[Mesa-dev] [PATCH 3/4] i965: Make room in the batch epilogue for three more pipe controls.

[Francisco Jerez]

Review carefully, it sucks to have to keep track of the number of
command packet dwords emitted in the batch epilogue manually.  The
MI_REPORT_PERF_COUNT_BATCH_DWORDS calculation was obviously wrong.
---
 src/mesa/drivers/dri/i965/brw_performance_monitor.c | 10 +++++-----
 src/mesa/drivers/dri/i965/intel_batchbuffer.h       | 10 +++++-----
 2 files changed, 10 insertions(+), 10 deletions(-)

diff --git a/src/mesa/drivers/dri/i965/brw_performance_monitor.c b/src/mesa/drivers/dri/i965/brw_performance_monitor.c
index cb1ab4c..f8e50e1 100644
--- a/src/mesa/drivers/dri/i965/brw_performance_monitor.c
+++ b/src/mesa/drivers/dri/i965/brw_performance_monitor.c
@@ -686,12 +686,12 @@ stop_oa_counters(struct brw_context *brw)
  * The amount of batch space it takes to emit an MI_REPORT_PERF_COUNT snapshot,
  * including the required PIPE_CONTROL flushes.
  *
- * Sandybridge is the worst case scenario: brw_emit_mi_flush
- * expands to three PIPE_CONTROLs which are 4 DWords each.  We have to flush
- * before and after MI_REPORT_PERF_COUNT, so multiply by two.  Finally, add
- * the 3 DWords for MI_REPORT_PERF_COUNT itself.
+ * Sandybridge is the worst case scenario: brw_emit_mi_flush expands to four
+ * PIPE_CONTROLs which are 5 DWords each.  We have to flush before and after
+ * MI_REPORT_PERF_COUNT, so multiply by two.  Finally, add the 3 DWords for
+ * MI_REPORT_PERF_COUNT itself.
  */
-#define MI_REPORT_PERF_COUNT_BATCH_DWORDS (2 * (3 * 4) + 3)
+#define MI_REPORT_PERF_COUNT_BATCH_DWORDS (2 * (4 * 5) + 3)
 
 /**
  * Emit an MI_REPORT_PERF_COUNT command packet.
diff --git a/src/mesa/drivers/dri/i965/intel_batchbuffer.h b/src/mesa/drivers/dri/i965/intel_batchbuffer.h
index aa1dc38..67e8e8f 100644
--- a/src/mesa/drivers/dri/i965/intel_batchbuffer.h
+++ b/src/mesa/drivers/dri/i965/intel_batchbuffer.h
@@ -21,13 +21,13 @@ extern "C" {
  *   - Gen4-5 record ending occlusion query values (4 * 4 = 16 bytes)
  *   - Disabling OA counters on Gen6+ (3 DWords = 12 bytes)
  *   - Ending MI_REPORT_PERF_COUNT on Gen5+, plus associated PIPE_CONTROLs:
- *     - Two sets of PIPE_CONTROLs, which become 3 PIPE_CONTROLs each on SNB,
- *       which are 5 DWords each ==> 2 * 3 * 5 * 4 = 120 bytes
+ *     - Two sets of PIPE_CONTROLs, which become 4 PIPE_CONTROLs each on SNB,
+ *       which are 5 DWords each ==> 2 * 4 * 5 * 4 = 160 bytes
  *     - 3 DWords for MI_REPORT_PERF_COUNT itself on Gen6+.  ==> 12 bytes.
  *       On Ironlake, it's 6 DWords, but we have some slack due to the lack of
  *       Sandybridge PIPE_CONTROL madness.
- *   - CC_STATE workaround on HSW (12 * 4 = 48 bytes)
- *     - 5 dwords for initial mi_flush
+ *   - CC_STATE workaround on HSW (17 * 4 = 68 bytes)
+ *     - 10 dwords for initial mi_flush
  *     - 2 dwords for CC state setup
  *     - 5 dwords for the required pipe control at the end
  *   - Restoring L3 configuration: (24 dwords = 96 bytes)
@@ -35,7 +35,7 @@ extern "C" {
  *     - 7 dwords for L3 configuration set-up.
  *     - 5 dwords for L3 atomic set-up (on HSW).
  */
-#define BATCH_RESERVED 248
+#define BATCH_RESERVED 308
 
 struct intel_batchbuffer;
 
-- 
2.9.0