Add Vulkan Support to RetroPie for Pi 4?

sergioad

How the development is going? I am hyped by the idea of Redream + Vulkan

dankcushions

Vulkan is not included in the RetroArch build and support for it is not present in Raspberry Pi OS.

we need to wait for the next version of raspberry pi OS, based on debian bullseye, which should include mesa with vulkan support. once that is there, retropie could update the script modules to include support for retroarch/appropriate emulators.

I am hyped by the idea of Redream + Vulkan

does redream support vulkan? i don't believe it does.

sergioad

@dankcushions thanks for the reply

sergioad

@dankcushions AFAIK it does

dankcushions

@sergioad it does not. confirmed by the dev on their discord many times, and the official website: https://redream.io/help

flycast supports vulkan, although apparently pi4 has issues: https://github.com/libretro/flycast/issues/983

sergioad

@dankcushions ok, thanks

bluestang

RPiOS "Bullseye" now has the 20.3.5 version of MESA and the Broadcom Vulkan driver is part of the mesa-vulkan-drivers pkg if installed through apt...

In upstream MESA, the rpi4 Vulkan driver is now 1.1 conformant but still lacks a few extensions and driver code improvements that 3D emulators would benefit from.

Nonetheless, it is an option available to the user to install through apt. Perhaps it might be time to make it an option in RetroPie?

mitu

@bluestang said in Add Vulkan Support to RetroPie for Pi 4?:

Perhaps it might be time to make it an option in RetroPie?

Does it work without xorg ? From my limited testing, it's not working with RetroArch in the current configuration (KMSDSM without xorg).

bluestang

@mitu said in Add Vulkan Support to RetroPie for Pi 4?:

@bluestang said in Add Vulkan Support to RetroPie for Pi 4?:

Perhaps it might be time to make it an option in RetroPie?

Does it work without xorg ? From my limited testing, it's not working with RetroArch in the current configuration (KMSDSM without xorg).

It should work in the KMS API, the extension VK_KHR_display allows it to do so.

What is your current configuration?

mitu

@bluestang Just a default RPI OS 11 install, with mesa-drivers-vulkan installed.

bluestang

@mitu said in Add Vulkan Support to RetroPie for Pi 4?:

@bluestang Just a default RPI OS 11 install, with mesa-drivers-vulkan installed.

Are you able to get vkcube running in KMS?

mitu

@bluestang No, it's not running. RetroArch seems to work from the desktop, but not without.

bluestang

@mitu said in Add Vulkan Support to RetroPie for Pi 4?:

@bluestang No, it's not running. RetroArch seems to work from the desktop, but not without.

Are you able to get RetroArch working in KMS just using GLES? I just made another fresh image (64-bit) and I can see your issues, I'm trying to pin down what the problem is because compiling MESA on my own I never had this issue before...

mitu

RetroArch works fine with the GL driver (using GLES3) and I know that Vulkan also works with KMS - using a 21.x version of Mesa and the included v3v Vulkan driver.
What I'm saying is that it doesn't appear to work (on KMS) with the Mesa out-of-the-box version included in the new RPI OS 'bullseye' release (20.3.5) .

bluestang

@mitu said in Add Vulkan Support to RetroPie for Pi 4?:

RetroArch works fine with the GL driver (using GLES3) and I know that Vulkan also works with KMS - using a 21.x version of Mesa and the included v3v Vulkan driver.
What I'm saying is that it doesn't appear to work (on KMS) with the Mesa out-of-the-box version included in the new RPI OS 'bullseye' release (20.3.5) .

Thanks for the clarification and you are right, Vulkan does not work in KMS with the included MESA (20.3.5) in Bullseye.

It was backported into MESA 21.0.0. How annoying...

bluestang

Updating this thread for informational purposes:

If you want to run the latest MESA git, there are several patches that are needed based on the 20.3.5 diff file from the RPiOS repository: (I kept the relevant changes that are not already upstreamed)

Without it, I ran into some display issues which I presume are the modifiers utilized in the newer libdrm pkgs.

