mesa/st: support lowering multi-planar YUV

Support multi-planar YUV for external EGLImage's (currently just in the
dma-buf import path) by lowering to multiple texture fetch's for each
plane and CSC in shader.

There was some discussion of alternative approaches for tracking the
additional UV or U/V planes:

  https://lists.freedesktop.org/archives/mesa-dev/2016-September/127832.html

They all seemed worse than pipe_resource::next

Signed-off-by: Rob Clark <[email protected]>

18 files changed, 359 insertions, 27 deletions

diff --git a/src/gallium/auxiliary/util/u_inlines.h b/src/gallium/auxiliary/util/u_inlines.h
index c2a0b08d151..b7b83135838 100644
--- a/src/gallium/auxiliary/util/u_inlines.h
+++ b/src/gallium/auxiliary/util/u_inlines.h
@@ -136,8 +136,10 @@ pipe_resource_reference(struct pipe_resource **ptr, struct pipe_resource *tex)
    struct pipe_resource *old_tex = *ptr;
 
    if (pipe_reference_described(&(*ptr)->reference, &tex->reference, 
-                                (debug_reference_descriptor)debug_describe_resource))
+                                (debug_reference_descriptor)debug_describe_resource)) {
+      pipe_resource_reference(&old_tex->next, NULL);
       old_tex->screen->resource_destroy(old_tex->screen, old_tex);
+   }
    *ptr = tex;
 }
 
diff --git a/src/gallium/include/pipe/p_state.h b/src/gallium/include/pipe/p_state.h
index ebd03379e94..415ea85a2ad 100644
--- a/src/gallium/include/pipe/p_state.h
+++ b/src/gallium/include/pipe/p_state.h
@@ -498,6 +498,12 @@ struct pipe_resource
 
    unsigned bind;            /**< bitmask of PIPE_BIND_x */
    unsigned flags;           /**< bitmask of PIPE_RESOURCE_FLAG_x */
+
+   /**
+    * For planar images, ie. YUV EGLImage external, etc, pointer to the
+    * next plane.
+    */
+   struct pipe_resource *next;
 };
 
 
diff --git a/src/gallium/include/state_tracker/st_api.h b/src/gallium/include/state_tracker/st_api.h
index 21d5177aa2f..06abfc56507 100644
--- a/src/gallium/include/state_tracker/st_api.h
+++ b/src/gallium/include/state_tracker/st_api.h
@@ -200,6 +200,9 @@ struct st_egl_image
    /* this is owned by the caller */
    struct pipe_resource *texture;
 
+   /* format only differs from texture->format for multi-planar (YUV): */
+   enum pipe_format format;
+
    unsigned level;
    unsigned layer;
 };
diff --git a/src/gallium/state_trackers/dri/dri2.c b/src/gallium/state_trackers/dri/dri2.c
index 43a5df11d0a..a22e7ee6ace 100644
--- a/src/gallium/state_trackers/dri/dri2.c
+++ b/src/gallium/state_trackers/dri/dri2.c
@@ -83,6 +83,21 @@ static int convert_fourcc(int format, int *dri_components_p)
       format = __DRI_IMAGE_FORMAT_GR88;
       dri_components = __DRI_IMAGE_COMPONENTS_RG;
       break;
+   /*
+    * For multi-planar YUV formats, we return the format of the first
+    * plane only.  Since there is only one caller which supports multi-
+    * planar YUV it gets to figure out the remaining planes on it's
+    * own.
+    */
+   case __DRI_IMAGE_FOURCC_YUV420:
+   case __DRI_IMAGE_FOURCC_YVU420:
+      format = __DRI_IMAGE_FORMAT_R8;
+      dri_components = __DRI_IMAGE_COMPONENTS_Y_U_V;
+      break;
+   case __DRI_IMAGE_FOURCC_NV12:
+      format = __DRI_IMAGE_FORMAT_R8;
+      dri_components = __DRI_IMAGE_COMPONENTS_Y_UV;
+      break;
    default:
       return -1;
    }
@@ -90,6 +105,11 @@ static int convert_fourcc(int format, int *dri_components_p)
    return format;
 }
 
