path_fiddle/fiddle_context_dawn.cpp

#include "fiddle_context.hpp"

#ifndef RIVE_DAWN

std::unique_ptr<FiddleContext> FiddleContext::MakeDawnPLS(FiddleContextOptions options)
{
    return nullptr;
}

#else

#include "dawn/native/DawnNative.h"
#include "dawn/dawn_proc.h"

#include "rive/pls/pls_factory.hpp"
#include "rive/pls/pls_renderer.hpp"
#include "rive/pls/webgpu/pls_render_context_webgpu_impl.hpp"

#include <array>
#include <thread>

using namespace rive;
using namespace rive::pls;

static void print_device_error(WGPUErrorType errorType, const char* message, void*)
{
    const char* errorTypeName = "";
    switch (errorType)
    {
        case WGPUErrorType_Validation:
            errorTypeName = "Validation";
            break;
        case WGPUErrorType_OutOfMemory:
            errorTypeName = "Out of memory";
            break;
        case WGPUErrorType_Unknown:
            errorTypeName = "Unknown";
            break;
        case WGPUErrorType_DeviceLost:
            errorTypeName = "Device lost";
            break;
        default:
            RIVE_UNREACHABLE();
            return;
    }
    printf("%s error: %s\n", errorTypeName, message);
}

static void device_lost_callback(WGPUDeviceLostReason reason, const char* message, void*)
{
    printf("device lost: %s\n", message);
}

static void device_log_callback(WGPULoggingType type, const char* message, void*)
{
    printf("Device log %s\n", message);
}

#ifdef __APPLE__
extern float GetDawnWindowBackingScaleFactor(GLFWwindow*, bool retina);
extern std::unique_ptr<wgpu::ChainedStruct> SetupDawnWindowAndGetSurfaceDescriptor(GLFWwindow*,
                                                                                   bool retina);
#else

#define GLFW_EXPOSE_NATIVE_WIN32
#include <GLFW/glfw3.h>
#include <GLFW/glfw3native.h>

static float GetDawnWindowBackingScaleFactor(GLFWwindow*, bool retina) { return 1; }

static std::unique_ptr<wgpu::ChainedStruct> SetupDawnWindowAndGetSurfaceDescriptor(
    GLFWwindow* window,
    bool retina)
{
    std::unique_ptr<wgpu::SurfaceDescriptorFromWindowsHWND> desc =
        std::make_unique<wgpu::SurfaceDescriptorFromWindowsHWND>();
    desc->hwnd = glfwGetWin32Window(window);
    desc->hinstance = GetModuleHandle(nullptr);
    return std::move(desc);
}
#endif

class FiddleContextDawnPLS : public FiddleContext
{
public:
    FiddleContextDawnPLS(FiddleContextOptions options) : m_options(options)
    {
        WGPUInstanceDescriptor instanceDescriptor{};
        instanceDescriptor.features.timedWaitAnyEnable = true;
        m_instance = std::make_unique<dawn::native::Instance>(&instanceDescriptor);

        wgpu::RequestAdapterOptions adapterOptions = {
            .powerPreference = wgpu::PowerPreference::HighPerformance,
        };

        // Get an adapter for the backend to use, and create the device.
        auto adapters = m_instance->EnumerateAdapters(&adapterOptions);
        wgpu::DawnAdapterPropertiesPowerPreference power_props{};
        wgpu::AdapterProperties adapterProperties{};
        adapterProperties.nextInChain = &power_props;
        // Find the first adapter which satisfies the adapterType requirement.
        auto isAdapterType = [&adapterProperties](const auto& adapter) -> bool {
            adapter.GetProperties(&adapterProperties);
            return adapterProperties.adapterType == wgpu::AdapterType::DiscreteGPU;
        };
        auto preferredAdapter = std::find_if(adapters.begin(), adapters.end(), isAdapterType);
        if (preferredAdapter == adapters.end())
        {
            fprintf(stderr, "Failed to find an adapter! Please try another adapter type.\n");
            return;
        }

        std::vector<const char*> enableToggleNames = {
            "allow_unsafe_apis",
            "turn_off_vsync",
            // "skip_validation",
        };
        std::vector<const char*> disabledToggleNames;

