[Mesa-dev] [RFC PATCH] radeonsi: set a per-buffer flag that disables inter-process sharing (v2)
zhoucm1
david1.zhou at amd.com
Fri Jul 21 03:14:12 UTC 2017
On 2017年07月20日 22:59, Marek Olšák wrote:
>
>
> On Jul 19, 2017 10:21 PM, "zhoucm1" <david1.zhou at amd.com
> <mailto:david1.zhou at amd.com>> wrote:
>
>
>
> On 2017年07月19日 23:34, Marek Olšák wrote:
>>
>>
>> On Jul 19, 2017 3:36 AM, "zhoucm1" <david1.zhou at amd.com
>> <mailto:david1.zhou at amd.com>> wrote:
>>
>>
>>
>> On 2017年07月19日 04:08, Marek Olšák wrote:
>>
>> From: Marek Olšák <marek.olsak at amd.com
>> <mailto:marek.olsak at amd.com>>
>>
>> For lower overhead in the CS ioctl.
>> Winsys allocators are not used with interprocess-sharable
>> resources.
>>
>> Hi Marek,
>>
>> Could I know from how your this way reduces overhead in CS
>> ioctl? reusing BO to short bo list?
>>
>>
>> The kernel part of the work hasn't been done yet. The idea is
>> that nonsharable buffers don't have to be revalidated by TTM,
> OK, Maybe I only can see the whole picture of this idea when you
> complete kernel part.
> Out of curious, why/how can nonsharable buffers be revalidated by
> TTM without exposing like amdgpu_bo_make_resident api?
>
>
> I think the idea is that all nonsharable buffers will be backed by the
> same reservation object, so TTM can skip buffer validation if no
> buffer has been moved. It's just an optimization for the current design.
>
>
> With mentioned in another thread, if we can expose make_resident
> api, we can remove bo_list, even we can remove reservation
> operation in CS ioctl.
> And now, I think our bo list is a very bad design,
> first, umd must create bo list for every command submission, this
> is a extra cpu overhead compared with traditional way.
> second, kernel also have to iterate the list, when bo list is too
> long, like OpenCL program, they always throw several thousands BOs
> to bo list, reservation must keep these thousands ww_mutex safe,
> CPU overhead is too big.
>
> So I strongly suggest we should expose make_resident api to user
> space. if cannot, I want to know any specific reason to see if we
> can solve it.
>
>
> Yeah, I think the BO list idea is likely to die sooner or later. It
> made sense for GL before bindless was a thing. Nowadays I don't see
> much value in it.
>
> MesaGL will keep tracking the BO list because it's a requirement for
> good GL performance (it determines whether to flush IBs before BO
> synchronization, it allows tracking fences for each BO, which are used
> to determine dependencies between IBs, and that all allows async SDMA
> and async compute for GL, which doesn't have separate queues).
>
> However, we don't need any BO list at the libdrm level and lower. I
> think a BO_CREATE flag that causes that the buffer is added to a
> kernel-side per-fd BO list would be sufficient. How the kernel manages
> its BO list should be its own implementation detail. Initially we can
> just move the current BO list management into the kernel.
I guess this idea will make bo list worse, which just decrease umd
effort, but increase kernel driver complication.
First, from your and Christian's comments, we can get this agreement
that bo list design is not a good way.
My proposal of exposing amdgpu_bo_make_resident is to replace bo list.
If we can make all needed bo resident, then we don't need to validate it
again in cs ioctl, then we don't need their reservation lock more. After
job pushed to scheduler, then we can un-resident BOs.
Even we can make it for VM bo, then we don't need to check vm update
again while done in va map ioctl.
If this is got done(eviction has been improved more), I cannot see any
obvious gap for performance.
What do you think of this proposal of exposing amdgpu_bo_make_resident
api to user space? Or any other idea we can discuss.
If you all agree with, I can volunteer to try with UMD guys.