+/* NOTE this probably isn't going to do the right thing for YUV images
+ * (but I think the same can be said for intel_query_image()).  I think
+ * only needed for exporting dmabuf's, so I think I won't loose much
+ * sleep over it.
+ */
 static int convert_to_fourcc(int format)
 {
    switch(format) {
@@ -762,14 +782,16 @@ dri2_lookup_egl_image(struct dri_screen *screen, void *handle)
 static __DRIimage *
 dri2_create_image_from_winsys(__DRIscreen *_screen,
                               int width, int height, int format,
-                              struct winsys_handle *whandle,
+                              int num_handles, struct winsys_handle *whandle,
                               void *loaderPrivate)
 {
    struct dri_screen *screen = dri_screen(_screen);
+   struct pipe_screen *pscreen = screen->base.screen;
    __DRIimage *img;
    struct pipe_resource templ;
    unsigned tex_usage;
    enum pipe_format pf;
+   int i;
 
    tex_usage = PIPE_BIND_RENDER_TARGET | PIPE_BIND_SAMPLER_VIEW;
 
@@ -783,19 +805,47 @@ dri2_create_image_from_winsys(__DRIscreen *_screen,
 
    memset(&templ, 0, sizeof(templ));
    templ.bind = tex_usage;
-   templ.format = pf;
    templ.target = screen->target;
    templ.last_level = 0;
-   templ.width0 = width;
-   templ.height0 = height;
    templ.depth0 = 1;
    templ.array_size = 1;
 
-   img->texture = screen->base.screen->resource_from_handle(screen->base.screen,
-         &templ, whandle, PIPE_HANDLE_USAGE_READ_WRITE);
-   if (!img->texture) {
-      FREE(img);
-      return NULL;
+   for (i = num_handles - 1; i >= 0; i--) {
+      struct pipe_resource *tex;
+
+      /* TODO: something a lot less ugly */
+      switch (i) {
+      case 0:
+         templ.width0 = width;
+         templ.height0 = height;
+         templ.format = pf;
+         break;
+      case 1:
+         templ.width0 = width / 2;
+         templ.height0 = height / 2;
+         templ.format = (num_handles == 2) ?
+               PIPE_FORMAT_RG88_UNORM :   /* NV12, etc */
+               PIPE_FORMAT_R8_UNORM;      /* I420, etc */
+         break;
+      case 2:
+         templ.width0 = width / 2;
+         templ.height0 = height / 2;
+         templ.format = PIPE_FORMAT_R8_UNORM;
+         break;
+      default:
+         unreachable("too many planes!");
+      }
+
+      tex = pscreen->resource_from_handle(pscreen,
+            &templ, &whandle[i], PIPE_HANDLE_USAGE_READ_WRITE);
+      if (!tex) {
+         pipe_resource_reference(&img->texture, NULL);
+         FREE(img);
+         return NULL;
+      }
+
+      tex->next = img->texture;
+      img->texture = tex;
    }
 
    img->level = 0;
@@ -826,7 +876,7 @@ dri2_create_image_from_name(__DRIscreen *_screen,
    whandle.stride = pitch * util_format_get_blocksize(pf);
 
    return dri2_create_image_from_winsys(_screen, width, height, format,
-                                        &whandle, loaderPrivate);
+                                        1, &whandle, loaderPrivate);
 }
 
 static __DRIimage *
@@ -836,12 +886,26 @@ dri2_create_image_from_fd(__DRIscreen *_screen,
                           int *offsets, unsigned *error,
                           int *dri_components, void *loaderPrivate)
 {
-   struct winsys_handle whandle;
+   struct winsys_handle whandles[3];
    int format;
    __DRIimage *img = NULL;
    unsigned err = __DRI_IMAGE_ERROR_SUCCESS;
+   int expected_num_fds, i;
 