Note: you must compile libdrm, libglvnd, and mesa for it to work.

diff --git a/src/gallium/drivers/v3d/driinfo_v3d.h b/src/gallium/drivers/v3d/driinfo_v3d.h
index 147ad0b49bd..a524a03e7e5 100644
--- a/src/gallium/drivers/v3d/driinfo_v3d.h
+++ b/src/gallium/drivers/v3d/driinfo_v3d.h
@@ -2,4 +2,6 @@
 
 DRI_CONF_SECTION_MISCELLANEOUS
    DRI_CONF_V3D_NONMSAA_TEXTURE_SIZE_LIMIT(false)
+   DRI_CONF_V3D_MAINTAIN_IGNORABLE_SCANOUT(false)
+   DRI_CONF_V3D_IGNORE_SCANOUT_USAGES_WITH_MODIFIERS(false)
 DRI_CONF_SECTION_END
diff --git a/src/gallium/drivers/v3d/meson.build b/src/gallium/drivers/v3d/meson.build
index b760ca5c025..1a537637b03 100644
--- a/src/gallium/drivers/v3d/meson.build
+++ b/src/gallium/drivers/v3d/meson.build
@@ -59,6 +59,16 @@ endif
 
 v3d_versions = ['33', '41']
 
+v3d_deps = [dep_v3dv3, dep_libdrm, dep_valgrind, idep_nir_headers]
+
+if with_platform_x11
+  v3d_deps += dep_xcb
+endif
+
+if with_platform_wayland
+  v3d_deps += dep_wayland_client
+endif
+
 per_version_libs = []
 foreach ver : v3d_versions
   per_version_libs += static_library(
@@ -70,7 +80,7 @@ foreach ver : v3d_versions
     ],
     c_args : [v3d_args, '-DV3D_VERSION=' + ver],
     gnu_symbol_visibility : 'hidden',
-    dependencies : [dep_v3dv3, dep_libdrm, dep_valgrind, idep_nir_headers],
+    dependencies : v3d_deps,
 )
 
 endforeach
@@ -93,7 +103,7 @@ libv3d = static_library(
   c_args : [v3d_args],
   cpp_args : [v3d_args],
   gnu_symbol_visibility : 'hidden',
-  dependencies : [dep_v3dv3, dep_libdrm, dep_valgrind, idep_nir_headers],
+  dependencies : v3d_deps,
   link_with: [per_version_libs],
 )
 
diff --git a/src/gallium/drivers/v3d/v3d_resource.c b/src/gallium/drivers/v3d/v3d_resource.c
index 0885a2cb80f..36b6ab34d71 100644
--- a/src/gallium/drivers/v3d/v3d_resource.c
+++ b/src/gallium/drivers/v3d/v3d_resource.c
@@ -436,7 +436,7 @@ v3d_resource_get_handle(struct pipe_screen *pscreen,
         case WINSYS_HANDLE_TYPE_SHARED:
                 return v3d_bo_flink(bo, &whandle->handle);
         case WINSYS_HANDLE_TYPE_KMS:
-                if (screen->ro) {
+                if (screen->ro && rsc->scanout) {
                         if (renderonly_get_handle(rsc->scanout, whandle)) {
                                 whandle->stride = rsc->slices[0].stride;
                                 return true;
@@ -750,6 +750,30 @@ v3d_resource_setup(struct pipe_screen *pscreen,
         return rsc;
 }
 
+static bool
+v3d_resource_should_scanout(struct pipe_screen *pscreen,
+                            const struct pipe_resource *tmpl,
+                            const uint64_t *modifiers,
+                            int count)
+{
+        struct v3d_screen *screen = v3d_screen(pscreen);
+
+        if (tmpl->bind & PIPE_BIND_SCANOUT) {
+                if (screen->maintain_ignorable_scanout)
+                        return true;
+                if (screen->has_x_session && screen->ignore_scanout_usages)
+                        return false;
+                return true;
+        }
+
+        if (tmpl->bind & PIPE_BIND_SHARED) {
+                if (screen->ignore_scanout_usages_with_modifiers)
+                        return false;
+                return true;
+        }
+        return false;
+}
+
 static struct pipe_resource *
 v3d_resource_create_with_modifiers(struct pipe_screen *pscreen,
                                    const struct pipe_resource *tmpl,
@@ -763,6 +787,8 @@ v3d_resource_create_with_modifiers(struct pipe_screen *pscreen,
         struct pipe_resource *prsc = &rsc->base;
         /* Use a tiled layout if we can, for better 3D performance. */
         bool should_tile = true;
+        bool should_scanout = v3d_resource_should_scanout(pscreen, tmpl,
+                                                          modifiers, count);
 
