On 2018-12-14 12:10, Joonas Lahtinen wrote:
Quoting Tvrtko Ursulin (2018-12-10 17:40:34)
On 09/11/2018 17:18, Tomasz Lis wrote:
The patch adds support of preempt-to-idle requesting by setting a proper
bit within Execlist Control Register, and receiving preemption result from
Context Status Buffer.

Preemption in previous gens required a special batch buffer to be executed,
so the Command Streamer never preempted to idle directly. In Icelake it is
possible, as there is a hardware mechanism to inform the kernel about
status of the preemption request.

This patch does not cover using the new preemption mechanism when GuC is
active.

The advantage of this new preemption path is that one less context switch is
needed, and returning information about preempion being complete is received
earlier. This leads to significant improvement in our IGT latency test.

Test performed: `gem_exec_latency --run-subtest render-preemption`, executed
100 times, on the same platform, same kernel, without and with this patch.
Then taken average of the execution latency times:

subcase               old preempt.    icl preempt.
render-render 853.2036        840.1176
render-bsd    2328.8708       2083.2576
render-blt    2080.1501       1852.0792
render-vebox  1553.5134       1428.762

Improvement observed:

subcase               improvement
render-render  1.53%
render-bsd    10.55%
render-blt    10.96%
render-vebox   8.03%
Who can explain what do the parts other than render-render mean? At
least I can make sense of render-render - measure how long it takes for
one context to preempt another, but render-$other draws a blank for me.
How are engines pre-empting one another?
These cases submit low priority spin buffer to 'render' ring, and then on high priority context they submit two buffers which only write current timestamp to a known location: first one to 'render', and second to the other engine. Submission to the other engine with the same context makes sure that 'render' gets back to spin buffer after each iteration. The two high priority tasks are not parallelized, because they announce they write to the same object.

The time taken to do all operations in one iteration is then displayed. So the time includes:
- preempting spin buffer
- executing timestamp write from within 'render'
- executing timestamp write from within other ring (while at the same time, spin buffer gets back to execution on render)

If I'm not mistaken with the above, it looks like the 'render-render' case is the worst one to measure performance increase - if both high priority tasks are executed on the same engine, there's a considerable chance that the hardware will get straight to another pair of them, without going though spin buffer and preempting it.

-Tomasz


But anyway, even if only the 1.53% improvement is the real one, FWIW
that's I think good enough to justify the patch. It is sufficiently
small and contained that I don't see a problem. So:

Acked-by: Tvrtko Ursulin <tvrtko.ursu...@intel.com>
According to Chris, the baseline measurements are off by a decade or so
compared to where they should be. This might be attributed to execution
on frequency locked parts?

Would it be worthy to repeat the numbers with some unlocked parts?

Regards, Joonas

Regards,

Tvrtko

v2: Added needs_preempt_context() change so that it is not created when
      preempt-to-idle is supported. (Chris)
      Updated setting HWACK flag so that it is cleared after
      preempt-to-dle. (Chris, Daniele)
      Updated to use I915_ENGINE_HAS_PREEMPTION flag. (Chris)

v3: Fixed needs_preempt_context() change. (Chris)
      Merged preemption trigger functions to one. (Chris)
      Fixed conyext state tonot assume COMPLETED_MASK after preemption,
      since idle-to-idle case will not have it set.

v4: Simplified needs_preempt_context() change. (Daniele)
      Removed clearing HWACK flag in idle-to-idle preempt. (Daniele)

v5: Renamed inject_preempt_context(). (Daniele)
      Removed duplicated GEM_BUG_ON() on HWACK (Daniele)

v6: Added performance test results.