Regards,
David Zhou
>
> Marek
>
>
>
>
>
> Regards,
> David Zhou
>
>> so it should remove a lot of kernel overhead and the BO list
>> remains the same.
>>
>> Marek
>>
>>
>>
>> Thanks,
>> David Zhou
>>
>>
>> v2: It shouldn't crash anymore, but the kernel will
>> reject the new flag.
>> ---
>> src/gallium/drivers/radeon/r600_buffer_common.c | 7 +++++
>> src/gallium/drivers/radeon/radeon_winsys.h | 20
>> +++++++++++---
>> src/gallium/winsys/amdgpu/drm/amdgpu_bo.c | 36
>> ++++++++++++++++---------
>> src/gallium/winsys/radeon/drm/radeon_drm_bo.c | 27
>> +++++++++++--------
>> 4 files changed, 62 insertions(+), 28 deletions(-)
>>
>> diff --git
>> a/src/gallium/drivers/radeon/r600_buffer_common.c
>> b/src/gallium/drivers/radeon/r600_buffer_common.c
>> index dd1c209..2747ac4 100644
>> --- a/src/gallium/drivers/radeon/r600_buffer_common.c
>> +++ b/src/gallium/drivers/radeon/r600_buffer_common.c
>> @@ -160,20 +160,27 @@ void
>> r600_init_resource_fields(struct r600_common_screen *rscreen,
>> }
>> /* Tiled textures are unmappable. Always put them
>> in VRAM. */
>> if ((res->b.b.target != PIPE_BUFFER &&
>> !rtex->surface.is_linear) ||
>> res->flags & R600_RESOURCE_FLAG_UNMAPPABLE) {
>> res->domains = RADEON_DOMAIN_VRAM;
>> res->flags |= RADEON_FLAG_NO_CPU_ACCESS |
>> RADEON_FLAG_GTT_WC;
>> }
>> + /* Only displayable single-sample textures can be
>> shared between
>> + * processes. */
>> + if (res->b.b.target == PIPE_BUFFER ||
>> + res->b.b.nr_samples >= 2 ||
>> + rtex->surface.micro_tile_mode !=
>> RADEON_MICRO_MODE_DISPLAY)
>> + res->flags |=
>> RADEON_FLAG_NO_INTERPROCESS_SHARING;
>> +
>> /* If VRAM is just stolen system memory, allow
>> both VRAM and
>> * GTT, whichever has free space. If a buffer is
>> evicted from
>> * VRAM to GTT, it will stay there.
>> *
>> * DRM 3.6.0 has good BO move throttling, so we
>> can allow VRAM-only
>> * placements even with a low amount of stolen VRAM.
>> */
>> if (!rscreen->info.has_dedicated_vram &&
>> (rscreen->info.drm_major < 3 || rscreen->info.drm_minor <
>> 6) &&
>> res->domains == RADEON_DOMAIN_VRAM) {
>> diff --git a/src/gallium/drivers/radeon/radeon_winsys.h
>> b/src/gallium/drivers/radeon/radeon_winsys.h
>> index 351edcd..0abcb56 100644
>> --- a/src/gallium/drivers/radeon/radeon_winsys.h
>> +++ b/src/gallium/drivers/radeon/radeon_winsys.h
>> @@ -47,20 +47,21 @@ enum radeon_bo_domain { /* bitfield */
>> RADEON_DOMAIN_GTT = 2,
>> RADEON_DOMAIN_VRAM = 4,
>> RADEON_DOMAIN_VRAM_GTT = RADEON_DOMAIN_VRAM |
>> RADEON_DOMAIN_GTT
>> };
>> enum radeon_bo_flag { /* bitfield */
>> RADEON_FLAG_GTT_WC = (1 << 0),
>> RADEON_FLAG_NO_CPU_ACCESS = (1 << 1),
>> RADEON_FLAG_NO_SUBALLOC = (1 << 2),
>> RADEON_FLAG_SPARSE = (1 << 3),
>> + RADEON_FLAG_NO_INTERPROCESS_SHARING = (1 << 4),
>> };
>> enum radeon_bo_usage { /* bitfield */
>> RADEON_USAGE_READ = 2,
>> RADEON_USAGE_WRITE = 4,
>> RADEON_USAGE_READWRITE = RADEON_USAGE_READ |
>> RADEON_USAGE_WRITE,
>> /* The winsys ensures that the CS submission will
>> be scheduled after
>> * previously flushed CSs referencing this BO in a
>> conflicting way.