-   if (num_fds != 1) {
+   switch (fourcc) {
+   case __DRI_IMAGE_FOURCC_YUV420:
+   case __DRI_IMAGE_FOURCC_YVU420:
+      expected_num_fds = 3;
+      break;
+   case __DRI_IMAGE_FOURCC_NV12:
+      expected_num_fds = 2;
+      break;
+   default:
+      expected_num_fds = 1;
+      break;
+   }
+
+   if (num_fds != expected_num_fds) {
       err = __DRI_IMAGE_ERROR_BAD_MATCH;
       goto exit;
    }
@@ -852,19 +916,30 @@ dri2_create_image_from_fd(__DRIscreen *_screen,
       goto exit;
    }
 
-   if (fds[0] < 0) {
-      err = __DRI_IMAGE_ERROR_BAD_ALLOC;
-      goto exit;
+   memset(whandles, 0, sizeof(whandles));
+
+   for (i = 0; i < num_fds; i++) {
+      if (fds[i] < 0) {
+         err = __DRI_IMAGE_ERROR_BAD_ALLOC;
+         goto exit;
+      }
+
+      whandles[i].type = DRM_API_HANDLE_TYPE_FD;
+      whandles[i].handle = (unsigned)fds[i];
+      whandles[i].stride = (unsigned)strides[i];
+      whandles[i].offset = (unsigned)offsets[i];
    }
 
-   memset(&whandle, 0, sizeof(whandle));
-   whandle.type = DRM_API_HANDLE_TYPE_FD;
-   whandle.handle = (unsigned)fds[0];
-   whandle.stride = (unsigned)strides[0];
-   whandle.offset = (unsigned)offsets[0];
+   if (fourcc == __DRI_IMAGE_FOURCC_YVU420) {
+      /* convert to YUV420 by swapping 2nd and 3rd planes: */
+      struct winsys_handle tmp = whandles[1];
+      whandles[1] = whandles[2];
+      whandles[2] = tmp;
+      fourcc = __DRI_IMAGE_FOURCC_YUV420;
+   }
 
    img = dri2_create_image_from_winsys(_screen, width, height, format,
-                                       &whandle, loaderPrivate);
+                                       num_fds, whandles, loaderPrivate);
    if(img == NULL)
       err = __DRI_IMAGE_ERROR_BAD_ALLOC;
 
@@ -1067,7 +1142,7 @@ dri2_from_names(__DRIscreen *screen, int width, int height, int format,
    whandle.offset = offsets[0];
 
    img = dri2_create_image_from_winsys(screen, width, height, format,
-                                       &whandle, loaderPrivate);
+                                       1, &whandle, loaderPrivate);
    if (img == NULL)
       return NULL;
 
diff --git a/src/gallium/state_trackers/dri/dri_screen.c b/src/gallium/state_trackers/dri/dri_screen.c
index 79bcb5abce7..aa0ad09ce58 100644
--- a/src/gallium/state_trackers/dri/dri_screen.c
+++ b/src/gallium/state_trackers/dri/dri_screen.c
@@ -334,6 +334,17 @@ dri_get_egl_image(struct st_manager *smapi,
 
    stimg->texture = NULL;
    pipe_resource_reference(&stimg->texture, img->texture);
+   switch (img->dri_components) {
+   case __DRI_IMAGE_COMPONENTS_Y_U_V:
+      stimg->format = PIPE_FORMAT_IYUV;
+      break;
+   case __DRI_IMAGE_COMPONENTS_Y_UV:
+      stimg->format = PIPE_FORMAT_NV12;
+      break;
+   default:
+      stimg->format = img->texture->format;
+      break;
+   }
    stimg->level = img->level;
    stimg->layer = img->layer;
 
diff --git a/src/mesa/main/mtypes.h b/src/mesa/main/mtypes.h
index 85aeb1e0df3..ee811ff5ddf 100644
--- a/src/mesa/main/mtypes.h
+++ b/src/mesa/main/mtypes.h
@@ -45,6 +45,7 @@
 #include "compiler/shader_enums.h"
 #include "main/formats.h"       /* MESA_FORMAT_COUNT */
 #include "compiler/glsl/list.h"
+#include "util/bitscan.h"
 
