diff options
Diffstat (limited to 'drivers/gpu')
-rw-r--r-- | drivers/gpu/nvgpu/common/linux/channel.c | 22 | ||||
-rw-r--r-- | drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.c | 15 | ||||
-rw-r--r-- | drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.h | 3 |
3 files changed, 39 insertions, 1 deletions
diff --git a/drivers/gpu/nvgpu/common/linux/channel.c b/drivers/gpu/nvgpu/common/linux/channel.c index a725cd6b..35fb3023 100644 --- a/drivers/gpu/nvgpu/common/linux/channel.c +++ b/drivers/gpu/nvgpu/common/linux/channel.c | |||
@@ -720,7 +720,8 @@ int gk20a_submit_channel_gpfifo(struct channel_gk20a *c, | |||
720 | return -EINVAL; | 720 | return -EINVAL; |
721 | 721 | ||
722 | if ((flags & (NVGPU_SUBMIT_GPFIFO_FLAGS_FENCE_WAIT | | 722 | if ((flags & (NVGPU_SUBMIT_GPFIFO_FLAGS_FENCE_WAIT | |
723 | NVGPU_SUBMIT_GPFIFO_FLAGS_FENCE_GET)) && | 723 | NVGPU_SUBMIT_GPFIFO_FLAGS_FENCE_GET | |
724 | NVGPU_SUBMIT_GPFIFO_FLAGS_USER_FENCE_UPDATE)) && | ||
724 | !fence) | 725 | !fence) |
725 | return -EINVAL; | 726 | return -EINVAL; |
726 | 727 | ||
@@ -757,6 +758,16 @@ int gk20a_submit_channel_gpfifo(struct channel_gk20a *c, | |||
757 | (g->can_railgate && !c->deterministic) || | 758 | (g->can_railgate && !c->deterministic) || |
758 | !skip_buffer_refcounting; | 759 | !skip_buffer_refcounting; |
759 | 760 | ||
761 | /* | ||
762 | * If User is adding increments to the pushbuffer and doing all job | ||
763 | * tracking, then no need for kernel tracking here | ||
764 | * User should ensure that all pre-requisites for fast submit are met | ||
765 | * Fail the submit if that's not the case | ||
766 | */ | ||
767 | if (need_job_tracking && | ||
768 | (flags & NVGPU_SUBMIT_GPFIFO_FLAGS_USER_FENCE_UPDATE)) | ||
769 | return -EINVAL; | ||
770 | |||
760 | if (need_job_tracking) { | 771 | if (need_job_tracking) { |
761 | bool need_sync_framework = false; | 772 | bool need_sync_framework = false; |
762 | 773 | ||
@@ -868,6 +879,15 @@ int gk20a_submit_channel_gpfifo(struct channel_gk20a *c, | |||
868 | goto clean_up; | 879 | goto clean_up; |
869 | } | 880 | } |
870 | 881 | ||
882 | if (flags & NVGPU_SUBMIT_GPFIFO_FLAGS_USER_FENCE_UPDATE) { | ||
883 | /* | ||
884 | * User space adds increments in the pushbuffer, so just | ||
885 | * handle the threshold book keeping in kernel by adding | ||
886 | * number of syncpoint increments to threshold | ||
887 | */ | ||
888 | c->sync->add_user_incrs(c->sync, fence->value); | ||
889 | } | ||
890 | |||
871 | if (need_job_tracking) { | 891 | if (need_job_tracking) { |
872 | err = channel_gk20a_alloc_job(c, &job); | 892 | err = channel_gk20a_alloc_job(c, &job); |
873 | if (err) | 893 | if (err) |
diff --git a/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.c b/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.c index e965a329..18d61faa 100644 --- a/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.c +++ b/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.c | |||
@@ -301,6 +301,13 @@ static u64 gk20a_channel_syncpt_address(struct gk20a_channel_sync *s) | |||
301 | return sp->syncpt_buf.gpu_va; | 301 | return sp->syncpt_buf.gpu_va; |
302 | } | 302 | } |
303 | 303 | ||
304 | static u32 gk20a_channel_add_user_incrs(struct gk20a_channel_sync *s, u32 val) | ||
305 | { | ||
306 | struct gk20a_channel_syncpt *sp = | ||
307 | container_of(s, struct gk20a_channel_syncpt, ops); | ||
308 | return nvgpu_nvhost_syncpt_incr_max_ext(sp->nvhost_dev, sp->id, val); | ||
