| Message ID | 20220922143209.535323-2-riana.tauro@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Add SLPC selftest live_slpc_power | expand |
On 9/22/2022 7:32 AM, Riana Tauro wrote: > A fundamental assumption is that at lower frequencies, > not only do we run slower, but we save power compared to > higher frequencies. > live_slpc_power checks if running at low frequency saves power > > Signed-off-by: Riana Tauro <riana.tauro@intel.com> > --- > drivers/gpu/drm/i915/gt/selftest_slpc.c | 116 ++++++++++++++++++++++-- > 1 file changed, 107 insertions(+), 9 deletions(-) > > diff --git a/drivers/gpu/drm/i915/gt/selftest_slpc.c b/drivers/gpu/drm/i915/gt/selftest_slpc.c > index f8a1d27df272..f22f091d2844 100644 > --- a/drivers/gpu/drm/i915/gt/selftest_slpc.c > +++ b/drivers/gpu/drm/i915/gt/selftest_slpc.c > @@ -11,7 +11,8 @@ > enum test_type { > VARY_MIN, > VARY_MAX, > - MAX_GRANTED > + MAX_GRANTED, > + SLPC_POWER, > }; > > static int slpc_set_min_freq(struct intel_guc_slpc *slpc, u32 freq) > @@ -41,6 +42,42 @@ static int slpc_set_max_freq(struct intel_guc_slpc *slpc, u32 freq) > return ret; > } > > +static int slpc_set_freq(struct intel_gt *gt, u32 freq) > +{ > + int err; > + struct intel_guc_slpc *slpc = >->uc.guc.slpc; > + > + err = slpc_set_max_freq(slpc, freq); > + if (err) { > + pr_err("Unable to update max freq"); > + return err; > + } > + > + err = slpc_set_min_freq(slpc, freq); > + if (err) { > + pr_err("Unable to update min freq"); > + return err; > + } > + > + return intel_rps_read_actual_frequency(>->rps); The return value here is overloaded (either -ERR or frequency). Can we just return the error status here and query the act_freq in the caller instead? > +} > + > +static u64 measure_slpc_power_at(struct intel_gt *gt, int *freq) Name is a little misleading, maybe slpc_measure_power_at() ? > +{ > + u64 x[5]; > + int i; > + > + *freq = slpc_set_freq(gt, *freq); Here, we can check for return code and then query for act_freq. > + for (i = 0; i < 5; i++) > + x[i] = __measure_power(5); > + *freq = (*freq + intel_rps_read_actual_frequency(>->rps)) / 2; > + > + /* A simple triangle filter for better result stability */ > + sort(x, 5, sizeof(*x), cmp_u64, NULL); > + > + return div_u64(x[1] + 2 * x[2] + x[3], 4); we are duplicating code from selftest_rps here, is it possible to add a helper instead (like __measure_power())? > +} > + > static int vary_max_freq(struct intel_guc_slpc *slpc, struct intel_rps *rps, > u32 *max_act_freq) > { > @@ -113,6 +150,52 @@ static int vary_min_freq(struct intel_guc_slpc *slpc, struct intel_rps *rps, > return err; > } > > +static int slpc_power(struct intel_gt *gt, struct intel_engine_cs *engine) > +{ > + struct intel_guc_slpc *slpc = >->uc.guc.slpc; > + struct { > + u64 power; > + int freq; > + } min, max; > + int err = 0; > + > + /* > + * Our fundamental assumption is that running at lower frequency > + * actually saves power. Let's see if our RAPL measurement support supports* > + * that theory. > + */ > + if (!librapl_supported(gt->i915)) > + return 0; > + > + min.freq = slpc->min_freq; > + min.power = measure_slpc_power_at(gt, &min.freq); > + > + max.freq = slpc->rp0_freq; > + max.power = measure_slpc_power_at(gt, &max.freq); > + > + pr_info("%s: min:%llumW @ %uMHz, max:%llumW @ %uMHz\n", > + engine->name, > + min.power, min.freq, > + max.power, max.freq); > + > + if (10 * min.freq >= 9 * max.freq) { > + pr_notice("Could not control frequency, ran at [%uMHz, %uMhz]\n", > + min.freq, max.freq); > + } > + > + if (11 * min.power > 10 * max.power) { > + pr_err("%s: did not conserve power when setting lower frequency!