Bspec: 18922
Cc: Joonas Lahtinen <joonas.lahti...@linux.intel.com>
Cc: Chris Wilson <ch...@chris-wilson.co.uk>
Cc: Daniele Ceraolo Spurio <daniele.ceraolospu...@intel.com>
Cc: Michal Winiarski <michal.winiar...@intel.com>
Cc: Mika Kuoppala <mika.kuopp...@intel.com>
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospu...@intel.com>
Signed-off-by: Tomasz Lis <tomasz....@intel.com>
---
   drivers/gpu/drm/i915/i915_drv.h          |   2 +
   drivers/gpu/drm/i915/i915_gem_context.c  |   3 +-
   drivers/gpu/drm/i915/i915_pci.c          |   3 +-
   drivers/gpu/drm/i915/intel_device_info.h |   1 +
   drivers/gpu/drm/i915/intel_lrc.c         | 109 
+++++++++++++++++++++----------
   drivers/gpu/drm/i915/intel_lrc.h         |   1 +
   6 files changed, 84 insertions(+), 35 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 08d25aa..d2cc9f1 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -2579,6 +2579,8 @@ intel_info(const struct drm_i915_private *dev_priv)
               ((dev_priv)->info.has_logical_ring_elsq)
   #define HAS_LOGICAL_RING_PREEMPTION(dev_priv) \
               ((dev_priv)->info.has_logical_ring_preemption)
+#define HAS_HW_PREEMPT_TO_IDLE(dev_priv) \
+             ((dev_priv)->info.has_hw_preempt_to_idle)
#define HAS_EXECLISTS(dev_priv) HAS_LOGICAL_RING_CONTEXTS(dev_priv) diff --git a/drivers/gpu/drm/i915/i915_gem_context.c b/drivers/gpu/drm/i915/i915_gem_context.c
index b97963d..10b1d61 100644
--- a/drivers/gpu/drm/i915/i915_gem_context.c
+++ b/drivers/gpu/drm/i915/i915_gem_context.c
@@ -529,7 +529,8 @@ static void init_contexts(struct drm_i915_private *i915)
static bool needs_preempt_context(struct drm_i915_private *i915)
   {
-     return HAS_LOGICAL_RING_PREEMPTION(i915);
+     return HAS_LOGICAL_RING_PREEMPTION(i915) &&
+            !HAS_HW_PREEMPT_TO_IDLE(i915);
   }
int i915_gem_contexts_init(struct drm_i915_private *dev_priv)
diff --git a/drivers/gpu/drm/i915/i915_pci.c b/drivers/gpu/drm/i915/i915_pci.c
index 4ccab83..82125cf 100644
--- a/drivers/gpu/drm/i915/i915_pci.c
+++ b/drivers/gpu/drm/i915/i915_pci.c
@@ -600,7 +600,8 @@ static const struct intel_device_info intel_cannonlake_info 
= {
                          TRANSCODER_DSI0_OFFSET, TRANSCODER_DSI1_OFFSET}, \
       GEN(11), \
       .ddb_size = 2048, \
-     .has_logical_ring_elsq = 1
+     .has_logical_ring_elsq = 1, \
+     .has_hw_preempt_to_idle = 1
static const struct intel_device_info intel_icelake_11_info = {
       GEN11_FEATURES,
diff --git a/drivers/gpu/drm/i915/intel_device_info.h 
b/drivers/gpu/drm/i915/intel_device_info.h
index 86ce1db..a2ee278 100644
--- a/drivers/gpu/drm/i915/intel_device_info.h
+++ b/drivers/gpu/drm/i915/intel_device_info.h
@@ -104,6 +104,7 @@ enum intel_ppgtt {
       func(has_logical_ring_contexts); \
       func(has_logical_ring_elsq); \
       func(has_logical_ring_preemption); \
+     func(has_hw_preempt_to_idle); \
       func(has_overlay); \
       func(has_pooled_eu); \
       func(has_psr); \
diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index 08fd9b1..26b7062 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -155,6 +155,7 @@
   #define GEN8_CTX_STATUS_ACTIVE_IDLE (1 << 3)
   #define GEN8_CTX_STATUS_COMPLETE    (1 << 4)
   #define GEN8_CTX_STATUS_LITE_RESTORE        (1 << 15)
+#define GEN11_CTX_STATUS_PREEMPT_IDLE        (1 << 29)
#define GEN8_CTX_STATUS_COMPLETED_MASK \
        (GEN8_CTX_STATUS_COMPLETE | GEN8_CTX_STATUS_PREEMPTED)
@@ -500,29 +501,49 @@ static void port_assign(struct execlist_port *port, 
struct i915_request *rq)
