[Intel-gfx] [PATCH 4/4] drm/i915/selftests: Split RPS frequency measurement
Mika Kuoppala
mika.kuoppala at linux.intel.com
Mon Apr 20 16:51:09 UTC 2020
Chris Wilson <chris at chris-wilson.co.uk> writes:
> Split the frequency measurement into two modes, so that we can judge the
> impact of the llc setup on top of the pure CS frequency scaling.
>
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> ---
> drivers/gpu/drm/i915/gt/selftest_gt_pm.c | 3 +-
> drivers/gpu/drm/i915/gt/selftest_rps.c | 157 ++++++++++++++++++++++-
> drivers/gpu/drm/i915/gt/selftest_rps.h | 3 +-
> 3 files changed, 154 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/gt/selftest_gt_pm.c b/drivers/gpu/drm/i915/gt/selftest_gt_pm.c
> index de3eaef40596..9855e6f0ce7c 100644
> --- a/drivers/gpu/drm/i915/gt/selftest_gt_pm.c
> +++ b/drivers/gpu/drm/i915/gt/selftest_gt_pm.c
> @@ -54,7 +54,8 @@ int intel_gt_pm_live_selftests(struct drm_i915_private *i915)
> static const struct i915_subtest tests[] = {
> SUBTEST(live_rc6_manual),
> SUBTEST(live_rps_control),
> - SUBTEST(live_rps_frequency),
> + SUBTEST(live_rps_frequency_cs),
> + SUBTEST(live_rps_frequency_srm),
> SUBTEST(live_rps_power),
> SUBTEST(live_rps_interrupt),
> SUBTEST(live_gt_resume),
> diff --git a/drivers/gpu/drm/i915/gt/selftest_rps.c b/drivers/gpu/drm/i915/gt/selftest_rps.c
> index 19fa6a561de3..dbca673519a2 100644
> --- a/drivers/gpu/drm/i915/gt/selftest_rps.c
> +++ b/drivers/gpu/drm/i915/gt/selftest_rps.c
> @@ -33,6 +33,7 @@ static int cmp_u64(const void *A, const void *B)
> static struct i915_vma *
> create_spin_counter(struct intel_engine_cs *engine,
> struct i915_address_space *vm,
> + bool srm,
> u32 **cancel,
> u32 **counter)
> {
> @@ -91,10 +92,12 @@ create_spin_counter(struct intel_engine_cs *engine,
> *cs++ = MI_MATH_ADD;
> *cs++ = MI_MATH_STORE(MI_MATH_REG(COUNT), MI_MATH_REG_ACCU);
>
> - *cs++ = MI_STORE_REGISTER_MEM_GEN8;
> - *cs++ = i915_mmio_reg_offset(CS_GPR(COUNT));
> - *cs++ = lower_32_bits(vma->node.start + 1000 * sizeof(*cs));
> - *cs++ = upper_32_bits(vma->node.start + 1000 * sizeof(*cs));
> + if (srm) {
> + *cs++ = MI_STORE_REGISTER_MEM_GEN8;
> + *cs++ = i915_mmio_reg_offset(CS_GPR(COUNT));
> + *cs++ = lower_32_bits(vma->node.start + 1000 * sizeof(*cs));
> + *cs++ = upper_32_bits(vma->node.start + 1000 * sizeof(*cs));
> + }
>
> *cs++ = MI_BATCH_BUFFER_START_GEN8;
> *cs++ = lower_32_bits(vma->node.start + loop * sizeof(*cs));
> @@ -103,7 +106,7 @@ create_spin_counter(struct intel_engine_cs *engine,
> i915_gem_object_flush_map(obj);
>
> *cancel = base + loop;
> - *counter = memset32(base + 1000, 0, 1);
> + *counter = srm ? memset32(base + 1000, 0, 1) : NULL;
Ok hmm you want a blowout on wrong usage. Fair enough.
> return vma;
> }
>
> @@ -301,12 +304,152 @@ static u64 measure_frequency_at(struct intel_rps *rps, u32 *cntr, int *freq)
> return div_u64(x[1] + 2 * x[2] + x[3], 4);
> }
>
> +static u64 __measure_cs_frequency(struct intel_engine_cs *engine,
> + int duration_ms)
> +{
> + u64 dc, dt;
> +
> + dt = ktime_get();
> + dc = intel_uncore_read_fw(engine->uncore, CS_GPR(0));
> + usleep_range(1000 * duration_ms, 2000 * duration_ms);
> + dc = intel_uncore_read_fw(engine->uncore, CS_GPR(0)) - dc;
Ok, again you measure active engine and getwaway with this.
On the basis that you want minimal impact on the observed
engine,
Reviewed-by: Mika Kuoppala <mika.kuoppala at linux.intel.com>
> + dt = ktime_get() - dt;
> +
> + return div64_u64(1000 * 1000 * dc, dt);
> +}
> +
> +static u64 measure_cs_frequency_at(struct intel_rps *rps,
> + struct intel_engine_cs *engine,
> + int *freq)
> +{
> + u64 x[5];
> + int i;
> +
> + *freq = rps_set_check(rps, *freq);
> + for (i = 0; i < 5; i++)
> + x[i] = __measure_cs_frequency(engine, 2);
> + *freq = (*freq + read_cagf(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 bool scaled_within(u64 x, u64 y, u32 f_n, u32 f_d)
> {
> return f_d * x > f_n * y && f_n * x < f_d * y;
> }
>
> -int live_rps_frequency(void *arg)
> +int live_rps_frequency_cs(void *arg)
> +{
> + void (*saved_work)(struct work_struct *wrk);
> + struct intel_gt *gt = arg;
> + struct intel_rps *rps = >->rps;
> + struct intel_engine_cs *engine;
> + enum intel_engine_id id;
> + int err = 0;
> +
> + /*
> + * The premise is that the GPU does change freqency at our behest.