 
 #ifdef __cplusplus
@@ -1929,6 +1930,7 @@ struct gl_program
    GLbitfield TexturesUsed[MAX_COMBINED_TEXTURE_IMAGE_UNITS];  /**< TEXTURE_x_BIT bitmask */
    GLbitfield SamplersUsed;   /**< Bitfield of which samplers are used */
    GLbitfield ShadowSamplers; /**< Texture units used for shadow sampling. */
+   GLbitfield ExternalSamplersUsed; /**< Texture units used for samplerExternalOES */
 
    GLboolean UsesGather; /**< Does this program use gather4 at all? */
 
@@ -2460,6 +2462,20 @@ struct gl_linked_shader
    struct gl_shader_info info;
 };
 
+static inline GLbitfield gl_external_samplers(struct gl_linked_shader *shader)
+{
+   GLbitfield external_samplers = 0;
+   GLbitfield mask = shader->active_samplers;
+
+   while (mask) {
+      int idx = u_bit_scan(&mask);
+      if (shader->SamplerTargets[idx] == TEXTURE_EXTERNAL_INDEX)
+         external_samplers |= (1 << idx);
+   }
+
+   return external_samplers;
+}
+
 /**
  * A GLSL shader object.
  */
diff --git a/src/mesa/program/ir_to_mesa.cpp b/src/mesa/program/ir_to_mesa.cpp
index 530c00fa785..debc18d95c9 100644
--- a/src/mesa/program/ir_to_mesa.cpp
+++ b/src/mesa/program/ir_to_mesa.cpp
@@ -2921,6 +2921,7 @@ get_mesa_program(struct gl_context *ctx,
 
    prog->SamplersUsed = shader->active_samplers;
    prog->ShadowSamplers = shader->shadow_samplers;
+   prog->ExternalSamplersUsed = gl_external_samplers(shader);
    _mesa_update_shader_textures_used(shader_program, prog);
 
    /* Set the gl_FragDepth layout. */
diff --git a/src/mesa/state_tracker/st_atom_sampler.c b/src/mesa/state_tracker/st_atom_sampler.c
index 6b36ac74aad..065df6dd006 100644
--- a/src/mesa/state_tracker/st_atom_sampler.c
+++ b/src/mesa/state_tracker/st_atom_sampler.c
@@ -243,13 +243,13 @@ update_shader_samplers(struct st_context *st,
                        struct pipe_sampler_state *samplers,
                        unsigned *num_samplers)
 {
+   GLbitfield samplers_used = prog->SamplersUsed;
+   GLbitfield free_slots = ~prog->SamplersUsed;
+   GLbitfield external_samplers_used = prog->ExternalSamplersUsed;
    GLuint unit;
-   GLbitfield samplers_used;
    const GLuint old_max = *num_samplers;
    const struct pipe_sampler_state *states[PIPE_MAX_SAMPLERS];
 
-   samplers_used = prog->SamplersUsed;
-
    if (*num_samplers == 0 && samplers_used == 0x0)
       return;
 
@@ -275,6 +275,41 @@ update_shader_samplers(struct st_context *st,
       }
    }
 
