From e1879aa0bca46bb630e738208521a978eceeed72 Mon Sep 17 00:00:00 2001
From: wang-bin <wbsecg1@gmail.com>
Date: Sat, 27 Jun 2015 02:15:46 +0800
Subject: [PATCH] a convenience func to make host video frame from gpu

---
 src/QtAV/VideoFrame.h        | 23 ++++++++++++++-
 src/VideoFrame.cpp           | 55 ++++++++++++++++++++++++++++++++++++
 src/utils/CopyFrame_SSE2.cpp |  1 +
 3 files changed, 78 insertions(+), 1 deletion(-)

diff --git a/src/QtAV/VideoFrame.h b/src/QtAV/VideoFrame.h
index bfbf1b5d..37d2f1e7 100644
--- a/src/QtAV/VideoFrame.h
+++ b/src/QtAV/VideoFrame.h
@@ -35,6 +35,20 @@ class Q_AV_EXPORT VideoFrame : public Frame
 {
     Q_DECLARE_PRIVATE(VideoFrame)
 public:
+    /*!
+     * \brief fromGPU
+     * Make a VideoFrame with data on host memory from GPU resource
+     * \param fmt video format of GPU resource
+     * \param width frame width
+     * \param height frame height
+     * \param surface_h surface height. Can be greater than visual frame height because of alignment
+     * \param src CPU accessible address of frame planes on GPU. src[0] must be valid. src[i>0] will be filled depending on pixel format, pitch and surface_h if it's NULL.
+     * \param pitch plane pitch on GPU. pitch[0] must be valid. pitch[i>0] will be filled depending on pixel format, pitch[0] and surface_h if it's NULL.
+     * \param optimized try to use SIMD to copy from GPU. otherwise use memcpy
+     * \param swapUV
+     */
+    static VideoFrame fromGPU(const VideoFormat& fmt, int width, int height, int surface_h, quint8 *src[], int pitch[], bool optimized = true, bool swapUV = false);
+
     VideoFrame();
     //must set planes and linesize manually
     VideoFrame(int width, int height, const VideoFormat& format);
@@ -90,6 +104,13 @@ public:
      * \param roi NOT implemented!
      */
     QImage toImage(QImage::Format fmt = QImage::Format_ARGB32, const QSize& dstSize = QSize(), const QRectF& roi = QRect()) const;
+    /*!
+     * \brief to
+     * The result frame data is always on host memory.
+     * \param pixfmt target pixel format
+     * \param dstSize target frame size
+     * \param roi interested region of source frame
+     */
     VideoFrame to(VideoFormat::PixelFormat pixfmt, const QSize& dstSize = QSize(), const QRectF& roi = QRect()) const;
     VideoFrame to(const VideoFormat& fmt, const QSize& dstSize = QSize(), const QRectF& roi = QRect()) const;
     /*!
@@ -129,7 +150,7 @@ public:
     void setEq(int brightness, int contrast, int saturation);
     /*!
      * \brief convert
-     * return a frame with a given format from a given source frame
+     * return a frame with a given format from a given source frame. The result frame data is always on host memory.
      */
     VideoFrame convert(const VideoFrame& frame, const VideoFormat& fmt) const;
     VideoFrame convert(const VideoFrame& frame, VideoFormat::PixelFormat fmt) const;
diff --git a/src/VideoFrame.cpp b/src/VideoFrame.cpp
index 7116916b..82f9a0e8 100644
--- a/src/VideoFrame.cpp
+++ b/src/VideoFrame.cpp
@@ -26,6 +26,7 @@
 #include <QtCore/QSharedPointer>
 #include <QtGui/QImage>
 #include "QtAV/private/AVCompat.h"
+#include "utils/GPUMemCopy.h"
 #include "utils/Logger.h"
 
 // FF_API_PIX_FMT
@@ -44,6 +45,60 @@ static const struct RegisterMetaTypes
 } _registerMetaTypes;
 }
 
+VideoFrame VideoFrame::fromGPU(const VideoFormat& fmt, int width, int height, int surface_h, quint8 *src[], int pitch[], bool optimized, bool swapUV)
+{
+    Q_ASSERT(src[0] && pitch[0] > 0 && "VideoFrame::fromGPU: src[0] and pitch[0] must be set");
+    const int nb_planes = fmt.planeCount();
+    const int chroma_pitch = nb_planes > 1 ? fmt.bytesPerLine(pitch[0], 1) : 0;
+    const int chroma_h = fmt.chromaHeight(surface_h);
+    int h[] = { surface_h, 0, 0};
+    for (int i = 1; i < nb_planes; ++i) {
+        h[i] = chroma_h;
+        // set chroma address and pitch if not set
+        if (pitch[i] <= 0)
+            pitch[i] = chroma_pitch;
+        if (!src[i])
+            src[i] = src[i-1] + pitch[i-1]*h[i-1];
+    }
+    if (swapUV) {
+        std::swap(src[1], src[2]);
+        std::swap(pitch[1], pitch[2]);
+    }
+    VideoFrame frame;
+    if (optimized) {
+        qDebug("simd");
+        int yuv_size = 0;
+        for (int i = 0; i < nb_planes; ++i) {
+            yuv_size += pitch[i]*h[i];
+        }
+        // additional 15 bytes to ensure 16 bytes aligned
+        QByteArray buf(15 + yuv_size, 0);
+        const int offset_16 = (16 - ((uintptr_t)buf.data() & 0x0f)) & 0x0f;
+        // plane 1, 2... is aligned?
+        uchar* plane_ptr = (uchar*)buf.data() + offset_16;
+        QVector<uchar*> dst(nb_planes, 0);
+        for (int i = 0; i < nb_planes; ++i) {
+            dst[i] = plane_ptr;
+            // TODO: add VideoFormat::planeWidth/Height() ?
+            // pitch instead of surface_width
+            plane_ptr += pitch[i] * h[i];
+            gpu_memcpy(dst[i], src[i], pitch[i]*h[i]);
+
+        }
+        frame = VideoFrame(buf, width, height, fmt);
+        frame.setBits(dst);
+        frame.setBytesPerLine(pitch);
+    } else {
+        frame = VideoFrame(width, height, fmt);
+        frame.setBits(src);
+        frame.setBytesPerLine(pitch);
+        // TODO: why clone is faster()?
+        // TODO: buffer pool and create VideoFrame when needed to avoid copy? also for other va
+        frame = frame.clone();
+    }
+    return frame;
+}
+
 class VideoFramePrivate : public FramePrivate
 {
     Q_DISABLE_COPY(VideoFramePrivate)
diff --git a/src/utils/CopyFrame_SSE2.cpp b/src/utils/CopyFrame_SSE2.cpp
index 294379fa..fe00a10f 100644
--- a/src/utils/CopyFrame_SSE2.cpp
+++ b/src/utils/CopyFrame_SSE2.cpp
@@ -19,6 +19,7 @@
     Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
 ******************************************************************************/
 #include <stdint.h> //intptr_t
+#include <string.h>
 #include <emmintrin.h>