radeonsi: Inhibit clock-gating for perf counters.

Otherwise most counters return 0. Should be much more user friendly
than having to totally disable clock-gating on the kernel cmdline.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5972>

diff --git a/src/gallium/drivers/radeonsi/si_perfcounter.c b/src/gallium/drivers/radeonsi/si_perfcounter.c
index 8825926064dafee2a612a15c4acff3926c99f30b..d2f0db81d2c88ac2e3d1866786d25101fe88e5cf 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_perfcounter.c
+++ b/src/gallium/drivers/radeonsi/si_perfcounter.c
@@ -919,6 +919,17 @@ static void si_pc_query_destroy(struct si_context *sctx, struct si_query *squery
    FREE(query);
 }
 
+static void si_inhibit_clockgating(struct si_context *sctx, bool inhibit)
+{
+   if (sctx->chip_class >= GFX10) {
+      radeon_set_uconfig_reg(sctx->gfx_cs, R_037390_RLC_PERFMON_CLK_CNTL,
+                            S_037390_PERFMON_CLOCK_STATE(inhibit));
+   } else if (sctx->chip_class >= GFX8) {
+      radeon_set_uconfig_reg(sctx->gfx_cs, R_0372FC_RLC_PERFMON_CLK_CNTL,
+                            S_0372FC_PERFMON_CLOCK_STATE(inhibit));
+   }
+}
+
 static void si_pc_query_resume(struct si_context *sctx, struct si_query *squery)
 /*
                                    struct si_query_hw *hwquery,
@@ -935,6 +946,8 @@ static void si_pc_query_resume(struct si_context *sctx, struct si_query *squery)
    if (query->shaders)
       si_pc_emit_shaders(sctx, query->shaders);
 
+   si_inhibit_clockgating(sctx, true);
+
    for (struct si_query_group *group = query->groups; group; group = group->next) {
       struct si_pc_block *block = group->block;
 
@@ -986,6 +999,8 @@ static void si_pc_query_suspend(struct si_context *sctx, struct si_query *squery
    }
 
    si_pc_emit_instance(sctx, -1, -1);
+
+   si_inhibit_clockgating(sctx, false);
 }
 
 static bool si_pc_query_begin(struct si_context *ctx, struct si_query *squery)