       port_set(port, port_pack(i915_request_get(rq), port_count(port)));
   }
-static void inject_preempt_context(struct intel_engine_cs *engine)
+static void execlist_send_preempt_to_idle(struct intel_engine_cs *engine)
   {
       struct intel_engine_execlists *execlists = &engine->execlists;
-     struct intel_context *ce =
-             to_intel_context(engine->i915->preempt_context, engine);
-     unsigned int n;
+     GEM_TRACE("%s\n", engine->name);
- GEM_BUG_ON(execlists->preempt_complete_status !=
-                upper_32_bits(ce->lrc_desc));
+     if (HAS_HW_PREEMPT_TO_IDLE(engine->i915)) {
+             /*
+              * hardware which HAS_HW_PREEMPT_TO_IDLE(), always also
+              * HAS_LOGICAL_RING_ELSQ(), so we can assume ctrl_reg is set
+              */
+             GEM_BUG_ON(execlists->ctrl_reg == NULL);
- /*
-      * Switch to our empty preempt context so
-      * the state of the GPU is known (idle).
-      */
-     GEM_TRACE("%s\n", engine->name);
-     for (n = execlists_num_ports(execlists); --n; )
-             write_desc(execlists, 0, n);
+             /*
+              * If we have hardware preempt-to-idle, we do not need to
+              * inject any job to the hardware. We only set a flag.
+              */
+             writel(EL_CTRL_PREEMPT_TO_IDLE, execlists->ctrl_reg);
+     } else {
+             struct intel_context *ce =
+                     to_intel_context(engine->i915->preempt_context, engine);
+             unsigned int n;
- write_desc(execlists, ce->lrc_desc, n);
+             GEM_BUG_ON(execlists->preempt_complete_status !=
+                        upper_32_bits(ce->lrc_desc));
+             GEM_BUG_ON((ce->lrc_reg_state[CTX_CONTEXT_CONTROL + 1] &
+                     _MASKED_BIT_ENABLE(CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT |
+                                        CTX_CTRL_ENGINE_CTX_SAVE_INHIBIT)) !=
+                     _MASKED_BIT_ENABLE(CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT |
+                                        CTX_CTRL_ENGINE_CTX_SAVE_INHIBIT));
- /* we need to manually load the submit queue */
-     if (execlists->ctrl_reg)
-             writel(EL_CTRL_LOAD, execlists->ctrl_reg);
+             /*
+              * Switch to our empty preempt context so
+              * the state of the GPU is known (idle).
+              */
+             for (n = execlists_num_ports(execlists); --n; )
+                     write_desc(execlists, 0, n);
+
+             write_desc(execlists, ce->lrc_desc, n);
+
+             /* we need to manually load the submit queue */
+             if (execlists->ctrl_reg)
+                     writel(EL_CTRL_LOAD, execlists->ctrl_reg);
+     }
execlists_clear_active(execlists, EXECLISTS_ACTIVE_HWACK);
       execlists_set_active(execlists, EXECLISTS_ACTIVE_PREEMPT);
@@ -595,7 +616,7 @@ static void execlists_dequeue(struct intel_engine_cs 
*engine)
                       return;
if (need_preempt(engine, last, execlists->queue_priority)) {
-                     inject_preempt_context(engine);
+                     execlist_send_preempt_to_idle(engine);
                       return;
               }
@@ -922,22 +943,43 @@ static void process_csb(struct intel_engine_cs *engine)
                         execlists->active);
status = buf[2 * head];
-             if (status & (GEN8_CTX_STATUS_IDLE_ACTIVE |
-                           GEN8_CTX_STATUS_PREEMPTED))
-                     execlists_set_active(execlists,
-                                          EXECLISTS_ACTIVE_HWACK);
