glx/dri3: add GPU offloading support.

The differences with DRI2 GPU offloading are:
a) There's no logic for GPU offloading needed in the Xserver

b) for DRI2, the card would render to a back buffer, and
the content would be copied to the front buffer (the same buffers
everytime). Here we can potentially use several back buffers and copy
to buffers with no tiling to share with X. We send them with the
Present extension.

That means than the DRI2 solution is forced to have tearings with GPU
offloading. In the ideal scenario, this DRI3 solution doesn't have this
problem.

However without dma-buf fences, a race can appear (if the card is slow
and the rendering hasn't finished before the server card reads the buffer),
and then old content is displayed. If a user hits this, he should probably
revert to the DRI2 solution (LIBGL_DRI3_DISABLE). Users with cards fast
enough seem to not hit this in practice (I have an Amd hd 7730m, and I
don't hit this, except if I force a low dpm mode)

c) for non-fullscreen apps, the DRI2 GPU offloading solution requires
compositing. This DRI3 solution doesn't have this requirement. Rendering
to a pixmap also works.

d) There is no need to have a DDX loaded for the secondary card.

V4: Fixes some piglit tests

Signed-off-by: Axel Davy <axel.davy@ens.fr>
Signed-off-by: Dave Airlie <airlied@redhat.com>

2 files changed, 188 insertions, 32 deletions

diff --git a/src/glx/dri3_glx.c b/src/glx/dri3_glx.c
index 3092bc1..e3fc4de 100644
--- a/src/glx/dri3_glx.c
+++ b/src/glx/dri3_glx.c
@@ -596,22 +596,44 @@ dri3_copy_sub_buffer(__GLXDRIdrawable *pdraw, int x, int y,
 {
    struct dri3_drawable *priv = (struct dri3_drawable *) pdraw;
    struct dri3_screen *psc = (struct dri3_screen *) pdraw->psc;
+   struct dri3_context *pcp = (struct dri3_context *) __glXGetCurrentContext();
    xcb_connection_t     *c = XGetXCBConnection(priv->base.psc->dpy);
-   struct dri3_buffer *back = dri3_back_buffer(priv);
+   struct dri3_buffer *back;
 
-   unsigned flags;
+   unsigned flags = __DRI2_FLUSH_DRAWABLE;
 
    /* Check we have the right attachments */
    if (!priv->have_back || priv->is_pixmap)
       return;
 
-   flags = __DRI2_FLUSH_DRAWABLE;
    if (flush)
       flags |= __DRI2_FLUSH_CONTEXT;
    dri3_flush(psc, priv, flags, __DRI2_THROTTLE_SWAPBUFFER);
 
+   back = dri3_back_buffer(priv);
    y = priv->height - y - height;
 
+   if (psc->is_different_gpu && (&pcp->base != &dummyContext) && pcp->base.psc == &psc->base) {
+      /* Update the linear buffer part of the back buffer
+       * for the dri3_copy_area operation
+       */
+      psc->image->blitImage(pcp->driContext,
+                            back->linear_buffer,
+                            back->image,
+                            0, 0, back->width,
+                            back->height,
+                            0, 0, back->width,
+                            back->height, __BLIT_FLAG_FLUSH);
+      /* We use blitImage to update our fake front,
+       */
+      if (priv->have_fake_front)
+         psc->image->blitImage(pcp->driContext,
+                               dri3_fake_front_buffer(priv)->image,
+                               back->image,
+                               x, y, width, height,
+                               x, y, width, height, __BLIT_FLAG_FLUSH);
+   }
+
    dri3_fence_reset(c, back);
    dri3_copy_area(c,
                   dri3_back_buffer(priv)->pixmap,
@@ -622,7 +644,7 @@ dri3_copy_sub_buffer(__GLXDRIdrawable *pdraw, int x, int y,
    /* Refresh the fake front (if present) after we just damaged the real
     * front.
     */
-   if (priv->have_fake_front) {
+   if (priv->have_fake_front && !psc->is_different_gpu) {
       dri3_fence_reset(c, dri3_fake_front_buffer(priv));
       dri3_copy_area(c,
                      dri3_back_buffer(priv)->pixmap,
@@ -655,25 +677,62 @@ dri3_copy_drawable(struct dri3_drawable *priv, Drawable dest, Drawable src)
 static void
 dri3_wait_x(struct glx_context *gc)
 {
+   struct dri3_context *pcp = (struct dri3_context *) gc;
    struct dri3_drawable *priv = (struct dri3_drawable *)
       GetGLXDRIDrawable(gc->currentDpy, gc->currentDrawable);
+   struct dri3_screen *psc;
+   struct dri3_buffer *front;
 
