[Mesa-dev] [PATCH] r600: Use DMA transfers in r600_copy_global_buffer

[Bruno Jimenez]

Hi,

I'm not sure if this will work. Imagine this case:

We  have an item in the pool, and we want to use
r600_resource_copy_region with it, for example because we want to demote
it. This will call r600_copy_global_buffer, and with your patch it will
call r600_compute_global_demote_or_alloc, which will again call
compute_memory_demote_item causing an infinite cycle.

Also, why are you reassigning src and dst in r600_copy_global_buffer?

- Bruno

On Sun, 2014-09-07 at 18:32 +0200, Niels Ole Salscheider wrote:
> Signed-off-by: Niels Ole Salscheider <niels_ole at salscheider-online.de>
> ---
>  src/gallium/drivers/r600/evergreen_compute.c | 27 ++++++++++++-------
>  src/gallium/drivers/r600/evergreen_compute.h |  1 +
>  src/gallium/drivers/r600/r600_blit.c         | 40 ++++++++++++++++------------
>  3 files changed, 41 insertions(+), 27 deletions(-)
> 
> diff --git a/src/gallium/drivers/r600/evergreen_compute.c b/src/gallium/drivers/r600/evergreen_compute.c
> index 38b78c7..b495868 100644
> --- a/src/gallium/drivers/r600/evergreen_compute.c
> +++ b/src/gallium/drivers/r600/evergreen_compute.c
> @@ -953,6 +953,22 @@ void r600_compute_global_buffer_destroy(
>  	free(res);
>  }
>  
> +void r600_compute_global_demote_or_alloc(
> +	struct compute_memory_pool *pool,
> +	struct compute_memory_item *item,
> +	struct pipe_context *ctx)
> +{
> +	if (is_item_in_pool(item)) {
> +		compute_memory_demote_item(pool, item, ctx);
> +	} else {
> +		if (item->real_buffer == NULL) {
> +			item->real_buffer = (struct r600_resource*)
> +					r600_compute_buffer_alloc_vram(pool->screen, item->size_in_dw * 4);
> +		}
> +	}
> +
> +}
> +
>  void *r600_compute_global_transfer_map(
>  	struct pipe_context *ctx_,
>  	struct pipe_resource *resource,
> @@ -970,16 +986,7 @@ void *r600_compute_global_transfer_map(
>  	struct pipe_resource *dst = NULL;
>  	unsigned offset = box->x;
>  
> -	if (is_item_in_pool(item)) {
> -		compute_memory_demote_item(pool, item, ctx_);
> -	}
> -	else {
> -		if (item->real_buffer == NULL) {
> -			item->real_buffer = (struct r600_resource*)
> -					r600_compute_buffer_alloc_vram(pool->screen, item->size_in_dw * 4);
> -		}
> -	}
> -
> +	r600_compute_global_demote_or_alloc(pool, item, ctx_);
>  	dst = (struct pipe_resource*)item->real_buffer;
>  
>  	if (usage & PIPE_TRANSFER_READ)
> diff --git a/src/gallium/drivers/r600/evergreen_compute.h b/src/gallium/drivers/r600/evergreen_compute.h
> index 4fb53a1..39bb854 100644
> --- a/src/gallium/drivers/r600/evergreen_compute.h
> +++ b/src/gallium/drivers/r600/evergreen_compute.h
> @@ -47,6 +47,7 @@ void evergreen_emit_cs_shader(struct r600_context *rctx, struct r600_atom * atom
>  
>  struct pipe_resource *r600_compute_global_buffer_create(struct pipe_screen *screen, const struct pipe_resource *templ);
>  void r600_compute_global_buffer_destroy(struct pipe_screen *screen, struct pipe_resource *res);
> +void r600_compute_global_demote_or_alloc(struct compute_memory_pool *pool, struct compute_memory_item *item, struct pipe_context *ctx);
>  void *r600_compute_global_transfer_map(
>  	struct pipe_context *ctx_,
>  	struct pipe_resource *resource,
> diff --git a/src/gallium/drivers/r600/r600_blit.c b/src/gallium/drivers/r600/r600_blit.c
> index f766e37..f6471cb 100644
> --- a/src/gallium/drivers/r600/r600_blit.c
> +++ b/src/gallium/drivers/r600/r600_blit.c
> @@ -21,6 +21,8 @@
>   * USE OR OTHER DEALINGS IN THE SOFTWARE.
>   */
>  #include "r600_pipe.h"
> +#include "compute_memory_pool.h"
> +#include "evergreen_compute.h"
>  #include "util/u_surface.h"
>  #include "util/u_format.h"
>  #include "evergreend.h"
> @@ -514,29 +516,33 @@ static void r600_copy_buffer(struct pipe_context *ctx, struct pipe_resource *dst
>   * into a single global resource (r600_screen::global_pool).  The means
>   * they don't have their own cs_buf handle, so they cannot be passed
>   * to r600_copy_buffer() and must be handled separately.
> - *
> - * XXX: It should be possible to implement this function using
> - * r600_copy_buffer() by passing the memory_pool resource as both src
> - * and dst and updating dstx and src_box to point to the correct offsets.
> - * This would likely perform better than the current implementation.
>   */
>  static void r600_copy_global_buffer(struct pipe_context *ctx,
>  				    struct pipe_resource *dst, unsigned
>  				    dstx, struct pipe_resource *src,
>  				    const struct pipe_box *src_box)
>  {
> -	struct pipe_box dst_box; struct pipe_transfer *src_pxfer,
> -	*dst_pxfer;
> -
> -	u_box_1d(dstx, src_box->width, &dst_box);
> -	void *src_ptr = ctx->transfer_map(ctx, src, 0, PIPE_TRANSFER_READ,
> -					  src_box, &src_pxfer);
> -	void *dst_ptr = ctx->transfer_map(ctx, dst, 0, PIPE_TRANSFER_WRITE,
> -					  &dst_box, &dst_pxfer);
> -	memcpy(dst_ptr, src_ptr, src_box->width);
> -
> -	ctx->transfer_unmap(ctx, src_pxfer);
> -	ctx->transfer_unmap(ctx, dst_pxfer);
> +	struct r600_context *rctx = (struct r600_context*)ctx;
> +	struct compute_memory_pool *pool = rctx->screen->global_pool;
> +
> +	if (src->bind & PIPE_BIND_GLOBAL) {
> +		struct r600_resource_global *rsrc =
> +			(struct r600_resource_global *)src;
> +		struct compute_memory_item *item = rsrc->chunk;
> +
> +		r600_compute_global_demote_or_alloc(pool, item, ctx);
> +		src = (struct pipe_resource*)item->real_buffer;
> +	}
> +	if (dst->bind & PIPE_BIND_GLOBAL) {
> +		struct r600_resource_global *rdst =
> +			(struct r600_resource_global *)dst;
> +		struct compute_memory_item *item = rdst->chunk;
> +
> +		r600_compute_global_demote_or_alloc(pool, item, ctx);
> +		dst = (struct pipe_resource*)item->real_buffer;
> +	}
> +
> +	r600_copy_buffer(ctx, dst, dstx, src, src_box);
>  }
>  
>  static void r600_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst,