vrend: apply CPU data reordering for all bgr* resources on GLES hosts

After cleaning up many of the previous BGRA hacks, it is clear now that
we must manually reorder all bgr* data en-route to/from the guest, but
only for GLES hosts.

GL hosts instead enjoy the use of the GL_BGRA pixel transfer format,
prompting the host driver to the the reordering internally, and only
when it deems it necessary.

Trying to use GL_BGRA_EXT for GLES hosts is somewhat possible with the
new GL_MESA_bgra extension, but it still struggles when multisampled
textures are involved. It is more robust to use the GL_RGBA pixel
transfer format for GLES hosts instead and handle reordering ourselves.

Signed-off-by: Ryan Neph <ryanneph@google.com>
Reviewed-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/725>
macos/master
Ryan Neph 2 years ago committed by Marge Bot
parent c1d7416555
commit 3a25c8c2ec
  1. 18
      src/vrend_renderer.c

@ -7935,9 +7935,8 @@ static int vrend_renderer_transfer_write_iov(struct vrend_context *ctx,
need_temp = true;
}
if (vrend_state.use_gles && vrend_format_is_bgra(res->base.format) &&
!vrend_resource_is_emulated_bgra(res))
need_temp = true;
if (vrend_state.use_gles && vrend_format_is_bgra(res->base.format))
need_temp = true;
if (vrend_state.use_core_profile == true &&
(res->y_0_top || (res->base.format == VIRGL_FORMAT_Z24X8_UNORM))) {
@ -8041,8 +8040,7 @@ static int vrend_renderer_transfer_write_iov(struct vrend_context *ctx,
/* GLES doesn't allow format conversions, which we need for BGRA resources with RGBA
* internal format. So we fallback to performing a CPU swizzle before uploading. */
if (vrend_state.use_gles && vrend_format_is_bgra(res->base.format) &&
!vrend_resource_is_emulated_bgra(res)) {
if (vrend_state.use_gles && vrend_format_is_bgra(res->base.format)) {
VREND_DEBUG(dbg_bgra, ctx, "manually swizzling bgra->rgba on upload since gles+bgra\n");
vrend_swizzle_data_bgra(send_size, data);
}
@ -8319,9 +8317,8 @@ static int vrend_transfer_send_readpixels(struct vrend_context *ctx,
if (num_iovs > 1 || separate_invert)
need_temp = 1;
if (vrend_state.use_gles && vrend_format_is_bgra(res->base.format) &&
!vrend_resource_is_emulated_bgra(res))
need_temp = true;
if (vrend_state.use_gles && vrend_format_is_bgra(res->base.format))
need_temp = true;
if (need_temp) {
send_size = util_format_get_nblocks(res->base.format, info->box->width, info->box->height) * info->box->depth * util_format_get_blocksize(res->base.format);
@ -8384,8 +8381,7 @@ static int vrend_transfer_send_readpixels(struct vrend_context *ctx,
* on upload and need to do the same on readback.
* The notable exception is externally-stored (GBM/EGL) BGR* resources, for which BGR*
* byte-ordering is used instead to match external access patterns. */
if (vrend_state.use_gles && vrend_format_is_bgra(res->base.format) &&
!vrend_resource_is_emulated_bgra(res)) {
if (vrend_state.use_gles && vrend_format_is_bgra(res->base.format)) {
VREND_DEBUG(dbg_bgra, ctx, "manually swizzling rgba->bgra on readback since gles+bgra\n");
vrend_swizzle_data_bgra(send_size, data);
}
@ -9064,7 +9060,7 @@ static void vrend_resource_copy_fallback(struct vrend_resource *src_res,
* On the contrary, externally-stored BGR* resources are assumed to remain in BGR* format at
* all times.
*/
if (vrend_format_is_bgra(dst_res->base.format) && !vrend_resource_is_emulated_bgra(dst_res))
if (vrend_state.use_gles && vrend_format_is_bgra(dst_res->base.format))
vrend_swizzle_data_bgra(total_size, tptr);
} else {
uint32_t read_chunk_size;

Loading…
Cancel
Save