    if (priv == NULL || !priv->have_fake_front)
       return;
 
-   dri3_copy_drawable(priv, dri3_fake_front_buffer(priv)->pixmap, priv->base.xDrawable);
+   psc = (struct dri3_screen *) priv->base.psc;
+   front = dri3_fake_front_buffer(priv);
+
+   dri3_copy_drawable(priv, front->pixmap, priv->base.xDrawable);
+
+   /* In the psc->is_different_gpu case, the linear buffer has been updated,
+    * but not yet the tiled buffer.
+    * Copy back to the tiled buffer we use for rendering.
+    * Note that we don't need flushing.
+    */
+   if (psc->is_different_gpu && (&pcp->base != &dummyContext) && pcp->base.psc == &psc->base)
+      psc->image->blitImage(pcp->driContext,
+                            front->image,
+                            front->linear_buffer,
+                            0, 0, front->width,
+                            front->height,
+                            0, 0, front->width,
+                            front->height, 0);
 }
 
 static void
 dri3_wait_gl(struct glx_context *gc)
 {
+   struct dri3_context *pcp = (struct dri3_context *) gc;
    struct dri3_drawable *priv = (struct dri3_drawable *)
       GetGLXDRIDrawable(gc->currentDpy, gc->currentDrawable);
+   struct dri3_screen *psc;
+   struct dri3_buffer *front;
 
    if (priv == NULL || !priv->have_fake_front)
       return;
 
-   dri3_copy_drawable(priv, priv->base.xDrawable, dri3_fake_front_buffer(priv)->pixmap);
+   psc = (struct dri3_screen *) priv->base.psc;
+   front = dri3_fake_front_buffer(priv);
+
+   /* In the psc->is_different_gpu case, we update the linear_buffer
+    * before updating the real front.
+    */
+   if (psc->is_different_gpu && (&pcp->base != &dummyContext) && pcp->base.psc == &psc->base)
+      psc->image->blitImage(pcp->driContext,
+                            front->linear_buffer,
+                            front->image,
+                            0, 0, front->width,
+                            front->height,
+                            0, 0, front->width,
+                            front->height, __BLIT_FLAG_FLUSH);
+   dri3_copy_drawable(priv, priv->base.xDrawable, front->pixmap);
 }
 
 /**
@@ -741,6 +800,7 @@ dri3_alloc_render_buffer(struct glx_screen *glx_screen, Drawable draw,
    struct dri3_screen *psc = (struct dri3_screen *) glx_screen;
    Display *dpy = glx_screen->dpy;
    struct dri3_buffer *buffer;
+   __DRIimage *pixmap_buffer;
    xcb_connection_t *c = XGetXCBConnection(dpy);
    xcb_pixmap_t pixmap;
    xcb_sync_fence_t sync_fence;
@@ -769,24 +829,47 @@ dri3_alloc_render_buffer(struct glx_screen *glx_screen, Drawable draw,
    if (!buffer->cpp)
       goto no_image;
 