         /* VBOs/PBOs are untiled (and 1 height). */
         if (tmpl->target == PIPE_BUFFER)
@@ -788,8 +814,8 @@ v3d_resource_create_with_modifiers(struct pipe_screen *pscreen,
         /* If using the old-school SCANOUT flag, we don't know what the screen
          * might support other than linear. Just force linear.
          */
-        if (tmpl->bind & PIPE_BIND_SCANOUT)
-                should_tile = false;
+        if ((tmpl->bind & PIPE_BIND_SCANOUT) && should_scanout)
+                 should_tile = false;
 
         /* No user-specified modifier; determine our own. */
         if (count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID) {
@@ -818,8 +844,8 @@ v3d_resource_create_with_modifiers(struct pipe_screen *pscreen,
          * We always allocate this way for SHARED, because get_handle will
          * need a resource on the display fd.
          */
-        if (screen->ro && (tmpl->bind & (PIPE_BIND_SCANOUT |
-                                         PIPE_BIND_SHARED))) {
+
+        if (screen->ro && should_scanout) {
                 struct winsys_handle handle;
                 struct pipe_resource scanout_tmpl = {
                         .target = prsc->target,
@@ -949,7 +975,7 @@ v3d_resource_from_handle(struct pipe_screen *pscreen,
                  }
         }
 
-        if (screen->ro) {
+        if (screen->ro && !rsc->tiled) {
                 /* Make sure that renderonly has a handle to our buffer in the
                  * display's fd, so that a later renderonly_get_handle()
                  * returns correct handles or GEM names.
@@ -995,7 +1021,9 @@ v3d_update_shadow_texture(struct pipe_context *pctx,
 
         assert(view->texture != pview->texture);
 
-        if (shadow->writes == orig->writes && orig->bo->private)
+        if (shadow->writes == orig->writes &&
+            orig->base.sync_status == 0 &&
+            (orig->bo->private || orig->base.sync_condition))
                 return;
 
         perf_debug("Updating %dx%d@%d shadow for linear texture\n",
@@ -1038,6 +1066,7 @@ v3d_update_shadow_texture(struct pipe_context *pctx,
         }
 
         shadow->writes = orig->writes;
+        orig->base.sync_status = 0;
 }
 
 static struct pipe_surface *
diff --git a/src/gallium/drivers/v3d/v3d_screen.c b/src/gallium/drivers/v3d/v3d_screen.c
index 1c72dc4af0d..1ac9e8fc1d9 100644
--- a/src/gallium/drivers/v3d/v3d_screen.c
+++ b/src/gallium/drivers/v3d/v3d_screen.c
@@ -47,6 +47,42 @@
 #include "compiler/v3d_compiler.h"
 #include "drm-uapi/drm_fourcc.h"
 