+   /* For any external samplers with multiplaner YUV, stuff the additional
+    * sampler states we need at the end.
+    *
+    * Just re-use the existing sampler-state from the primary slot.
+    */
+   while (unlikely(external_samplers_used)) {
+      GLuint unit = u_bit_scan(&external_samplers_used);
+      GLuint extra = 0;
+      struct st_texture_object *stObj =
+            st_get_texture_object(st->ctx, prog, unit);
+      struct pipe_sampler_state *sampler = samplers + unit;
+
+      if (!stObj)
+         continue;
+
+      switch (st_get_view_format(stObj)) {
+      case PIPE_FORMAT_NV12:
+         /* we need one additional sampler: */
+         extra = u_bit_scan(&free_slots);
+         states[extra] = sampler;
+         break;
+      case PIPE_FORMAT_IYUV:
+         /* we need two additional samplers: */
+         extra = u_bit_scan(&free_slots);
+         states[extra] = sampler;
+         extra = u_bit_scan(&free_slots);
+         states[extra] = sampler;
+         break;
+      default:
+         break;
+      }
+
+      *num_samplers = MAX2(*num_samplers, extra + 1);
+   }
+
    cso_set_samplers(st->cso_context, shader_stage, *num_samplers, states);
 }
 
diff --git a/src/mesa/state_tracker/st_atom_shader.c b/src/mesa/state_tracker/st_atom_shader.c
index 7a23469cede..2f700a2fc55 100644
--- a/src/mesa/state_tracker/st_atom_shader.c
+++ b/src/mesa/state_tracker/st_atom_shader.c
@@ -51,6 +51,7 @@
 #include "st_context.h"
 #include "st_atom.h"
 #include "st_program.h"
+#include "st_texture.h"
 
 
 /** Compress the fog function enums into a 2-bit value */
@@ -142,6 +143,8 @@ update_fp( struct st_context *st )
       }
    }
 
+   key.external = st_get_external_sampler_key(st, &stfp->Base.Base);
+
    st->fp_variant = st_get_fp_variant(st, stfp, &key);
 
    st_reference_fragprog(st, &st->fp, stfp);
diff --git a/src/mesa/state_tracker/st_atom_texture.c b/src/mesa/state_tracker/st_atom_texture.c
index 19df6622ef1..efc8c90ffae 100644
--- a/src/mesa/state_tracker/st_atom_texture.c
+++ b/src/mesa/state_tracker/st_atom_texture.c
@@ -408,6 +408,15 @@ update_single_texture(struct st_context *st,
       }
    }
 
+   switch (view_format) {
+   case PIPE_FORMAT_NV12:
+   case PIPE_FORMAT_IYUV:
+      view_format = PIPE_FORMAT_R8_UNORM;
+      break;
+   default:
+      break;
+   }
+
    *sampler_view =
       st_get_texture_sampler_view_from_stobj(st, stObj, view_format,
                                              glsl_version);
@@ -426,6 +435,8 @@ update_textures(struct st_context *st,
 {
    const GLuint old_max = *num_textures;
    GLbitfield samplers_used = prog->SamplersUsed;
+   GLbitfield free_slots = ~prog->SamplersUsed;
+   GLbitfield external_samplers_used = prog->ExternalSamplersUsed;
    GLuint unit;
    struct gl_shader_program *shader =
       st->ctx->_Shader->CurrentProgram[mesa_shader];
@@ -460,6 +471,53 @@ update_textures(struct st_context *st,
       pipe_sampler_view_reference(&(sampler_views[unit]), sampler_view);
    }
 
+   /* For any external samplers with multiplaner YUV, stuff the additional
+    * sampler views we need at the end.
+    *
+    * Trying to cache the sampler view in the stObj looks painful, so just
+    * re-create the sampler view for the extra planes each time.  Main use
+    * case is video playback (ie. fps games wouldn't be using this) so I
+    * guess no point to try to optimize this feature.
+    */
+   while (unlikely(external_samplers_used)) {
+      GLuint unit = u_bit_scan(&external_samplers_used);
+      GLuint extra = 0;
+      struct st_texture_object *stObj =
+            st_get_texture_object(st->ctx, prog, unit);
+      struct pipe_sampler_view tmpl;
+
+      if (!stObj)
+         continue;
+
+      /* use original view as template: */
+      tmpl = *sampler_views[unit];
+
+      switch (st_get_view_format(stObj)) {
+      case PIPE_FORMAT_NV12:
+         /* we need one additional R8G8 view: */
+         tmpl.format = PIPE_FORMAT_RG88_UNORM;
+         tmpl.swizzle_g = PIPE_SWIZZLE_Y;   /* tmpl from Y plane is R8 */
+         extra = u_bit_scan(&free_slots);
+         sampler_views[extra] =
+               st->pipe->create_sampler_view(st->pipe, stObj->pt->next, &tmpl);
+         break;
+      case PIPE_FORMAT_IYUV:
+         /* we need two additional R8 views: */
+         tmpl.format = PIPE_FORMAT_R8_UNORM;
+         extra = u_bit_scan(&free_slots);
+         sampler_views[extra] =
+               st->pipe->create_sampler_view(st->pipe, stObj->pt->next, &tmpl);
+         extra = u_bit_scan(&free_slots);
+         sampler_views[extra] =
+               st->pipe->create_sampler_view(st->pipe, stObj->pt->next->next, &tmpl);
+         break;
+      default:
+         break;
+      }
+
+      *num_textures = MAX2(*num_textures, extra + 1);
+   }
+
    cso_set_sampler_views(st->cso_context,
                          shader_stage,
                          *num_textures,
diff --git a/src/mesa/state_tracker/st_cb_eglimage.c b/src/mesa/state_tracker/st_cb_eglimage.c
index 1782d154dd6..7bea5653e46 100644
--- a/src/mesa/state_tracker/st_cb_eglimage.c
+++ b/src/mesa/state_tracker/st_cb_eglimage.c
@@ -119,6 +119,24 @@ st_bind_surface(struct gl_context *ctx, GLenum target,
 
    texFormat = st_pipe_format_to_mesa_format(ps->format);
 
+   /* TODO RequiredTextureImageUnits should probably be reset back
+    * to 1 somewhere if different texture is bound??
+    */
+   if (texFormat == MESA_FORMAT_NONE) {
+      switch (ps->format) {
+      case PIPE_FORMAT_NV12:
+         texFormat = MESA_FORMAT_R_UNORM8;
+         texObj->RequiredTextureImageUnits = 2;
+         break;
+      case PIPE_FORMAT_IYUV:
+         texFormat = MESA_FORMAT_R_UNORM8;
+         texObj->RequiredTextureImageUnits = 3;
+         break;
+      default:
+         unreachable("bad YUV format!");
+      }
+   }
+
    _mesa_init_teximage_fields(ctx, texImage,
                               ps->width, ps->height, 1, 0, internalFormat,
                               texFormat);
diff --git a/src/mesa/state_tracker/st_context.c b/src/mesa/state_tracker/st_context.c
index 81b3387626c..b9dc0c67da3 100644
--- a/src/mesa/state_tracker/st_context.c
+++ b/src/mesa/state_tracker/st_context.c
@@ -255,11 +255,16 @@ void st_invalidate_state(struct gl_context * ctx, GLbitfield new_state)
       st->active_states = st_get_active_states(ctx);
    }
 
-   if (new_state & _NEW_TEXTURE)
+   if (new_state & _NEW_TEXTURE) {
       st->dirty |= st->active_states &
                    (ST_NEW_SAMPLER_VIEWS |
                     ST_NEW_SAMPLERS |
                     ST_NEW_IMAGE_UNITS);
+      if (ctx->FragmentProgram._Current &&
+          ctx->FragmentProgram._Current->Base.ExternalSamplersUsed) {
+         st->dirty |= ST_NEW_FS_STATE;
+      }
+   }
 
    if (new_state & _NEW_PROGRAM_CONSTANTS)
       st->dirty |= st->active_states & ST_NEW_CONSTANTS;
diff --git a/src/mesa/state_tracker/st_glsl_to_nir.cpp b/src/mesa/state_tracker/st_glsl_to_nir.cpp
index 5b5fad44f18..f43460ba67a 100644
--- a/src/mesa/state_tracker/st_glsl_to_nir.cpp
+++ b/src/mesa/state_tracker/st_glsl_to_nir.cpp
@@ -423,6 +423,7 @@ st_nir_get_mesa_program(struct gl_context *ctx,
 
    prog->SamplersUsed = shader->active_samplers;
    prog->ShadowSamplers = shader->shadow_samplers;
+   prog->ExternalSamplersUsed = gl_external_samplers(shader);
    _mesa_update_shader_textures_used(shader_program, prog);
 
