etnaviv: Bypass BGRA-internal optimization for shared resources

The rb-swap rework (commit a5766e75e48) uses A8R8G8B8 as the texture
format for RB_SWAP pipe formats and performs R<->B conversion during
transfer blits. This breaks dmabuf import/export: external producers
write RGBA bytes, but A8R8G8B8 interprets them as BGRA; transfer blits
incorrectly swap R<->B on data already in the correct byte order.

For shared/imported resources, bypass the optimization: use the native
texture format (A8B8G8R8) so the sampler reads RGBA bytes correctly in
both the descriptor and state-based texture paths, skip the R<->B
swizzle in BLT and RS transfer blits since shared resources already
store data in standard byte order, and re-enable the texture shadow
shortcut for shared RB_SWAP resources since the BLT swap is no longer
applied.

Internal-only resources continue using the BGRA-internal optimization
unchanged.

Fixes: a5766e75e48 ("etnaviv: Use BGRA-internal texture format with BLT/RS R/B swizzle")
Signed-off-by: Christian Gmeiner <cgmeiner@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/40029>

src/gallium/drivers/etnaviv/etnaviv_blt.c

@@ -791,7 +791,8 @@ etna_try_blt_blit(struct pipe_context *pctx,
       /* For transfer blits of RB_SWAP formats, apply R<->B swizzle on the
        * linear side to convert between GPU-internal BGRA and CPU RGBA. */
       if (ctx->in_transfer_blit &&
-          translate_pe_format_rb_swap(blit_info->src.format)) {
+          translate_pe_format_rb_swap(blit_info->src.format) &&
+          !src->shared && !dst->shared) {
          bool src_linear = src->layout == ETNA_LAYOUT_LINEAR;
          bool dst_linear = dst->layout == ETNA_LAYOUT_LINEAR;
 

src/gallium/drivers/etnaviv/etnaviv_format.c

@@ -395,6 +395,21 @@ translate_pe_format_rb_swap(enum pipe_format fmt)
    return formats[fmt].pe & PE_FORMAT_RB_SWAP;
 }
 
+/* For shared resources with RB_SWAP formats, remaps the HW texture format
+ * back to the one matching the actual byte order in memory (e.g., A8B8G8R8
+ * for RGBA data). Normally we use A8R8G8B8 to match PE-internal BGRA byte
+ * order, but shared resources store data in the standard byte order.
+ */
+uint32_t
+remap_texture_format_rb_swap(uint32_t format)
+{
+   switch (format) {
+   case TEXTURE_FORMAT_A8R8G8B8: return TEXTURE_FORMAT_A8B8G8R8;
+   case TEXTURE_FORMAT_X8R8G8B8: return TEXTURE_FORMAT_X8B8G8R8;
+   default: return format;
+   }
+}
+
 enum pipe_format
 translate_pe_internal_format(enum pipe_format fmt)
 {

src/gallium/drivers/etnaviv/etnaviv_format.h

@@ -56,6 +56,9 @@ translate_pe_format(enum pipe_format fmt);
 int
 translate_pe_format_rb_swap(enum pipe_format fmt);
 
+uint32_t
+remap_texture_format_rb_swap(uint32_t format);
+
 enum pipe_format
 translate_pe_internal_format(enum pipe_format fmt);
 

src/gallium/drivers/etnaviv/etnaviv_rs.c

@@ -884,7 +884,8 @@ etna_try_rs_blit(struct pipe_context *pctx,
       .downsample_x = downsample_x,
       .downsample_y = downsample_y,
       .swap_rb = ctx->in_transfer_blit &&
-                 translate_pe_format_rb_swap(blit_info->src.format),
+                 translate_pe_format_rb_swap(blit_info->src.format) &&
+                 !src->shared && !dst->shared,
       .dither = {0xffffffff, 0xffffffff}, // XXX dither when going from 24 to 16 bit?
       .clear_mode = VIVS_RS_CLEAR_CONTROL_MODE_DISABLED,
       .width = width,

src/gallium/drivers/etnaviv/etnaviv_texture_desc.c

@@ -137,7 +137,14 @@ etna_create_sampler_view_desc(struct pipe_context *pctx, struct pipe_resource *p
    const struct util_format_description *desc = util_format_description(so->format);
    struct etna_sampler_view_desc *sv = CALLOC_STRUCT(etna_sampler_view_desc);
    struct etna_context *ctx = etna_context(pctx);
-   const uint32_t format = translate_texture_format(so->format);
+   uint32_t format = translate_texture_format(so->format);
+
+   /* Shared resources store data in standard byte order (e.g., RGBA for
+    * R8G8B8A8_UNORM). Use the native texture format to read them correctly,
+    * bypassing the BGRA-internal optimization used for internal resources. */
+   if (etna_resource(prsc)->shared && translate_pe_format_rb_swap(so->format))
+      format = remap_texture_format_rb_swap(format);
+
    const bool ext = !!(format & EXT_FORMAT);
    const bool astc = !!(format & ASTC_FORMAT);
    const uint32_t swiz = get_texture_swiz(so->format, so->swizzle_r,

src/gallium/drivers/etnaviv/etnaviv_texture_state.c

@@ -164,7 +164,14 @@ etna_create_sampler_view_state(struct pipe_context *pctx, struct pipe_resource *
    struct etna_sampler_view *sv = CALLOC_STRUCT(etna_sampler_view);
    struct etna_context *ctx = etna_context(pctx);
    struct etna_screen *screen = ctx->screen;
-   const uint32_t format = translate_texture_format(so->format);
+   uint32_t format = translate_texture_format(so->format);
+
+   /* Shared resources store data in standard byte order (e.g., RGBA for
+    * R8G8B8A8_UNORM). Use the native texture format to read them correctly,
+    * bypassing the BGRA-internal optimization used for internal resources. */
+   if (etna_resource(prsc)->shared && translate_pe_format_rb_swap(so->format))
+      format = remap_texture_format_rb_swap(format);
+
    const bool ext = !!(format & EXT_FORMAT);
    const bool astc = !!(format & ASTC_FORMAT);
    const bool srgb = util_format_is_srgb(so->format);

src/gallium/drivers/etnaviv/etnaviv_transfer.c

@@ -347,7 +347,7 @@ etna_texture_map(struct pipe_context *pctx, struct pipe_resource *prsc,
    }
 
    if (rsc->texture && !etna_resource_newer(rsc, etna_resource(rsc->texture)) &&
-       !translate_pe_format_rb_swap(prsc->format)) {
+       (!translate_pe_format_rb_swap(prsc->format) || rsc->shared)) {
       /* We have a texture resource which is the same age or newer than the
        * render resource. Use the texture resource, which avoids bouncing
        * pixels between the two resources, and we can de-tile it in s/w. */