NGX_Parameter_PerfQualityValue) == expectedQuality);
     *handle = reinterpret_cast<NVSDK_NGX_Handle*>(2);
     ++creations;
     return NVSDK_NGX_Result_Success;
@@ -121,7 +123,12 @@ int main()
     for (unsigned color : {0, 1})
     for (unsigned motion : {0, 1})
     for (unsigned depth : {0, 1, 2})
+    for (const auto mode : {
+             std::pair{kDlssRrQualityBalanced, NVSDK_NGX_PerfQuality_Value_Balanced},
+             std::pair{kDlssRrQualityPerformance, NVSDK_NGX_PerfQuality_Value_MaxPerf},
+             std::pair{kDlssRrQualityUltraPerformance, NVSDK_NGX_PerfQuality_Value_UltraPerformance}})
     {
+        expectedQuality = mode.second;
         Parameters params;
         NgxContext context(L"", L"", nullptr);
         context.parameters = &params;
@@ -145,7 +152,7 @@ int main()
         NgxFrameConstants frame{};
         auto evaluate = [&]() {
             return nvidia_dlss_ngx_evaluate(&context, reinterpret_cast<VkCommandBuffer>(3),
-                kDlssRrQualityBalanced, &frame, &resources);
+                mode.first, &frame, &resources);
         };
         for (const auto scale : {std::pair{1.0f, 1.0f}, std::pair{64.0f, -32.0f},
                  std::pair{0.5f, 0.0f}, std::pair{0.0f, -0.5f}, std::pair{0.0f, -0.0f}})
@@ -188,7 +195,7 @@ int main()
         }
         assert(evaluations == before && creations == created);
     }
-    assert(creations == 12 && evaluations == 120);
+    assert(creations == 36 && evaluations == 360);
     assert(!validConfig({2, 0, 0}));
     assert(!validConfig({0, 2, 0}));
     assert(!validConfig({0, 0, 3}));

crates/nvidia-dlss/src/lib.rs

@@ -90,6 +90,8 @@ pub enum DlssRrQuality {
     Quality = 1,
     /// Native-resolution anti-aliasing; render and output dimensions must be equal.
     Dlaa = 2,
+    Performance = 3,
+    UltraPerformance = 4,
 }
 
 /// Images for NGX evaluation. All inputs, including depth and the reflection guide,
@@ -567,6 +569,8 @@ mod tests {
         assert_eq!(DlssRrQuality::Balanced as u32, 0);
         assert_eq!(DlssRrQuality::Quality as u32, 1);
         assert_eq!(DlssRrQuality::Dlaa as u32, 2);
+        assert_eq!(DlssRrQuality::Performance as u32, 3);
+        assert_eq!(DlssRrQuality::UltraPerformance as u32, 4);
 
         assert_eq!(size_of::<VulkanImage>(), 64);
         assert_eq!(align_of::<VulkanImage>(), 8);

crates/nvidia-nrd/CHANGELOG.md

@@ -0,0 +1,25 @@
+# nvidia-nrd Changelog
+
+## 0.2.0 (Unreleased)
+
+- **Breaking:** Update the optional `vk-graph` adapter from 0.14.7 to 0.15.
+
+Publication is pending the crates.io release of `vk-graph` 0.15 and completion
+of the release gates.
+
+## 0.1.1 (2026-09-10)
+
+- Fix repeated native compilation in downstream projects by watching SDK source
+  files individually instead of recursively watching the entire SDK directory.
+
+## 0.1.0 (2026-09-05)
+
+Published `nvidia-nrd` to crates.io. Initial extraction from private codebase.
+
+- NRD RELAX SH on x86-64 Linux/Windows and Apple silicon macOS.
+- Default-enabled optional `vk-graph` 0.14.7 adapter and graph-independent Vulkan APIs.
+- SDK-free documentation builds; no NVIDIA SDK payloads vendored in the crate.
+
+The declared minimum Rust version is 1.95, including default-feature dependencies.
+Native SDK components retain their upstream licenses separately from these
+MIT OR Apache-2.0 bindings.

crates/nvidia-nrd/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "nvidia-nrd"
-version = "0.1.1"
+version = "0.2.0"
 edition.workspace = true
 rust-version.workspace = true
 license.workspace = true

crates/nvidia-nrd/native/nrd.cpp

@@ -57,6 +57,10 @@ struct NrdFrame {
     float denoising_range;
     uint32_t width;
     uint32_t height;
+    uint32_t resource_width;
+    uint32_t resource_height;
+    uint32_t previous_width;
+    uint32_t previous_height;
     uint32_t frame_index;
     uint32_t reset;
     struct {
@@ -75,7 +79,7 @@ struct NrdFrame {
 
 static_assert(sizeof(NrdImage) == 16);
 static_assert(sizeof(NrdResources) == 176);
-static_assert(sizeof(NrdFrame) == 344);
+static_assert(sizeof(NrdFrame) == 360);
 
 struct Context {
     nrd::Integration integration;
@@ -189,7 +193,13 @@ extern "C" int32_t nvidia_nrd_evaluate(
 
     try {
         Context& context = *static_cast<Context*>(opaque);
-        int32_t result = recreate(context, frame->width, frame->height);
+        if (frame->width == 0 || frame->height == 0
+            || frame->previous_width == 0 || frame->previous_height == 0
+            || frame->width > frame->resource_width || frame->height > frame->resource_height
+            || frame->previous_width > frame->resource_width || frame->previous_height > frame->resource_height)
+            return -2;
+
+        int32_t result = recreate(context, frame->resource_width, frame->resource_height);
         if (result != 0)
             return result;
 
