[Nouveau] [kernel PATCH] drm/nouveau: add NOUVEAU_GEM_RELOC_HILO

[Francisco Jerez]

Marcin Slusarz <marcin.slusarz at gmail.com> writes:

> add new relocation flag which allows to halving number of relocation
> entries on >= nv50 cards
>
I suspect that this change is unnecessary. If you're actually seeing a
performance improvement (I guess you are?), it's very likely to be
masking the actual root of the problem -- last time I checked the nv50
gallium pipe driver was rather careless about reloc re-submission (it
re-submits the whole reloc chain once for every draw call instead of
once for every pushbuf flush), and nouveau_reloc_emit() and friends are
very unoptimized.

I'm betting that using HILO or not will make pretty much no difference
once those two places are fixed :).

> Signed-off-by: Marcin Slusarz <marcin.slusarz at gmail.com>
> ---
>  drivers/gpu/drm/nouveau/nouveau_gem.c |   22 +++++++++++++++-------
>  include/drm/nouveau_drm.h             |    1 +
>  2 files changed, 16 insertions(+), 7 deletions(-)
>
> diff --git a/drivers/gpu/drm/nouveau/nouveau_gem.c b/drivers/gpu/drm/nouveau/nouveau_gem.c
> index 9a1fdcf..2e21412 100644
> --- a/drivers/gpu/drm/nouveau/nouveau_gem.c
> +++ b/drivers/gpu/drm/nouveau/nouveau_gem.c
> @@ -507,6 +507,7 @@ nouveau_gem_pushbuf_reloc_apply(struct drm_device *dev,
>  		struct drm_nouveau_gem_pushbuf_bo *b;
>  		struct nouveau_bo *nvbo;
>  		uint32_t data;
> +		uint32_t data2;
>  
>  		if (unlikely(r->bo_index > req->nr_buffers)) {
>  			NV_ERROR(dev, "reloc bo index invalid\n");
> @@ -544,17 +545,22 @@ nouveau_gem_pushbuf_reloc_apply(struct drm_device *dev,
>  
>  		if (r->flags & NOUVEAU_GEM_RELOC_LOW)
>  			data = b->presumed.offset + r->data;
> -		else
> -		if (r->flags & NOUVEAU_GEM_RELOC_HIGH)
> +		else if (r->flags & NOUVEAU_GEM_RELOC_HIGH)
>  			data = (b->presumed.offset + r->data) >> 32;
> -		else
> +		else if (r->flags & NOUVEAU_GEM_RELOC_HILO) {
> +			data = (b->presumed.offset + r->data) >> 32;
> +			data2 = b->presumed.offset + r->data;
> +		} else
>  			data = r->data;
>  
>  		if (r->flags & NOUVEAU_GEM_RELOC_OR) {
> -			if (b->presumed.domain == NOUVEAU_GEM_DOMAIN_GART)
> -				data |= r->tor;
> -			else
> -				data |= r->vor;
> +			if (b->presumed.domain == NOUVEAU_GEM_DOMAIN_GART) {
> +				data  |= r->tor;
> +				data2 |= r->tor;
> +			} else {
> +				data  |= r->vor;
> +				data2 |= r->vor;
> +			}
>  		}
>  
>  		spin_lock(&nvbo->bo.lock);
> @@ -566,6 +572,8 @@ nouveau_gem_pushbuf_reloc_apply(struct drm_device *dev,
>  		}
>  
>  		nouveau_bo_wr32(nvbo, r->reloc_bo_offset >> 2, data);
> +		if (r->flags & NOUVEAU_GEM_RELOC_HILO)
> +			nouveau_bo_wr32(nvbo, (r->reloc_bo_offset >> 2) + 1, data2);
>  	}
>  
>  	kfree(reloc);
> diff --git a/include/drm/nouveau_drm.h b/include/drm/nouveau_drm.h
> index bc5590b..126ea8a 100644
> --- a/include/drm/nouveau_drm.h
> +++ b/include/drm/nouveau_drm.h
> @@ -137,6 +137,7 @@ struct drm_nouveau_gem_pushbuf_bo {
>  #define NOUVEAU_GEM_RELOC_LOW  (1 << 0)
>  #define NOUVEAU_GEM_RELOC_HIGH (1 << 1)
>  #define NOUVEAU_GEM_RELOC_OR   (1 << 2)
> +#define NOUVEAU_GEM_RELOC_HILO (1 << 3)
>  #define NOUVEAU_GEM_MAX_RELOCS 1024
>  struct drm_nouveau_gem_pushbuf_reloc {
>  	uint32_t reloc_bo_index;
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 229 bytes
Desc: not available
URL: <http://lists.freedesktop.org/archives/nouveau/attachments/20110112/50b0fd07/attachment.pgp>