+#ifdef HAVE_WAYLAND_PLATFORM
+#include <wayland-client.h>
+#endif
+
+#ifdef HAVE_X11_PLATFORM
+#include <xcb/xcb.h>
+#endif
+
+static bool
+check_x_session()
+{
+        bool xcb_connection = false;
+
+#ifdef HAVE_WAYLAND_PLATFORM
+        struct wl_display *display;
+
+        display = wl_display_connect(NULL);
+
+        if (display) {
+                wl_display_disconnect(display);
+                return xcb_connection;
+        }
+#endif
+
+#ifdef HAVE_X11_PLATFORM
+        xcb_connection_t *conn;
+
+        conn = xcb_connect(NULL, NULL);
+
+        if (!xcb_connection_has_error(conn))
+                xcb_connection = true;
+        xcb_disconnect(conn);
+#endif
+        return xcb_connection;
+}
+
 static const char *
 v3d_screen_get_name(struct pipe_screen *pscreen)
 {
@@ -828,6 +864,28 @@ v3d_screen_create(int fd, const struct pipe_screen_config *config,
         screen->has_cache_flush =
                 v3d_has_feature(screen, DRM_V3D_PARAM_SUPPORTS_CACHE_FLUSH);
         screen->has_perfmon = v3d_has_feature(screen, DRM_V3D_PARAM_SUPPORTS_PERFMON);
+        
+        screen->has_x_session = check_x_session();
+        
+        screen->ignore_scanout_usages = getenv("V3D_IGNORE_SCANOUT_USAGES");
+
+        const char *ignore_scanout_usages_with_modifiers_name =
+                "v3d_ignore_scanout_usages_with_modifiers";
+        screen->ignore_scanout_usages_with_modifiers =
+                driCheckOption(config->options,
+                               ignore_scanout_usages_with_modifiers_name,
+                               DRI_BOOL) &&
+                driQueryOptionb(config->options,
+                                ignore_scanout_usages_with_modifiers_name);
+
+        const char *maintain_ignorable_scanout_name =
+                "v3d_maintain_ignorable_scanout";
+        screen->maintain_ignorable_scanout =
+                driCheckOption(config->options,
+                               maintain_ignorable_scanout_name,
+                               DRI_BOOL) &&
+                driQueryOptionb(config->options,
+                                maintain_ignorable_scanout_name);
 
         v3d_fence_init(screen);
 
diff --git a/src/gallium/drivers/v3d/v3d_screen.h b/src/gallium/drivers/v3d/v3d_screen.h
index 9bf2a065487..afdce3f3c73 100644
--- a/src/gallium/drivers/v3d/v3d_screen.h
+++ b/src/gallium/drivers/v3d/v3d_screen.h
@@ -82,6 +82,12 @@ struct v3d_screen {
         bool has_cache_flush;
         bool has_perfmon;
         bool nonmsaa_texture_size_limit;
+        bool ignore_scanout_usages;
+        bool ignore_scanout_usages_with_modifiers;
+        bool maintain_ignorable_scanout;
+        
+        /* Are we running in an X session? */
+        bool has_x_session;
 
         struct v3d_simulator_file *sim_file;
 };
diff --git a/src/gallium/include/pipe/p_state.h b/src/gallium/include/pipe/p_state.h
index c96a387bab8..708bb2fad4f 100644
--- a/src/gallium/include/pipe/p_state.h
+++ b/src/gallium/include/pipe/p_state.h
@@ -572,6 +572,10 @@ struct pipe_resource
    unsigned usage:8;         /**< PIPE_USAGE_x (not a bitmask) */
    unsigned bind;            /**< bitmask of PIPE_BIND_x */
    unsigned flags;           /**< bitmask of PIPE_RESOURCE_FLAG_x */
+   
+   /* Hack for avoiding sync on v3d */
+   unsigned sync_condition;
+   unsigned sync_status;
 
    /**
     * For planar images, ie. YUV EGLImage external, etc, pointer to the
diff --git a/src/glx/dri_common.h b/src/glx/dri_common.h
index 455e8541ed8..1fe78e73b75 100644
--- a/src/glx/dri_common.h
+++ b/src/glx/dri_common.h
@@ -56,6 +56,10 @@ extern struct glx_config *driConvertConfigs(const __DRIcoreExtension * core,
                                            const __DRIconfig ** configs);
 
 extern void driDestroyConfigs(const __DRIconfig **configs);
+ 
+#ifndef __GLXDRIdrawable
+typedef struct __GLXDRIdrawableRec __GLXDRIdrawable;
+#endif
 
