u_vbuf_mgr: cleanup original vs real vertex buffer arrays

It can now override both buffer offsets and strides in additions to resources.
Overriding buffer offsets was kinda hackish and could cause issues with
non-native vertex formats.

src/gallium/auxiliary/util/u_vbuf_mgr.c

@@ -63,8 +63,6 @@ struct u_vbuf_mgr_priv {
    void *saved_ve, *fallback_ve;
    boolean ve_binding_lock;
 
-   unsigned saved_buffer_offset[PIPE_MAX_ATTRIBS];
-
    boolean any_user_vbs;
    boolean incompatible_vb_layout;
 };
@@ -127,9 +125,11 @@ void u_vbuf_mgr_destroy(struct u_vbuf_mgr *mgrb)
    struct u_vbuf_mgr_priv *mgr = (struct u_vbuf_mgr_priv*)mgrb;
    unsigned i;
 
-   for (i = 0; i < mgr->b.nr_real_vertex_buffers; i++) {
+   for (i = 0; i < mgr->b.nr_vertex_buffers; i++) {
       pipe_resource_reference(&mgr->b.vertex_buffer[i].buffer, NULL);
-      pipe_resource_reference(&mgr->b.real_vertex_buffer[i], NULL);
+   }
+   for (i = 0; i < mgr->b.nr_real_vertex_buffers; i++) {
+      pipe_resource_reference(&mgr->b.real_vertex_buffer[i].buffer, NULL);
    }
 
    translate_cache_destroy(mgr->translate_cache);
@@ -266,9 +266,9 @@ u_vbuf_translate_begin(struct u_vbuf_mgr_priv *mgr,
    if (mgr->translate_vb_slot != ~0) {
       /* Setup the new vertex buffer. */
       pipe_resource_reference(
-            &mgr->b.real_vertex_buffer[mgr->translate_vb_slot], out_buffer);
-      mgr->b.vertex_buffer[mgr->translate_vb_slot].buffer_offset = out_offset;
-      mgr->b.vertex_buffer[mgr->translate_vb_slot].stride = key.output_stride;
+            &mgr->b.real_vertex_buffer[mgr->translate_vb_slot].buffer, out_buffer);
+      mgr->b.real_vertex_buffer[mgr->translate_vb_slot].buffer_offset = out_offset;
+      mgr->b.real_vertex_buffer[mgr->translate_vb_slot].stride = key.output_stride;
 
       /* Setup new vertex elements. */
       for (i = 0; i < mgr->ve->count; i++) {
@@ -312,7 +312,7 @@ static void u_vbuf_translate_end(struct u_vbuf_mgr_priv *mgr)
    mgr->fallback_ve = NULL;
 
    /* Delete the now-unused VBO. */
-   pipe_resource_reference(&mgr->b.real_vertex_buffer[mgr->translate_vb_slot],
+   pipe_resource_reference(&mgr->b.real_vertex_buffer[mgr->translate_vb_slot].buffer,
                            NULL);
    mgr->b.nr_real_vertex_buffers = mgr->b.nr_vertex_buffers;
 }
@@ -473,8 +473,13 @@ void u_vbuf_mgr_set_vertex_buffers(struct u_vbuf_mgr *mgrb,
       const struct pipe_vertex_buffer *vb = &bufs[i];
 
       pipe_resource_reference(&mgr->b.vertex_buffer[i].buffer, vb->buffer);
-      pipe_resource_reference(&mgr->b.real_vertex_buffer[i], NULL);
-      mgr->saved_buffer_offset[i] = vb->buffer_offset;
+      pipe_resource_reference(&mgr->b.real_vertex_buffer[i].buffer, NULL);
+
+      mgr->b.real_vertex_buffer[i].buffer_offset =
+      mgr->b.vertex_buffer[i].buffer_offset = vb->buffer_offset;
+
+      mgr->b.real_vertex_buffer[i].stride =
+      mgr->b.vertex_buffer[i].stride = vb->stride;
 
       if (!vb->buffer) {
          continue;
@@ -485,16 +490,15 @@ void u_vbuf_mgr_set_vertex_buffers(struct u_vbuf_mgr *mgrb,
          continue;
       }
 
-      pipe_resource_reference(&mgr->b.real_vertex_buffer[i], vb->buffer);
+      pipe_resource_reference(&mgr->b.real_vertex_buffer[i].buffer, vb->buffer);
    }
 