-   buffer->image = (*psc->image->createImage) (psc->driScreen,
-                                               width, height,
-                                               format,
-                                               __DRI_IMAGE_USE_SHARE|__DRI_IMAGE_USE_SCANOUT,
-                                               buffer);
-
-
-   if (!buffer->image)
-      goto no_image;
+   if (!psc->is_different_gpu) {
+      buffer->image = (*psc->image->createImage) (psc->driScreen,
+                                                  width, height,
+                                                  format,
+                                                  __DRI_IMAGE_USE_SHARE |
+                                                  __DRI_IMAGE_USE_SCANOUT,
+                                                  buffer);
+      pixmap_buffer = buffer->image;
+
+      if (!buffer->image)
+         goto no_image;
+   } else {
+      buffer->image = (*psc->image->createImage) (psc->driScreen,
+                                                  width, height,
+                                                  format,
+                                                  0,
+                                                  buffer);
+
+      if (!buffer->image)
+         goto no_image;
+
+      buffer->linear_buffer = (*psc->image->createImage) (psc->driScreen,
+                                                          width, height,
+                                                          format,
+                                                          __DRI_IMAGE_USE_SHARE |
+                                                          __DRI_IMAGE_USE_LINEAR,
+                                                          buffer);
+      pixmap_buffer = buffer->linear_buffer;
+
+      if (!buffer->linear_buffer)
+         goto no_linear_buffer;
+   }
 
    /* X wants the stride, so ask the image for it
     */
-   if (!(*psc->image->queryImage)(buffer->image, __DRI_IMAGE_ATTRIB_STRIDE, &stride))
+   if (!(*psc->image->queryImage)(pixmap_buffer, __DRI_IMAGE_ATTRIB_STRIDE, &stride))
       goto no_buffer_attrib;
 
    buffer->pitch = stride;
 
-   if (!(*psc->image->queryImage)(buffer->image, __DRI_IMAGE_ATTRIB_FD, &buffer_fd))
+   if (!(*psc->image->queryImage)(pixmap_buffer, __DRI_IMAGE_ATTRIB_FD, &buffer_fd))
       goto no_buffer_attrib;
 
    xcb_dri3_pixmap_from_buffer(c,
@@ -817,7 +900,10 @@ dri3_alloc_render_buffer(struct glx_screen *glx_screen, Drawable draw,
    return buffer;
 
 no_buffer_attrib:
-   (*psc->image->destroyImage)(buffer->image);
+   (*psc->image->destroyImage)(pixmap_buffer);
+no_linear_buffer:
+   if (psc->is_different_gpu)
+      (*psc->image->destroyImage)(buffer->image);
 no_image:
    free(buffer);
 no_buffer:
@@ -843,6 +929,8 @@ dri3_free_render_buffer(struct dri3_drawable *pdraw, struct dri3_buffer *buffer)
    xcb_sync_destroy_fence(c, buffer->sync_fence);
    xshmfence_unmap_shm(buffer->shm_fence);
    (*psc->image->destroyImage)(buffer->image);
+   if (buffer->linear_buffer)
+      (*psc->image->destroyImage)(buffer->linear_buffer);
    free(buffer);
 }
 