-             if (status & GEN8_CTX_STATUS_ACTIVE_IDLE)
-                     execlists_clear_active(execlists,
-                                            EXECLISTS_ACTIVE_HWACK);
-
-             if (!(status & GEN8_CTX_STATUS_COMPLETED_MASK))
-                     continue;
+             /*
+              * Check if preempted from idle to idle directly.
+              * The STATUS_IDLE_ACTIVE flag is used to mark
+              * such transition.
+              */
+             if ((status & GEN8_CTX_STATUS_IDLE_ACTIVE) &&
+                  (status & GEN11_CTX_STATUS_PREEMPT_IDLE)) {
- /* We should never get a COMPLETED | IDLE_ACTIVE! */
-             GEM_BUG_ON(status & GEN8_CTX_STATUS_IDLE_ACTIVE);
+                     /*
+                      * We could not have COMPLETED anything
+                      * if we were idle before preemption.
+                      */
+                     GEM_BUG_ON(status & GEN8_CTX_STATUS_COMPLETED_MASK);
+             } else {
+                     if (status & (GEN8_CTX_STATUS_IDLE_ACTIVE |
+                                   GEN8_CTX_STATUS_PREEMPTED))
+                             execlists_set_active(execlists,
+                                                  EXECLISTS_ACTIVE_HWACK);
+
+                     if (status & GEN8_CTX_STATUS_ACTIVE_IDLE)
+                             execlists_clear_active(execlists,
+                                                    EXECLISTS_ACTIVE_HWACK);
+
+                     if (!(status & GEN8_CTX_STATUS_COMPLETED_MASK))
+                             continue;
- if (status & GEN8_CTX_STATUS_COMPLETE &&
-                 buf[2*head + 1] == execlists->preempt_complete_status) {
+                     /* We should never get a COMPLETED | IDLE_ACTIVE! */
+                     GEM_BUG_ON(status & GEN8_CTX_STATUS_IDLE_ACTIVE);
+             }
+
+             /*
+              * Check if preempted to real idle, either directly or
+              * the preemptive context already finished executing
+              */
+             if ((status & GEN11_CTX_STATUS_PREEMPT_IDLE) ||
+                 (status & GEN8_CTX_STATUS_COMPLETE &&
+                 buf[2*head + 1] == execlists->preempt_complete_status)) {
                       GEM_TRACE("%s preempt-idle\n", engine->name);
                       complete_preempt_context(execlists);
                       continue;
@@ -2150,7 +2192,8 @@ void intel_execlists_set_default_submission(struct 
intel_engine_cs *engine)
       engine->unpark = NULL;
engine->flags |= I915_ENGINE_SUPPORTS_STATS;
-     if (engine->i915->preempt_context)
+     if (engine->i915->preempt_context ||
+         HAS_HW_PREEMPT_TO_IDLE(engine->i915))
               engine->flags |= I915_ENGINE_HAS_PREEMPTION;
engine->i915->caps.scheduler =
diff --git a/drivers/gpu/drm/i915/intel_lrc.h b/drivers/gpu/drm/i915/intel_lrc.h
index f5a5502..871901a 100644
--- a/drivers/gpu/drm/i915/intel_lrc.h
+++ b/drivers/gpu/drm/i915/intel_lrc.h
@@ -43,6 +43,7 @@
   #define RING_EXECLIST_SQ_CONTENTS(engine)   _MMIO((engine)->mmio_base + 
0x510)
   #define RING_EXECLIST_CONTROL(engine)               
_MMIO((engine)->mmio_base + 0x550)
   #define       EL_CTRL_LOAD                          (1 << 0)
+#define        EL_CTRL_PREEMPT_TO_IDLE               (1 << 1)
/* The docs specify that the write pointer wraps around after 5h, "After status
    * is written out to the last available status QW at offset 5h, this pointer

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

Reply via email to