[Intel-xe] [PATCH v2 4/4] drm/xe: Correlate engine and cpu timestamps with better accuracy
Souza, Jose
jose.souza at intel.com
Fri Sep 1 17:58:45 UTC 2023
On Mon, 2023-08-14 at 15:37 -0700, Umesh Nerlige Ramappa wrote:
> Perf measurements rely on CPU and engine timestamps to correlate
> events of interest across these time domains. Current mechanisms get
> these timestamps separately and the calculated delta between these
> timestamps lack enough accuracy.
>
> To improve the accuracy of these time measurements to within a few us,
> add a query that returns the engine and cpu timestamps captured as
> close to each other as possible.
>
> Mesa MR: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24591
>
> v2:
> - Fix kernel-doc warnings (CI)
> - Document input params and group them together (Jose)
> - s/cs/engine/ (Jose)
> - Remove padding in the query (Ashutosh)
>
> Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa at intel.com>
> ---
> drivers/gpu/drm/xe/xe_query.c | 138 ++++++++++++++++++++++++++++++++++
> include/uapi/drm/xe_drm.h | 93 ++++++++++++++++++-----
> 2 files changed, 212 insertions(+), 19 deletions(-)
>
> diff --git a/drivers/gpu/drm/xe/xe_query.c b/drivers/gpu/drm/xe/xe_query.c
> index b9d565264ceb..af083f57a618 100644
> --- a/drivers/gpu/drm/xe/xe_query.c
> +++ b/drivers/gpu/drm/xe/xe_query.c
> @@ -6,10 +6,12 @@
> #include "xe_query.h"
>
> #include <linux/nospec.h>
> +#include <linux/sched/clock.h>
>
> #include <drm/ttm/ttm_placement.h>
> #include <drm/xe_drm.h>
>
> +#include "regs/xe_engine_regs.h"
> #include "xe_bo.h"
> #include "xe_device.h"
> #include "xe_exec_queue.h"
> @@ -17,6 +19,7 @@
> #include "xe_gt.h"
> #include "xe_guc_hwconfig.h"
> #include "xe_macros.h"
> +#include "xe_mmio.h"
> #include "xe_ttm_vram_mgr.h"
>
> static const u16 xe_to_user_engine_class[] = {
> @@ -27,6 +30,14 @@ static const u16 xe_to_user_engine_class[] = {
> [XE_ENGINE_CLASS_COMPUTE] = DRM_XE_ENGINE_CLASS_COMPUTE,
> };
>
> +static const enum xe_engine_class user_to_xe_engine_class[] = {
> + [DRM_XE_ENGINE_CLASS_RENDER] = XE_ENGINE_CLASS_RENDER,
> + [DRM_XE_ENGINE_CLASS_COPY] = XE_ENGINE_CLASS_COPY,
> + [DRM_XE_ENGINE_CLASS_VIDEO_DECODE] = XE_ENGINE_CLASS_VIDEO_DECODE,
> + [DRM_XE_ENGINE_CLASS_VIDEO_ENHANCE] = XE_ENGINE_CLASS_VIDEO_ENHANCE,
> + [DRM_XE_ENGINE_CLASS_COMPUTE] = XE_ENGINE_CLASS_COMPUTE,
> +};
> +
> static size_t calc_hw_engine_info_size(struct xe_device *xe)
> {
> struct xe_hw_engine *hwe;
> @@ -45,6 +56,132 @@ static size_t calc_hw_engine_info_size(struct xe_device *xe)
> return i * sizeof(struct drm_xe_engine_class_instance);
> }
>
> +typedef u64 (*__ktime_func_t)(void);
> +static __ktime_func_t __clock_id_to_func(clockid_t clk_id)
> +{
> + /*
> + * Use logic same as the perf subsystem to allow user to select the
> + * reference clock id to be used for timestamps.