\n", > + engine->name); > + err = -EINVAL; > + } > + > + /* Restore min/max frequencies */ > + slpc_set_max_freq(slpc, slpc->rp0_freq); > + slpc_set_min_freq(slpc, slpc->min_freq); > + > + return err; > +} > + > static int max_granted_freq(struct intel_guc_slpc *slpc, struct intel_rps *rps, u32 *max_act_freq) > { > struct intel_gt *gt = rps_to_gt(rps); > @@ -233,17 +316,23 @@ static int run_test(struct intel_gt *gt, int test_type) > > err = max_granted_freq(slpc, rps, &max_act_freq); > break; > + > + case SLPC_POWER: > + err = slpc_power(gt, engine); > + break; > } > > - pr_info("Max actual frequency for %s was %d\n", > - engine->name, max_act_freq); > + if (test_type != SLPC_POWER) { > + pr_info("Max actual frequency for %s was %d\n", > + engine->name, max_act_freq); > > - /* Actual frequency should rise above min */ > - if (max_act_freq <= slpc_min_freq) { > - pr_err("Actual freq did not rise above min\n"); > - pr_err("Perf Limit Reasons: 0x%x\n", > - intel_uncore_read(gt->uncore, GT0_PERF_LIMIT_REASONS)); > - err = -EINVAL; > + /* Actual frequency should rise above min */ > + if (max_act_freq <= slpc_min_freq) { > + pr_err("Actual freq did not rise above min\n"); > + pr_err("Perf Limit Reasons: 0x%x\n", > + intel_uncore_read(gt->uncore, GT0_PERF_LIMIT_REASONS)); > + err = -EINVAL; > + } > } > > igt_spinner_end(&spin); > @@ -292,12 +381,21 @@ static int live_slpc_max_granted(void *arg) > return run_test(gt, MAX_GRANTED); > } > > +static int live_slpc_power(void *arg) > +{ > + struct drm_i915_private *i915 = arg; > + struct intel_gt *gt = to_gt(i915); > + > + return run_test(gt, SLPC_POWER); > +} > + > int intel_slpc_live_selftests(struct drm_i915_private *i915) > { > static const struct i915_subtest tests[] = { > SUBTEST(live_slpc_vary_max), > SUBTEST(live_slpc_vary_min), > SUBTEST(live_slpc_max_granted), > + SUBTEST(live_slpc_power), Thanks, Vinay. > }; > > if (intel_gt_is_wedged(to_gt(i915)))
diff --git a/drivers/gpu/drm/i915/gt/selftest_slpc.c b/drivers/gpu/drm/i915/gt/selftest_slpc.c index f8a1d27df272..f22f091d2844 100644 --- a/drivers/gpu/drm/i915/gt/selftest_slpc.c +++ b/drivers/gpu/drm/i915/gt/selftest_slpc.c @@ -11,7 +11,8 @@ enum test_type { VARY_MIN, VARY_MAX, - MAX_GRANTED + MAX_GRANTED, + SLPC_POWER, }; static int slpc_set_min_freq(struct intel_guc_slpc *slpc, u32 freq) @@ -41,6 +42,42 @@ static int slpc_set_max_freq(struct intel_guc_slpc *slpc, u32 freq) return ret; } +static int slpc_set_freq(struct intel_gt *gt, u32 freq) +{ + int err; + struct intel_guc_slpc *slpc = >->uc.guc.slpc; + + err = slpc_set_max_freq(slpc, freq); + if (err) { + pr_err("Unable to update max freq"); + return err; + } + + err = slpc_set_min_freq(slpc, freq); + if (err) { + pr_err("Unable to update min freq"); + return err; + } + + return intel_rps_read_actual_frequency(>->rps); +} + +static u64 measure_slpc_power_at(struct intel_gt *gt, int *freq) +{ + u64 x[5]; + int i; + + *freq = slpc_set_freq(gt, *freq); + for (i = 0; i < 5; i++) + x[i] = __measure_power(5); + *freq = (*freq + intel_rps_read_actual_frequency(>->rps)) / 2; + + /* A simple triangle filter for better result stability */ + sort(x, 