> + * Let's check there is a correspondence between the requested
> + * frequency, the actual frequency, and the observed clock rate.
> + */
> +
> + if (!rps->enabled || rps->max_freq <= rps->min_freq)
> + return 0;
> +
> + if (INTEL_GEN(gt->i915) < 8) /* for CS simplicity */
> + return 0;
> +
> + intel_gt_pm_wait_for_idle(gt);
> + saved_work = rps->work.func;
> + rps->work.func = dummy_rps_work;
> +
> + for_each_engine(engine, gt, id) {
> + struct i915_request *rq;
> + struct i915_vma *vma;
> + u32 *cancel, *cntr;
> + struct {
> + u64 count;
> + int freq;
> + } min, max;
> +
> + vma = create_spin_counter(engine,
> + engine->kernel_context->vm, false,
> + &cancel, &cntr);
> + if (IS_ERR(vma)) {
> + err = PTR_ERR(vma);
> + break;
> + }
> +
> + rq = intel_engine_create_kernel_request(engine);
> + if (IS_ERR(rq)) {
> + err = PTR_ERR(rq);
> + goto err_vma;
> + }
> +
> + i915_vma_lock(vma);
> + err = i915_request_await_object(rq, vma->obj, false);
> + if (!err)
> + err = i915_vma_move_to_active(vma, rq, 0);
> + if (!err)
> + err = rq->engine->emit_bb_start(rq,
> + vma->node.start,
> + PAGE_SIZE, 0);
> + i915_vma_unlock(vma);
> + i915_request_add(rq);
> + if (err)
> + goto err_vma;
> +
> + if (wait_for(intel_uncore_read(engine->uncore, CS_GPR(0)),
> + 10)) {
> + pr_err("%s: timed loop did not start\n",
> + engine->name);
> + goto err_vma;
> + }
> +
> + min.freq = rps->min_freq;
> + min.count = measure_cs_frequency_at(rps, engine, &min.freq);
> +
> + max.freq = rps->max_freq;
> + max.count = measure_cs_frequency_at(rps, engine, &max.freq);
> +
> + pr_info("%s: min:%lluKHz @ %uMHz, max:%lluKHz @ %uMHz [%d%%]\n",
> + engine->name,
> + min.count, intel_gpu_freq(rps, min.freq),
> + max.count, intel_gpu_freq(rps, max.freq),
> + (int)DIV64_U64_ROUND_CLOSEST(100 * min.freq * max.count,
> + max.freq * min.count));
> +
> + if (!scaled_within(max.freq * min.count,
> + min.freq * max.count,
> + 2, 3)) {
> + pr_err("%s: CS did not scale with frequency! scaled min:%llu, max:%llu\n",
> + engine->name,
> + max.freq * min.count,
> + min.freq * max.count);
> + err = -EINVAL;
> + }
> +
> +err_vma:
> + *cancel = MI_BATCH_BUFFER_END;
> + i915_gem_object_unpin_map(vma->obj);
> + i915_vma_unpin(vma);
> + i915_vma_put(vma);
> +
> + if (igt_flush_test(gt->i915))
> + err = -EIO;
> + if (err)
> + break;
> + }
> +
> + intel_gt_pm_wait_for_idle(gt);
> + rps->work.func = saved_work;
> +
> + return err;
> +}
> +
> +int live_rps_frequency_srm(void *arg)
> {
> void (*saved_work)(struct work_struct *wrk);
> struct intel_gt *gt = arg;
> @@ -341,7 +484,7 @@ int live_rps_frequency(void *arg)
> } min, max;
>
> vma = create_spin_counter(engine,
> - engine->kernel_context->vm,
> + engine->kernel_context->vm, true,
> &cancel, &cntr);
> if (IS_ERR(vma)) {
> err = PTR_ERR(vma);
> diff --git a/drivers/gpu/drm/i915/gt/selftest_rps.h b/drivers/gpu/drm/i915/gt/selftest_rps.h
> index be0bf8e3f639..22e46c5341c5 100644
> --- a/drivers/gpu/drm/i915/gt/selftest_rps.h
> +++ b/drivers/gpu/drm/i915/gt/selftest_rps.h
> @@ -7,7 +7,8 @@
> #define SELFTEST_RPS_H
>
> int live_rps_control(void *arg);
> -int live_rps_frequency(void *arg);
> +int live_rps_frequency_cs(void *arg);
> +int live_rps_frequency_srm(void *arg);
> int live_rps_interrupt(void *arg);
> int live_rps_power(void *arg);
>
> --
> 2.20.1
>
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/intel-gfx
More information about the Intel-gfx
mailing list