Start getting the readback api into shape

examples/rhi/offscreen_vulkan/main.cpp

@@ -103,8 +103,15 @@ int main(int argc, char **argv)
         r->beginPass(rt, cb, { 0, 1, 0, 1 }, { 1, 0 }, nullptr);
         r->endPass(cb);
 
+        QRhiReadbackDescription rb(tex);
+        QRhiReadbackResult rbResult;
+        r->readback(cb, rb, &rbResult);
+
         qDebug("Submit and wait");
         r->endAndWaitOffscreenFrame();
+
+        // here it's simple, no need to bother with the completed callback of rbResult
+        qDebug() << rbResult.data.size();
     }
 
     rt->releaseAndDestroy();

src/rhi/qrhi.cpp

@@ -487,9 +487,9 @@ void QRhi::drawIndexed(QRhiCommandBuffer *cb, quint32 indexCount,
     d->drawIndexed(cb, indexCount, instanceCount, firstIndex, vertexOffset, firstInstance);
 }
 
-void QRhi::readback(QRhiCommandBuffer *cb, QRhiReadback *rb)
+void QRhi::readback(QRhiCommandBuffer *cb, const QRhiReadbackDescription &rb, QRhiReadbackResult *result)
 {
-    d->readback(cb, rb);
+    d->readback(cb, rb, result);
 }
 
 QVector<int> QRhi::supportedSampleCounts() const

src/rhi/qrhi.h

@@ -45,6 +45,7 @@
 #include <QVector>
 #include <QImage>
 #include <QtShaderTools/QBakedShader>
+#include <functional>
 
 QT_BEGIN_NAMESPACE
 
@@ -280,12 +281,22 @@ Q_DECLARE_TYPEINFO(QRhiTextureUploadDescription::Layer::MipLevel, Q_MOVABLE_TYPE
 Q_DECLARE_TYPEINFO(QRhiTextureUploadDescription::Layer, Q_MOVABLE_TYPE);
 Q_DECLARE_TYPEINFO(QRhiTextureUploadDescription, Q_MOVABLE_TYPE);
 
-struct Q_RHI_EXPORT QRhiReadback
+struct Q_RHI_EXPORT QRhiReadbackDescription
 {
-    // ###
+    QRhiReadbackDescription() { }
+    QRhiReadbackDescription(QRhiTexture *texture_) : texture(texture_) { }
+    QRhiTexture *texture = nullptr;
+    int layer = 0;
+    int level = 0;
 };
 
-Q_DECLARE_TYPEINFO(QRhiReadback, Q_MOVABLE_TYPE);
+Q_DECLARE_TYPEINFO(QRhiReadbackDescription, Q_MOVABLE_TYPE);
+
+struct Q_RHI_EXPORT QRhiReadbackResult
+{
+    std::function<void()> completed = nullptr;
+    QByteArray data;
+}; // non-movable due to the std::function
 
 class Q_RHI_EXPORT QRhiResource
 {
@@ -980,12 +991,13 @@ public:
       The typical use case is to use it in completely offscreen applications,
       e.g. to generate image sequences by rendering and reading back without
       ever showing a window.
+          QRhiReadbackResult rbResult;
           QRhiCommandBuffer *cb; // not owned
           beginOffscreenFrame(&cb);
           // ... the usual, set up a QRhiTextureRenderTarget, beginPass-endPass, etc.
-          readback(cb, someTexture, rb);
+          readback(cb, rb, &rbResult);
           endAndWaitOffscreenFrame();
-          // the results are available in rb
+          // image data available in rbResult
      */
     FrameOpResult beginOffscreenFrame(QRhiCommandBuffer **cb);
     FrameOpResult endAndWaitOffscreenFrame();
@@ -1037,8 +1049,8 @@ public:
                      quint32 instanceCount = 1, quint32 firstIndex = 0,
                      qint32 vertexOffset = 0, quint32 firstInstance = 0);
 
