[Mesa-dev] [PATCH] etnaviv: keep track of buffer valid ranges

[Lucas Stach]

Am Donnerstag, den 19.10.2017, 07:59 +0200 schrieb Christian Gmeiner:
> This allows a write to proceed to an uninitialized part of a buffer
> even when the GPU is using the previously-initialized portions.
> Same is done for freedreno, nouveau and radeon.
> 
> > Signed-off-by: Christian Gmeiner <christian.gmeiner at gmail.com>
> ---
>  src/gallium/drivers/etnaviv/etnaviv_resource.c |  3 +++
>  src/gallium/drivers/etnaviv/etnaviv_resource.h |  4 ++++
>  src/gallium/drivers/etnaviv/etnaviv_transfer.c | 22 +++++++++++++++++++---
>  3 files changed, 26 insertions(+), 3 deletions(-)
> 
> diff --git a/src/gallium/drivers/etnaviv/etnaviv_resource.c b/src/gallium/drivers/etnaviv/etnaviv_resource.c
> index d6cccd2dbb..4eb49bf936 100644
> --- a/src/gallium/drivers/etnaviv/etnaviv_resource.c
> +++ b/src/gallium/drivers/etnaviv/etnaviv_resource.c
> @@ -268,6 +268,7 @@ etna_resource_alloc(struct pipe_screen *pscreen, unsigned layout,
>  
>     pipe_reference_init(&rsc->base.reference, 1);
>     list_inithead(&rsc->list);
> +   util_range_init(&rsc->valid_buffer_range);
>  
>     size = setup_miptree(rsc, paddingX, paddingY, msaa_xscale, msaa_yscale);
>  
> @@ -458,6 +459,7 @@ etna_resource_destroy(struct pipe_screen *pscreen, struct pipe_resource *prsc)
>     if (rsc->scanout)
>        renderonly_scanout_destroy(rsc->scanout, etna_screen(pscreen)->ro);
>  
> +   util_range_destroy(&rsc->valid_buffer_range);
>     list_delinit(&rsc->list);
>  
>     pipe_resource_reference(&rsc->texture, NULL);
> @@ -494,6 +496,7 @@ etna_resource_from_handle(struct pipe_screen *pscreen,
>  
>     pipe_reference_init(&prsc->reference, 1);
>     list_inithead(&rsc->list);
> +   util_range_init(&rsc->valid_buffer_range);
>     prsc->screen = pscreen;
>  
>     rsc->bo = etna_screen_bo_from_handle(pscreen, handle, &level->stride);
> diff --git a/src/gallium/drivers/etnaviv/etnaviv_resource.h b/src/gallium/drivers/etnaviv/etnaviv_resource.h
> index 0b135e2373..705c1d1af6 100644
> --- a/src/gallium/drivers/etnaviv/etnaviv_resource.h
> +++ b/src/gallium/drivers/etnaviv/etnaviv_resource.h
> @@ -31,6 +31,7 @@
>  #include "etnaviv_tiling.h"
>  #include "pipe/p_state.h"
>  #include "util/list.h"
> +#include "util/u_range.h"
>  
>  struct pipe_screen;
>  
> @@ -73,6 +74,9 @@ struct etna_resource {
>  
>     struct etna_resource_level levels[ETNA_NUM_LOD];
>  
> +   /* buffer range that has been initialized */
> +   struct util_range valid_buffer_range;
> +
>     /* When we are rendering to a texture, we need a differently tiled resource */
>     struct pipe_resource *texture;
>     /*
> diff --git a/src/gallium/drivers/etnaviv/etnaviv_transfer.c b/src/gallium/drivers/etnaviv/etnaviv_transfer.c
> index 6c1edd4835..33f3b058fa 100644
> --- a/src/gallium/drivers/etnaviv/etnaviv_transfer.c
> +++ b/src/gallium/drivers/etnaviv/etnaviv_transfer.c
> @@ -126,6 +126,10 @@ etna_transfer_unmap(struct pipe_context *pctx, struct pipe_transfer *ptrans)
>     if (!trans->rsc && !(ptrans->usage & PIPE_TRANSFER_UNSYNCHRONIZED))
>        etna_bo_cpu_fini(rsc->bo);
>  
> +   util_range_add(&rsc->valid_buffer_range,
> +                  ptrans->box.x,
> +                  ptrans->box.x + ptrans->box.width);
> +
>     pipe_resource_reference(&trans->rsc, NULL);
>     pipe_resource_reference(&ptrans->resource, NULL);
>     slab_free(&ctx->transfer_pool, trans);
> @@ -283,7 +287,14 @@ etna_transfer_map(struct pipe_context *pctx, struct pipe_resource *prsc,
>      * Pull resources into the CPU domain. Only skipped for unsynchronized
>      * transfers without a temporary resource.
>      */
> -   if (trans->rsc || !(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
> +   if ((usage & PIPE_TRANSFER_WRITE) &&
> +         prsc->target == PIPE_BUFFER &&
> +         !util_ranges_intersect(&rsc->valid_buffer_range,
> +                                box->x, box->x + box->width)) {
> +         /* We are trying to write to a previously uninitialized range. No need
> +          * to wait.
> +          */

This unbalances the cpu_prep/fini in the map/unmap path. This isn't
allowed and the kernel will start to reject this in the near future.

Also you need to differentiate between transfers with and without a
temporary resource (trans->rsc), as the valid range of the temporary
might be different from the base resource, especially if we need to
read back from the base.

Regards,
Lucas

> +    } else if (trans->rsc || !(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
>        uint32_t prep_flags = 0;
>  
>        if (usage & PIPE_TRANSFER_READ)
> @@ -360,10 +371,15 @@ fail_prep:
>  
>  static void
>  etna_transfer_flush_region(struct pipe_context *pctx,
> -                           struct pipe_transfer *transfer,
> +                           struct pipe_transfer *ptrans,
>                             const struct pipe_box *box)
>  {
> -   /* NOOP for now */
> +   struct etna_resource *rsc = etna_resource(ptrans->resource);
> +
> +   if (ptrans->resource->target == PIPE_BUFFER)
> +      util_range_add(&rsc->valid_buffer_range,
> +                     ptrans->box.x + box->x,
> +                     ptrans->box.x + box->x + box->width);
>  }
>  
>  void