    _mesa_reference_program(ctx, &shader->Program, prog);
diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
index 507a7826e38..81ea233a8ec 100644
--- a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
+++ b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
@@ -4347,6 +4347,10 @@ count_resources(glsl_to_tgsi_visitor *v, gl_program *prog)
             }
          }
       }
+
+      if (inst->tex_target == TEXTURE_EXTERNAL_INDEX)
+         prog->ExternalSamplersUsed |= 1 << inst->sampler.index;
+
       if (inst->buffer.file != PROGRAM_UNDEFINED && (
                 is_resource_instruction(inst->op) ||
                 inst->op == TGSI_OPCODE_STORE)) {
diff --git a/src/mesa/state_tracker/st_manager.c b/src/mesa/state_tracker/st_manager.c
index e2da054d2a5..fece5d5dffd 100644
--- a/src/mesa/state_tracker/st_manager.c
+++ b/src/mesa/state_tracker/st_manager.c
@@ -845,6 +845,7 @@ st_manager_get_egl_image_surface(struct st_context *st, void *eglimg)
       return NULL;
 
    u_surface_default_template(&surf_tmpl, stimg.texture);
+   surf_tmpl.format = stimg.format;
    surf_tmpl.u.tex.level = stimg.level;
    surf_tmpl.u.tex.first_layer = stimg.layer;
    surf_tmpl.u.tex.last_layer = stimg.layer;
diff --git a/src/mesa/state_tracker/st_program.c b/src/mesa/state_tracker/st_program.c
index 2a4edfa7b7a..91887dc27db 100644
--- a/src/mesa/state_tracker/st_program.c
+++ b/src/mesa/state_tracker/st_program.c
@@ -53,6 +53,7 @@
 #include "st_cb_bitmap.h"
 #include "st_cb_drawpixels.h"
 #include "st_context.h"
+#include "st_tgsi_lower_yuv.h"
 #include "st_program.h"
 #include "st_mesa_to_tgsi.h"
 #include "st_atifs_to_tgsi.h"
@@ -1024,8 +1025,23 @@ st_create_fp_variant(struct st_context *st,
          NIR_PASS_V(tgsi.ir.nir, nir_lower_drawpixels, &options);
       }
 
+      if (unlikely(key->external.lower_nv12 || key->external.lower_iyuv)) {
+         nir_lower_tex_options options = {0};
+         options.lower_y_uv_external = key->external.lower_nv12;
+         options.lower_y_u_v_external = key->external.lower_iyuv;
+         NIR_PASS_V(tgsi.ir.nir, nir_lower_tex, &options);
+      }
+
       st_finalize_nir(st, &stfp->Base.Base, tgsi.ir.nir);
 
+      if (unlikely(key->external.lower_nv12 || key->external.lower_iyuv)) {
+         /* This pass needs to happen *after* nir_lower_sampler */
+         NIR_PASS_V(tgsi.ir.nir, st_nir_lower_tex_src_plane,
+                    ~stfp->Base.Base.SamplersUsed,
+                    key->external.lower_nv12,
+                    key->external.lower_iyuv);
+      }
+
       variant->driver_shader = pipe->create_fs_state(pipe, &tgsi);
       variant->key = *key;
 
@@ -1122,6 +1138,25 @@ st_create_fp_variant(struct st_context *st,
          fprintf(stderr, "mesa: cannot create a shader for glDrawPixels\n");
    }
 
