]> git.baikalelectronics.ru Git - kernel.git/commitdiff
drm/i915/selftests: Verify frequency scaling with RPS
authorChris Wilson <chris@chris-wilson.co.uk>
Mon, 20 Apr 2020 17:27:34 +0000 (18:27 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Mon, 20 Apr 2020 19:08:06 +0000 (20:08 +0100)
One of the core tenents of reclocking the GPU is that its throughput
scales with the clock frequency. We can observe this by incrementing a
loop counter on the GPU, and compare the different execution rates at
the notional RPS frequencies.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200420172739.11620-1-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/gt/selftest_gt_pm.c
drivers/gpu/drm/i915/gt/selftest_rps.c
drivers/gpu/drm/i915/gt/selftest_rps.h

index 0141c334f2ac5e0257f741d73819cc79c03d5ff6..4b2733967c4220e8b1d211934690adbbb8b6b931 100644 (file)
@@ -53,8 +53,9 @@ int intel_gt_pm_live_selftests(struct drm_i915_private *i915)
 {
        static const struct i915_subtest tests[] = {
                SUBTEST(live_rc6_manual),
-               SUBTEST(live_rps_interrupt),
+               SUBTEST(live_rps_frequency),
                SUBTEST(live_rps_power),
+               SUBTEST(live_rps_interrupt),
                SUBTEST(live_gt_resume),
        };
 
index 360f56aa4b826c0c09681edc7c852c1715276d7d..b1a435db1edcee50ded0baa12023f346a5f47190 100644 (file)
@@ -6,6 +6,7 @@
 #include <linux/sort.h>
 
 #include "intel_engine_pm.h"
+#include "intel_gpu_commands.h"
 #include "intel_gt_pm.h"
 #include "intel_rc6.h"
 #include "selftest_rps.h"
@@ -17,6 +18,242 @@ static void dummy_rps_work(struct work_struct *wrk)
 {
 }
 
+static int cmp_u64(const void *A, const void *B)
+{
+       const u64 *a = A, *b = B;
+
+       if (a < b)
+               return -1;
+       else if (a > b)
+               return 1;
+       else
+               return 0;
+}
+
+static struct i915_vma *
+create_spin_counter(struct intel_engine_cs *engine,
+                   struct i915_address_space *vm,
+                   u32 **cancel,
+                   u32 **counter)
+{
+       enum {
+               COUNT,
+               INC,
+               __NGPR__,
+       };
+#define CS_GPR(x) GEN8_RING_CS_GPR(engine->mmio_base, x)
+       struct drm_i915_gem_object *obj;
+       struct i915_vma *vma;
+       u32 *base, *cs;
+       int loop, i;
+       int err;
+
+       obj = i915_gem_object_create_internal(vm->i915, 4096);
+       if (IS_ERR(obj))
+               return ERR_CAST(obj);
+
+       vma = i915_vma_instance(obj, vm, NULL);
+       if (IS_ERR(vma)) {
+               i915_gem_object_put(obj);
+               return vma;
+       }
+
+       err = i915_vma_pin(vma, 0, 0, PIN_USER);
+       if (err) {
+               i915_vma_put(vma);
+               return ERR_PTR(err);
+       }
+
+       base = i915_gem_object_pin_map(obj, I915_MAP_WC);
+       if (IS_ERR(base)) {
+               i915_gem_object_put(obj);
+               return ERR_CAST(base);
+       }
+       cs = base;
+
+       *cs++ = MI_LOAD_REGISTER_IMM(__NGPR__ * 2);
+       for (i = 0; i < __NGPR__; i++) {
+               *cs++ = i915_mmio_reg_offset(CS_GPR(i));
+               *cs++ = 0;
+               *cs++ = i915_mmio_reg_offset(CS_GPR(i)) + 4;
+               *cs++ = 0;
+       }
+
+       *cs++ = MI_LOAD_REGISTER_IMM(1);
+       *cs++ = i915_mmio_reg_offset(CS_GPR(INC));
+       *cs++ = 1;
+
+       loop = cs - base;
+
+       *cs++ = MI_MATH(4);
+       *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(COUNT));
+       *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(INC));
+       *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));
+
+       *cs++ = MI_BATCH_BUFFER_START_GEN8;
+       *cs++ = lower_32_bits(vma->node.start + loop * sizeof(*cs));
+       *cs++ = upper_32_bits(vma->node.start + loop * sizeof(*cs));
+
+       i915_gem_object_flush_map(obj);
+
+       *cancel = base + loop;
+       *counter = memset32(base + 1000, 0, 1);
+       return vma;
+}
+
+static u64 __measure_frequency(u32 *cntr, int duration_ms)
+{
+       u64 dc, dt;
+
+       dt = ktime_get();
+       dc = READ_ONCE(*cntr);
+       usleep_range(1000 * duration_ms, 2000 * duration_ms);
+       dc = READ_ONCE(*cntr) - dc;
+       dt = ktime_get() - dt;
+
+       return div64_u64(1000 * 1000 * dc, dt);
+}
+
+static u64 measure_frequency_at(struct intel_rps *rps, u32 *cntr, int *freq)
+{
+       u64 x[5];
+       int i;
+
+       mutex_lock(&rps->lock);
+       GEM_BUG_ON(!rps->active);
+       intel_rps_set(rps, *freq);
+       mutex_unlock(&rps->lock);
+
+       msleep(20); /* more than enough time to stabilise! */
+
+       for (i = 0; i < 5; i++)
+               x[i] = __measure_frequency(cntr, 2);
+       *freq = read_cagf(rps);
+
+       /* 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)
+{
+       void (*saved_work)(struct work_struct *wrk);
+       struct intel_gt *gt = arg;
+       struct intel_rps *rps = &gt->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,
+                                         &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(READ_ONCE(*cntr), 10)) {
+                       pr_err("%s: timed loop did not start\n",
+                              engine->name);
+                       goto err_vma;
+               }
+
+               min.freq = rps->min_freq;
+               min.count = measure_frequency_at(rps, cntr, &min.freq);
+
+               max.freq = rps->max_freq;
+               max.count = measure_frequency_at(rps, cntr, &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,
+                                  1, 2)) {
+                       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;
+}
+
 static void sleep_for_ei(struct intel_rps *rps, int timeout_us)
 {
        /* Flush any previous EI */
@@ -248,18 +485,6 @@ static u64 __measure_power(int duration_ms)
        return div64_u64(1000 * 1000 * dE, dt);
 }
 
-static int cmp_u64(const void *A, const void *B)
-{
-       const u64 *a = A, *b = B;
-
-       if (a < b)
-               return -1;
-       else if (a > b)
-               return 1;
-       else
-               return 0;
-}
-
 static u64 measure_power_at(struct intel_rps *rps, int freq)
 {
        u64 x[5];
index cad515a7f0e5b0231e8dcc1ff3f4d4b7eb8e6e06..07c2bddf88992a80ac2158c4d1d0a2dce1107666 100644 (file)
@@ -6,6 +6,7 @@
 #ifndef SELFTEST_RPS_H
 #define SELFTEST_RPS_H
 
+int live_rps_frequency(void *arg);
 int live_rps_interrupt(void *arg);
 int live_rps_power(void *arg);