Merge pull request #11789 from Kelebek1/spirv_shift_right

Manually robust on Maxwell and earlier

src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp

@@ -111,16 +111,33 @@ Id GetCbuf(EmitContext& ctx, Id result_type, Id UniformDefinitions::*member_ptr,
     } else if (element_size > 1) {
         const u32 log2_element_size{static_cast<u32>(std::countr_zero(element_size))};
         const Id shift{ctx.Const(log2_element_size)};
-        buffer_offset = ctx.OpShiftRightArithmetic(ctx.U32[1], ctx.Def(offset), shift);
+        buffer_offset = ctx.OpShiftRightLogical(ctx.U32[1], ctx.Def(offset), shift);
     } else {
         buffer_offset = ctx.Def(offset);
     }
     if (!binding.IsImmediate()) {
         return ctx.OpFunctionCall(result_type, indirect_func, ctx.Def(binding), buffer_offset);
     }
+
     const Id cbuf{ctx.cbufs[binding.U32()].*member_ptr};
     const Id access_chain{ctx.OpAccessChain(uniform_type, cbuf, ctx.u32_zero_value, buffer_offset)};
-    return ctx.OpLoad(result_type, access_chain);
+    const Id val = ctx.OpLoad(result_type, access_chain);
+
+    if (offset.IsImmediate() || !ctx.profile.has_broken_robust) {
+        return val;
+    }
+
+    const auto is_float = UniformDefinitions::IsFloat(member_ptr);
+    const auto num_elements = UniformDefinitions::NumElements(member_ptr);
+    const std::array zero_vec{
+        is_float ? ctx.Const(0.0f) : ctx.Const(0u),
+        is_float ? ctx.Const(0.0f) : ctx.Const(0u),
+        is_float ? ctx.Const(0.0f) : ctx.Const(0u),
+        is_float ? ctx.Const(0.0f) : ctx.Const(0u),
+    };
+    const Id cond = ctx.OpULessThanEqual(ctx.TypeBool(), buffer_offset, ctx.Const(0xFFFFu));
+    const Id zero = ctx.OpCompositeConstruct(result_type, std::span(zero_vec.data(), num_elements));
+    return ctx.OpSelect(result_type, cond, val, zero);
 }
 
 Id GetCbufU32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
@@ -138,7 +155,7 @@ Id GetCbufElement(EmitContext& ctx, Id vector, const IR::Value& offset, u32 inde
         const u32 element{(offset.U32() / 4) % 4 + index_offset};
         return ctx.OpCompositeExtract(ctx.U32[1], vector, element);
     }
-    const Id shift{ctx.OpShiftRightArithmetic(ctx.U32[1], ctx.Def(offset), ctx.Const(2u))};
+    const Id shift{ctx.OpShiftRightLogical(ctx.U32[1], ctx.Def(offset), ctx.Const(2u))};
     Id element{ctx.OpBitwiseAnd(ctx.U32[1], shift, ctx.Const(3u))};
     if (index_offset > 0) {
         element = ctx.OpIAdd(ctx.U32[1], element, ctx.Const(index_offset));

src/shader_recompiler/backend/spirv/spirv_emit_context.h

@@ -64,6 +64,42 @@ struct UniformDefinitions {
     Id F32{};
     Id U32x2{};
     Id U32x4{};
+
+    constexpr static size_t NumElements(Id UniformDefinitions::*member_ptr) {
+        if (member_ptr == &UniformDefinitions::U8) {
+            return 1;
+        }
+        if (member_ptr == &UniformDefinitions::S8) {
+            return 1;
+        }
+        if (member_ptr == &UniformDefinitions::U16) {
+            return 1;
+        }
+        if (member_ptr == &UniformDefinitions::S16) {
+            return 1;
+        }
+        if (member_ptr == &UniformDefinitions::U32) {
+            return 1;
+        }
+        if (member_ptr == &UniformDefinitions::F32) {
+            return 1;
+        }
+        if (member_ptr == &UniformDefinitions::U32x2) {
+            return 2;
+        }
+        if (member_ptr == &UniformDefinitions::U32x4) {
+            return 4;
+        }
+        ASSERT(false);
+        return 1;
+    }
+
+    constexpr static bool IsFloat(Id UniformDefinitions::*member_ptr) {
+        if (member_ptr == &UniformDefinitions::F32) {
+            return true;
+        }
+        return false;
+    }
 };
 
 struct StorageTypeDefinition {

src/shader_recompiler/profile.h

@@ -9,7 +9,6 @@ namespace Shader {
 
 struct Profile {
     u32 supported_spirv{0x00010000};
-
     bool unified_descriptor_binding{};
     bool support_descriptor_aliasing{};
     bool support_int8{};
@@ -82,6 +81,9 @@ struct Profile {
     bool has_broken_spirv_subgroup_mask_vector_extract_dynamic{};
 
     u32 gl_max_compute_smem_size{};
+
+    /// Maxwell and earlier nVidia architectures have broken robust support
+    bool has_broken_robust{};
 };
 
 } // namespace Shader

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -356,7 +356,11 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, const Device& device
         .has_broken_fp16_float_controls = driver_id == VK_DRIVER_ID_NVIDIA_PROPRIETARY,
         .ignore_nan_fp_comparisons = false,
         .has_broken_spirv_subgroup_mask_vector_extract_dynamic =
-            driver_id == VK_DRIVER_ID_QUALCOMM_PROPRIETARY};
+            driver_id == VK_DRIVER_ID_QUALCOMM_PROPRIETARY,
+        .has_broken_robust =
+            device.IsNvidia() && device.GetNvidiaArch() <= NvidiaArchitecture::Arch_Maxwell,
+    };
+
     host_info = Shader::HostTranslateInfo{
         .support_float64 = device.IsFloat64Supported(),
         .support_float16 = device.IsFloat16Supported(),

src/video_core/vulkan_common/vulkan_device.cpp

@@ -83,15 +83,6 @@ constexpr std::array VK_FORMAT_A4B4G4R4_UNORM_PACK16{
 
 } // namespace Alternatives
 
-enum class NvidiaArchitecture {
-    KeplerOrOlder,
-    Maxwell,
-    Pascal,
-    Volta,
-    Turing,
-    AmpereOrNewer,
-};
-
 template <typename T>
 void SetNext(void**& next, T& data) {
     *next = &data;
@@ -326,9 +317,9 @@ NvidiaArchitecture GetNvidiaArchitecture(vk::PhysicalDevice physical,
         if (shading_rate_props.primitiveFragmentShadingRateWithMultipleViewports) {
             // Only Ampere and newer support this feature
             // TODO: Find a way to differentiate Ampere and Ada
-            return NvidiaArchitecture::AmpereOrNewer;
+            return NvidiaArchitecture::Arch_AmpereOrNewer;
         }
-        return NvidiaArchitecture::Turing;
+        return NvidiaArchitecture::Arch_Turing;
     }
 
     if (exts.contains(VK_EXT_BLEND_OPERATION_ADVANCED_EXTENSION_NAME)) {
@@ -340,7 +331,7 @@ NvidiaArchitecture GetNvidiaArchitecture(vk::PhysicalDevice physical,
         physical_properties.pNext = &advanced_blending_props;
         physical.GetProperties2(physical_properties);
         if (advanced_blending_props.advancedBlendMaxColorAttachments == 1) {
-            return NvidiaArchitecture::Maxwell;
+            return NvidiaArchitecture::Arch_Maxwell;
         }
 
         if (exts.contains(VK_EXT_CONSERVATIVE_RASTERIZATION_EXTENSION_NAME)) {
@@ -350,13 +341,13 @@ NvidiaArchitecture GetNvidiaArchitecture(vk::PhysicalDevice physical,
             physical_properties.pNext = &conservative_raster_props;
             physical.GetProperties2(physical_properties);
             if (conservative_raster_props.degenerateLinesRasterized) {
-                return NvidiaArchitecture::Volta;
+                return NvidiaArchitecture::Arch_Volta;
             }
-            return NvidiaArchitecture::Pascal;
+            return NvidiaArchitecture::Arch_Pascal;
         }
     }
 