        WGPUDawnTogglesDescriptor toggles = {
            .chain =
                {
                    .next = nullptr,
                    .sType = WGPUSType_DawnTogglesDescriptor,
                },
            .enabledToggleCount = enableToggleNames.size(),
            .enabledToggles = enableToggleNames.data(),
            .disabledToggleCount = disabledToggleNames.size(),
            .disabledToggles = disabledToggleNames.data(),
        };

        std::vector<WGPUFeatureName> requiredFeatures = {
            // WGPUFeatureName_IndirectFirstInstance,
            // WGPUFeatureName_ShaderF16,
            // WGPUFeatureName_BGRA8UnormStorage,
            // WGPUFeatureName_Float32Filterable,
            // WGPUFeatureName_DawnInternalUsages,
            // WGPUFeatureName_DawnMultiPlanarFormats,
            // WGPUFeatureName_DawnNative,
            // WGPUFeatureName_ImplicitDeviceSynchronization,
            WGPUFeatureName_SurfaceCapabilities,
            // WGPUFeatureName_TransientAttachments,
            // WGPUFeatureName_DualSourceBlending,
            // WGPUFeatureName_Norm16TextureFormats,
            // WGPUFeatureName_HostMappedPointer,
            // WGPUFeatureName_ChromiumExperimentalReadWriteStorageTexture,
        };

        WGPUDeviceDescriptor deviceDesc = {
            .nextInChain = reinterpret_cast<WGPUChainedStruct*>(&toggles),
            .requiredFeatureCount = requiredFeatures.size(),
            .requiredFeatures = requiredFeatures.data(),
        };

        m_backendDevice = preferredAdapter->CreateDevice(&deviceDesc);
        DawnProcTable backendProcs = dawn::native::GetProcs();
        dawnProcSetProcs(&backendProcs);
        backendProcs.deviceSetUncapturedErrorCallback(m_backendDevice, print_device_error, nullptr);
        backendProcs.deviceSetDeviceLostCallback(m_backendDevice, device_lost_callback, nullptr);
        backendProcs.deviceSetLoggingCallback(m_backendDevice, device_log_callback, nullptr);
        m_device = wgpu::Device::Acquire(m_backendDevice);
        m_queue = m_device.GetQueue();
        m_plsContext =
            PLSRenderContextWebGPUImpl::MakeContext(m_device,
                                                    m_queue,
                                                    PLSRenderContextWebGPUImpl::ContextOptions());
    }

    float dpiScale(GLFWwindow* window) const override
    {
        return GetDawnWindowBackingScaleFactor(window, m_options.retinaDisplay);
    }

    Factory* factory() override { return m_plsContext.get(); }

    rive::pls::PLSRenderContext* plsContextOrNull() override { return m_plsContext.get(); }

    void onSizeChanged(GLFWwindow* window, int width, int height, uint32_t sampleCount) override
    {
        DawnProcTable backendProcs = dawn::native::GetProcs();

        // Create the swapchain
        auto surfaceChainedDesc =
            SetupDawnWindowAndGetSurfaceDescriptor(window, m_options.retinaDisplay);
        WGPUSurfaceDescriptor surfaceDesc = {
            .nextInChain = reinterpret_cast<WGPUChainedStruct*>(surfaceChainedDesc.get()),
        };
        WGPUSurface surface = backendProcs.instanceCreateSurface(m_instance->Get(), &surfaceDesc);

        WGPUSwapChainDescriptor swapChainDesc = {
            .usage = WGPUTextureUsage_RenderAttachment,
            .format = WGPUTextureFormat_BGRA8Unorm,
            .width = static_cast<uint32_t>(width),
            .height = static_cast<uint32_t>(height),
            .presentMode = WGPUPresentMode_Immediate, // No vsync.
        };
        if (m_options.enableReadPixels)
        {
            swapChainDesc.usage |= WGPUTextureUsage_CopySrc;
        }

        WGPUSwapChain backendSwapChain =
            backendProcs.deviceCreateSwapChain(m_backendDevice, surface, &swapChainDesc);
        m_swapchain = wgpu::SwapChain::Acquire(backendSwapChain);

        m_renderTarget = m_plsContext->static_impl_cast<PLSRenderContextWebGPUImpl>()
                             ->makeRenderTarget(wgpu::TextureFormat::BGRA8Unorm, width, height);
        m_pixelReadBuff = {};
    }

    void toggleZoomWindow() override {}

    std::unique_ptr<Renderer> makeRenderer(int width, int height) override
    {
        return std::make_unique<PLSRenderer>(m_plsContext.get());
    }