+   if (unlikely(key->external.lower_nv12 || key->external.lower_iyuv)) {
+      const struct tgsi_token *tokens;
+
+      /* samplers inserted would conflict, but this should be unpossible: */
+      assert(!(key->bitmap || key->drawpixels));
+
+      tokens = st_tgsi_lower_yuv(tgsi.tokens,
+                                 ~stfp->Base.Base.SamplersUsed,
+                                 key->external.lower_nv12,
+                                 key->external.lower_iyuv);
+      if (tokens) {
+         if (tgsi.tokens != stfp->tgsi.tokens)
+            tgsi_free_tokens(tgsi.tokens);
+         tgsi.tokens = tokens;
+      } else {
+         fprintf(stderr, "mesa: cannot create a shader for samplerExternalOES\n");
+      }
+   }
+
    if (ST_DEBUG & DEBUG_TGSI) {
       tgsi_dump(tgsi.tokens, 0);
       debug_printf("\n");
diff --git a/src/mesa/state_tracker/st_program.h b/src/mesa/state_tracker/st_program.h
index 8e11bf02fa9..ea55d476ff0 100644
--- a/src/mesa/state_tracker/st_program.h
+++ b/src/mesa/state_tracker/st_program.h
@@ -39,6 +39,7 @@
 #include "program/program.h"
 #include "pipe/p_state.h"
 #include "st_context.h"
+#include "st_texture.h"
 #include "st_glsl_to_tgsi.h"
 
 
@@ -48,6 +49,40 @@ extern "C" {
 
 #define ST_DOUBLE_ATTRIB_PLACEHOLDER 0xffffffff
 
+struct st_external_sampler_key
+{
+   GLuint lower_nv12;             /**< bitmask of 2 plane YUV samplers */
+   GLuint lower_iyuv;             /**< bitmask of 3 plane YUV samplers */
+};
+
+static inline struct st_external_sampler_key
+st_get_external_sampler_key(struct st_context *st, struct gl_program *prog)
+{
+   unsigned mask = prog->ExternalSamplersUsed;
+   struct st_external_sampler_key key;
+
+   memset(&key, 0, sizeof(key));
+
+   while (unlikely(mask)) {
+      unsigned unit = u_bit_scan(&mask);
+      struct st_texture_object *stObj =
+            st_get_texture_object(st->ctx, prog, unit);
+
+      switch (st_get_view_format(stObj)) {
+      case PIPE_FORMAT_NV12:
+         key.lower_nv12 |= (1 << unit);
+         break;
+      case PIPE_FORMAT_IYUV:
+         key.lower_iyuv |= (1 << unit);
+         break;
+      default:
+         break;
+      }
+   }
+
+   return key;
+}
+
 /** Fragment program variant key */
 struct st_fp_variant_key
 {
@@ -72,6 +107,8 @@ struct st_fp_variant_key
 
    /** needed for ATI_fragment_shader */
    char texture_targets[MAX_NUM_FRAGMENT_REGISTERS_ATI];
+
+   struct st_external_sampler_key external;
 };
 
 
diff --git a/src/mesa/state_tracker/st_texture.h b/src/mesa/state_tracker/st_texture.h
index b2ddc14ddfc..d2c4f3f880c 100644
--- a/src/mesa/state_tracker/st_texture.h
+++ b/src/mesa/state_tracker/st_texture.h
@@ -170,6 +170,27 @@ st_create_texture_sampler_view(struct pipe_context *pipe,
                                                 texture->format);
 }
 
+static inline struct st_texture_object *
+st_get_texture_object(struct gl_context *ctx,
+                      const struct gl_program *prog,
+                      unsigned unit)
+{
+   const GLuint texUnit = prog->SamplerUnits[unit];
+   struct gl_texture_object *texObj = ctx->Texture.Unit[texUnit]._Current;
+
+   if (!texObj)
+      return NULL;
+
+   return st_texture_object(texObj);
+}
+
+static inline enum pipe_format
+st_get_view_format(struct st_texture_object *stObj)
+{
+   if (!stObj)
+      return PIPE_FORMAT_NONE;
+   return stObj->surface_based ? stObj->surface_format : stObj->pt->format;
+}
 
 
 extern struct pipe_resource *