[PATCH] drm/amdgpu: reserve GDS resources on the gfx ring for gfx10
Koenig, Christian
Christian.Koenig at amd.com
Wed Jul 17 15:21:39 UTC 2019
Am 17.07.19 um 16:54 schrieb Marek Olšák:
On Wed., Jul. 17, 2019, 10:43 Christian König, <ckoenig.leichtzumerken at gmail.com<mailto:ckoenig.leichtzumerken at gmail.com>> wrote:
Am 17.07.19 um 16:27 schrieb Marek Olšák:
On Wed., Jul. 17, 2019, 03:15 Christian König, <ckoenig.leichtzumerken at gmail.com<mailto:ckoenig.leichtzumerken at gmail.com>> wrote:
Am 17.07.19 um 02:06 schrieb Marek Olšák:
> From: Marek Olšák <marek.olsak at amd.com<mailto:marek.olsak at amd.com>>
>
> Hopefully we'll only use 1 gfx ring, because otherwise we'd have to have
> separate GDS buffers for each gfx ring.
>
> This is a workaround to ensure stability of transform feedback. Shaders hang
> waiting for a GDS instruction (ds_sub, not ds_ordered_count).
>
> The disadvantage is that compute IBs might get a different VMID,
> because now gfx always has GDS and compute doesn't.
So far compute is only using GWS, but I don't think that those
reservations are a good idea in general.
How severe is the ENOMEM problem you see with using an explicit GWS
allocation?
I guess you mean GDS or OA.
Yeah, just a typo. Compute is using GWS and we want to use GDS and OA here.
There is no ENOMEM, it just hangs. I don't know why. The shader is waiting for ds_sub and can't continue, but GDS is idle.
Well could it be because we don't correctly handle non zero offsets or stuff like that?
I don't know what you mean.
I wanted to try to leak some GDS/OA during allocation to narrow down what is going wrong here.
Does it work with this hack when you don't allocate GDS/OA from the start? (Just allocate it twice or something like this).
It's only allocated once by the kernel with this hack.
Well my question is why does that actually help?
I mean is it a bug in the memory management or some hardware issue that we can't switch GDS/OS?
I think we need to find the root cause of the problem and if it's related to switching GDS/OS what the actual restrictions are.
Christian.
Marek
Christian.
Marek
Regards,
Christian.
>
> Signed-off-by: Marek Olšák <marek.olsak at amd.com<mailto:marek.olsak at amd.com>>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu.h | 1 +
> drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 10 ++++++++++
> drivers/gpu/drm/amd/amdgpu/amdgpu_gds.h | 6 ++++++
> drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 20 ++++++++++++++++++++
> 4 files changed, 37 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> index 4b514a44184c..cbd55d061b72 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> @@ -456,20 +456,21 @@ struct amdgpu_cs_parser {
> struct drm_file *filp;
> struct amdgpu_ctx *ctx;
>
> /* chunks */
> unsigned nchunks;
> struct amdgpu_cs_chunk *chunks;
>
> /* scheduler job object */
> struct amdgpu_job *job;
> struct drm_sched_entity *entity;
> + unsigned hw_ip;
>
> /* buffer objects */
> struct ww_acquire_ctx ticket;
> struct amdgpu_bo_list *bo_list;
> struct amdgpu_mn *mn;
> struct amdgpu_bo_list_entry vm_pd;
> struct list_head validated;
> struct dma_fence *fence;
> uint64_t bytes_moved_threshold;
> uint64_t bytes_moved_vis_threshold;
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
> index c691df6f7a57..9125cd69a124 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
> @@ -678,20 +678,28 @@ static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
> if (r)
> goto error_validate;
>
> amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
> p->bytes_moved_vis);
>
> gds = p->bo_list->gds_obj;
> gws = p->bo_list->gws_obj;
> oa = p->bo_list->oa_obj;
>
> + if (p->hw_ip == AMDGPU_HW_IP_GFX) {
> + /* Only gfx10 allocates these. */
> + if (!gds)
> + gds = p->adev->gds.gds_gfx_bo;
> + if (!oa)
> + oa = p->adev->gds.oa_gfx_bo;
> + }
> +
> amdgpu_bo_list_for_each_entry(e, p->bo_list) {
> struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo<http://tv.bo>);
>
> /* Make sure we use the exclusive slot for shared BOs */
> if (bo->prime_shared_count)
> e->tv.num_shared = 0;
> e->bo_va = amdgpu_vm_bo_find(vm, bo);
> }
>
> if (gds) {
> @@ -954,20 +962,22 @@ static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