-    // Must only be called outside a begin-endPass section.
-    void readback(QRhiCommandBuffer *cb, QRhiReadback *rb);
+    // Cannot be mixed with beginPass-endPass.
+    void readback(QRhiCommandBuffer *cb, const QRhiReadbackDescription &rb, QRhiReadbackResult *result);
 
     QVector<int> supportedSampleCounts() const;
 

src/rhi/qrhi_p.h

@@ -113,7 +113,7 @@ public:
                              quint32 instanceCount, quint32 firstIndex,
                              qint32 vertexOffset, quint32 firstInstance) = 0;
 
-    virtual void readback(QRhiCommandBuffer *cb, QRhiReadback *rb) = 0;
+    virtual void readback(QRhiCommandBuffer *cb, const QRhiReadbackDescription &rb, QRhiReadbackResult *result) = 0;
 
     virtual QVector<int> supportedSampleCounts() const = 0;
     virtual int ubufAlignment() const = 0;

src/rhi/qrhid3d11.cpp

@@ -453,10 +453,11 @@ void QRhiD3D11::drawIndexed(QRhiCommandBuffer *cb, quint32 indexCount,
     cbD->commands.append(cmd);
 }
 
-void QRhiD3D11::readback(QRhiCommandBuffer *cb, QRhiReadback *rb)
+void QRhiD3D11::readback(QRhiCommandBuffer *cb, const QRhiReadbackDescription &rb, QRhiReadbackResult *result)
 {
     Q_UNUSED(cb);
     Q_UNUSED(rb);
+    Q_UNUSED(result);
 }
 
 QRhi::FrameOpResult QRhiD3D11::beginFrame(QRhiSwapChain *swapChain)

src/rhi/qrhid3d11_p.h

@@ -405,7 +405,7 @@ public:
                      quint32 instanceCount, quint32 firstIndex,
                      qint32 vertexOffset, quint32 firstInstance) override;
 
-    void readback(QRhiCommandBuffer *cb, QRhiReadback *rb) override;
+    void readback(QRhiCommandBuffer *cb, const QRhiReadbackDescription &rb, QRhiReadbackResult *result) override;
 
     QVector<int> supportedSampleCounts() const override;
     int ubufAlignment() const override;

src/rhi/qrhigles2.cpp

@@ -366,10 +366,11 @@ void QRhiGles2::drawIndexed(QRhiCommandBuffer *cb, quint32 indexCount,
     cbD->commands.append(cmd);
 }
 
-void QRhiGles2::readback(QRhiCommandBuffer *cb, QRhiReadback *rb)
+void QRhiGles2::readback(QRhiCommandBuffer *cb, const QRhiReadbackDescription &rb, QRhiReadbackResult *result)
 {
     Q_UNUSED(cb);
     Q_UNUSED(rb);
+    Q_UNUSED(result);
 }
 
 QRhi::FrameOpResult QRhiGles2::beginFrame(QRhiSwapChain *swapChain)

src/rhi/qrhigles2_p.h

@@ -363,7 +363,7 @@ public:
                      quint32 instanceCount, quint32 firstIndex,
                      qint32 vertexOffset, quint32 firstInstance) override;
 
-    void readback(QRhiCommandBuffer *cb, QRhiReadback *rb) override;
+    void readback(QRhiCommandBuffer *cb, const QRhiReadbackDescription &rb, QRhiReadbackResult *result) override;
 
     QVector<int> supportedSampleCounts() const override;
     int ubufAlignment() const override;

src/rhi/qrhimetal.mm

@@ -496,10 +496,11 @@ void QRhiMetal::drawIndexed(QRhiCommandBuffer *cb, quint32 indexCount,
       baseInstance: firstInstance];
 }
 
-void QRhiMetal::readback(QRhiCommandBuffer *cb, QRhiReadback *rb)
+void QRhiMetal::readback(QRhiCommandBuffer *cb, const QRhiReadbackDescription &rb, QRhiReadbackResult *result)
 {
     Q_UNUSED(cb);
     Q_UNUSED(rb);
+    Q_UNUSED(result);
 }
 