309 | } | ||
310 | |||
304 | static void gk20a_channel_syncpt_destroy(struct gk20a_channel_sync *s) | 311 | static void gk20a_channel_syncpt_destroy(struct gk20a_channel_sync *s) |
305 | { | 312 | { |
306 | struct gk20a_channel_syncpt *sp = | 313 | struct gk20a_channel_syncpt *sp = |
@@ -353,6 +360,7 @@ gk20a_channel_syncpt_create(struct channel_gk20a *c) | |||
353 | sp->ops.signal_timeline = gk20a_channel_syncpt_signal_timeline; | 360 | sp->ops.signal_timeline = gk20a_channel_syncpt_signal_timeline; |
354 | sp->ops.syncpt_id = gk20a_channel_syncpt_id; | 361 | sp->ops.syncpt_id = gk20a_channel_syncpt_id; |
355 | sp->ops.syncpt_address = gk20a_channel_syncpt_address; | 362 | sp->ops.syncpt_address = gk20a_channel_syncpt_address; |
363 | sp->ops.add_user_incrs = gk20a_channel_add_user_incrs; | ||
356 | sp->ops.destroy = gk20a_channel_syncpt_destroy; | 364 | sp->ops.destroy = gk20a_channel_syncpt_destroy; |
357 | 365 | ||
358 | return &sp->ops; | 366 | return &sp->ops; |
@@ -878,6 +886,12 @@ static u64 gk20a_channel_semaphore_syncpt_address(struct gk20a_channel_sync *s) | |||
878 | return 0; | 886 | return 0; |
879 | } | 887 | } |
880 | 888 | ||
889 | static u32 gk20a_channel_semaphore_add_user_incrs(struct gk20a_channel_sync *s, | ||
890 | u32 val) | ||
891 | { | ||
892 | return 0; | ||
893 | } | ||
894 | |||
881 | static void gk20a_channel_semaphore_destroy(struct gk20a_channel_sync *s) | 895 | static void gk20a_channel_semaphore_destroy(struct gk20a_channel_sync *s) |
882 | { | 896 | { |
883 | struct gk20a_channel_semaphore *sema = | 897 | struct gk20a_channel_semaphore *sema = |
@@ -930,6 +944,7 @@ gk20a_channel_semaphore_create(struct channel_gk20a *c) | |||
930 | sema->ops.signal_timeline = gk20a_channel_semaphore_signal_timeline; | 944 | sema->ops.signal_timeline = gk20a_channel_semaphore_signal_timeline; |
931 | sema->ops.syncpt_id = gk20a_channel_semaphore_syncpt_id; | 945 | sema->ops.syncpt_id = gk20a_channel_semaphore_syncpt_id; |
932 | sema->ops.syncpt_address = gk20a_channel_semaphore_syncpt_address; | 946 | sema->ops.syncpt_address = gk20a_channel_semaphore_syncpt_address; |
947 | sema->ops.add_user_incrs = gk20a_channel_semaphore_add_user_incrs; | ||
933 | sema->ops.destroy = gk20a_channel_semaphore_destroy; | 948 | sema->ops.destroy = gk20a_channel_semaphore_destroy; |
934 | 949 | ||
935 | return &sema->ops; | 950 | return &sema->ops; |
diff --git a/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.h b/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.h index fe1d8526..c80ebd38 100644 --- a/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.h +++ b/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.h | |||
@@ -105,6 +105,9 @@ struct gk20a_channel_sync { | |||
105 | /* Returns the sync point address of sync point or 0 if not supported */ | 105 | /* Returns the sync point address of sync point or 0 if not supported */ |
106 | u64 (*syncpt_address)(struct gk20a_channel_sync *s); | 106 | u64 (*syncpt_address)(struct gk20a_channel_sync *s); |
107 | 107 | ||
108 | /* Handle user added increments in the push buffer */ | ||
109 | u32 (*add_user_incrs)(struct gk20a_channel_sync *s, u32 val); | ||
110 | |||
108 | /* Free the resources allocated by gk20a_channel_sync_create. */ | 111 | /* Free the resources allocated by gk20a_channel_sync_create. */ |
109 | void (*destroy)(struct gk20a_channel_sync *s); | 112 | void (*destroy)(struct gk20a_channel_sync *s); |
110 | }; | 113 | }; |