aboutsummaryrefslogtreecommitdiffstats
path: root/src/gallium
diff options
context:
space:
mode:
Diffstat (limited to 'src/gallium')
-rw-r--r--src/gallium/drivers/ilo/ilo_gpe_gen6.c15
1 files changed, 14 insertions, 1 deletions
diff --git a/src/gallium/drivers/ilo/ilo_gpe_gen6.c b/src/gallium/drivers/ilo/ilo_gpe_gen6.c
index 536dbf8872d..89566b72a04 100644
--- a/src/gallium/drivers/ilo/ilo_gpe_gen6.c
+++ b/src/gallium/drivers/ilo/ilo_gpe_gen6.c
@@ -767,7 +767,20 @@ gen6_emit_3DSTATE_VERTEX_BUFFERS(const struct ilo_dev_info *dev,
if (vb->buffer && vb->stride <= 2048) {
const struct ilo_buffer *buf = ilo_buffer(vb->buffer);
const uint32_t start_offset = vb->buffer_offset;
- const uint32_t end_offset = buf->bo_size - 1;
+ /*
+ * 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;
dw |= vb->stride << BRW_VB0_PITCH_SHIFT;