libdrm_intel: Add API for execbuf pad to size functionality
RFC 1/1 libdrm_intel: Add API for execbuf pad to size functionality
From: Tvrtko Ursulin <[email protected]>
New kernels add the ability to pad objects to specified size at execbuf time.
Add the drm_intel_bo_pad_to_size API via which this padded size can be set.
--- include/drm/i915_drm.h | 5 +++-- intel/intel_bufmgr.c | 6 ++++++ intel/intel_bufmgr.h | 2 ++ intel/intel_bufmgr_gem.c | 23 +++++++++++++++++++++-- intel/intel_bufmgr_priv.h | 17 +++++++++++++++++ 5 files changed, 49 insertions(+), 4 deletions(-)
include/drm/i915_drm.h +3 -2
@@ -662,10 +662,11 @@ struct drm_i915_gem_exec_object2 {#define EXEC_OBJECT_NEEDS_FENCE (1<<0)#define EXEC_OBJECT_NEEDS_GTT (1<<1)#define EXEC_OBJECT_WRITE (1<<2) -#define __EXEC_OBJECT_UNKNOWN_FLAGS -(EXEC_OBJECT_WRITE<<1) +#define EXEC_OBJECT_PAD_TO_SIZE (1<<3) +#define __EXEC_OBJECT_UNKNOWN_FLAGS -(EXEC_OBJECT_PAD_TO_SIZE<<1)__u64 flags; -__u64 rsvd1; +__u64 pad_to_size;__u64 rsvd2; };
intel/intel_bufmgr.c +6 -0
@@ -299,6 +299,12 @@ drm_intel_bo_references(drm_intel_bo *bo, drm_intel_bo *target_bo) } drm_public int +drm_intel_bo_pad_to_size (drm_intel_bo *bo,uint64_t pad_to_size) +{ +return bo->bufmgr->bo_pad_to_size (bo, pad_to_size); +} + +drm_public intdrm_intel_get_pipe_from_crtc_id (drm_intel_bufmgr *bufmgr,int crtc_id) {if (bufmgr->get_pipe_from_crtc_id)
intel/intel_bufmgr.h +2 -0
@@ -165,6 +165,8 @@ int drm_intel_bo_disable_reuse(drm_intel_bo *bo);int drm_intel_bo_is_reusable (drm_intel_bo *bo);int drm_intel_bo_references (drm_intel_bo *bo,drm_intel_bo *target_bo); +int drm_intel_bo_pad_to_size (drm_intel_bo *bo,uint64_t pad_to_size); +/* drm_intel_bufmgr_gem.c */ drm_intel_bufmgr *drm_intel_bufmgr_gem_init (int fd,int batch_size); drm_intel_bo *drm_intel_bo_gem_create_from_name(drm_intel_bufmgr *bufmgr,
intel/intel_bufmgr_gem.c +21 -2
@@ -255,6 +255,12 @@ struct _drm_intel_bo_gem {drm_intel_aub_annotation *aub_annotations;unsigned aub_annotation_count; + +/** +* Size to pad the object to. +* +*/ +uint64_t pad_to_size; }; static unsigned int
@@ -504,8 +510,8 @@ drm_intel_add_validate_buffer2(drm_intel_bo *bo, int need_fence) bufmgr_gem->exec2_objects[index].alignment =0 ; bufmgr_gem->exec2_objects[index].offset =0 ; bufmgr_gem->exec_bos[index] = bo; - bufmgr_gem->exec2_objects[index].flags =0 ; - bufmgr_gem->exec2_objects[index].rsvd1 =0 ; + bufmgr_gem->exec2_objects[index].flags = bo_gem->pad_to_size ?EXEC_OBJECT_PAD_TO_SIZE :0 ; + bufmgr_gem->exec2_objects[index].pad_to_size = bo_gem->pad_to_size; bufmgr_gem->exec2_objects[index].rsvd2 =0 ;if (need_fence) { bufmgr_gem->exec2_objects[index].flags |=
@@ -1198,6 +1204,8 @@ drm_intel_gem_bo_unreference_final(drm_intel_bo *bo, time_t time)DBG ("bo_unreference final: %d (%s)\n" , bo_gem->gem_handle, bo_gem->name); + bo_gem->pad_to_size =0 ; +/* release memory associated with this object */ if (bo_gem->reloc_target_info) {free (bo_gem->reloc_target_info);
