[Mesa-dev] [PATCH v2] r600: Use DMA transfers in r600_copy_global_buffer

[Niels Ole Salscheider]

On Sunday 28 September 2014, 17:44:53, Bruno Jimenez wrote:
> Hi,
> 
> Sorry for not answering until now, but I have had some personal issues
> (changing university, moving to another city...)
> 
> As you said, this is used from clover's resource::copy, which is used by
> clEnqueueCopyBuffer if I remember correctly (and understand correctly
> clover) If it doesn't regress any piglit test then it has my R-b :)
> 
> Thanks a lot!
> Bruno

Hi,

no problem, I have been a bit busy with my thesis anyway (I have to hand it in 
on Tuesday)...

You are right, it is used from clEnqueueCopyBuffer - and it does not regress 
any piglit tests for me.
Can someone with write access please push this?

Thanks,

Ole

> On Mon, 2014-09-08 at 20:10 +0200, Niels Ole Salscheider wrote:
> > v2: Do not demote items that are already in the pool
> > 
> > Signed-off-by: Niels Ole Salscheider <niels_ole at salscheider-online.de>
> > ---
> > 
> >  src/gallium/drivers/r600/evergreen_compute.h |  1 +
> >  src/gallium/drivers/r600/r600_blit.c         | 59
> >  ++++++++++++++++++++-------- 2 files changed, 43 insertions(+), 17
> >  deletions(-)
> > 
> > diff --git a/src/gallium/drivers/r600/evergreen_compute.h
> > b/src/gallium/drivers/r600/evergreen_compute.h index 4fb53a1..e4d3a38
> > 100644
> > --- a/src/gallium/drivers/r600/evergreen_compute.h
> > +++ b/src/gallium/drivers/r600/evergreen_compute.h
> > @@ -45,6 +45,7 @@ void evergreen_init_atom_start_compute_cs(struct
> > r600_context *rctx);> 
> >  void evergreen_init_compute_state_functions(struct r600_context *rctx);
> >  void evergreen_emit_cs_shader(struct r600_context *rctx, struct r600_atom
> >  * atom);> 
> > +struct r600_resource* r600_compute_buffer_alloc_vram(struct r600_screen
> > *screen, unsigned size);> 
> >  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_transfer_map(
> > 
> > diff --git a/src/gallium/drivers/r600/r600_blit.c
> > b/src/gallium/drivers/r600/r600_blit.c index f766e37..b334a75 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 +516,52 @@ 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;
> > +	struct pipe_box new_src_box = *src_box;
> > +
> > +	if (src->bind & PIPE_BIND_GLOBAL) {
> > +		struct r600_resource_global *rsrc =
> > +			(struct r600_resource_global *)src;
> > +		struct compute_memory_item *item = rsrc->chunk;
> > +
> > +		if (is_item_in_pool(item)) {
> > +			new_src_box.x += 4 * item->start_in_dw;
> > +			src = (struct pipe_resource *)pool->bo;
> > +		} else {
> > +			if (item->real_buffer == NULL) {
> > +				item->real_buffer = (struct r600_resource*)
> > +					r600_compute_buffer_alloc_vram(pool->screen,
> > +								       item->size_in_dw * 4);
> > +			}
> > +			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;
> > +
> > +		if (is_item_in_pool(item)) {
> > +			dstx += 4 * item->start_in_dw;
> > +			dst = (struct pipe_resource *)pool->bo;
> > +		} else {
> > +			if (item->real_buffer == NULL) {
> > +				item->real_buffer = (struct r600_resource*)
> > +					r600_compute_buffer_alloc_vram(pool->screen,
> > +								       item->size_in_dw * 4);
> > +			}
> > +			dst = (struct pipe_resource*)item->real_buffer;
> > +		}
> > +	}
> > +
> > +	r600_copy_buffer(ctx, dst, dstx, src, &new_src_box);
> > 
> >  }
> >  
> >  static void r600_clear_buffer(struct pipe_context *ctx, struct
> >  pipe_resource *dst,