>> */
>> @@ -685,28 +686,33 @@ static inline enum radeon_bo_domain
>> radeon_domain_from_heap(enum radeon_heap hea
>> default:
>> assert(0);
>> return (enum radeon_bo_domain)0;
>> }
>> }
>> static inline unsigned radeon_flags_from_heap(enum
>> radeon_heap heap)
>> {
>> switch (heap) {
>> case RADEON_HEAP_VRAM_NO_CPU_ACCESS:
>> - return RADEON_FLAG_GTT_WC |
>> RADEON_FLAG_NO_CPU_ACCESS;
>> + return RADEON_FLAG_GTT_WC |
>> + RADEON_FLAG_NO_CPU_ACCESS |
>> + RADEON_FLAG_NO_INTERPROCESS_SHARING;
>> +
>> case RADEON_HEAP_VRAM:
>> case RADEON_HEAP_VRAM_GTT:
>> case RADEON_HEAP_GTT_WC:
>> - return RADEON_FLAG_GTT_WC;
>> + return RADEON_FLAG_GTT_WC |
>> + RADEON_FLAG_NO_INTERPROCESS_SHARING;
>> +
>> case RADEON_HEAP_GTT:
>> default:
>> - return 0;
>> + return RADEON_FLAG_NO_INTERPROCESS_SHARING;
>> }
>> }
>> /* The pb cache bucket is chosen to minimize pb_cache
>> misses.
>> * It must be between 0 and 3 inclusive.
>> */
>> static inline unsigned
>> radeon_get_pb_cache_bucket_index(enum radeon_heap heap)
>> {
>> switch (heap) {
>> case RADEON_HEAP_VRAM_NO_CPU_ACCESS:
>> @@ -724,22 +730,28 @@ static inline unsigned
>> radeon_get_pb_cache_bucket_index(enum radeon_heap heap)
>> /* Return the heap index for winsys allocators, or -1
>> on failure. */
>> static inline int radeon_get_heap_index(enum
>> radeon_bo_domain domain,
>> enum radeon_bo_flag flags)
>> {
>> /* VRAM implies WC (write combining) */
>> assert(!(domain & RADEON_DOMAIN_VRAM) || flags &
>> RADEON_FLAG_GTT_WC);
>> /* NO_CPU_ACCESS implies VRAM only. */
>> assert(!(flags & RADEON_FLAG_NO_CPU_ACCESS) ||
>> domain == RADEON_DOMAIN_VRAM);
>> + /* Resources with interprocess sharing don't use
>> any winsys allocators. */
>> + if (!(flags & RADEON_FLAG_NO_INTERPROCESS_SHARING))
>> + return -1;
>> +
>> /* Unsupported flags: NO_SUBALLOC, SPARSE. */
>> - if (flags & ~(RADEON_FLAG_GTT_WC |
>> RADEON_FLAG_NO_CPU_ACCESS))
>> + if (flags & ~(RADEON_FLAG_GTT_WC |
>> + RADEON_FLAG_NO_CPU_ACCESS |
>> + RADEON_FLAG_NO_INTERPROCESS_SHARING))
>> return -1;
>> switch (domain) {
>> case RADEON_DOMAIN_VRAM:
>> if (flags & RADEON_FLAG_NO_CPU_ACCESS)
>> return RADEON_HEAP_VRAM_NO_CPU_ACCESS;
>> else
>> return RADEON_HEAP_VRAM;
>> case RADEON_DOMAIN_VRAM_GTT:
>> return RADEON_HEAP_VRAM_GTT;
>> diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
>> b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
>> index 97bbe23..06b8198 100644
>> --- a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
>> +++ b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
>> @@ -31,20 +31,24 @@
>> #include "amdgpu_cs.h"
>> #include "os/os_time.h"
>> #include "state_tracker/drm_driver.h"