-   for (; i < mgr->b.nr_real_vertex_buffers; i++) {
+   for (i = count; i < mgr->b.nr_vertex_buffers; i++) {
       pipe_resource_reference(&mgr->b.vertex_buffer[i].buffer, NULL);
-      pipe_resource_reference(&mgr->b.real_vertex_buffer[i], NULL);
    }
-
-   memcpy(mgr->b.vertex_buffer, bufs,
-          sizeof(struct pipe_vertex_buffer) * count);
+   for (i = count; i < mgr->b.nr_real_vertex_buffers; i++) {
+      pipe_resource_reference(&mgr->b.real_vertex_buffer[i].buffer, NULL);
+   }
 
    mgr->b.nr_vertex_buffers = count;
    mgr->b.nr_real_vertex_buffers = count;
@@ -540,17 +544,17 @@ u_vbuf_upload_buffers(struct u_vbuf_mgr_priv *mgr,
 
          u_upload_data(mgr->b.uploader, first, size,
                        u_vbuf_resource(vb->buffer)->user_ptr + first,
-                       &vb->buffer_offset,
-                       &mgr->b.real_vertex_buffer[index],
+                       &mgr->b.real_vertex_buffer[index].buffer_offset,
+                       &mgr->b.real_vertex_buffer[index].buffer,
                        &flushed);
 
-         vb->buffer_offset -= first;
+         mgr->b.real_vertex_buffer[index].buffer_offset -= first;
 
          uploaded[index] = TRUE;
          if (flushed)
             retval |= U_VBUF_UPLOAD_FLUSHED;
       } else {
-         assert(mgr->b.real_vertex_buffer[index]);
+         assert(mgr->b.real_vertex_buffer[index].buffer);
       }
    }
 
@@ -634,13 +638,6 @@ u_vbuf_mgr_draw_begin(struct u_vbuf_mgr *mgrb,
 void u_vbuf_mgr_draw_end(struct u_vbuf_mgr *mgrb)
 {
    struct u_vbuf_mgr_priv *mgr = (struct u_vbuf_mgr_priv*)mgrb;
-   unsigned i;
-
-   /* buffer offsets were modified in u_vbuf_upload_buffers */
-   if (mgr->any_user_vbs) {
-      for (i = 0; i < mgr->b.nr_vertex_buffers; i++)
-         mgr->b.vertex_buffer[i].buffer_offset = mgr->saved_buffer_offset[i];
-   }
 
    if (mgr->fallback_ve) {
       u_vbuf_translate_end(mgr);

src/gallium/auxiliary/util/u_vbuf_mgr.h

@@ -64,7 +64,7 @@ struct u_vbuf_mgr {
    /* Contains only real vertex buffers.
     * Hardware drivers should use real_vertex_buffers[i]
     * instead of vertex_buffers[i].buffer. */
-   struct pipe_resource *real_vertex_buffer[PIPE_MAX_ATTRIBS];
+   struct pipe_vertex_buffer real_vertex_buffer[PIPE_MAX_ATTRIBS];
    int nr_real_vertex_buffers;
 
    /* Precomputed max_index for hardware vertex buffers. */

src/gallium/drivers/r300/r300_emit.c

@@ -823,8 +823,7 @@ void r300_emit_textures_state(struct r300_context *r300,
 void r300_emit_vertex_arrays(struct r300_context* r300, int offset,
                              boolean indexed, int instance_id)
 {
-    struct pipe_vertex_buffer *vbuf = r300->vbuf_mgr->vertex_buffer;
-    struct pipe_resource **valid_vbuf = r300->vbuf_mgr->real_vertex_buffer;
+    struct pipe_vertex_buffer *vbuf = r300->vbuf_mgr->real_vertex_buffer;
     struct pipe_vertex_element *velem = r300->velems->velem;
     struct r300_resource *buf;
     int i;
@@ -862,7 +861,7 @@ void r300_emit_vertex_arrays(struct r300_context* r300, int offset,
         }
 
         for (i = 0; i < vertex_array_count; i++) {
-            buf = r300_resource(valid_vbuf[velem[i].vertex_buffer_index]);
+            buf = r300_resource(vbuf[velem[i].vertex_buffer_index].buffer);
             OUT_CS_RELOC(buf);
         }
     } else {
@@ -914,7 +913,7 @@ void r300_emit_vertex_arrays(struct r300_context* r300, int offset,
         }
 
         for (i = 0; i < vertex_array_count; i++) {
-            buf = r300_resource(valid_vbuf[velem[i].vertex_buffer_index]);
+            buf = r300_resource(vbuf[velem[i].vertex_buffer_index].buffer);
             OUT_CS_RELOC(buf);
         }
     }
@@ -1222,15 +1221,17 @@ validate:
                                 r300_resource(r300->vbo)->domain, 0);
     /* ...vertex buffers for HWTCL path... */
     if (do_validate_vertex_buffers && r300->vertex_arrays_dirty) {
-        struct pipe_resource **buf = r300->vbuf_mgr->real_vertex_buffer;
-        struct pipe_resource **last = r300->vbuf_mgr->real_vertex_buffer +
+        struct pipe_vertex_buffer *vbuf = r300->vbuf_mgr->real_vertex_buffer;
+        struct pipe_vertex_buffer *last = r300->vbuf_mgr->real_vertex_buffer +
                                       r300->vbuf_mgr->nr_real_vertex_buffers;
-        for (; buf != last; buf++) {
-            if (!*buf)
+        struct pipe_resource *buf;
+        for (; vbuf != last; vbuf++) {
+            buf = vbuf->buffer;
+            if (!buf)
                 continue;
 