    void begin(PLSRenderContext::FrameDescriptor&& frameDescriptor) override
    {
        assert(m_swapchain.GetCurrentTexture().GetWidth() == m_renderTarget->width());
        assert(m_swapchain.GetCurrentTexture().GetHeight() == m_renderTarget->height());
        m_renderTarget->setTargetTextureView(m_swapchain.GetCurrentTextureView());
        frameDescriptor.renderTarget = m_renderTarget;
        m_plsContext->beginFrame(std::move(frameDescriptor));
    }

    void end(GLFWwindow* window, std::vector<uint8_t>* pixelData) final
    {
        m_plsContext->flush();

        if (pixelData != nullptr)
        {
            // Read back pixels from the framebuffer!
            uint32_t w = m_renderTarget->width();
            uint32_t h = m_renderTarget->height();
            uint32_t rowBytesInReadBuff = math::round_up_to_multiple_of<256>(w * 4);

            // Create a buffer to receive the pixels.
            if (!m_pixelReadBuff)
            {
                wgpu::BufferDescriptor buffDesc{
                    .usage = wgpu::BufferUsage::MapRead | wgpu::BufferUsage::CopyDst,
                    .size = h * rowBytesInReadBuff,
                };
                m_pixelReadBuff = m_device.CreateBuffer(&buffDesc);
            }
            assert(m_pixelReadBuff.GetSize() == h * rowBytesInReadBuff);

            // Blit the framebuffer into m_pixelReadBuff.
            wgpu::CommandEncoder readEncoder = m_device.CreateCommandEncoder();
            wgpu::ImageCopyTexture srcTexture = {
                .texture = m_swapchain.GetCurrentTexture(),
                .origin = {0, 0, 0},
            };
            wgpu::ImageCopyBuffer dstBuffer = {
                .layout =
                    {
                        .offset = 0,
                        .bytesPerRow = rowBytesInReadBuff,
                    },
                .buffer = m_pixelReadBuff,
            };
            wgpu::Extent3D copySize = {
                .width = w,
                .height = h,
            };
            readEncoder.CopyTextureToBuffer(&srcTexture, &dstBuffer, &copySize);

            wgpu::CommandBuffer commands = readEncoder.Finish();
            m_queue.Submit(1, &commands);

            // Request a mapping of m_pixelReadBuff and wait for it to complete.
            bool mapped = false;
            m_pixelReadBuff.MapAsync(
                wgpu::MapMode::Read,
                0,
                h * rowBytesInReadBuff,
                [](WGPUBufferMapAsyncStatus status, void* mapped) {
                    if (status != WGPUBufferMapAsyncStatus_Success)
                    {
                        fprintf(stderr, "failed to map m_pixelReadBuff\n");
                        exit(-1);
                    }
                    *reinterpret_cast<bool*>(mapped) = true;
                },
                &mapped);
            while (!mapped)
            {
                // Spin until the GPU is finished with m_pixelReadBuff and we can read it.
                std::this_thread::sleep_for(std::chrono::milliseconds(1));
                tick();
            }

            // Copy the image data from m_pixelReadBuff to pixelData.
            pixelData->resize(h * w * 4);
            const uint8_t* pixelReadBuffData =
                reinterpret_cast<const uint8_t*>(m_pixelReadBuff.GetConstMappedRange());
            for (size_t y = 0; y < h; ++y)
            {
                // Flip Y.
                const uint8_t* src = &pixelReadBuffData[(h - y - 1) * rowBytesInReadBuff];
                size_t row = y * w * 4;
                for (size_t x = 0; x < w * 4; x += 4)
                {
                    // BGBRA -> RGBA.
                    (*pixelData)[row + x + 0] = src[x + 2];
                    (*pixelData)[row + x + 1] = src[x + 1];
                    (*pixelData)[row + x + 2] = src[x + 0];
                    (*pixelData)[row + x + 3] = src[x + 3];
                }
            }
            m_pixelReadBuff.Unmap();
        }

        m_swapchain.Present();
    }

    void tick() override { m_device.Tick(); }

private:
    const FiddleContextOptions m_options;
    WGPUDevice m_backendDevice = {};
    wgpu::Device m_device = {};
    wgpu::Queue m_queue = {};
    wgpu::SwapChain m_swapchain = {};
    std::unique_ptr<dawn::native::Instance> m_instance;
    std::unique_ptr<PLSRenderContext> m_plsContext;
    rcp<PLSRenderTargetWebGPU> m_renderTarget;
    wgpu::Buffer m_pixelReadBuff;
};

std::unique_ptr<FiddleContext> FiddleContext::MakeDawnPLS(FiddleContextOptions options)
{
    return std::make_unique<FiddleContextDawnPLS>(options);
}

#endif