[PATCH V6] drm/xe/uapi: Use hint for guc to set GT frequency
Souza, Jose
jose.souza at intel.com
Fri Jan 24 18:52:21 UTC 2025
On Fri, 2025-01-24 at 18:25 +0000, Souza, Jose wrote:
> On Fri, 2025-01-24 at 19:03 +0530, Tejas Upadhyay wrote:
> > Allow user to provide a low latency hint. When set,
> > KMD sends a hint to GuC which results in special handling for that process.
> > SLPC will ramp the GT frequency aggressively every time
> > it switches to this process.
> >
> > We need to enable the use of SLPC Compute strategy during init, but
> > it will apply only to processes that set this bit during process
> > creation.
> >
> > Improvement with this approach as below:
> >
> > Before,
> >
> > :~$ NEOReadDebugKeys=1 EnableDirectSubmission=0 clpeak --kernel-latency
> > Platform: Intel(R) OpenCL Graphics
> > Device: Intel(R) Graphics [0xe20b]
> > Driver version : 24.52.0 (Linux x64)
> > Compute units : 160
> > Clock frequency : 2850 MHz
> > Kernel launch latency : 283.16 us
> >
> > After,
> >
> > :~$ NEOReadDebugKeys=1 EnableDirectSubmission=0 clpeak --kernel-latency
> > Platform: Intel(R) OpenCL Graphics
> > Device: Intel(R) Graphics [0xe20b]
> > Driver version : 24.52.0 (Linux x64)
> > Compute units : 160
> > Clock frequency : 2850 MHz
> >
> > Kernel launch latency : 63.38 us
> >
> > Link to UMD PR : https://github.com/intel/compute-runtime/pull/794
Another UMD MR https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33214
> >
> > V6:
> > - init flags, remove redundant flags check (MAuld)
> > V5:
> > - Move uapi doc to documentation and GuC ABI specific change (Rodrigo)
> > - Modify logic to restrict exec queue flags (MAuld)
> > V4:
> > - To make it clear, dont use exec queue word (Vinay)
> > - Correct typo in description of flag (Jose/Vinay)
> > - rename set_strategy api and replace ctx with exec queue(Vinay)
> > - Start with 0th bit to indentify user flags (Jose)
> > V3:
> > - Conver user flag to kernel internal flag and use (Oak)
> > - Support query config for use to check kernel support (Jose)
> > - Dont need to take runtime pm (Vinay)
> > V2:
> > - DRM_XE_EXEC_QUEUE_LOW_LATENCY_HINT 1 is already planned for other hint(Szymon)
> > - Add motivation to description (Lucas)
> >
> > Signed-off-by: Tejas Upadhyay <tejas.upadhyay at intel.com>
> > ---
> > Documentation/gpu/drm-uapi.rst | 18 ++++++++++++++++++
> > drivers/gpu/drm/xe/abi/guc_actions_slpc_abi.h | 3 +++
> > drivers/gpu/drm/xe/xe_exec_queue.c | 10 +++++++---
> > drivers/gpu/drm/xe/xe_exec_queue_types.h | 2 ++
> > drivers/gpu/drm/xe/xe_guc_pc.c | 16 ++++++++++++++++
> > drivers/gpu/drm/xe/xe_guc_submit.c | 8 ++++++++
> > drivers/gpu/drm/xe/xe_query.c | 3 ++-
> > include/uapi/drm/xe_drm.h | 7 ++++++-
> > 8 files changed, 62 insertions(+), 5 deletions(-)
> >
> > diff --git a/Documentation/gpu/drm-uapi.rst b/Documentation/gpu/drm-uapi.rst
> > index b75cc9a70d1f..234c38ef26e3 100644
> > --- a/Documentation/gpu/drm-uapi.rst
> > +++ b/Documentation/gpu/drm-uapi.rst
> > @@ -583,3 +583,21 @@ dma-buf interoperability
> >
> > Please see Documentation/userspace-api/dma-buf-alloc-exchange.rst for
> > information on how dma-buf is integrated and exposed within DRM.
> > +
> > +Low latency hint by user
> > +========================
> > +
> > +Allow users to provide a hint to kernel for cases demanding low latency profile.