>> #include <amdgpu_drm.h>
>> #include <xf86drm.h>
>> #include <stdio.h>
>> #include <inttypes.h>
>> +#ifndef AMDGPU_GEM_CREATE_NO_INTERPROCESS_SHARING
>> +#define AMDGPU_GEM_CREATE_NO_INTERPROCESS_SHARING (1 << 6)
>> +#endif
>> +
>> /* Set to 1 for verbose output showing committed sparse
>> buffer ranges. */
>> #define DEBUG_SPARSE_COMMITS 0
>> struct amdgpu_sparse_backing_chunk {
>> uint32_t begin, end;
>> };
>> static struct pb_buffer *
>> amdgpu_bo_create(struct radeon_winsys *rws,
>> uint64_t size,
>> @@ -395,20 +399,22 @@ static struct amdgpu_winsys_bo
>> *amdgpu_create_bo(struct amdgpu_winsys *ws,
>> if (initial_domain & RADEON_DOMAIN_VRAM)
>> request.preferred_heap |= AMDGPU_GEM_DOMAIN_VRAM;
>> if (initial_domain & RADEON_DOMAIN_GTT)
>> request.preferred_heap |= AMDGPU_GEM_DOMAIN_GTT;
>> if (flags & RADEON_FLAG_NO_CPU_ACCESS)
>> request.flags |= AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
>> if (flags & RADEON_FLAG_GTT_WC)
>> request.flags |= AMDGPU_GEM_CREATE_CPU_GTT_USWC;
>> + if (flags & RADEON_FLAG_NO_INTERPROCESS_SHARING)
>> + request.flags |=
>> AMDGPU_GEM_CREATE_NO_INTERPROCESS_SHARING;
>> r = amdgpu_bo_alloc(ws->dev, &request, &buf_handle);
>> if (r) {
>> fprintf(stderr, "amdgpu: Failed to allocate a
>> buffer:\n");
>> fprintf(stderr, "amdgpu: size : %"PRIu64"
>> bytes\n", size);
>> fprintf(stderr, "amdgpu: alignment : %u bytes\n",
>> alignment);
>> fprintf(stderr, "amdgpu: domains : %u\n",
>> initial_domain);
>> goto error_bo_alloc;
>> }
>> @@ -1127,21 +1133,21 @@ static void
>> amdgpu_buffer_set_metadata(struct pb_buffer *_buf,
>> static struct pb_buffer *
>> amdgpu_bo_create(struct radeon_winsys *rws,
>> uint64_t size,
>> unsigned alignment,
>> enum radeon_bo_domain domain,
>> enum radeon_bo_flag flags)
>> {
>> struct amdgpu_winsys *ws = amdgpu_winsys(rws);
>> struct amdgpu_winsys_bo *bo;
>> - unsigned usage = 0, pb_cache_bucket;
>> + unsigned usage = 0, pb_cache_bucket = 0;
>> /* VRAM implies WC. This is not optional. */
>> assert(!(domain & RADEON_DOMAIN_VRAM) || flags &
>> RADEON_FLAG_GTT_WC);
>> /* NO_CPU_ACCESS is valid with VRAM only. */
>> assert(domain == RADEON_DOMAIN_VRAM || !(flags &
>> RADEON_FLAG_NO_CPU_ACCESS));
>> /* Sub-allocate small buffers from slabs. */
>> if (!(flags & (RADEON_FLAG_NO_SUBALLOC |
>> RADEON_FLAG_SPARSE)) &&
>> size <= (1 << AMDGPU_SLAB_MAX_SIZE_LOG2) &&
>> @@ -1182,48 +1188,52 @@ no_slab:
>> /* This flag is irrelevant for the cache. */
>> flags &= ~RADEON_FLAG_NO_SUBALLOC;
>> /* Align size to page size. This is the minimum
>> alignment for normal
>> * BOs. Aligning this here helps the cached bufmgr.