@@ -1118,7 +1206,9 @@ dri3_get_buffer(__DRIdrawable *driDrawable,
                 enum dri3_buffer_type buffer_type,
                 void *loaderPrivate)
 {
+   struct dri3_context *pcp = (struct dri3_context *) __glXGetCurrentContext();
    struct dri3_drawable *priv = loaderPrivate;
+   struct dri3_screen *psc = (struct dri3_screen *) priv->base.psc;
    xcb_connection_t     *c = XGetXCBConnection(priv->base.psc->dpy);
    struct dri3_buffer      *buffer;
    int                  buf_id;
@@ -1154,14 +1244,24 @@ dri3_get_buffer(__DRIdrawable *driDrawable,
       switch (buffer_type) {
       case dri3_buffer_back:
          if (buffer) {
-            dri3_fence_reset(c, new_buffer);
-            dri3_fence_await(c, buffer);
-            dri3_copy_area(c,
-                           buffer->pixmap,
-                           new_buffer->pixmap,
-                           dri3_drawable_gc(priv),
-                           0, 0, 0, 0, priv->width, priv->height);
+            if (!buffer->linear_buffer) {
+               dri3_fence_reset(c, new_buffer);
+               dri3_fence_await(c, buffer);
+               dri3_copy_area(c,
+                              buffer->pixmap,
+                              new_buffer->pixmap,
+                              dri3_drawable_gc(priv),
+                              0, 0, 0, 0, priv->width, priv->height);
             dri3_fence_trigger(c, new_buffer);
+            } else if ((&pcp->base != &dummyContext) && pcp->base.psc == &psc->base) {
+               psc->image->blitImage(pcp->driContext,
+                                     new_buffer->image,
+                                     buffer->image,
+                                     0, 0, priv->width,
+                                     priv->height,
+                                     0, 0, priv->width,
+                                     priv->height, 0);
+            }
             dri3_free_render_buffer(priv, buffer);
          }
          break;
@@ -1173,6 +1273,17 @@ dri3_get_buffer(__DRIdrawable *driDrawable,
                         dri3_drawable_gc(priv),
                         0, 0, 0, 0, priv->width, priv->height);
          dri3_fence_trigger(c, new_buffer);
+
+         if (new_buffer->linear_buffer && (&pcp->base != &dummyContext) && pcp->base.psc == &psc->base) {
+            dri3_fence_await(c, new_buffer);
+            psc->image->blitImage(pcp->driContext,
+                                  new_buffer->image,
+                                  new_buffer->linear_buffer,
+                                  0, 0, priv->width,
+                                  priv->height,
+                                  0, 0, priv->width,
+                                  priv->height, 0);
+         }
          break;
       }
       buffer = new_buffer;
@@ -1235,6 +1346,7 @@ dri3_get_buffers(__DRIdrawable *driDrawable,
                  struct __DRIimageList *buffers)
 {
    struct dri3_drawable *priv = loaderPrivate;
+   struct dri3_screen *psc = (struct dri3_screen *) priv->base.psc;
    struct dri3_buffer   *front, *back;
 
    buffers->image_mask = 0;
@@ -1252,7 +1364,15 @@ dri3_get_buffers(__DRIdrawable *driDrawable,
       buffer_mask |= __DRI_IMAGE_BUFFER_FRONT;
 
    if (buffer_mask & __DRI_IMAGE_BUFFER_FRONT) {
-      if (priv->is_pixmap)
+      /* All pixmaps are owned by the server gpu.
+       * When we use a different gpu, we can't use the pixmap
+       * as buffer since it is potentially tiled a way
+       * our device can't understand. In this case, use
+       * a fake front buffer. Hopefully the pixmap
+       * content will get synced with the fake front
+       * buffer.
+       */
+      if (priv->is_pixmap && !psc->is_different_gpu)
          front = dri3_get_pixmap_buffer(driDrawable,
                                         format,
                                         dri3_buffer_front,
@@ -1286,7 +1406,7 @@ dri3_get_buffers(__DRIdrawable *driDrawable,
    if (front) {
       buffers->image_mask |= __DRI_IMAGE_BUFFER_FRONT;
       buffers->front = front->image;
-      priv->have_fake_front = !priv->is_pixmap;
+      priv->have_fake_front = psc->is_different_gpu || !priv->is_pixmap;
    }
 
    if (back) {
@@ -1327,6 +1447,7 @@ static int64_t
 dri3_swap_buffers(__GLXDRIdrawable *pdraw, int64_t target_msc, int64_t divisor,
                   int64_t remainder, Bool flush)
 {
+   struct dri3_context *pcp = (struct dri3_context *) __glXGetCurrentContext();
    struct dri3_drawable *priv = (struct dri3_drawable *) pdraw;
    struct dri3_screen *psc = (struct dri3_screen *) priv->base.psc;
    Display *dpy = priv->base.psc->dpy;
@@ -1340,6 +1461,26 @@ dri3_swap_buffers(__GLXDRIdrawable *pdraw, int64_t target_msc, int64_t divisor,
    dri3_flush(psc, priv, flags, __DRI2_THROTTLE_SWAPBUFFER);
 