 extern __GLXDRIdrawable *
 driFetchDrawable(struct glx_context *gc, GLXDrawable glxDrawable);
diff --git a/src/mesa/main/mtypes.h b/src/mesa/main/mtypes.h
index 6de28f39112..ece789148d8 100644
--- a/src/mesa/main/mtypes.h
+++ b/src/mesa/main/mtypes.h
@@ -1021,6 +1021,9 @@ struct gl_texture_object
    /** GL_ARB_bindless_texture */
    struct util_dynarray SamplerHandles;
    struct util_dynarray ImageHandles;
+   
+   /* Hack for avoiding sync on v3d */
+   GLboolean SyncCondition;
 };
 
 
diff --git a/src/mesa/main/texparam.c b/src/mesa/main/texparam.c
index 3a3ea8aae94..eb1cb561012 100644
--- a/src/mesa/main/texparam.c
+++ b/src/mesa/main/texparam.c
@@ -275,6 +275,13 @@ set_tex_parameteri(struct gl_context *ctx,
    }
 
    switch (pname) {
+   case GL_SYNC_CONDITION:
+      if (!!texObj->SyncCondition == !!params[0])
+         return GL_FALSE;
+      texObj->SyncCondition = !!params[0];
+      return GL_TRUE;
+   case GL_SYNC_STATUS:
+      return GL_TRUE;
    case GL_TEXTURE_MIN_FILTER:
       if (!_mesa_target_allows_setting_sampler_parameters(texObj->Target))
          goto invalid_dsa;
diff --git a/src/mesa/state_tracker/st_cb_texture.c b/src/mesa/state_tracker/st_cb_texture.c
index 599d2099d6f..9fbab50c1d7 100644
--- a/src/mesa/state_tracker/st_cb_texture.c
+++ b/src/mesa/state_tracker/st_cb_texture.c
@@ -3620,6 +3620,12 @@ st_TexParameter(struct gl_context *ctx,
        */
       st_texture_release_all_sampler_views(st, stObj);
       break;
+   case GL_SYNC_CONDITION:
+      stObj->pt->sync_condition = texObj->SyncCondition;
+      break;
+   case GL_SYNC_STATUS:
+      stObj->pt->sync_status = 1;
+      break;
    default:
       ; /* nothing */
    }
diff --git a/src/util/00-mesa-defaults.conf b/src/util/00-mesa-defaults.conf
index 7e155d3e690..f0f3cadfb02 100644
--- a/src/util/00-mesa-defaults.conf
+++ b/src/util/00-mesa-defaults.conf
@@ -625,6 +625,7 @@ TODO: document the other workarounds.
         <application name="mutter" executable="mutter">
             <option name="adaptive_sync" value="false" />
             <option name="v3d_nonmsaa_texture_size_limit" value="true" />
+            <option name="v3d_maintain_ignorable_scanout" value="true" />
         </application>
         <application name="muffin" executable="muffin">
             <option name="adaptive_sync" value="false" />
@@ -676,6 +677,7 @@ TODO: document the other workarounds.
         </application>
         <application name="Xorg" executable="Xorg">
             <option name="v3d_nonmsaa_texture_size_limit" value="true" />
+            <option name="v3d_ignore_scanout_usages_with_modifiers" value="true" />
         </application>
 
         <application name="gfxbench" executable="testfw_app">
diff --git a/src/util/driconf.h b/src/util/driconf.h
index 24726e04bd3..9f2ab7581c6 100644
--- a/src/util/driconf.h
+++ b/src/util/driconf.h
@@ -480,6 +480,14 @@
    DRI_CONF_OPT_B(v3d_nonmsaa_texture_size_limit, def, \
                   "Report the non-MSAA-only texture size limit")
 
+#define DRI_CONF_V3D_IGNORE_SCANOUT_USAGES_WITH_MODIFIERS(def) \
+   DRI_CONF_OPT_B(v3d_ignore_scanout_usages_with_modifiers, def, \
+                  "Ignore SCANOUT usage on resource allocations with modifiers.")
+
+#define DRI_CONF_V3D_MAINTAIN_IGNORABLE_SCANOUT(def)   \
+   DRI_CONF_OPT_B(v3d_maintain_ignorable_scanout, def, \
+                  "Maintain SCANOUT usage on resource allocations when the environment allows ignoring SCANOUT usage.")
+
 /**
  * \brief virgl specific configuration options
  */