buffer update opt: Add new API for getting uniform buffer offset alignment (#9347)

filament/backend/include/private/backend/DriverAPI.inc

@@ -366,6 +366,7 @@ DECL_DRIVER_API_SYNCHRONOUS_N(void, setupExternalImage, void*, image)
 DECL_DRIVER_API_SYNCHRONOUS_N(backend::TimerQueryResult, getTimerQueryValue, backend::TimerQueryHandle, tqh, uint64_t*, elapsedTime)
 DECL_DRIVER_API_SYNCHRONOUS_N(bool, isWorkaroundNeeded, backend::Workaround, workaround)
 DECL_DRIVER_API_SYNCHRONOUS_0(backend::FeatureLevel, getFeatureLevel)
+DECL_DRIVER_API_SYNCHRONOUS_0(size_t, getUniformBufferOffsetAlignment)
 
 /*
  * Updating driver objects

filament/backend/src/metal/MetalBuffer.h

@@ -19,6 +19,7 @@
 
 #include "MetalContext.h"
 #include "MetalFlags.h"
+#include "MetalUtils.h"
 
 #include <backend/DriverEnums.h>
 #include <backend/platforms/PlatformMetal.h>
@@ -260,18 +261,8 @@ public:
     // In practice, MetalRingBuffer is used for argument buffers, which are kept in the constant
     // address space. Constant buffers have specific alignment requirements when specifying an
     // offset.
-#if defined(FILAMENT_IOS)
-#if TARGET_OS_SIMULATOR
-    // The iOS simulator has differing alignment requirements.
-    static constexpr auto METAL_CONSTANT_BUFFER_OFFSET_ALIGNMENT = 256;
-#else
-    static constexpr auto METAL_CONSTANT_BUFFER_OFFSET_ALIGNMENT = 4;
-#endif  // TARGET_OS_SIMULATOR
-#else
-    static constexpr auto METAL_CONSTANT_BUFFER_OFFSET_ALIGNMENT = 32;
-#endif
     static inline auto computeSlotSize(MTLSizeAndAlign layout) {
-         return align(align(layout.size, layout.align), METAL_CONSTANT_BUFFER_OFFSET_ALIGNMENT);
+         return align(align(layout.size, layout.align), getUniformBufferOffsetAlignment());
     }
 
     MetalRingBuffer(id<MTLDevice> device, MTLResourceOptions options, MTLSizeAndAlign layout,

filament/backend/src/metal/MetalDriver.mm

@@ -29,6 +29,7 @@
 #include "MetalHandles.h"
 #include "MetalState.h"
 #include "MetalTimerQuery.h"
+#include "MetalUtils.h"
 
 #include <backend/platforms/PlatformMetal.h>
 #include <backend/platforms/PlatformMetal-ObjC.h>
@@ -1252,6 +1253,10 @@ size_t MetalDriver::getMaxArrayTextureLayers() {
     return 256;
 }
 
+size_t MetalDriver::getUniformBufferOffsetAlignment() {
+    return ::filament::backend::getUniformBufferOffsetAlignment();
+}
+
 void MetalDriver::updateIndexBuffer(Handle<HwIndexBuffer> ibh, BufferDescriptor&& data,
         uint32_t byteOffset) {
     FILAMENT_CHECK_PRECONDITION(data.buffer)

filament/backend/src/metal/MetalUtils.h

@@ -50,6 +50,22 @@ id<MTLTexture> createTextureViewWithLodRange(id<MTLTexture> texture, NSUInteger
  */
 id<MTLTexture> createTextureViewWithSingleSlice(id<MTLTexture> texture, NSUInteger slice);
 
+/**
+ * Returns the required alignment for uniform buffer offsets.
+ */
+constexpr size_t getUniformBufferOffsetAlignment() {
+#if defined(FILAMENT_IOS)
+#if TARGET_OS_SIMULATOR
+    // The iOS simulator has differing alignment requirements.
+    return 256;
+#else
+    return 4;
+#endif  // TARGET_OS_SIMULATOR
+#else
+    return 32;
+#endif
+}
+
 } // namespace filament::backend
 
 #endif //TNT_FILAMENT_DRIVER_METALUTILS_H

filament/backend/src/metal/MetalUtils.mm

@@ -68,4 +68,5 @@ id<MTLTexture> createTextureViewWithSingleSlice(id<MTLTexture> texture, NSUInteg
                                            slices:NSMakeRange(slice, 1)];
 }
 
+
 } // namespace filament::backend

filament/backend/src/noop/NoopDriver.cpp

@@ -274,6 +274,10 @@ size_t NoopDriver::getMaxArrayTextureLayers() {
     return 256u;
 }
 
+size_t NoopDriver::getUniformBufferOffsetAlignment() {
+    return 256u;
+}
+
 void NoopDriver::updateIndexBuffer(Handle<HwIndexBuffer> ibh, BufferDescriptor&& p,
         uint32_t byteOffset) {
     scheduleDestroy(std::move(p));

filament/backend/src/opengl/OpenGLDriver.cpp

@@ -2666,6 +2666,10 @@ size_t OpenGLDriver::getMaxArrayTextureLayers() {
     return mContext.gets.max_array_texture_layers;
 }
 
+size_t OpenGLDriver::getUniformBufferOffsetAlignment() {
+    return mContext.gets.uniform_buffer_offset_alignment;
+}
+
 // ------------------------------------------------------------------------------------------------
 // Swap chains
 // ------------------------------------------------------------------------------------------------

filament/backend/src/vulkan/VulkanDriver.cpp

@@ -1390,7 +1390,8 @@ uint8_t VulkanDriver::getMaxDrawBuffers() {
 }
 
 size_t VulkanDriver::getMaxUniformBufferSize() {
-    return mContext.getPhysicalDeviceLimits().maxUniformBufferRange;
+    return std::max(mContext.getPhysicalDeviceLimits().maxUniformBufferRange,
+            static_cast<uint32_t>(mContext.getPhysicalDeviceLimits().nonCoherentAtomSize));
 }
 
 size_t VulkanDriver::getMaxTextureSize(SamplerType type) {
@@ -1410,6 +1411,10 @@ size_t VulkanDriver::getMaxArrayTextureLayers() {
     return mContext.getPhysicalDeviceLimits().maxImageArrayLayers;
 }
 
+size_t VulkanDriver::getUniformBufferOffsetAlignment() {
+    return mContext.getPhysicalDeviceLimits().minUniformBufferOffsetAlignment;
+}
+
 void VulkanDriver::setVertexBufferObject(Handle<HwVertexBuffer> vbh, uint32_t index,
         Handle<HwBufferObject> boh) {
     auto vb = resource_ptr<VulkanVertexBuffer>::cast(&mResourceManager, vbh);

filament/backend/src/webgpu/WebGPUDriver.cpp

@@ -841,6 +841,10 @@ size_t WebGPUDriver::getMaxArrayTextureLayers() {
     return mDeviceLimits.maxTextureArrayLayers;
 }
 
+size_t WebGPUDriver::getUniformBufferOffsetAlignment(){
+    return mDeviceLimits.minUniformBufferOffsetAlignment;
+}
+
 void WebGPUDriver::updateIndexBuffer(Handle<HwIndexBuffer> indexBufferHandle,
         BufferDescriptor&& bufferDescriptor, const uint32_t byteOffset) {
     // make sure command elements (draws, etc.) prior to the buffer update are processed before the