>> Especially small BOs,
>> * like constant/uniform buffers, can benefit from
>> better and more reuse.
>> */
>> size = align64(size, ws->info.gart_page_size);
>> alignment = align(alignment, ws->info.gart_page_size);
>> - int heap = radeon_get_heap_index(domain, flags);
>> - assert(heap >= 0 && heap < RADEON_MAX_CACHED_HEAPS);
>> - usage = 1 << heap; /* Only set one usage bit for each
>> heap. */
>> + bool use_reusable_pool = flags &
>> RADEON_FLAG_NO_INTERPROCESS_SHARING;
>> - pb_cache_bucket =
>> radeon_get_pb_cache_bucket_index(heap);
>> - assert(pb_cache_bucket <
>> ARRAY_SIZE(ws->bo_cache.buckets));
>> + if (use_reusable_pool) {
>> + int heap = radeon_get_heap_index(domain, flags);
>> + assert(heap >= 0 && heap < RADEON_MAX_CACHED_HEAPS);
>> + usage = 1 << heap; /* Only set one usage bit for
>> each heap. */
>> - /* Get a buffer from the cache. */
>> - bo = (struct amdgpu_winsys_bo*)
>> - pb_cache_reclaim_buffer(&ws->bo_cache, size, alignment,
>> usage,
>> - pb_cache_bucket);
>> - if (bo)
>> - return &bo->base;
>> + pb_cache_bucket =
>> radeon_get_pb_cache_bucket_index(heap);
>> + assert(pb_cache_bucket <
>> ARRAY_SIZE(ws->bo_cache.buckets));
>> +
>> + /* Get a buffer from the cache. */
>> + bo = (struct amdgpu_winsys_bo*)
>> + pb_cache_reclaim_buffer(&ws->bo_cache, size, alignment,
>> usage,
>> + pb_cache_bucket);
>> + if (bo)
>> + return &bo->base;
>> + }
>> /* Create a new one. */
>> bo = amdgpu_create_bo(ws, size, alignment, usage,
>> domain, flags,
>> pb_cache_bucket);
>> if (!bo) {
>> /* Clear the cache and try again. */
>> pb_slabs_reclaim(&ws->bo_slabs);
>> pb_cache_release_all_buffers(&ws->bo_cache);
>> bo = amdgpu_create_bo(ws, size, alignment, usage,
>> domain, flags,
>> pb_cache_bucket);
>> if (!bo)
>> return NULL;
>> }
>> - bo->u.real.use_reusable_pool = true;
>> + bo->u.real.use_reusable_pool = use_reusable_pool;
>> return &bo->base;
>> }
>> static struct pb_buffer *amdgpu_bo_from_handle(struct
>> radeon_winsys *rws,
>> struct winsys_handle *whandle,
>> unsigned *stride,
>> unsigned *offset)
>> {
>> struct amdgpu_winsys *ws = amdgpu_winsys(rws);
>> struct amdgpu_winsys_bo *bo;
>> diff --git
>> a/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
>> b/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
>> index 8027a5f..15e9d38 100644
>> --- a/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
>> +++ b/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
>> @@ -907,21 +907,21 @@ static void
>> radeon_bo_set_metadata(struct pb_buffer *_buf,
>> static struct pb_buffer *
>> radeon_winsys_bo_create(struct radeon_winsys *rws,
>> uint64_t size,
>> unsigned alignment,
>> enum radeon_bo_domain domain,
>> enum radeon_bo_flag flags)
>> {
>> struct radeon_drm_winsys *ws = radeon_drm_winsys(rws);
>> struct radeon_bo *bo;
>> - unsigned usage = 0, pb_cache_bucket;
>> + unsigned usage = 0, pb_cache_bucket = 0;
>> assert(!(flags & RADEON_FLAG_SPARSE)); /* not
>> supported */
>> /* Only 32-bit sizes are supported. */
>> if (size > UINT_MAX)
>> return NULL;
>> /* VRAM implies WC. This is not optional. */
>> if (domain & RADEON_DOMAIN_VRAM)
>> flags |= RADEON_FLAG_GTT_WC;
>> @@ -962,46 +962,51 @@ no_slab:
>> /* This flag is irrelevant for the cache. */
>> flags &= ~RADEON_FLAG_NO_SUBALLOC;
>> /* Align size to page size. This is the minimum
>> alignment for normal
>> * BOs. Aligning this here helps the cached bufmgr.