@@ -205,10 +215,10 @@ extern "C" int32_t nvidia_nrd_evaluate(
         common.resourceSize[1] = context.height;
         common.resourceSizePrev[0] = context.width;
         common.resourceSizePrev[1] = context.height;
-        common.rectSize[0] = context.width;
-        common.rectSize[1] = context.height;
-        common.rectSizePrev[0] = context.width;
-        common.rectSizePrev[1] = context.height;
+        common.rectSize[0] = static_cast<uint16_t>(frame->width);
+        common.rectSize[1] = static_cast<uint16_t>(frame->height);
+        common.rectSizePrev[0] = static_cast<uint16_t>(frame->previous_width);
+        common.rectSizePrev[1] = static_cast<uint16_t>(frame->previous_height);
         common.denoisingRange = frame->denoising_range;
         common.frameIndex = frame->frame_index;
         common.accumulationMode = frame->reset ? nrd::AccumulationMode::RESTART : nrd::AccumulationMode::CONTINUE;

crates/nvidia-nrd/src/lib.rs

@@ -37,6 +37,24 @@ pub const SDK_VERSION: &str = "4.17.3";
 #[cfg(any(nvidia_nrd_native, test))]
 const VULKAN_1_4: u32 = vk::make_api_version(0, 1, 4, 0);
 
+#[cfg(any(nvidia_nrd_native, test))]
+fn validate_rectangles(
+    resource: [u32; 2],
+    current: [u32; 2],
+    previous: [u32; 2],
+) -> anyhow::Result<()> {
+    anyhow::ensure!(
+        (0..2).all(|axis| resource[axis] > 0
+            && resource[axis] <= u32::from(u16::MAX)
+            && current[axis] > 0
+            && current[axis] <= resource[axis]
+            && previous[axis] > 0
+            && previous[axis] <= resource[axis]),
+        "nrd active rectangles must fit nonzero u16 resource dimensions"
+    );
+    Ok(())
+}
+
 #[derive(Clone, Copy, Debug)]
 #[repr(C)]
 pub struct Frame {
@@ -50,6 +68,12 @@ pub struct Frame {
     pub denoising_range: f32,
     pub width: u32,
     pub height: u32,
+    /// Fixed allocation dimensions shared by every input and output image.
+    pub resource_width: u32,
+    pub resource_height: u32,
+    /// Previous frame's active rectangle, for dynamic-resolution reprojection.
+    pub previous_width: u32,
+    pub previous_height: u32,
     pub frame_index: u32,
     pub reset: u32,
     pub settings: RelaxSettings,
@@ -184,6 +208,26 @@ impl Resources {
 mod tests {
     use super::*;
 
+    #[test]
+    fn dynamic_rectangles_must_fit_the_fixed_allocation() {
+        for (current, previous) in [
+            ([960, 540], [427, 240]),
+            ([427, 240], [960, 540]),
+            ([693, 390], [691, 389]),
+        ] {
+            assert!(validate_rectangles([960, 540], current, previous).is_ok());
+        }
+        for (resource, current, previous) in [
+            ([960, 540], [0, 240], [960, 540]),
+            ([960, 540], [961, 540], [960, 540]),
+            ([960, 540], [427, 240], [960, 541]),
+            ([0, 540], [1, 240], [1, 240]),
+            ([65536, 540], [427, 240], [427, 240]),
+        ] {
+            assert!(validate_rectangles(resource, current, previous).is_err());
+        }
+    }
+
     #[test]
     fn rejects_unsupported_api_versions_without_ffi() {
         for version in [
@@ -277,7 +321,7 @@ mod tests {
         assert_eq!(size_of::<Image>(), 16);
         assert_eq!(size_of::<Resources>(), 176);
         assert_eq!(size_of::<RelaxSettings>(), 40);
-        assert_eq!(size_of::<Frame>(), 344);
+        assert_eq!(size_of::<Frame>(), 360);
     }
 
     #[test]

crates/nvidia-nrd/src/native.rs

@@ -1,13 +1,13 @@
-#[cfg(feature = "vk-graph")]
-use vk_graph::driver::device::Device;
-
 use {
-    crate::{Frame, Resources},
+    crate::{Frame, Resources, validate_rectangles},
     anyhow::Context as _,
     ash::vk::{self, Handle as _},
     std::{ffi::c_void, ptr::NonNull, sync::Mutex},
 };
 
+#[cfg(feature = "vk-graph")]
+use vk_graph::driver::device::Device;
+
 static API_LOCK: Mutex<()> = Mutex::new(());
 
 unsafe extern "C" {
@@ -165,6 +165,11 @@ impl Nrd {
         resources: &Resources,
     ) -> anyhow::Result<()> {
         resources.validate_layouts()?;
+        validate_rectangles(
+            [frame.resource_width, frame.resource_height],
+            [frame.width, frame.height],
+            [frame.previous_width, frame.previous_height],
+        )?;
 
         let context = self.context.context("nrd has been shut down")?;
         let _guard = API_LOCK

crates/nvidia-omm/CHANGELOG.md

@@ -0,0 +1,12 @@
+# nvidia-omm Changelog
+
+## 0.1.0 (2026-09-05)
+
+Published `nvidia-omm` to crates.io. Initial extraction from private codebase.
+
+- Native CPU baking on x86-64 Linux/Windows and Intel/Apple silicon macOS.
+- SDK-free documentation builds; no NVIDIA SDK payloads vendored in the crate.
+
+The declared minimum Rust version is 1.95, including default-feature dependencies.
+Native SDK components retain their upstream licenses separately from these
+MIT OR Apache-2.0 bindings.

crates/nvidia-sdk/CHANGELOG.md

@@ -0,0 +1 @@
+../../CHANGELOG.md