@@ -1855,6 +1863,16 @@ do_bo_emit_reloc(drm_intel_bo *bo, uint32_t offset, }static int +drm_intel_gem_bo_pad_to_size (drm_intel_bo *bo,uint64_t pad_to_size) +{ +drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; + + bo_gem->pad_to_size = bo->size < pad_size ? pad_size :0 ; + +return 0 ; +} + +static int drm_intel_gem_bo_emit_reloc (drm_intel_bo *bo,uint32_t offset,drm_intel_bo *target_bo,uint32_t target_offset,uint32_t read_domains,uint32_t write_domain)
@@ -3603,6 +3621,7 @@ drm_intel_bufmgr_gem_init(int fd, int batch_size) bufmgr_gem->bufmgr.bo_subdata = drm_intel_gem_bo_subdata; bufmgr_gem->bufmgr.bo_get_subdata = drm_intel_gem_bo_get_subdata; bufmgr_gem->bufmgr.bo_wait_rendering = drm_intel_gem_bo_wait_rendering; + bufmgr_gem->bufmgr.bo_pad_to_size = drm_intel_gem_bo_pad_to_size; bufmgr_gem->bufmgr.bo_emit_reloc = drm_intel_gem_bo_emit_reloc; bufmgr_gem->bufmgr.bo_emit_reloc_fence = drm_intel_gem_bo_emit_reloc_fence; bufmgr_gem->bufmgr.bo_pin = drm_intel_gem_bo_pin;
intel/intel_bufmgr_priv.h +17 -0
@@ -152,6 +152,23 @@ struct _drm_intel_bufmgr {void (*destroy) (drm_intel_bufmgr *bufmgr);/** +* Sets buffer total padded size when buffer is used by the GPU. +* +* This enables dynamic padding to be added without using any backing +* storage. For example handling GPU padding requirements for buffers +* allocated by an entity unaware of the same. +* +* Set padded size remains active until reset (to zero or actual object +* size). +* +* Returns 0 on success or an error code. +* +* \param bo Buffer to set total padded size for +* \param pad_to_size Total size in bytes of object plus padding +*/ +int (*bo_pad_to_size) (drm_intel_bo *bo,uint64_t pad_to_size); + +/** * Add relocation entry in reloc_buf, which will be updated with the * target buffer's real offset on on command submission. *
Signature · 6 lines
2.3.2 _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 5 lines · Tvrtko Ursulin
From: Tvrtko Ursulin <[email protected]>
New kernels add the ability to pad objects to specified size at execbuf time.
Add the
drm_intel_bo_pad_to_sizeAPI via which this padded size can be set.
Looks good, exactly what I had in mind.
intel/intel_bufmgr_gem.c do_bo_emit_reloc(drm_intel_bo *bo, uint32_t offset, · line 1870
Quoted 6 lines
static int
intel/intel_bufmgr_gem.c +5
-0+drm_intel_gem_bo_pad_to_size(drm_intel_bo *bo, uint64_t pad_to_size) +{ +drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; + + bo_gem->pad_to_size = bo->size < pad_size ? pad_size :0 ;
Probably best to return -EINVAL here rather than silently drop the request. Knowing our callers, they will ignore the error anyway.
One thing I was considering inside the kernel was to allow the GTT allocation to be smaller than the object. I don't have a good usecase though, because with a subobject you are likely to want to use all of it at some point, so you will likely find it all bound anyway. (And I don't think the penalty for the occasional rebind is that onerous.) -Chris
Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
RFC v2 1/1 libdrm_intel: Add API for execbuf pad to size functionality
In reply to Tvrtko UrsulinFrom: Tvrtko Ursulin <[email protected]>
New kernels add the ability to pad objects to specified size at execbuf time.
Add the drm_intel_bo_pad_to_size API via which this padded size can be set.