    back = priv->buffers[DRI3_BACK_ID(priv->cur_back)];
+   if (psc->is_different_gpu && back) {
+      /* Update the linear buffer before presenting the pixmap */
+      psc->image->blitImage(pcp->driContext,
+                            back->linear_buffer,
+                            back->image,
+                            0, 0, back->width,
+                            back->height,
+                            0, 0, back->width,
+                            back->height, __BLIT_FLAG_FLUSH);
+      /* Update the fake front */
+      if (priv->have_fake_front)
+         psc->image->blitImage(pcp->driContext,
+                               priv->buffers[DRI3_FRONT_ID]->image,
+                               back->image,
+                               0, 0, priv->width,
+                               priv->height,
+                               0, 0, priv->width,
+                               priv->height, __BLIT_FLAG_FLUSH);
+   }
+
    dri3_flush_present_events(priv);
 
    if (back && !priv->is_pixmap) {
@@ -1376,7 +1517,7 @@ dri3_swap_buffers(__GLXDRIdrawable *pdraw, int64_t target_msc, int64_t divisor,
        * to reset the fence and make future users block until
        * the X server is done copying the bits
        */
-      if (priv->have_fake_front) {
+      if (priv->have_fake_front && !psc->is_different_gpu) {
          dri3_fence_reset(c, priv->buffers[DRI3_FRONT_ID]);
          dri3_copy_area(c,
                         back->pixmap,
@@ -1592,7 +1733,12 @@ dri3_bind_extensions(struct dri3_screen *psc, struct glx_display * priv,
                                  "GLX_EXT_create_context_es2_profile");
 
    for (i = 0; extensions[i]; i++) {
-      if ((strcmp(extensions[i]->name, __DRI_TEX_BUFFER) == 0)) {
+      /* when on a different gpu than the server, the server pixmaps
+       * can have a tiling mode we can't read. Thus we can't create
+       * a texture from them.
+       */
+      if (!psc->is_different_gpu &&
+         (strcmp(extensions[i]->name, __DRI_TEX_BUFFER) == 0)) {
          psc->texBuffer = (__DRItexBufferExtension *) extensions[i];
          __glXEnableDirectExtension(&psc->base, "GLX_EXT_texture_from_pixmap");
       }
@@ -1676,6 +1822,8 @@ dri3_create_screen(int screen, struct glx_display * priv)
 
       return NULL;
    }
+
+   psc->fd = loader_get_user_preferred_fd(psc->fd, &psc->is_different_gpu);
    deviceName = NULL;
 
    driverName = loader_get_driver_for_fd(psc->fd, 0);
@@ -1735,9 +1883,15 @@ dri3_create_screen(int screen, struct glx_display * priv)
       goto handle_error;
    }
 
-   if (!psc->texBuffer || psc->texBuffer->base.version < 2 ||
-       !psc->texBuffer->setTexBuffer2)
-   {
+   if (psc->is_different_gpu && psc->image->base.version < 9) {
+      ErrorMessageF("Different GPU, but image extension version 9 or later not found\n");
+      goto handle_error;
+   }
+
+   if (!psc->is_different_gpu && (
+       !psc->texBuffer || psc->texBuffer->base.version < 2 ||
+       !psc->texBuffer->setTexBuffer2
+       )) {
       ErrorMessageF("Version 2 or later of texBuffer extension not found\n");
       goto handle_error;
    }
diff --git a/src/glx/dri3_priv.h b/src/glx/dri3_priv.h
index 6894886..c0e35ee 100644
--- a/src/glx/dri3_priv.h
+++ b/src/glx/dri3_priv.h
@@ -72,6 +72,7 @@ enum dri3_buffer_type {
 
 struct dri3_buffer {
    __DRIimage   *image;
+   __DRIimage   *linear_buffer;
    uint32_t     pixmap;
 
    /* Synchronization between the client and X server is done using an
@@ -135,6 +136,7 @@ struct dri3_screen {
 
    void *driver;
    int fd;
+   int is_different_gpu;
 
    Bool show_fps;
 };