ethosu: Add performance counter debug output

Add simple performance counter support as debug output. This is enough
to measure NPU cycles for networks.

Signed-off-by: Rob Herring (Arm) <robh@kernel.org>
Reviewed-by: Tomeu Vizoso <tomeu@tomeuvizoso.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/40269>

include/drm-uapi/ethosu_accel.h

@@ -43,6 +43,11 @@ enum drm_ethosu_ioctl_id {
 
 	/** @DRM_ETHOSU_SUBMIT: Submit a job and BOs to run. */
 	DRM_ETHOSU_SUBMIT,
+
+	DRM_ETHOSU_PERFMON_CREATE,
+	DRM_ETHOSU_PERFMON_DESTROY,
+	DRM_ETHOSU_PERFMON_GET_VALUES,
+	DRM_ETHOSU_PERFMON_SET_GLOBAL,
 };
 
 /**
@@ -79,7 +84,9 @@ struct drm_ethosu_npu_info {
 	__u32 config;
 
 	__u32 sram_size;
+	__u32 pmu_counters;
 };
+
 /**
  * struct drm_ethosu_dev_query - Arguments passed to DRM_ETHOSU_IOCTL_DEV_QUERY
  */
@@ -171,7 +178,6 @@ struct drm_ethosu_bo_wait {
 	__s64 timeout_ns;	/* absolute */
 };
 
-
 struct drm_ethosu_cmdstream_bo_create {
 	/* Size of the data argument. */
 	__u32 size;
@@ -220,10 +226,52 @@ struct drm_ethosu_submit {
 	/** Input: Number of jobs passed in. */
 	__u32 job_count;
 
-	/** Reserved, must be zero. */
+	/** Input: Id returned by DRM_ETHOSU_PERFMON_CREATE */
-	__u32 pad;
+	__u32 perfmon_id;
 };
 
+#define DRM_ETHOSU_MAX_PERF_EVENT_COUNTERS	8
+#define DRM_ETHOSU_MAX_PERF_COUNTERS \
+	(DRM_ETHOSU_MAX_PERF_EVENT_COUNTERS + 1)
+
+struct drm_ethosu_perfmon_create {
+	__u32 id;
+	__u32 ncounters;
+	__u16 counters[DRM_ETHOSU_MAX_PERF_EVENT_COUNTERS];
+};
+
+struct drm_ethosu_perfmon_destroy {
+	__u32 id;
+};
+
+/*
+ * Returns the values of the performance counters tracked by this
+ * perfmon (as an array of (ncounters + 1) u64 values).
+ *
+ * No implicit synchronization is performed, so the user has to
+ * guarantee that any jobs using this perfmon have already been
+ * completed.
+ */
+struct drm_ethosu_perfmon_get_values {
+	__u32 id;
+	__u32 pad;
+	__u64 values_ptr;
+};
+
+#define DRM_ETHOSU_PERFMON_CLEAR_GLOBAL    0x0001
+
+/**
+ * struct drm_ethosu_perfmon_set_global - ioctl to define a global performance
+ * monitor
+ *
+ * The global performance monitor will be used for all jobs. If a global
+ * performance monitor is defined, jobs with a self-defined performance
+ * monitor won't be allowed.
+ */
+struct drm_ethosu_perfmon_set_global {
+	__u32 flags;
+	__u32 id;
+};
 
 /**
  * DRM_IOCTL_ETHOSU() - Build a ethosu IOCTL number
@@ -253,6 +301,14 @@ enum {
 		DRM_IOCTL_ETHOSU(WR, CMDSTREAM_BO_CREATE, cmdstream_bo_create),
 	DRM_IOCTL_ETHOSU_SUBMIT =
 		DRM_IOCTL_ETHOSU(WR, SUBMIT, submit),
+	DRM_IOCTL_ETHOSU_PERFMON_CREATE =
+		DRM_IOCTL_ETHOSU(WR, PERFMON_CREATE, perfmon_create),
+	DRM_IOCTL_ETHOSU_PERFMON_DESTROY =
+		DRM_IOCTL_ETHOSU(WR, PERFMON_DESTROY, perfmon_destroy),
+	DRM_IOCTL_ETHOSU_PERFMON_GET_VALUES =
+		DRM_IOCTL_ETHOSU(WR, PERFMON_GET_VALUES, perfmon_get_values),
+	DRM_IOCTL_ETHOSU_PERFMON_SET_GLOBAL =
+		DRM_IOCTL_ETHOSU(WR, PERFMON_SET_GLOBAL, perfmon_set_global),
 };
 
 #if defined(__cplusplus)

src/gallium/drivers/ethosu/ethosu_device.c

@@ -23,6 +23,7 @@ static const struct debug_named_value ethosu_debug_options[] = {
    {"disable_nhcwb16", ETHOSU_DBG_DISABLE_NHCWB16, "Disable NHCWB16"},
    {"disable_sram", ETHOSU_DBG_DISABLE_SRAM, "Disable SRAM"},
    {"force_u85", ETHOSU_DBG_FORCE_U85, "Force U85 behavior even on U65 hardware"},
+   {"dump_perf", ETHOSU_DBG_DUMP_PERF, "Dump performance counters for each submit"},
    DEBUG_NAMED_VALUE_END};
 