-            r300->rws->cs_add_reloc(r300->cs, r300_resource(*buf)->cs_buf,
-                                    r300_resource(*buf)->domain, 0);
+            r300->rws->cs_add_reloc(r300->cs, r300_resource(buf)->cs_buf,
+                                    r300_resource(buf)->domain, 0);
         }
     }
     /* ...and index buffer for HWTCL path. */

src/gallium/drivers/r300/r300_render.c

@@ -146,7 +146,7 @@ static void r300_emit_draw_init(struct r300_context *r300, unsigned mode,
 static void r300_split_index_bias(struct r300_context *r300, int index_bias,
                                   int *buffer_offset, int *index_offset)
 {
-    struct pipe_vertex_buffer *vb, *vbufs = r300->vbuf_mgr->vertex_buffer;
+    struct pipe_vertex_buffer *vb, *vbufs = r300->vbuf_mgr->real_vertex_buffer;
     struct pipe_vertex_element *velem = r300->velems->velem;
     unsigned i, size;
     int max_neg_bias;
@@ -338,7 +338,7 @@ static boolean immd_is_good_idea(struct r300_context *r300,
         vbi = velem->vertex_buffer_index;
 
         if (!checked[vbi]) {
-            buf = r300->vbuf_mgr->real_vertex_buffer[vbi];
+            buf = r300->vbuf_mgr->real_vertex_buffer[vbi].buffer;
 
             if ((r300_resource(buf)->domain != RADEON_DOMAIN_GTT)) {
                 return FALSE;
@@ -389,13 +389,13 @@ static void r300_draw_arrays_immediate(struct r300_context *r300,
         velem = &r300->velems->velem[i];
         size[i] = r300->velems->format_size[i] / 4;
         vbi = velem->vertex_buffer_index;
-        vbuf = &r300->vbuf_mgr->vertex_buffer[vbi];
+        vbuf = &r300->vbuf_mgr->real_vertex_buffer[vbi];
         stride[i] = vbuf->stride / 4;
 
         /* Map the buffer. */
         if (!map[vbi]) {
             map[vbi] = (uint32_t*)r300->rws->buffer_map(
-                r300_resource(r300->vbuf_mgr->real_vertex_buffer[vbi])->buf,
+                r300_resource(vbuf->buffer)->buf,
                 r300->cs, PIPE_TRANSFER_READ | PIPE_TRANSFER_UNSYNCHRONIZED);
             map[vbi] += (vbuf->buffer_offset / 4) + stride[i] * info->start;
         }
@@ -423,7 +423,7 @@ static void r300_draw_arrays_immediate(struct r300_context *r300,
         vbi = r300->velems->velem[i].vertex_buffer_index;
 
         if (map[vbi]) {
-            r300->rws->buffer_unmap(r300_resource(r300->vbuf_mgr->real_vertex_buffer[vbi])->buf);
+            r300->rws->buffer_unmap(r300_resource(r300->vbuf_mgr->real_vertex_buffer[vbi].buffer)->buf);
             map[vbi] = NULL;
         }
     }

src/gallium/drivers/r600/r600_state_common.c

@@ -506,13 +506,13 @@ static void r600_vertex_buffer_update(struct r600_pipe_context *rctx)
 			/* one resource per vertex elements */
 			unsigned vbuffer_index;
 			vbuffer_index = rctx->vertex_elements->elements[i].vertex_buffer_index;
-			vertex_buffer = &rctx->vbuf_mgr->vertex_buffer[vbuffer_index];
-			rbuffer = (struct r600_resource*)rctx->vbuf_mgr->real_vertex_buffer[vbuffer_index];
+			vertex_buffer = &rctx->vbuf_mgr->real_vertex_buffer[vbuffer_index];
+			rbuffer = (struct r600_resource*)vertex_buffer->buffer;
 			offset = rctx->vertex_elements->vbuffer_offset[i];
 		} else {
 			/* bind vertex buffer once */
-			vertex_buffer = &rctx->vbuf_mgr->vertex_buffer[i];
-			rbuffer = (struct r600_resource*)rctx->vbuf_mgr->real_vertex_buffer[i];
+			vertex_buffer = &rctx->vbuf_mgr->real_vertex_buffer[i];
+			rbuffer = (struct r600_resource*)vertex_buffer->buffer;
 			offset = 0;
 		}
 		if (vertex_buffer == NULL || rbuffer == NULL)