>> Especially small BOs,
>> * like constant/uniform buffers, can benefit from
>> better and more reuse.
>> */
>> size = align(size, ws->info.gart_page_size);
>> alignment = align(alignment, ws->info.gart_page_size);
>> - int heap = radeon_get_heap_index(domain, flags);
>> - assert(heap >= 0 && heap < RADEON_MAX_CACHED_HEAPS);
>> - usage = 1 << heap; /* Only set one usage bit for
>> each heap. */
>> + bool use_reusable_pool = flags &
>> RADEON_FLAG_NO_INTERPROCESS_SHARING;
>> - pb_cache_bucket =
>> radeon_get_pb_cache_bucket_index(heap);
>> - assert(pb_cache_bucket <
>> ARRAY_SIZE(ws->bo_cache.buckets));
>> + /* Shared resources don't use cached heaps. */
>> + if (use_reusable_pool) {
>> + int heap = radeon_get_heap_index(domain, flags);
>> + assert(heap >= 0 && heap < RADEON_MAX_CACHED_HEAPS);
>> + usage = 1 << heap; /* Only set one usage bit for
>> each heap. */
>> - bo =
>> radeon_bo(pb_cache_reclaim_buffer(&ws->bo_cache, size,
>> alignment,
>> - usage, pb_cache_bucket));
>> - if (bo)
>> - return &bo->base;
>> + pb_cache_bucket =
>> radeon_get_pb_cache_bucket_index(heap);
>> + assert(pb_cache_bucket <
>> ARRAY_SIZE(ws->bo_cache.buckets));
>> +
>> + bo =
>> radeon_bo(pb_cache_reclaim_buffer(&ws->bo_cache, size,
>> alignment,
>> + usage, pb_cache_bucket));
>> + if (bo)
>> + return &bo->base;
>> + }
>> bo = radeon_create_bo(ws, size, alignment, usage,
>> domain, flags,
>> pb_cache_bucket);
>> if (!bo) {
>> /* Clear the cache and try again. */
>> if (ws->info.has_virtual_memory)
>> pb_slabs_reclaim(&ws->bo_slabs);
>> pb_cache_release_all_buffers(&ws->bo_cache);
>> bo = radeon_create_bo(ws, size, alignment,
>> usage, domain, flags,
>> pb_cache_bucket);
>> if (!bo)
>> return NULL;
>> }
>> - bo->u.real.use_reusable_pool = true;
>> + bo->u.real.use_reusable_pool = use_reusable_pool;
>> mtx_lock(&ws->bo_handles_mutex);
>> util_hash_table_set(ws->bo_handles,
>> (void*)(uintptr_t)bo->handle, bo);
>> mtx_unlock(&ws->bo_handles_mutex);
>> return &bo->base;
>> }
>> static struct pb_buffer
>> *radeon_winsys_bo_from_ptr(struct radeon_winsys *rws,
>> void *pointer, uint64_t size)
>>
>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20170721/ecd522b4/attachment-0001.html>
More information about the amd-gfx
mailing list