v2: Reject padded size smaller than object size. (Chris Wilson)
--- include/drm/i915_drm.h | 5 +++-- intel/intel_bufmgr.c | 6 ++++++ intel/intel_bufmgr.h | 2 ++ intel/intel_bufmgr_gem.c | 27 +++++++++++++++++++++++++-- intel/intel_bufmgr_priv.h | 17 +++++++++++++++++ 5 files changed, 53 insertions(+), 4 deletions(-)
include/drm/i915_drm.h +3 -2
@@ -662,10 +662,11 @@ struct drm_i915_gem_exec_object2 {#define EXEC_OBJECT_NEEDS_FENCE (1<<0)#define EXEC_OBJECT_NEEDS_GTT (1<<1)#define EXEC_OBJECT_WRITE (1<<2) -#define __EXEC_OBJECT_UNKNOWN_FLAGS -(EXEC_OBJECT_WRITE<<1) +#define EXEC_OBJECT_PAD_TO_SIZE (1<<3) +#define __EXEC_OBJECT_UNKNOWN_FLAGS -(EXEC_OBJECT_PAD_TO_SIZE<<1)__u64 flags; -__u64 rsvd1; +__u64 pad_to_size;__u64 rsvd2; };
intel/intel_bufmgr.c +6 -0
@@ -299,6 +299,12 @@ drm_intel_bo_references(drm_intel_bo *bo, drm_intel_bo *target_bo) } drm_public int +drm_intel_bo_pad_to_size (drm_intel_bo *bo,uint64_t pad_to_size) +{ +return bo->bufmgr->bo_pad_to_size (bo, pad_to_size); +} + +drm_public intdrm_intel_get_pipe_from_crtc_id (drm_intel_bufmgr *bufmgr,int crtc_id) {if (bufmgr->get_pipe_from_crtc_id)
intel/intel_bufmgr.h +2 -0
@@ -165,6 +165,8 @@ int drm_intel_bo_disable_reuse(drm_intel_bo *bo);int drm_intel_bo_is_reusable (drm_intel_bo *bo);int drm_intel_bo_references (drm_intel_bo *bo,drm_intel_bo *target_bo); +int drm_intel_bo_pad_to_size (drm_intel_bo *bo,uint64_t pad_to_size); +/* drm_intel_bufmgr_gem.c */ drm_intel_bufmgr *drm_intel_bufmgr_gem_init (int fd,int batch_size); drm_intel_bo *drm_intel_bo_gem_create_from_name(drm_intel_bufmgr *bufmgr,
intel/intel_bufmgr_gem.c +25 -2
@@ -255,6 +255,12 @@ struct _drm_intel_bo_gem {drm_intel_aub_annotation *aub_annotations;unsigned aub_annotation_count; + +/** +* Size to pad the object to. +* +*/ +uint64_t pad_to_size; }; static unsigned int
@@ -504,8 +510,8 @@ drm_intel_add_validate_buffer2(drm_intel_bo *bo, int need_fence) bufmgr_gem->exec2_objects[index].alignment =0 ; bufmgr_gem->exec2_objects[index].offset =0 ; bufmgr_gem->exec_bos[index] = bo; - bufmgr_gem->exec2_objects[index].flags =0 ; - bufmgr_gem->exec2_objects[index].rsvd1 =0 ; + bufmgr_gem->exec2_objects[index].flags = bo_gem->pad_to_size ?EXEC_OBJECT_PAD_TO_SIZE :0 ; + bufmgr_gem->exec2_objects[index].pad_to_size = bo_gem->pad_to_size; bufmgr_gem->exec2_objects[index].rsvd2 =0 ;if (need_fence) { bufmgr_gem->exec2_objects[index].flags |=
@@ -1198,6 +1204,8 @@ drm_intel_gem_bo_unreference_final(drm_intel_bo *bo, time_t time)DBG ("bo_unreference final: %d (%s)\n" , bo_gem->gem_handle, bo_gem->name); + bo_gem->pad_to_size =0 ; +/* release memory associated with this object */ if (bo_gem->reloc_target_info) {free (bo_gem->reloc_target_info);
@@ -1855,6 +1863,20 @@ do_bo_emit_reloc(drm_intel_bo *bo, uint32_t offset, }static int +drm_intel_gem_bo_pad_to_size (drm_intel_bo *bo,uint64_t pad_to_size) +{ +drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; + +if (pad_to_size < bo->size) +return -EINVAL ; + +if (pad_to_size > bo->size) + bo_gem->pad_to_size = pad_to_size; + +return 0 ; +} + +static int drm_intel_gem_bo_emit_reloc (drm_intel_bo *bo,uint32_t offset,drm_intel_bo *target_bo,uint32_t target_offset,uint32_t read_domains,uint32_t write_domain)
@@ -3603,6 +3625,7 @@ drm_intel_bufmgr_gem_init(int fd, int batch_size) bufmgr_gem->bufmgr.bo_subdata = drm_intel_gem_bo_subdata; bufmgr_gem->bufmgr.bo_get_subdata = drm_intel_gem_bo_get_subdata; bufmgr_gem->bufmgr.bo_wait_rendering = drm_intel_gem_bo_wait_rendering; + bufmgr_gem->bufmgr.bo_pad_to_size = drm_intel_gem_bo_pad_to_size; bufmgr_gem->bufmgr.bo_emit_reloc = drm_intel_gem_bo_emit_reloc; bufmgr_gem->bufmgr.bo_emit_reloc_fence = drm_intel_gem_bo_emit_reloc_fence; bufmgr_gem->bufmgr.bo_pin = drm_intel_gem_bo_pin;