> struct drm_amdgpu_cs_chunk_ib *chunk_ib;
> struct drm_sched_entity *entity;
>
> chunk = &parser->chunks[i];
> ib = &parser->job->ibs[j];
> chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
>
> if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
> continue;
>
> + parser->hw_ip = chunk_ib->ip_type;
> +
> if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
> (amdgpu_mcbp || amdgpu_sriov_vf(adev))) {
> if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
> if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
> ce_preempt++;
> else
> de_preempt++;
> }
>
> /* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gds.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_gds.h
> index df8a23554831..0943b8e1d97e 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gds.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gds.h
> @@ -26,20 +26,26 @@
>
> struct amdgpu_ring;
> struct amdgpu_bo;
>
> struct amdgpu_gds {
> uint32_t gds_size;
> uint32_t gws_size;
> uint32_t oa_size;
> uint32_t gds_compute_max_wave_id;
> uint32_t vgt_gs_max_wave_id;
> +
> + /* Reserved OA for the gfx ring. (gfx10) */
> + uint32_t gds_gfx_reservation_size;
> + uint32_t oa_gfx_reservation_size;
> + struct amdgpu_bo *gds_gfx_bo;
> + struct amdgpu_bo *oa_gfx_bo;
> };
>
> struct amdgpu_gds_reg_offset {
> uint32_t mem_base;
> uint32_t mem_size;
> uint32_t gws;
> uint32_t oa;
> };
>
> #endif /* __AMDGPU_GDS_H__ */
> diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c
> index 618291df659b..3952754c04ff 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c
> @@ -1314,20 +1314,33 @@ static int gfx_v10_0_sw_init(void *handle)
>
> kiq = &adev->gfx.kiq;
> r = amdgpu_gfx_kiq_init_ring(adev, &kiq->ring, &kiq->irq);
> if (r)
> return r;
>
> r = amdgpu_gfx_mqd_sw_init(adev, sizeof(struct v10_compute_mqd));
> if (r)
> return r;
>
> + /* allocate reserved GDS resources for transform feedback */
> + r = amdgpu_bo_create_kernel(adev, adev->gds.gds_gfx_reservation_size,
> + 4, AMDGPU_GEM_DOMAIN_GDS,
> + &adev->gds.gds_gfx_bo, NULL, NULL);
> + if (r)
> + return r;
> +
> + r = amdgpu_bo_create_kernel(adev, adev->gds.oa_gfx_reservation_size,
> + 1, AMDGPU_GEM_DOMAIN_OA,
> + &adev->gds.oa_gfx_bo, NULL, NULL);
> + if (r)
> + return r;
> +
> /* allocate visible FB for rlc auto-loading fw */
> if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) {
> r = gfx_v10_0_rlc_backdoor_autoload_buffer_init(adev);
> if (r)
> return r;
> }
>
> adev->gfx.ce_ram_size = F32_CE_PROGRAM_RAM_SIZE;
>
> gfx_v10_0_gpu_early_init(adev);
> @@ -1354,20 +1367,23 @@ static void gfx_v10_0_me_fini(struct amdgpu_device *adev)
> amdgpu_bo_free_kernel(&adev->gfx.me.me_fw_obj,
> &adev->gfx.me.me_fw_gpu_addr,
> (void **)&adev->gfx.me.me_fw_ptr);
> }
>
> static int gfx_v10_0_sw_fini(void *handle)
> {
> int i;
> struct amdgpu_device *adev = (struct amdgpu_device *)handle;
>
> + amdgpu_bo_free_kernel(&adev->gds.gds_gfx_bo, NULL, NULL);
> + amdgpu_bo_free_kernel(&adev->gds.oa_gfx_bo, NULL, NULL);
> +
> for (i = 0; i < adev->gfx.num_gfx_rings; i++)
> amdgpu_ring_fini(&adev->gfx.gfx_ring[i]);
> for (i = 0; i < adev->gfx.num_compute_rings; i++)
> amdgpu_ring_fini(&adev->gfx.compute_ring[i]);
>
> amdgpu_gfx_mqd_sw_fini(adev);
> amdgpu_gfx_kiq_free_ring(&adev->gfx.kiq.ring, &adev->gfx.kiq.irq);
> amdgpu_gfx_kiq_fini(adev);
>
> gfx_v10_0_pfp_fini(adev);
> @@ -5181,20 +5197,24 @@ static void gfx_v10_0_set_gds_init(struct amdgpu_device *adev)
> case CHIP_NAVI10:
> default:
> adev->gds.gds_size = 0x10000;
> adev->gds.gds_compute_max_wave_id = 0x4ff;
> adev->gds.vgt_gs_max_wave_id = 0x3ff;
> break;
> }
>
> adev->gds.gws_size = 64;
> adev->gds.oa_size = 16;
> +
> + /* Reserved for transform feedback. */
> + adev->gds.gds_gfx_reservation_size = 256;
> + adev->gds.oa_gfx_reservation_size = 4;
> }
>
> static void gfx_v10_0_set_user_wgp_inactive_bitmap_per_sh(struct amdgpu_device *adev,
> u32 bitmap)
> {
> u32 data;
>
> if (!bitmap)
> return;
>
_______________________________________________
amd-gfx mailing list
amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org>
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20190717/1afa06b6/attachment-0001.html>
More information about the amd-gfx
mailing list