5, sizeof(*x), cmp_u64, NULL); + + return div_u64(x[1] + 2 * x[2] + x[3], 4); +} + static int vary_max_freq(struct intel_guc_slpc *slpc, struct intel_rps *rps, u32 *max_act_freq) { @@ -113,6 +150,52 @@ static int vary_min_freq(struct intel_guc_slpc *slpc, struct intel_rps *rps, return err; } +static int slpc_power(struct intel_gt *gt, struct intel_engine_cs *engine) +{ + struct intel_guc_slpc *slpc = >->uc.guc.slpc; + struct { + u64 power; + int freq; + } min, max; + int err = 0; + + /* + * Our fundamental assumption is that running at lower frequency + * actually saves power. Let's see if our RAPL measurement support + * that theory. + */ + if (!librapl_supported(gt->i915)) + return 0; + + min.freq = slpc->min_freq; + min.power = measure_slpc_power_at(gt, &min.freq); + + max.freq = slpc->rp0_freq; + max.power = measure_slpc_power_at(gt, &max.freq); + + pr_info("%s: min:%llumW @ %uMHz, max:%llumW @ %uMHz\n", + engine->name, + min.power, min.freq, + max.power, max.freq); + + if (10 * min.freq >= 9 * max.freq) { + pr_notice("Could not control frequency, ran at [%uMHz, %uMhz]\n", + min.freq, max.freq); + } + + if (11 * min.power > 10 * max.power) { + pr_err("%s: did not conserve power when setting lower frequency!\n", + engine->name); + err = -EINVAL; + } + + /* Restore min/max frequencies */ + slpc_set_max_freq(slpc, slpc->rp0_freq); + slpc_set_min_freq(slpc, slpc->min_freq); + + return err; +} + static int max_granted_freq(struct intel_guc_slpc *slpc, struct intel_rps *rps, u32 *max_act_freq) { struct intel_gt *gt = rps_to_gt(rps); @@ -233,17 +316,23 @@ static int run_test(struct intel_gt *gt, int test_type) err = max_granted_freq(slpc, rps, &max_act_freq); break; + + case SLPC_POWER: + err = slpc_power(gt, engine); + break; } - pr_info("Max actual frequency for %s was %d\n", - engine->name, max_act_freq); + if (test_type != SLPC_POWER) { + pr_info("Max actual frequency for %s was %d\n", + engine->name, max_act_freq); - /* Actual frequency should rise above min */ - if (max_act_freq <= slpc_min_freq) { - pr_err("Actual freq did not rise above min\n"); - pr_err("Perf Limit Reasons: 0x%x\n", - intel_uncore_read(gt->uncore, GT0_PERF_LIMIT_REASONS)); - err = -EINVAL; + /* Actual frequency should rise above min */ + if (max_act_freq <= slpc_min_freq) { + pr_err("Actual freq did not rise above min\n"); + pr_err("Perf Limit Reasons: 0x%x\n", + intel_uncore_read(gt->uncore, GT0_PERF_LIMIT_REASONS)); + err = -EINVAL; + } } igt_spinner_end(&spin); @@ -292,12 +381,21 @@ static int live_slpc_max_granted(void *arg) return run_test(gt, MAX_GRANTED); } +static int live_slpc_power(void *arg) +{ + struct drm_i915_private *i915 = arg; + struct intel_gt *gt = to_gt(i915); + + return run_test(gt, SLPC_POWER); +} + int intel_slpc_live_selftests(struct drm_i915_private *i915) { static const struct i915_subtest tests[] = { SUBTEST(live_slpc_vary_max), SUBTEST(live_slpc_vary_min), SUBTEST(live_slpc_max_granted), + SUBTEST(live_slpc_power), }; if (intel_gt_is_wedged(to_gt(i915)))
A fundamental assumption is that at lower frequencies, not only do we run slower, but we save power compared to higher frequencies. live_slpc_power checks if running at low frequency saves power Signed-off-by: Riana Tauro <riana.tauro@intel.com> --- drivers/gpu/drm/i915/gt/selftest_slpc.c | 116 ++++++++++++++++++++++-- 1 file changed, 107 insertions(+), 9 deletions(-)