 QRhi::FrameOpResult QRhiMetal::beginFrame(QRhiSwapChain *swapChain)

src/rhi/qrhimetal_p.h

@@ -294,7 +294,7 @@ public:
                      quint32 instanceCount, quint32 firstIndex,
                      qint32 vertexOffset, quint32 firstInstance) override;
 
-    void readback(QRhiCommandBuffer *cb, QRhiReadback *rb) override;
+    void readback(QRhiCommandBuffer *cb, const QRhiReadbackDescription &rb, QRhiReadbackResult *result) override;
 
     QVector<int> supportedSampleCounts() const override;
     int ubufAlignment() const override;

src/rhi/qrhivulkan.cpp

@@ -1384,11 +1384,7 @@ QRhi::FrameOpResult QRhiVulkan::endAndWaitOffscreenFrame()
     submitInfo.pWaitDstStageMask = &psf;
 
     err = df->vkQueueSubmit(gfxQueue, 1, &submitInfo, ofr.cmdFence);
-    if (err == VK_SUCCESS) {
-        // wait for completion
-        df->vkWaitForFences(dev, 1, &ofr.cmdFence, VK_TRUE, UINT64_MAX);
-        df->vkResetFences(dev, 1, &ofr.cmdFence);
-    } else {
+    if (err != VK_SUCCESS) {
         if (checkDeviceLost(err))
             return QRhi::FrameOpDeviceLost;
         else
@@ -1396,6 +1392,14 @@ QRhi::FrameOpResult QRhiVulkan::endAndWaitOffscreenFrame()
         return QRhi::FrameOpError;
     }
 
+    // wait for completion
+    df->vkWaitForFences(dev, 1, &ofr.cmdFence, VK_TRUE, UINT64_MAX);
+    df->vkResetFences(dev, 1, &ofr.cmdFence);
+
+    // Here we know that executing the host-side reads for this (or any
+    // previous) frame is safe since we waited for completion above.
+    finishActiveReadbacks(true);
+
     return QRhi::FrameOpSuccess;
 }
 
@@ -1426,6 +1430,15 @@ void QRhiVulkan::prepareNewFrame(QRhiCommandBuffer *cb)
     Q_ASSERT(!inFrame);
     inFrame = true;
 
+    // Now is the time to do things for frame N-F, where N is the current one,
+    // F is QVK_FRAMES_IN_FLIGHT, because only here it is guaranteed that that
+    // frame has completed on the GPU (due to the fence wait in beginFrame). To
+    // decide if something is safe to handle now a simple "lastActiveFrameSlot
+    // == currentFrameSlot" is sufficient (remember that e.g. with F==2
+    // currentFrameSlot goes 0, 1, 0, 1, 0, ...)
+
+    finishActiveReadbacks();
+
     executeDeferredReleases();
 
     QRHI_RES(QVkCommandBuffer, cb)->resetState();
@@ -1972,6 +1985,27 @@ void QRhiVulkan::executeDeferredReleases(bool forced)
     }
 }
 
+void QRhiVulkan::finishActiveReadbacks(bool forced)
+{
+    QVarLengthArray<std::function<void()>, 4> completedCallbacks;
+
+    for (int i = activeReadbacks.count() - 1; i >= 0; --i) {
+        const QRhiVulkan::ActiveReadback &aRb(activeReadbacks[i]);
+        if (forced || currentFrameSlot == aRb.activeFrameSlot || aRb.activeFrameSlot < 0) {
+            // map and memcpy
+            // destroy temp buffer
+
+            if (aRb.result->completed)
+                completedCallbacks.append(aRb.result->completed);
+
+            activeReadbacks.removeAt(i);
+        }
+    }
+
+    for (auto f : completedCallbacks)
+        f();
+}
+
 static struct {
     VkSampleCountFlagBits mask;
     int count;
@@ -2315,10 +2349,20 @@ void QRhiVulkan::drawIndexed(QRhiCommandBuffer *cb, quint32 indexCount,
     df->vkCmdDrawIndexed(QRHI_RES(QVkCommandBuffer, cb)->cb, indexCount, instanceCount, firstIndex, vertexOffset, firstInstance);
 }
 