intel/intel_bufmgr_priv.h +17 -0
@@ -152,6 +152,23 @@ struct _drm_intel_bufmgr {void (*destroy) (drm_intel_bufmgr *bufmgr);/** +* Sets buffer total padded size when buffer is used by the GPU. +* +* This enables dynamic padding to be added without using any backing +* storage. For example handling GPU padding requirements for buffers +* allocated by an entity unaware of the same. +* +* Set padded size remains active until reset (to zero or actual object +* size). +* +* Returns 0 on success or an error code. +* +* \param bo Buffer to set total padded size for +* \param pad_to_size Total size in bytes of object plus padding +*/ +int (*bo_pad_to_size) (drm_intel_bo *bo,uint64_t pad_to_size); + +/** * Add relocation entry in reloc_buf, which will be updated with the * target buffer's real offset on on command submission. *
Signature · 6 lines
2.3.2 _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
intel/intel_bufmgr_gem.c do_bo_emit_reloc(drm_intel_bo *bo, uint32_t offset, · line 1874
Quoted 10 lines · Tvrtko Ursulin
static int
intel/intel_bufmgr_gem.c +9
-0+drm_intel_gem_bo_pad_to_size(drm_intel_bo *bo, uint64_t pad_to_size) +{ +drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; + +if (pad_to_size < bo->size) +return -EINVAL ; + +if (pad_to_size > bo->size) + bo_gem->pad_to_size = pad_to_size;
Thinking too hard here :)
This fails to restore a previously padded object to bo->size, and we need to allow pad_to_size = 0.
static intdrm_intel_gem_bo_pad_to_size(drm_intel_bo *bo, uint64_t pad_to_size)
{
drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
if (pad_to_size && pad_to_size < bo->size)
return -EINVAL;
bo_gem->pad_to_size = pad_to_size;
return 0;
}-Chris
Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
RFC v3 1/1 libdrm_intel: Add API for execbuf pad to size functionality
In reply to Tvrtko UrsulinFrom: Tvrtko Ursulin <[email protected]>
New kernels add the ability to pad objects to specified size at execbuf time.
Add the drm_intel_bo_pad_to_size API via which this padded size can be set.
v2: Reject padded size smaller than object size. (Chris Wilson)
v3: Fixed silly mistake in v2. (Chris Wilson)
--- include/drm/i915_drm.h | 5 +++-- intel/intel_bufmgr.c | 6 ++++++ intel/intel_bufmgr.h | 2 ++ intel/intel_bufmgr_gem.c | 25 +++++++++++++++++++++++-- intel/intel_bufmgr_priv.h | 17 +++++++++++++++++ 5 files changed, 51 insertions(+), 4 deletions(-)
include/drm/i915_drm.h +3 -2
@@ -662,10 +662,11 @@ struct drm_i915_gem_exec_object2 {#define EXEC_OBJECT_NEEDS_FENCE (1<<0)#define EXEC_OBJECT_NEEDS_GTT (1<<1)#define EXEC_OBJECT_WRITE (1<<2) -#define __EXEC_OBJECT_UNKNOWN_FLAGS -(EXEC_OBJECT_WRITE<<1) +#define EXEC_OBJECT_PAD_TO_SIZE (1<<3) +#define __EXEC_OBJECT_UNKNOWN_FLAGS -(EXEC_OBJECT_PAD_TO_SIZE<<1)__u64 flags; -__u64 rsvd1; +__u64 pad_to_size;__u64 rsvd2; };
intel/intel_bufmgr.c +6 -0
@@ -299,6 +299,12 @@ drm_intel_bo_references(drm_intel_bo *bo, drm_intel_bo *target_bo) } drm_public int +drm_intel_bo_pad_to_size (drm_intel_bo *bo,uint64_t pad_to_size) +{ +return bo->bufmgr->bo_pad_to_size (bo, pad_to_size); +} + +drm_public intdrm_intel_get_pipe_from_crtc_id (drm_intel_bufmgr *bufmgr,int crtc_id) {if (bufmgr->get_pipe_from_crtc_id)
intel/intel_bufmgr.h +2 -0
@@ -165,6 +165,8 @@ int drm_intel_bo_disable_reuse(drm_intel_bo *bo);int drm_intel_bo_is_reusable (drm_intel_bo *bo);int drm_intel_bo_references (drm_intel_bo *bo,drm_intel_bo *target_bo); +int drm_intel_bo_pad_to_size (drm_intel_bo *bo,uint64_t pad_to_size); +/* drm_intel_bufmgr_gem.c */ drm_intel_bufmgr *drm_intel_bufmgr_gem_init (int fd,int batch_size); drm_intel_bo *drm_intel_bo_gem_create_from_name(drm_intel_bufmgr *bufmgr,
