komorebi/nix/linux/dev/virglrenderer-video-memory-upload.patch

virglrenderer: upload the picture to encode through memory

virglrenderer shares a video buffer with the host's VA-API driver as a dma-buf: it
exports the VA surface (vaExportSurfaceHandle) and imports it into EGL to copy the
guest's picture in. The d3d12 driver on WSL 2 has no dma-bufs, so the export fails
("export surface failed, err = 0x6"), nothing is uploaded and the encoder returns an
empty bitstream.

When the first export fails, read the planes of the guest's picture back from their
textures (glReadPixels, into the planes of a VAImage) and write them to the surface
through vaDeriveImage(), as FFmpeg's hwupload does; where the driver cannot derive an
image, a vaCreateImage() image is copied with vaPutImage(). On the d3d12 driver the
vaPutImage() copy came out with an all-zero luma plane (the chroma was right), where
the derived image encodes exactly: a luma ramp through the encoder comes back as
43 125 207 for 43 125 207. Where the export works nothing changes: it is tried first,
and given up on only after it has failed.

diff --git a/src/vrend/virgl_video.c b/src/vrend/virgl_video.c
index dc1e8cd..8b47837 100644
--- a/src/vrend/virgl_video.c
+++ b/src/vrend/virgl_video.c
@@ -438,12 +438,100 @@ static void destroy_video_dma_buf(struct virgl_video_dma_buf *dmabuf)
     }
 }
 
+/*
+ * Upload the picture to encode through memory, for a host that cannot export
+ * a video buffer as a dma-buf: the d3d12 driver on WSL 2 has no dma-bufs, and
+ * vaExportSurfaceHandle() fails on it. The callback reads the planes of the
+ * guest's picture into a VAImage, which vaPutImage() copies to the surface,
+ * as FFmpeg's hwupload does.
+ */
+static void encode_upload_picture_cpu(struct virgl_video_codec *codec,
+                                      struct virgl_video_buffer *buffer)
+{
+    VAStatus va_stat;
+    VAImageFormat fmt;
+    VAImage image;
+    void *data = NULL;
+    uint8_t *dst[3] = { NULL, NULL, NULL };
+    unsigned pitches[3] = { 0, 0, 0 };
+    unsigned i;
+    int ret;
+    bool derived = false;
+
+    if (!callbacks || !callbacks->encode_read_picture)
+        return;
+
+    memset(&fmt, 0, sizeof(fmt));
+    fmt.byte_order = VA_LSB_FIRST;
+    fmt.bits_per_pixel = 12;
+    switch (buffer->format) {
+    case PIPE_FORMAT_NV12:
+        fmt.fourcc = VA_FOURCC_NV12;
+        break;
+    case PIPE_FORMAT_NV21:
+        fmt.fourcc = VA_FOURCC_NV21;
+        break;
+    default:
+        virgl_error("upload through memory: pipe format %d not supported\n",
+                    buffer->format);
+        return;
+    }
+
+    /*
+     * Write to the surface itself when the driver can show it as an image, as FFmpeg's
+     * hwupload does. On the d3d12 driver the picture of an image made with vaCreateImage() and
+     * copied by vaPutImage() came out with an all-zero luma plane (the chroma was right),
+     * where the same planes written through vaDeriveImage() encode exactly.
+     */
+    va_stat = vaDeriveImage(va_dpy, buffer->va_sfc, &image);
+    if (VA_STATUS_SUCCESS == va_stat) {
+        derived = true;
+    } else {
+        va_stat = vaCreateImage(va_dpy, &fmt, buffer->width, buffer->height, &image);
+        if (VA_STATUS_SUCCESS != va_stat) {
+            virgl_error("create image failed, err = 0x%x\n", va_stat);
+            return;
+        }
+    }
+
+    va_stat = vaMapBuffer(va_dpy, image.buf, &data);
+    if (VA_STATUS_SUCCESS != va_stat || !data) {
+        virgl_error("map image failed, err = 0x%x\n", va_stat);
+        goto destroy_image;
+    }
+
+    for (i = 0; i < image.num_planes && i < ARRAY_SIZE(dst); i++) {
+        dst[i] = (uint8_t *)data + image.offsets[i];
+        pitches[i] = image.pitches[i];
+    }
+
+    ret = callbacks->encode_read_picture(codec, buffer, image.num_planes,
+                                         dst, pitches);
+    vaUnmapBuffer(va_dpy, image.buf);
+    if (ret) {
+        virgl_error("read picture failed, err = %d\n", ret);
+        goto destroy_image;
+    }
+
+    if (!derived) {
+        va_stat = vaPutImage(va_dpy, buffer->va_sfc, image.image_id,
+                             0, 0, buffer->width, buffer->height,
+                             0, 0, buffer->width, buffer->height);
+        if (VA_STATUS_SUCCESS != va_stat)
+            virgl_error("put image failed, err = 0x%x\n", va_stat);
+    }
+
+destroy_image:
+    vaDestroyImage(va_dpy, image.image_id);
+}
+
 static void encode_upload_picture(struct virgl_video_codec *codec,
                                   struct virgl_video_buffer *buffer)
 {
     VAStatus va_stat;
+    static bool dmabuf_unsupported;
 
-    if (!callbacks || !callbacks->encode_upload_picture)
+    if (!callbacks)
         return;
 
     va_stat = vaSyncSurface(va_dpy, buffer->va_sfc);
@@ -452,11 +540,20 @@ static void encode_upload_picture(struct virgl_video_codec *codec,
         return;
     }
 
-    if (!buffer->dmabuf)
+    /* once a surface cannot be exported none can: go through memory */
+    if (!buffer->dmabuf && !dmabuf_unsupported && callbacks->encode_upload_picture) {
         buffer->dmabuf = export_video_dma_buf(buffer, VIRGL_VIDEO_DMABUF_WRITE_ONLY);
+        if (!buffer->dmabuf) {
+            dmabuf_unsupported = true;
+            virgl_warn("cannot export a video buffer as a dma-buf, "
+                       "uploading the pictures to encode through memory\n");
+        }
+    }
 
     if (buffer->dmabuf)
         callbacks->encode_upload_picture(codec, buffer->dmabuf);
+    else
+        encode_upload_picture_cpu(codec, buffer);
 }
 
 static void encode_completed(struct virgl_video_codec *codec,
diff --git a/src/vrend/virgl_video.h b/src/vrend/virgl_video.h
index e954559..c39d33d 100644
--- a/src/vrend/virgl_video.h
+++ b/src/vrend/virgl_video.h
@@ -115,6 +115,16 @@ struct virgl_video_callbacks {
     void (*encode_upload_picture)(struct virgl_video_codec *codec,
                                   const struct virgl_video_dma_buf *dmabuf);
 
+    /* Read the picture data to be encoded into memory, plane by plane, for
+     * a host whose video buffers cannot be exported as dma-bufs (the d3d12
+     * driver of WSL 2): plane i is written to dst[i], pitches[i] bytes a row.
+     * Returns 0 when the planes are filled. */
+    int (*encode_read_picture)(struct virgl_video_codec *codec,
+                               struct virgl_video_buffer *buffer,
+                               unsigned num_planes,
+                               uint8_t * const *dst,
+                               const unsigned *pitches);
+
     /* Callback when encoding is complete, used to download the encoded data
      * and reference picture */
     void (*encode_completed)(struct virgl_video_codec *codec,
diff --git a/src/vrend/vrend_video.c b/src/vrend/vrend_video.c
index 57341d8..dc30c55 100644
--- a/src/vrend/vrend_video.c
+++ b/src/vrend/vrend_video.c
@@ -293,6 +293,58 @@ static void vrend_video_enocde_upload_picture(
     sync_video_buffer_to_dmabuf(buf, dmabuf);
 }
 
+/*
+ * The picture to encode is the planes of a video buffer the guest rendered to, which are
+ * textures here (R8 for the luma, RG8 for the chroma of NV12). Read them into the planes of a
+ * VAImage, for a host that cannot share its video buffer as a dma-buf.
+ */
+static int vrend_video_encode_read_picture(
+                                struct virgl_video_codec *codec,
+                                struct virgl_video_buffer *vbuf,
+                                unsigned num_planes,
+                                uint8_t * const *dst,
+                                const unsigned *pitches)
+{
+    struct vrend_video_buffer *buf = vrend_video_buffer(vbuf);
+    int ret = 0;
+
+    (void)codec;
+
+    if (!buf)
+        return -1;
+
+    glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+    glPixelStorei(GL_PACK_ALIGNMENT, 1);
+
+    for (unsigned i = 0; i < num_planes && i < buf->num_planes; i++) {
+        struct vrend_video_plane *plane = &buf->planes[i];
+        struct vrend_resource *res;
+        /* NV12 and NV21 have a plane of one byte a pixel, and one of two */
+        const unsigned bytes_per_pixel = i ? 2 : 1;
+
+        res = vrend_renderer_ctx_res_lookup(buf->ctx->ctx, plane->res_handle);
+        if (!res) {
+            virgl_error("%s: res %d not found\n", __func__, plane->res_handle);
+            ret = -1;
+            continue;
+        }
+
+        glBindFramebuffer(GL_READ_FRAMEBUFFER, plane->framebuffer);
+        glFramebufferTexture2D(GL_READ_FRAMEBUFFER, GL_COLOR_ATTACHMENT0,
+                               GL_TEXTURE_2D, res->gl_id, 0);
+
+        glPixelStorei(GL_PACK_ROW_LENGTH, pitches[i] / bytes_per_pixel);
+        glReadPixels(0, 0, res->base.width0, res->base.height0,
+                     i ? GL_RG : GL_RED, GL_UNSIGNED_BYTE, dst[i]);
+    }
+
+    glPixelStorei(GL_PACK_ROW_LENGTH, 0);
+    glPixelStorei(GL_PACK_ALIGNMENT, 4);
+    glBindFramebuffer(GL_FRAMEBUFFER, 0);
+
+    return ret;
+}
+
 static void vrend_video_encode_completed(
                                 struct virgl_video_codec *codec,
                                 const struct virgl_video_dma_buf *src_buf,
@@ -349,6 +401,7 @@ static void vrend_video_encode_completed(
 static struct virgl_video_callbacks video_callbacks = {
     .decode_completed           = vrend_video_decode_completed,
     .encode_upload_picture      = vrend_video_enocde_upload_picture,
+    .encode_read_picture        = vrend_video_encode_read_picture,
     .encode_completed           = vrend_video_encode_completed,
 };