-void QRhiVulkan::readback(QRhiCommandBuffer *cb, QRhiReadback *rb)
+void QRhiVulkan::readback(QRhiCommandBuffer *cb, const QRhiReadbackDescription &rb, QRhiReadbackResult *result)
 {
     Q_UNUSED(cb);
-    Q_UNUSED(rb);
+
+    Q_ASSERT(inFrame && !inPass);
+
+    // create a host visible buffer
+    // add a vkCmdCopyImageToBuffer
+
+    ActiveReadback aRb;
+    aRb.activeFrameSlot = currentFrameSlot;
+    aRb.desc = rb;
+    aRb.result = result;
+    activeReadbacks.append(aRb);
 }
 
 static inline VkBufferUsageFlagBits toVkBufferUsage(QRhiBuffer::UsageFlags usage)

src/rhi/qrhivulkan_p.h

@@ -356,7 +356,7 @@ public:
                      quint32 instanceCount, quint32 firstIndex,
                      qint32 vertexOffset, quint32 firstInstance) override;
 
-    void readback(QRhiCommandBuffer *cb, QRhiReadback *rb) override;
+    void readback(QRhiCommandBuffer *cb, const QRhiReadbackDescription &rb, QRhiReadbackResult *result) override;
 
     QVector<int> supportedSampleCounts() const override;
     int ubufAlignment() const override;
@@ -403,6 +403,7 @@ public:
     void activateTextureRenderTarget(QRhiCommandBuffer *cb, QRhiTextureRenderTarget *rt);
     void deactivateTextureRenderTarget(QRhiCommandBuffer *cb, QRhiTextureRenderTarget *rt);
     void executeDeferredReleases(bool forced = false);
+    void finishActiveReadbacks(bool forced = false);
 
     void bufferBarrier(QRhiCommandBuffer *cb, QRhiBuffer *buf);
     void imageBarrier(QRhiCommandBuffer *cb, QRhiTexture *tex,
@@ -415,7 +416,6 @@ public:
     // in case they changed in the meantime.
     void updateShaderResourceBindings(QRhiShaderResourceBindings *srb, int descSetIdx = -1);
 
-    QRhi *q;
     QVulkanInstance *inst = nullptr;
     QWindow *maybeWindow = nullptr;
     bool importedDevPoolQueue = false;
@@ -464,6 +464,13 @@ public:
         VkFence cmdFence = VK_NULL_HANDLE;
     } ofr;
 
+    struct ActiveReadback {
+        int activeFrameSlot = -1;
+        QRhiReadbackDescription desc;
+        QRhiReadbackResult *result;
+    };
+    QVector<ActiveReadback> activeReadbacks;
+
     struct DeferredReleaseEntry {
         enum Type {
             Pipeline,
@@ -526,6 +533,7 @@ public:
 
 Q_DECLARE_TYPEINFO(QRhiVulkan::DescriptorPoolData, Q_MOVABLE_TYPE);
 Q_DECLARE_TYPEINFO(QRhiVulkan::DeferredReleaseEntry, Q_MOVABLE_TYPE);
+Q_DECLARE_TYPEINFO(QRhiVulkan::ActiveReadback, Q_MOVABLE_TYPE);
 
 QT_END_NAMESPACE
 

todo.txt

@@ -2,6 +2,7 @@ gl, mtl: compressed textures
 gl, mtl: srgb (tex, swapchain buf)
 rhi without a window, fully offscreen
 readbacks
+some wait-gpu stuff for waiting for readback results inside a begin-endFrame f.ex.?
 multi window? (multi swapchain) -> trouble
 vk: rendering hangs sometimes when minimize and back on some systems?
 mtl: cubemaps