intel/intel_bufmgr_gem.c +23 -2
@@ -255,6 +255,12 @@ struct _drm_intel_bo_gem {drm_intel_aub_annotation *aub_annotations;unsigned aub_annotation_count; + +/** +* Size to pad the object to. +* +*/ +uint64_t pad_to_size; }; static unsigned int
@@ -504,8 +510,8 @@ drm_intel_add_validate_buffer2(drm_intel_bo *bo, int need_fence) bufmgr_gem->exec2_objects[index].alignment =0 ; bufmgr_gem->exec2_objects[index].offset =0 ; bufmgr_gem->exec_bos[index] = bo; - bufmgr_gem->exec2_objects[index].flags =0 ; - bufmgr_gem->exec2_objects[index].rsvd1 =0 ; + bufmgr_gem->exec2_objects[index].flags = bo_gem->pad_to_size ?EXEC_OBJECT_PAD_TO_SIZE :0 ; + bufmgr_gem->exec2_objects[index].pad_to_size = bo_gem->pad_to_size; bufmgr_gem->exec2_objects[index].rsvd2 =0 ;if (need_fence) { bufmgr_gem->exec2_objects[index].flags |=
@@ -1198,6 +1204,8 @@ drm_intel_gem_bo_unreference_final(drm_intel_bo *bo, time_t time)DBG ("bo_unreference final: %d (%s)\n" , bo_gem->gem_handle, bo_gem->name); + bo_gem->pad_to_size =0 ; +/* release memory associated with this object */ if (bo_gem->reloc_target_info) {free (bo_gem->reloc_target_info);
@@ -1855,6 +1863,18 @@ do_bo_emit_reloc(drm_intel_bo *bo, uint32_t offset, }static int +drm_intel_gem_bo_pad_to_size (drm_intel_bo *bo,uint64_t pad_to_size) +{ +drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; + +if (pad_to_size && pad_to_size < bo->size) +return -EINVAL ; + + bo_gem->pad_to_size = pad_to_size; +return 0 ; +} + +static int drm_intel_gem_bo_emit_reloc (drm_intel_bo *bo,uint32_t offset,drm_intel_bo *target_bo,uint32_t target_offset,uint32_t read_domains,uint32_t write_domain)
@@ -3603,6 +3623,7 @@ drm_intel_bufmgr_gem_init(int fd, int batch_size) bufmgr_gem->bufmgr.bo_subdata = drm_intel_gem_bo_subdata; bufmgr_gem->bufmgr.bo_get_subdata = drm_intel_gem_bo_get_subdata; bufmgr_gem->bufmgr.bo_wait_rendering = drm_intel_gem_bo_wait_rendering; + bufmgr_gem->bufmgr.bo_pad_to_size = drm_intel_gem_bo_pad_to_size; bufmgr_gem->bufmgr.bo_emit_reloc = drm_intel_gem_bo_emit_reloc; bufmgr_gem->bufmgr.bo_emit_reloc_fence = drm_intel_gem_bo_emit_reloc_fence; bufmgr_gem->bufmgr.bo_pin = drm_intel_gem_bo_pin;
intel/intel_bufmgr_priv.h +17 -0
@@ -152,6 +152,23 @@ struct _drm_intel_bufmgr {void (*destroy) (drm_intel_bufmgr *bufmgr);/** +* Sets buffer total padded size when buffer is used by the GPU. +* +* This enables dynamic padding to be added without using any backing +* storage. For example handling GPU padding requirements for buffers +* allocated by an entity unaware of the same. +* +* Set padded size remains active until reset (to zero or actual object +* size). +* +* Returns 0 on success or an error code. +* +* \param bo Buffer to set total padded size for +* \param pad_to_size Total size in bytes of object plus padding +*/ +int (*bo_pad_to_size) (drm_intel_bo *bo,uint64_t pad_to_size); + +/** * Add relocation entry in reloc_buf, which will be updated with the * target buffer's real offset on on command submission. *
Signature · 6 lines
2.3.2 _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 11 lines · Tvrtko Ursulin
From: Tvrtko Ursulin <[email protected]>
New kernels add the ability to pad objects to specified size at execbuf time.
Add the
drm_intel_bo_pad_to_sizeAPI via which this padded size can be set.v2: Reject padded size smaller than object size. (Chris Wilson)
v3: Fixed silly mistake in v2. (Chris Wilson)Signed-off-by Tvrtko Ursulin <[email protected]>Cc Chris Wilson <[email protected]>
-Chris
Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx