Fixes for Web

Source/Engine/Graphics/Textures/StreamingTexture.cpp

@@ -383,12 +383,17 @@ protected:
         _streamingTexture->GetOwner()->GetMipData(absoluteMipIndex, data);
         if (data.IsInvalid())
             return Result::MissingData;
+        PixelFormat dataFormat = _streamingTexture->GetHeader()->Format;
 
         // Cache data
         const int32 arraySize = texture->ArraySize();
         uint32 rowPitch, slicePitch;
         if (!_streamingTexture->GetOwner()->GetMipDataCustomPitch(absoluteMipIndex, rowPitch, slicePitch))
-            texture->ComputePitch(_mipIndex, rowPitch, slicePitch);
+        {
+            int32 mipWidth, mipHeight;
+            texture->GetMipSize(_mipIndex, mipWidth, mipHeight);
+            RenderTools::ComputePitch(dataFormat, mipWidth, mipHeight, rowPitch, slicePitch);
+        }
         _data.Link(data);
 
         // Update all array slices
@@ -396,9 +401,9 @@ protected:
         int32 dataPerSlice = data.Length() / arraySize; // In most cases it's a slice pitch, except when using transcoding (eg. Basis), then each slice has to use the same amount of memory
         for (int32 arrayIndex = 0; arrayIndex < arraySize; arrayIndex++)
         {
-            if (TextureTool::UpdateTexture(context->GPU, texture, arrayIndex, _mipIndex, Span<byte>(dataSource, dataPerSlice), rowPitch, slicePitch, _streamingTexture->GetHeader()->Format))
+            if (TextureTool::UpdateTexture(context->GPU, texture, arrayIndex, _mipIndex, Span<byte>(dataSource, dataPerSlice), rowPitch, slicePitch, dataFormat))
             {
-                LOG(Error, "Cannot stream mip {} of texture {} (format: {})", _mipIndex, texture->ToString(), ScriptingEnum::ToString(_streamingTexture->GetHeader()->Format));
+                LOG(Error, "Cannot stream mip {} of texture {} (format: {})", _mipIndex, texture->ToString(), ScriptingEnum::ToString(dataFormat));
                 return Result::Failed;
             }
             dataSource += dataPerSlice;

Source/Engine/GraphicsDevice/WebGPU/GPUContextWebGPU.cpp

@@ -1090,12 +1090,13 @@ void GPUContextWebGPU::FlushBindGroup()
     {
         auto descriptors = _pipelineState->BindGroupDescriptors[groupIndex];
         key.Layout = _pipelineState->BindGroupLayouts[groupIndex];
-        if (!descriptors || !key.Layout)
-            continue;
 
         // Build descriptors
         uint32 dynamicOffsetsCount = 0;
-        BuildBindGroup(groupIndex, *descriptors, key, dynamicOffsets, dynamicOffsetsCount);
+        if (descriptors)
+            BuildBindGroup(groupIndex, *descriptors, key, dynamicOffsets, dynamicOffsetsCount);
+        else
+            key.EntriesCount = 0;
 
         // Bind group
         WGPUBindGroup bindGroup = _pipelineState->GetBindGroup(key);

Source/Engine/GraphicsDevice/WebGPU/GPUDeviceWebGPU.cpp

@@ -302,12 +302,11 @@ bool GPUDeviceWebGPU::Init()
             limits.maxComputeWorkgroupSizeY >= 256 &&
             limits.maxComputeWorkgroupSizeZ >= 8 &&
             limits.maxBufferSize >= 64 * 1024 * 1024; // 64MB
-        Limits.HasInstancing = true;
+        Limits.HasInstancing = features.Contains(WGPUFeatureName_PrimitiveIndex); // TODO: investigate why Firefox doesn't draw instanced draws (this check is kind of a hack)
         Limits.HasDrawIndirect = true;
         Limits.HasDepthAsSRV = true;
         Limits.HasReadOnlyDepth = true;
         Limits.HasDepthClip = features.Contains(WGPUFeatureName_DepthClipControl);
-        Limits.HasReadOnlyDepth = true;
         Limits.MaximumSamplerAnisotropy = 4;
         Limits.MaximumTexture1DSize = limits.maxTextureDimension1D;
         Limits.MaximumTexture2DSize = limits.maxTextureDimension2D;

Source/Engine/GraphicsDevice/WebGPU/GraphicsDeviceWebGPU.Build.cs

@@ -16,7 +16,7 @@ public class GraphicsDeviceWebGPU : GraphicsDeviceBaseModule
     /// 1 - via Asyncify (causes the WASM to be much larger) <br/>
     /// 2 - via JSPI (experimental) <br/>
     /// </summary>
-    public int WithAsyncify = 2;
+    public static int WithAsyncify = 2;
 
     /// <inheritdoc />
     public override void Setup(BuildOptions options)

Source/Engine/Main/Main.Build.cs

@@ -77,6 +77,7 @@ public class Main : EngineModule
             break;
         case TargetPlatform.Web:
             options.SourcePaths.Add(Path.Combine(FolderPath, "Web"));
+            options.PrivateDefinitions.Add("WEB_LOOP_MODE=" + GraphicsDeviceWebGPU.WithAsyncify);
             break;
         default: throw new InvalidPlatformException(options.Platform.Target);
         }

