+ /*
+ * As noted in ilo_translate_format(), we treat some 3-component
+ * formats as 4-component formats to work around hardware
+ * limitations. Imagine the case where the vertex buffer holds a
+ * single PIPE_FORMAT_R16G16B16_FLOAT vertex, and buf->bo_size is 6.
+ * The hardware would not be able to fetch it because the vertex
+ * buffer is expected to hold a PIPE_FORMAT_R16G16B16A16_FLOAT vertex
+ * and that takes at least 8 bytes.
+ *
+ * For the workaround to work, we query the physical size, which is
+ * page aligned, to calculate end_offset so that the last vertex has
+ * a better chance to be fetched.
+ */
+ const uint32_t end_offset = intel_bo_get_size(buf->bo) - 1;