-    return NvidiaArchitecture::KeplerOrOlder;
+    return NvidiaArchitecture::Arch_KeplerOrOlder;
 }
 
 std::vector<const char*> ExtensionListForVulkan(
@@ -436,6 +427,10 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
         throw vk::Exception(VK_ERROR_INCOMPATIBLE_DRIVER);
     }
 
+    if (is_nvidia) {
+        nvidia_arch = GetNvidiaArchitecture(physical, supported_extensions);
+    }
+
     SetupFamilies(surface);
     const auto queue_cis = GetDeviceQueueCreateInfos();
 
@@ -532,11 +527,11 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
 
     if (is_nvidia) {
         const u32 nv_major_version = (properties.properties.driverVersion >> 22) & 0x3ff;
-        const auto arch = GetNvidiaArchitecture(physical, supported_extensions);
+        const auto arch = GetNvidiaArch();
-        if (arch >= NvidiaArchitecture::AmpereOrNewer) {
+        if (arch >= NvidiaArchitecture::Arch_AmpereOrNewer) {
             LOG_WARNING(Render_Vulkan, "Ampere and newer have broken float16 math");
             features.shader_float16_int8.shaderFloat16 = false;
-        } else if (arch <= NvidiaArchitecture::Volta) {
+        } else if (arch <= NvidiaArchitecture::Arch_Volta) {
             if (nv_major_version < 527) {
                 LOG_WARNING(Render_Vulkan, "Volta and older have broken VK_KHR_push_descriptor");
                 RemoveExtension(extensions.push_descriptor, VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME);
@@ -686,8 +681,8 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
             RemoveExtension(extensions.push_descriptor, VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME);
         }
     } else if (extensions.push_descriptor && is_nvidia) {
-        const auto arch = GetNvidiaArchitecture(physical, supported_extensions);
+        const auto arch = GetNvidiaArch();
-        if (arch <= NvidiaArchitecture::Pascal) {
+        if (arch <= NvidiaArchitecture::Arch_Pascal) {
             LOG_WARNING(Render_Vulkan,
                         "Pascal and older architectures have broken VK_KHR_push_descriptor");
             RemoveExtension(extensions.push_descriptor, VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME);

src/video_core/vulkan_common/vulkan_device.h

@@ -177,6 +177,15 @@ enum class FormatType { Linear, Optimal, Buffer };
 /// Subgroup size of the guest emulated hardware (Nvidia has 32 threads per subgroup).
 const u32 GuestWarpSize = 32;
 
+enum class NvidiaArchitecture {
+    Arch_KeplerOrOlder,
+    Arch_Maxwell,
+    Arch_Pascal,
+    Arch_Volta,
+    Arch_Turing,
+    Arch_AmpereOrNewer,
+};
+
 /// Handles data specific to a physical device.
 class Device {
 public:
@@ -670,6 +679,14 @@ public:
         return false;
     }
 
+    bool IsNvidia() const noexcept {
+        return properties.driver.driverID == VK_DRIVER_ID_NVIDIA_PROPRIETARY;
+    }
+
+    NvidiaArchitecture GetNvidiaArch() const noexcept {
+        return nvidia_arch;
+    }
+
 private:
     /// Checks if the physical device is suitable and configures the object state
     /// with all necessary info about its properties.
@@ -788,6 +805,7 @@ private:
     bool supports_conditional_barriers{};      ///< Allows barriers in conditional control flow.
     u64 device_access_memory{};                ///< Total size of device local memory in bytes.
     u32 sets_per_pool{};                       ///< Sets per Description Pool
+    NvidiaArchitecture nvidia_arch{NvidiaArchitecture::Arch_AmpereOrNewer};
 
     // Telemetry parameters
     std::set<std::string, std::less<>> supported_extensions; ///< Reported Vulkan extensions.