> > +
> > +user need to indicate low latency hint with flag while creating exec queue
> > +as mentioned below,
> > +
> > + struct drm_xe_exec_queue_create exec_queue_create = {
> > + .flags = DRM_XE_EXEC_QUEUE_LOW_LATENCY_HINT
> > + .extensions = 0,
> > + .vm_id = vm,
> > + .num_bb_per_exec = 1,
> > + .num_eng_per_bb = 1,
> > + .instances = to_user_pointer(&instance),
> > + };
> > + ioctl(fd, DRM_IOCTL_XE_EXEC_QUEUE_CREATE, &exec_queue_create);
> > diff --git a/drivers/gpu/drm/xe/abi/guc_actions_slpc_abi.h b/drivers/gpu/drm/xe/abi/guc_actions_slpc_abi.h
> > index 85abe4f09ae2..b28c8fa061f7 100644
> > --- a/drivers/gpu/drm/xe/abi/guc_actions_slpc_abi.h
> > +++ b/drivers/gpu/drm/xe/abi/guc_actions_slpc_abi.h
> > @@ -174,6 +174,9 @@ struct slpc_task_state_data {
> > };
> > } __packed;
> >
> > +#define SLPC_CTX_FREQ_REQ_IS_COMPUTE REG_BIT(28)
> > +#define SLPC_OPTIMIZED_STRATEGY_COMPUTE REG_BIT(0)
> > +
> > struct slpc_shared_data_header {
> > /* Total size in bytes of this shared buffer. */
> > u32 size;
> > diff --git a/drivers/gpu/drm/xe/xe_exec_queue.c b/drivers/gpu/drm/xe/xe_exec_queue.c
> > index 8948f50ee58f..0c116faaccf7 100644
> > --- a/drivers/gpu/drm/xe/xe_exec_queue.c
> > +++ b/drivers/gpu/drm/xe/xe_exec_queue.c
> > @@ -549,11 +549,12 @@ int xe_exec_queue_create_ioctl(struct drm_device *dev, void *data,
> > struct xe_tile *tile;
> > struct xe_exec_queue *q = NULL;
> > u32 logical_mask;
> > + u32 flags = 0;
> > u32 id;
> > u32 len;
> > int err;
> >
> > - if (XE_IOCTL_DBG(xe, args->flags) ||
> > + if (XE_IOCTL_DBG(xe, args->flags & ~DRM_XE_EXEC_QUEUE_LOW_LATENCY_HINT) ||
> > XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1]))
> > return -EINVAL;
> >
> > @@ -570,6 +571,9 @@ int xe_exec_queue_create_ioctl(struct drm_device *dev, void *data,
> > if (XE_IOCTL_DBG(xe, eci[0].gt_id >= xe->info.gt_count))
> > return -EINVAL;
> >
> > + if (args->flags & DRM_XE_EXEC_QUEUE_LOW_LATENCY_HINT)
> > + flags |= EXEC_QUEUE_FLAG_LOW_LATENCY;
> > +
> > if (eci[0].engine_class == DRM_XE_ENGINE_CLASS_VM_BIND) {
> > if (XE_IOCTL_DBG(xe, args->width != 1) ||
> > XE_IOCTL_DBG(xe, args->num_placements != 1) ||
> > @@ -578,8 +582,8 @@ int xe_exec_queue_create_ioctl(struct drm_device *dev, void *data,
> >
> > for_each_tile(tile, xe, id) {
> > struct xe_exec_queue *new;
> > - u32 flags = EXEC_QUEUE_FLAG_VM;
> >
> > + flags |= EXEC_QUEUE_FLAG_VM;
> > if (id)
> > flags |= EXEC_QUEUE_FLAG_BIND_ENGINE_CHILD;
> >
> > @@ -626,7 +630,7 @@ int xe_exec_queue_create_ioctl(struct drm_device *dev, void *data,
> > }
> >
> > q = xe_exec_queue_create(xe, vm, logical_mask,
> > - args->width, hwe, 0,
> > + args->width, hwe, flags,
> > args->extensions);
> > up_read(&vm->lock);
> > xe_vm_put(vm);
> > diff --git a/drivers/gpu/drm/xe/xe_exec_queue_types.h b/drivers/gpu/drm/xe/xe_exec_queue_types.h
> > index 5af5419cec7a..b6fdc7ca687e 100644
> > --- a/drivers/gpu/drm/xe/xe_exec_queue_types.h
> > +++ b/drivers/gpu/drm/xe/xe_exec_queue_types.h
> > @@ -85,6 +85,8 @@ struct xe_exec_queue {
> > #define EXEC_QUEUE_FLAG_BIND_ENGINE_CHILD BIT(3)
> > /* kernel exec_queue only, set priority to highest level */
> > #define EXEC_QUEUE_FLAG_HIGH_PRIORITY BIT(4)
> > +/* flag to indicate low latency hint to guc */
> > +#define EXEC_QUEUE_FLAG_LOW_LATENCY BIT(5)
> >
> > /**
> > * @flags: flags for this exec queue, should statically setup aside from ban
> > diff --git a/drivers/gpu/drm/xe/xe_guc_pc.c b/drivers/gpu/drm/xe/xe_guc_pc.c
> > index 44b5211066ef..e926f89e4ce9 100644
> > --- a/drivers/gpu/drm/xe/xe_guc_pc.c
> > +++ b/drivers/gpu/drm/xe/xe_guc_pc.c
> > @@ -1014,6 +1014,17 @@ static int slpc_set_policies(struct xe_guc_pc *pc)
> > return 0;
> > }
> >
> > +static int pc_action_set_strategy(struct xe_guc_pc *pc, u32 val)
> > +{
> > + int ret = 0;
> > +
> > + ret = pc_action_set_param(pc,
> > + SLPC_PARAM_STRATEGIES,
> > + val);
> > +
> > + return ret;
> > +}
> > +
> > /**
> > * xe_guc_pc_start - Start GuC's Power Conservation component
> > * @pc: Xe_GuC_PC instance
> > @@ -1077,6 +1088,11 @@ int xe_guc_pc_start(struct xe_guc_pc *pc)
> > }
> >
> > ret = pc_action_setup_gucrc(pc, GUCRC_FIRMWARE_CONTROL);
> > + if (ret)
> > + goto out;
> > +
> > + /* Enable SLPC Optimized Strategy for compute */
> > + ret = pc_action_set_strategy(pc, SLPC_OPTIMIZED_STRATEGY_COMPUTE);
> >
> > out:
> > xe_force_wake_put(gt_to_fw(gt), fw_ref);
> > diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c b/drivers/gpu/drm/xe/xe_guc_submit.c
> > index 913c74d6e2ae..f2ce3086838c 100644
> > --- a/drivers/gpu/drm/xe/xe_guc_submit.c
> > +++ b/drivers/gpu/drm/xe/xe_guc_submit.c
> > @@ -15,6 +15,7 @@
> > #include <drm/drm_managed.h>
> >
> > #include "abi/guc_actions_abi.h"
> > +#include "abi/guc_actions_slpc_abi.h"
> > #include "abi/guc_klvs_abi.h"
> > #include "regs/xe_lrc_layout.h"
> > #include "xe_assert.h"
> > @@ -400,6 +401,7 @@ static void __guc_exec_queue_policy_add_##func(struct exec_queue_policy *policy,
> > MAKE_EXEC_QUEUE_POLICY_ADD(execution_quantum, EXECUTION_QUANTUM)
> > MAKE_EXEC_QUEUE_POLICY_ADD(preemption_timeout, PREEMPTION_TIMEOUT)
> > MAKE_EXEC_QUEUE_POLICY_ADD(priority, SCHEDULING_PRIORITY)
> > +MAKE_EXEC_QUEUE_POLICY_ADD(slpc_exec_queue_freq_req, SLPM_GT_FREQUENCY)
> > #undef MAKE_EXEC_QUEUE_POLICY_ADD
> >
> > static const int xe_exec_queue_prio_to_guc[] = {
> > @@ -414,14 +416,20 @@ static void init_policies(struct xe_guc *guc, struct xe_exec_queue *q)
> > struct exec_queue_policy policy;
> > enum xe_exec_queue_priority prio = q->sched_props.priority;
> > u32 timeslice_us = q->sched_props.timeslice_us;
> > + u32 slpc_exec_queue_freq_req = 0;
> > u32 preempt_timeout_us = q->sched_props.preempt_timeout_us;
> >
> > xe_gt_assert(guc_to_gt(guc), exec_queue_registered(q));
> >
> > + if (q->flags & EXEC_QUEUE_FLAG_LOW_LATENCY)
> > + slpc_exec_queue_freq_req |= SLPC_CTX_FREQ_REQ_IS_COMPUTE;
> > +
> > __guc_exec_queue_policy_start_klv(&policy, q->guc->id);
> > __guc_exec_queue_policy_add_priority(&policy, xe_exec_queue_prio_to_guc[prio]);
> > __guc_exec_queue_policy_add_execution_quantum(&policy, timeslice_us);
> > __guc_exec_queue_policy_add_preemption_timeout(&policy, preempt_timeout_us);
> > + __guc_exec_queue_policy_add_slpc_exec_queue_freq_req(&policy,
> > + slpc_exec_queue_freq_req);
> >
> > xe_guc_ct_send(&guc->ct, (u32 *)&policy.h2g,
> > __guc_exec_queue_policy_action_size(&policy), 0, 0);
> > diff --git a/drivers/gpu/drm/xe/xe_query.c b/drivers/gpu/drm/xe/xe_query.c
> > index c059639613f7..136280c0c6b9 100644
> > --- a/drivers/gpu/drm/xe/xe_query.c
> > +++ b/drivers/gpu/drm/xe/xe_query.c
> > @@ -334,7 +334,8 @@ static int query_config(struct xe_device *xe, struct drm_xe_device_query *query)
> > xe->info.devid | (xe->info.revid << 16);
> > if (xe_device_get_root_tile(xe)->mem.vram.usable_size)
> > config->info[DRM_XE_QUERY_CONFIG_FLAGS] =
> > - DRM_XE_QUERY_CONFIG_FLAG_HAS_VRAM;
> > + DRM_XE_QUERY_CONFIG_FLAG_HAS_VRAM |
> > + DRM_XE_QUERY_CONFIG_FLAG_HAS_LOW_LATENCY;
>
> is this feature only supported in discrete GPUs?
>
> > config->info[DRM_XE_QUERY_CONFIG_MIN_ALIGNMENT] =
> > xe->info.vram_flags & XE_VRAM_FLAGS_NEED64K ? SZ_64K : SZ_4K;
> > config->info[DRM_XE_QUERY_CONFIG_VA_BITS] = xe->info.va_bits;
> > diff --git a/include/uapi/drm/xe_drm.h b/include/uapi/drm/xe_drm.h
> > index e2160330ad01..decbb3b2df36 100644
> > --- a/include/uapi/drm/xe_drm.h
> > +++ b/include/uapi/drm/xe_drm.h
> > @@ -393,6 +393,8 @@ struct drm_xe_query_mem_regions {
> > *
> > * - %DRM_XE_QUERY_CONFIG_FLAG_HAS_VRAM - Flag is set if the device
> > * has usable VRAM
> > + * - %DRM_XE_QUERY_CONFIG_FLAG_HAS_LOW_LATENCY - Flag is set if the device
> > + * has low latency hint support
> > * - %DRM_XE_QUERY_CONFIG_MIN_ALIGNMENT - Minimal memory alignment
> > * required by this device, typically SZ_4K or SZ_64K
> > * - %DRM_XE_QUERY_CONFIG_VA_BITS - Maximum bits of a virtual address
> > @@ -409,6 +411,7 @@ struct drm_xe_query_config {
> > #define DRM_XE_QUERY_CONFIG_REV_AND_DEVICE_ID 0
> > #define DRM_XE_QUERY_CONFIG_FLAGS 1
> > #define DRM_XE_QUERY_CONFIG_FLAG_HAS_VRAM (1 << 0)
> > + #define DRM_XE_QUERY_CONFIG_FLAG_HAS_LOW_LATENCY (1 << 1)
> > #define DRM_XE_QUERY_CONFIG_MIN_ALIGNMENT 2
> > #define DRM_XE_QUERY_CONFIG_VA_BITS 3
> > #define DRM_XE_QUERY_CONFIG_MAX_EXEC_QUEUE_PRIORITY 4
> > @@ -1124,6 +1127,7 @@ struct drm_xe_vm_bind {
> > * .engine_class = DRM_XE_ENGINE_CLASS_RENDER,
> > * };
> > * struct drm_xe_exec_queue_create exec_queue_create = {
> > + * .flags = DRM_XE_EXEC_QUEUE_LOW_LATENCY_HINT
> > * .extensions = 0,
> > * .vm_id = vm,
> > * .num_bb_per_exec = 1,
> > @@ -1150,7 +1154,8 @@ struct drm_xe_exec_queue_create {
> > /** @vm_id: VM to use for this exec queue */
> > __u32 vm_id;
> >
> > - /** @flags: MBZ */
> > +#define DRM_XE_EXEC_QUEUE_LOW_LATENCY_HINT (1 << 0)
> > + /** @flags: flags to use for this exec queue */
> > __u32 flags;
> >
> > /** @exec_queue_id: Returned exec queue ID */
>
More information about the Intel-xe
mailing list