 DEBUG_GET_ONCE_FLAGS_OPTION(ethosu_debug, "ETHOSU_DEBUG", ethosu_debug_options, 0)
@@ -336,4 +337,4 @@ ethosu_ml_device_create(const char *spec)
    set_device_callbacks(device);
 
    return &device->base;
-}
+}

src/gallium/drivers/ethosu/ethosu_device.h

@@ -21,6 +21,7 @@ enum ethosu_dbg {
    ETHOSU_DBG_DISABLE_NHCWB16 = BITFIELD_BIT(3),
    ETHOSU_DBG_DISABLE_SRAM = BITFIELD_BIT(4),
    ETHOSU_DBG_FORCE_U85 = BITFIELD_BIT(5),
+   ETHOSU_DBG_DUMP_PERF = BITFIELD_BIT(6),
 };
 
 extern int ethosu_debug;

src/gallium/drivers/ethosu/ethosu_ml.c

@@ -320,6 +320,23 @@ prepare_for_submission(struct ethosu_subgraph *subgraph,
       }
    }
 
+   subgraph->perfmon_id = 0;
+   if (DBG_ENABLED(ETHOSU_DBG_DUMP_PERF)) {
+
+      struct drm_ethosu_perfmon_create perfmon_create = {
+         .counters = { 32, 35 }, /* npu-idle, npu-active */
+         .ncounters = 2,
+      };
+      ret = drmIoctl(screen->fd, DRM_IOCTL_ETHOSU_PERFMON_CREATE, &perfmon_create);
+      DBG("Perfmon create returned %d\n", ret);
+      if (ret == 0) {
+         subgraph->perfmon_id = perfmon_create.id;
+      } else {
+         DBG("Could not create perfmon: ret=%d errno=%d (%s)\n",
+             ret, errno, strerror(errno));
+      }
+   }
+
    DBG("subgraph->io_used %d\n", subgraph->io_used);
    subgraph->io_rsrc = pipe_buffer_create(pcontext->screen, 0,
                                           PIPE_USAGE_DEFAULT,
@@ -447,6 +464,7 @@ ethosu_ml_subgraph_invoke(struct pipe_context *pcontext,
 
    submit.jobs = (uintptr_t)&job;
    submit.job_count = 1;
+   submit.perfmon_id = subgraph->perfmon_id;
 
    if (DBG_ENABLED(ETHOSU_DBG_MSGS))
       clock_gettime(CLOCK_MONOTONIC_RAW, &start);
@@ -493,6 +511,25 @@ ethosu_ml_subgraph_read_outputs(struct pipe_context *pcontext,
 
       pipe_buffer_read(pcontext, subgraph->io_rsrc, output->offset, output->size, outputs[i]);
    }
+
+   if (DBG_ENABLED(ETHOSU_DBG_DUMP_PERF)) {
+      struct ethosu_screen *screen = ethosu_screen(pcontext->screen);
+      uint64_t values[9];
+      struct drm_ethosu_perfmon_get_values get_values = {
+         .id = subgraph->perfmon_id,
+         .values_ptr = (uintptr_t)values,
+      };
+      int ret;
+
+      ret = drmIoctl(screen->fd, DRM_IOCTL_ETHOSU_PERFMON_GET_VALUES, &get_values);
+      if (ret == 0) {
+         mesa_logi("PMU: cycles=%lu, npu-active=%lu, npu-idle=%lu\n",
+                   values[2], values[1], values[0]);
+      } else {
+         DBG("Could not read perfmon values: ret=%d errno=%d (%s)\n",
+             ret, errno, strerror(errno));
+      }
+   }
 }
 
 void
@@ -500,10 +537,10 @@ ethosu_ml_subgraph_destroy(struct pipe_ml_device *pdevice,
                            struct pipe_ml_subgraph *psubgraph)
 {
    struct ethosu_subgraph *subgraph = (struct ethosu_subgraph *)(psubgraph);
+   struct ethosu_screen *screen = subgraph->screen;
 
    if (subgraph->io_rsrc) {
       /* Post-submission state: cleanup DRM resources */
-      struct ethosu_screen *screen = subgraph->screen;
       struct drm_gem_close arg = {0};
       int ret;
 
@@ -521,6 +558,13 @@ ethosu_ml_subgraph_destroy(struct pipe_ml_device *pdevice,
       free(subgraph->coefs);
    }
 
+   if (DBG_ENABLED(ETHOSU_DBG_DUMP_PERF)) {
+      struct drm_ethosu_perfmon_destroy destroy = {
+         .id = subgraph->perfmon_id,
+      };
+      drmIoctl(screen->fd, DRM_IOCTL_ETHOSU_PERFMON_DESTROY, &destroy);
+   }
+
    util_dynarray_fini(&subgraph->tensors);
 
    free(subgraph);

src/gallium/drivers/ethosu/ethosu_ml.h

@@ -257,6 +257,7 @@ struct ethosu_subgraph {
    uint32_t *cmdstream;
    uint32_t *cursor;
    uint32_t cmdstream_bo;
+   uint32_t perfmon_id;
 
    struct pipe_resource *io_rsrc;
    unsigned io_used;