[PATCH] drm/amd/powerplay: disable ds and ulv for compute
Kenneth Feng
kenneth.feng at amd.com
Thu Nov 7 08:59:27 UTC 2019
This is to improve the performance in the compute mode
for vega10. For example, the original performance for a rocm
bandwidth test: 2G internal GPU copy, is about 99GB/s.
With the idle power features disabled dynamically, the porformance
is promoted to about 215GB/s.
Signed-off-by: Kenneth Feng <kenneth.feng at amd.com>
---
drivers/gpu/drm/amd/powerplay/hwmgr/vega10_hwmgr.c | 62 +++++++++++++++++++++-
1 file changed, 61 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/amd/powerplay/hwmgr/vega10_hwmgr.c b/drivers/gpu/drm/amd/powerplay/hwmgr/vega10_hwmgr.c
index 4ea63a2..c0dbb26 100644
--- a/drivers/gpu/drm/amd/powerplay/hwmgr/vega10_hwmgr.c
+++ b/drivers/gpu/drm/amd/powerplay/hwmgr/vega10_hwmgr.c
@@ -4940,6 +4940,59 @@ static int vega10_get_power_profile_mode(struct pp_hwmgr *hwmgr, char *buf)
return size;
}
+static int vega10_disable_power_features_for_compute_performance(struct pp_hwmgr *hwmgr, bool disable)
+{
+ struct vega10_hwmgr *data = hwmgr->backend;
+ uint32_t feature_mask = 0;
+
+ if (disable) {
+ feature_mask |= data->smu_features[GNLD_ULV].enabled ?
+ data->smu_features[GNLD_ULV].smu_feature_bitmap : 0;
+ feature_mask |= data->smu_features[GNLD_DS_GFXCLK].enabled ?
+ data->smu_features[GNLD_DS_GFXCLK].smu_feature_bitmap : 0;
+ feature_mask |= data->smu_features[GNLD_DS_SOCCLK].enabled ?
+ data->smu_features[GNLD_DS_SOCCLK].smu_feature_bitmap : 0;
+ feature_mask |= data->smu_features[GNLD_DS_LCLK].enabled ?
+ data->smu_features[GNLD_DS_LCLK].smu_feature_bitmap : 0;
+ feature_mask |= data->smu_features[GNLD_DS_DCEFCLK].enabled ?
+ data->smu_features[GNLD_DS_DCEFCLK].smu_feature_bitmap : 0;
+ } else {
+ feature_mask |= (!data->smu_features[GNLD_ULV].enabled) ?
+ data->smu_features[GNLD_ULV].smu_feature_bitmap : 0;
+ feature_mask |= (!data->smu_features[GNLD_DS_GFXCLK].enabled) ?
+ data->smu_features[GNLD_DS_GFXCLK].smu_feature_bitmap : 0;
+ feature_mask |= (!data->smu_features[GNLD_DS_SOCCLK].enabled) ?
+ data->smu_features[GNLD_DS_SOCCLK].smu_feature_bitmap : 0;
+ feature_mask |= (!data->smu_features[GNLD_DS_LCLK].enabled) ?
+ data->smu_features[GNLD_DS_LCLK].smu_feature_bitmap : 0;
+ feature_mask |= (!data->smu_features[GNLD_DS_DCEFCLK].enabled) ?
+ data->smu_features[GNLD_DS_DCEFCLK].smu_feature_bitmap : 0;
+ }
+
+ if (feature_mask)
+ PP_ASSERT_WITH_CODE(!vega10_enable_smc_features(hwmgr,
+ !disable, feature_mask),
+ "enable/disable power features for compute performance Failed!",
+ return -EINVAL);
+
+ if (disable) {
+ data->smu_features[GNLD_ULV].enabled = false;
+ data->smu_features[GNLD_DS_GFXCLK].enabled = false;
+ data->smu_features[GNLD_DS_SOCCLK].enabled = false;
+ data->smu_features[GNLD_DS_LCLK].enabled = false;
+ data->smu_features[GNLD_DS_DCEFCLK].enabled = false;
+ } else {
+ data->smu_features[GNLD_ULV].enabled = true;
+ data->smu_features[GNLD_DS_GFXCLK].enabled = true;
+ data->smu_features[GNLD_DS_SOCCLK].enabled = true;
+ data->smu_features[GNLD_DS_LCLK].enabled = true;
+ data->smu_features[GNLD_DS_DCEFCLK].enabled = true;
+ }
+
+ return 0;
+
+}
+
static int vega10_set_power_profile_mode(struct pp_hwmgr *hwmgr, long *input, uint32_t size)
{
struct vega10_hwmgr *data = hwmgr->backend;
@@ -4948,6 +5001,7 @@ static int vega10_set_power_profile_mode(struct pp_hwmgr *hwmgr, long *input, ui
uint8_t use_rlc_busy;
uint8_t min_active_level;
uint32_t power_profile_mode = input[size];
+ int ret = 0;
if (power_profile_mode == PP_SMC_POWER_PROFILE_CUSTOM) {
if (size != 0 && size != 4)
@@ -4975,11 +5029,17 @@ static int vega10_set_power_profile_mode(struct pp_hwmgr *hwmgr, long *input, ui
}
out:
+ if (power_profile_mode == PP_SMC_POWER_PROFILE_COMPUTE)
+ ret = vega10_disable_power_features_for_compute_performance(hwmgr, true);
+ else
+ ret = vega10_disable_power_features_for_compute_performance(hwmgr, false);
+
smum_send_msg_to_smc_with_parameter(hwmgr, PPSMC_MSG_SetWorkloadMask,
1 << power_profile_mode);
+
hwmgr->power_profile_mode = power_profile_mode;
- return 0;
+ return ret;
}
--
2.7.4
More information about the amd-gfx
mailing list