> + */
> + switch (clk_id) {
> + case CLOCK_MONOTONIC:
> + return &ktime_get_ns;
> + case CLOCK_MONOTONIC_RAW:
> + return &ktime_get_raw_ns;
> + case CLOCK_REALTIME:
> + return &ktime_get_real_ns;
> + case CLOCK_BOOTTIME:
> + return &ktime_get_boottime_ns;
> + case CLOCK_TAI:
> + return &ktime_get_clocktai_ns;
> + default:
> + return NULL;
> + }
> +}
> +
> +static void
> +__read_timestamps(struct xe_gt *gt,
> + struct xe_reg lower_reg,
> + struct xe_reg upper_reg,
> + u64 *engine_ts,
> + u64 *cpu_ts,
> + u64 *cpu_delta,
> + __ktime_func_t cpu_clock)
> +{
> + u32 upper, lower, old_upper, loop = 0;
> +
> + upper = xe_mmio_read32(gt, upper_reg);
> + do {
> + *cpu_delta = local_clock();
> + *cpu_ts = cpu_clock();
> + lower = xe_mmio_read32(gt, lower_reg);
> + *cpu_delta = local_clock() - *cpu_delta;
> + old_upper = upper;
> + upper = xe_mmio_read32(gt, upper_reg);
> + } while (upper != old_upper && loop++ < 2);
> +
> + *engine_ts = (u64)upper << 32 | lower;
> +}
> +
> +static int
> +query_engine_cycles(struct xe_device *xe,
> + struct drm_xe_device_query *query)
> +{
> + struct drm_xe_query_engine_cycles __user *query_ptr;
> + struct drm_xe_engine_class_instance *eci;
> + struct drm_xe_query_engine_cycles resp;
> + size_t size = sizeof(resp);
> + __ktime_func_t cpu_clock;
> + struct xe_hw_engine *hwe;
> + struct xe_gt *gt;
> +
> + if (query->size == 0) {
> + query->size = size;
> + return 0;
> + } else if (XE_IOCTL_DBG(xe, query->size != size)) {
> + return -EINVAL;
> + }
> +
> + query_ptr = u64_to_user_ptr(query->data);
> + if (copy_from_user(&resp, query_ptr, size))
> + return -EFAULT;
> +
> + cpu_clock = __clock_id_to_func(resp.clockid);
With the plans to remove DRM_IOCTL_XE_MMIO from Xe uAPI (https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/636);
Do you think is worthy the optimization of a resp.clockid value to skip CPU timestamp reads?
In some cases Mesa only needs GPU timestamp, so we could query DRM_XE_QUERY_CS_CYCLES with a clockid to ignore CPU timestamp and not waste some CPU
cycles...
Not sure if is a valid optimization. If not Mesa will continue to query with MONOTONIC and just ignore CPU timestamp.
> + if (!cpu_clock)
> + return -EINVAL;
> +
> + eci = &resp.eci;
> + if (eci->gt_id > XE_MAX_GT_PER_TILE)
> + return -EINVAL;
> +
> + gt = xe_device_get_gt(xe, eci->gt_id);
> + if (!gt)
> + return -EINVAL;
> +
> + if (eci->engine_class >= ARRAY_SIZE(user_to_xe_engine_class))
> + return -EINVAL;
> +
> + hwe = xe_gt_hw_engine(gt, user_to_xe_engine_class[eci->engine_class],
> + eci->engine_instance, true);
> + if (!hwe)
> + return -EINVAL;
> +
> + resp.engine_frequency = gt->info.clock_freq;
> +
> + xe_device_mem_access_get(xe);
> + xe_force_wake_get(gt_to_fw(gt), XE_FORCEWAKE_ALL);
> +
> + __read_timestamps(gt,
> + RING_TIMESTAMP(hwe->mmio_base),
> + RING_TIMESTAMP_UDW(hwe->mmio_base),
> + &resp.engine_cycles,
> + &resp.cpu_timestamp,
> + &resp.cpu_delta,
> + cpu_clock);
> +
> + xe_force_wake_put(gt_to_fw(gt), XE_FORCEWAKE_ALL);
> + xe_device_mem_access_put(xe);
> + resp.width = 36;
> +
> + /* Only write to the output fields of user query */
> + if (put_user(resp.engine_frequency, &query_ptr->engine_frequency))
> + return -EFAULT;
> +
> + if (put_user(resp.cpu_timestamp, &query_ptr->cpu_timestamp))
> + return -EFAULT;
> +
> + if (put_user(resp.cpu_delta, &query_ptr->cpu_delta))
> + return -EFAULT;
> +
> + if (put_user(resp.engine_cycles, &query_ptr->engine_cycles))
> + return -EFAULT;
> +
> + if (put_user(resp.width, &query_ptr->width))
> + return -EFAULT;
> +
> + return 0;
> +}
> +
> static int query_engines(struct xe_device *xe,
> struct drm_xe_device_query *query)
> {
> @@ -376,6 +513,7 @@ static int (* const xe_query_funcs[])(struct xe_device *xe,
> query_gts,
> query_hwconfig,
> query_gt_topology,
> + query_engine_cycles,
> };
>
> int xe_query_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
> diff --git a/include/uapi/drm/xe_drm.h b/include/uapi/drm/xe_drm.h
> index 53cd57342620..4c0c58ca63a3 100644
> --- a/include/uapi/drm/xe_drm.h
> +++ b/include/uapi/drm/xe_drm.h
> @@ -128,6 +128,25 @@ struct xe_user_extension {
> #define DRM_IOCTL_XE_WAIT_USER_FENCE DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_WAIT_USER_FENCE, struct drm_xe_wait_user_fence)
> #define DRM_IOCTL_XE_VM_MADVISE DRM_IOW(DRM_COMMAND_BASE + DRM_XE_VM_MADVISE, struct drm_xe_vm_madvise)
>
> +/** struct drm_xe_engine_class_instance - instance of an engine class */
> +struct drm_xe_engine_class_instance {
> +#define DRM_XE_ENGINE_CLASS_RENDER 0
> +#define DRM_XE_ENGINE_CLASS_COPY 1
> +#define DRM_XE_ENGINE_CLASS_VIDEO_DECODE 2
> +#define DRM_XE_ENGINE_CLASS_VIDEO_ENHANCE 3
> +#define DRM_XE_ENGINE_CLASS_COMPUTE 4
> + /*
> + * Kernel only class (not actual hardware engine class). Used for
> + * creating ordered queues of VM bind operations.