Source/Engine/Main/Web/main.cpp

@@ -7,17 +7,19 @@
 
 // Reference: https://github.com/kainino0x/webgpu-cross-platform-demo/blob/f5c69c6fccbb2584c1b6f9e559f9a41a38a9b5ad/main.cpp#L692-L704
 // Reference: https://github.com/kainino0x/webgpu-cross-platform-demo/blob/c26ea3e29ed9f73f9b39bddf7964b482ce3c6964/main.cpp#L737-L758
-#define WEB_LOOP_MODE 2 // 0 - default, 1 - Asyncify, 2 - JSPI
-#if WEB_LOOP_MODE != 0
+#ifndef WEB_LOOP_MODE // 0 - default, 1 - Asyncify, 2 - JSPI
+#error "Unknown WEB_LOOP_MODE"
+#endif
+#if WEB_LOOP_MODE == 2
 // Workaround for JSPI not working in emscripten_set_main_loop. Loosely based on this code:
 // https://github.com/emscripten-core/emscripten/issues/22493#issuecomment-2330275282
 // This code only works with JSPI is enabled.
 typedef bool (*FrameCallback)(); // If callback returns true, continues the loop.
-EM_JS(void, requestAnimationFrameLoopWithJSPI, (FrameCallback callback), {
+EM_JS(void, requestAnimationFrameLoopWithJSPI, (FrameCallback frame), {
 #if WEB_LOOP_MODE == 2
-    var callback = WebAssembly.promising(getWasmTableEntry(callback));
+    var callback = WebAssembly.promising(getWasmTableEntry(frame));
 #elif WEB_LOOP_MODE == 1
-    var callback = () = > globalThis['Module']['ccall']("callback", "boolean", [], [], { async: true });
+    var callback = () = > globalThis['Module']['ccall']("frame", "boolean", [], [], { async: true });
 #endif
     async function tick() {
         // Start the frame callback. 'await' means we won't call
@@ -31,7 +33,7 @@ EM_JS(void, requestAnimationFrameLoopWithJSPI, (FrameCallback callback), {
 
 class PlatformMain
 {
-#if WEB_LOOP_MODE == 0
+#if WEB_LOOP_MODE != 2
     static void Loop()
     {
         // Tick engine
@@ -75,7 +77,7 @@ public:
             return result;
 
         // Setup main loop to be called by Emscripten
-#if WEB_LOOP_MODE == 0
+#if WEB_LOOP_MODE != 2
         emscripten_set_main_loop(Loop, -1, false);
 #else
         requestAnimationFrameLoopWithJSPI(Loop);

Source/Engine/Tools/TextureTool/TextureTool.cpp

@@ -419,15 +419,15 @@ bool TextureTool::UpdateTexture(GPUContext* context, GPUTexture* texture, int32
         auto textureSampler = PixelFormatSampler::Get(textureFormat);
         if (!dataSampler || !textureSampler)
             return true;
-        auto bytesPerPixel = PixelFormatExtensions::SizeInBytes(textureFormat);
 
         int32 mipWidth, mipHeight, mipDepth;
         texture->GetMipSize(mipIndex, mipWidth, mipHeight, mipDepth);
 
-        auto tempRowPitch = mipWidth * bytesPerPixel;
+        auto tempRowPitch = mipWidth * textureSampler->PixelSize;
         auto tempSlicePitch = tempRowPitch * mipHeight;
         tempData.Resize(tempSlicePitch * mipDepth);
 
+        ASSERT(data.Length() / rowPitch >= mipHeight);
         for (int32 y = 0; y < mipHeight; y++)
         {
             for (int32 x = 0; x < mipWidth; x++)
@@ -514,7 +514,9 @@ PixelFormat TextureTool::GetTextureFormat(TextureFormatType textureType, PixelFo
         EnumHasAllFlags(GPUDevice::Instance->GetFormatFeatures(PixelFormat::R32_Float).Support, minSupport) &&
         PixelFormatSampler::Get(dataFormat))
         return PixelFormat::R32_Float;
-    if (dataFormat == PixelFormat::R16G16_UNorm && EnumHasAllFlags(GPUDevice::Instance->GetFormatFeatures(PixelFormat::R32G32_UInt).Support, minSupport))
+    if (dataFormat == PixelFormat::R16G16_UNorm && 
+        EnumHasAllFlags(GPUDevice::Instance->GetFormatFeatures(PixelFormat::R32G32_Float).Support, minSupport) &&
+        PixelFormatSampler::Get(dataFormat))
         return PixelFormat::R32G32_Float;
     if ((dataFormat == PixelFormat::R16G16B16A16_UNorm || dataFormat == PixelFormat::R16G16B16A16_Float) &&
         EnumHasAllFlags(GPUDevice::Instance->GetFormatFeatures(PixelFormat::R32G32B32A32_Float).Support, minSupport) &&