> + */
> +#define DRM_XE_ENGINE_CLASS_VM_BIND 5
> + __u16 engine_class;
> +
> + __u16 engine_instance;
> + __u16 gt_id;
> + __u16 rsvd;
> +};
> +
> /**
> * enum drm_xe_memory_class - Supported memory classes.
> */
> @@ -223,6 +242,60 @@ struct drm_xe_query_mem_region {
> __u64 reserved[6];
> };
>
> +/**
> + * struct drm_xe_query_engine_cycles - correlate CPU and GPU timestamps
> + *
> + * If a query is made with a struct drm_xe_device_query where .query is equal to
> + * DRM_XE_QUERY_CS_CYCLES, then the reply uses struct drm_xe_query_engine_cycles
> + * in .data. struct drm_xe_query_engine_cycles is allocated by the user and
> + * .data points to this allocated structure.
> + *
> + * The query returns the command streamer cycles and the frequency that can
> + * be used to calculate the command streamer timestamp. In addition the
> + * query returns a set of cpu timestamps that indicate when the command
> + * streamer cycle count was captured.
> + */
> +struct drm_xe_query_engine_cycles {
> + /**
> + * @eci: This is input by the user and is the engine for which command
> + * streamer cycles is queried.
> + */
> + struct drm_xe_engine_class_instance eci;
> +
> + /**
> + * @clockid: This is input by the user and is the reference clock id for
> + * CPU timestamp. For definition, see clock_gettime(2) and
> + * perf_event_open(2). Supported clock ids are CLOCK_MONOTONIC,
> + * CLOCK_MONOTONIC_RAW, CLOCK_REALTIME, CLOCK_BOOTTIME, CLOCK_TAI.
> + */
> + __s32 clockid;
> +
> + /** @width: Width of the engine cycle counter in bits. */
> + __u32 width;
> +
> + /**
> + * @engine_cycles: Command streamer cycles as read from the command streamer
> + * register at 0x358 offset.
> + */
> + __u64 engine_cycles;
> +
> + /** @engine_frequency: Frequency of the engine cycles in Hz. */
> + __u64 engine_frequency;
> +
> + /**
> + * @cpu_timestamp: CPU timestamp in ns. The timestamp is captured before
> + * reading the engine_cycles register using the reference clockid set by the
> + * user.
> + */
> + __u64 cpu_timestamp;
> +
> + /**
> + * @cpu_delta: Time delta in ns captured around reading the lower dword
> + * of the engine_cycles register.
> + */
> + __u64 cpu_delta;
> +};
> +
> /**
> * struct drm_xe_query_mem_usage - describe memory regions and usage
> *
> @@ -395,6 +468,7 @@ struct drm_xe_device_query {
> #define DRM_XE_DEVICE_QUERY_GTS 3
> #define DRM_XE_DEVICE_QUERY_HWCONFIG 4
> #define DRM_XE_DEVICE_QUERY_GT_TOPOLOGY 5
> +#define DRM_XE_QUERY_CS_CYCLES 6
> /** @query: The type of data to query */
> __u32 query;
>
> @@ -737,25 +811,6 @@ struct drm_xe_exec_queue_set_property {
> __u64 reserved[2];
> };
>
> -/** struct drm_xe_engine_class_instance - instance of an engine class */
> -struct drm_xe_engine_class_instance {
> -#define DRM_XE_ENGINE_CLASS_RENDER 0
> -#define DRM_XE_ENGINE_CLASS_COPY 1
> -#define DRM_XE_ENGINE_CLASS_VIDEO_DECODE 2
> -#define DRM_XE_ENGINE_CLASS_VIDEO_ENHANCE 3
> -#define DRM_XE_ENGINE_CLASS_COMPUTE 4
> - /*
> - * Kernel only class (not actual hardware engine class). Used for
> - * creating ordered queues of VM bind operations.
> - */
> -#define DRM_XE_ENGINE_CLASS_VM_BIND 5
> - __u16 engine_class;
> -
> - __u16 engine_instance;
> - __u16 gt_id;
> - __u16 rsvd;
> -};
> -
> struct drm_xe_exec_queue_create {
> #define XE_EXEC_QUEUE_EXTENSION_SET_PROPERTY 0
> /** @extensions: Pointer to the first extension struct, if any */
More information about the Intel-xe
mailing list