[Mesa-dev] [PATCH v2] i965: intel_texture_barrier reimplemented

[Francisco Jerez]

Hi Alejandro,

Alejandro Piñeiro <apinheiro at igalia.com> writes:

> Fixes:
> GL44-CTS.texture_barrier_ARB.same-texel-rw-multipass
>
> On Haswell, Broadwell and Skylake (note that in order to execute that
> test, it is needed to override GL and GLSL versions).
>
> On gen6 this test was already working without this change. It keeps
> working after it.
>
> This commit replaces the call to brw_emit_mi_flush for gen6+ with two
> calls to brw_emit_pipe_control_flush:
>
>  * The first one with RENDER_TARGET_FLUSH and CS_STALL set to initiate
>    a render cache flush after any concurrent rendering completes and
>    cause the CS to stop parsing commands until the render cache
>    becomes coherent with memory.
>
>  * The second one have TEXTURE_CACHE_INVALIDATE set (and no CS stall)
>    to clean up any stale data from the sampler caches before rendering
>    continues.
>
> Didn't touch gen4-5, basically because I don't have a way to test
> them.
>
> More info on commits:
> 0aa4f99f562a05880a779707cbcd46be459863bf
> 72473658c51d5e074ce219c1e6385a4cce29f467
>
> Thanks to Curro to help to tracking this down, as the root case was a
> hw race condition.
> ---
>
> The implementation is basically the same that the middle of function
> brw_render_cache_set_check_flush, with the exception that we don't
> check for a specific bo, neither call brw_render_cache_set_clear.
>
> Probably we could refactor that common part to an utility method, but
> I prefer to wait to get the review for this patch. Not sure also about
> the func name of that refactored method.
>
>  src/mesa/drivers/dri/i965/intel_tex.c | 22 +++++++++++++++++++++-
>  1 file changed, 21 insertions(+), 1 deletion(-)
>
> diff --git a/src/mesa/drivers/dri/i965/intel_tex.c b/src/mesa/drivers/dri/i965/intel_tex.c
> index cac33ac..695bc4f 100644
> --- a/src/mesa/drivers/dri/i965/intel_tex.c
> +++ b/src/mesa/drivers/dri/i965/intel_tex.c
> @@ -9,6 +9,7 @@
>  #include "intel_mipmap_tree.h"
>  #include "intel_tex.h"
>  #include "intel_fbo.h"
> +#include "intel_reg.h"
>  
>  #define FILE_DEBUG_FLAG DEBUG_TEXTURE
>  
> @@ -362,7 +363,26 @@ intel_texture_barrier(struct gl_context *ctx)
>  {
>     struct brw_context *brw = brw_context(ctx);
>  
> -   brw_emit_mi_flush(brw);
> +   if (brw->gen >= 6) {
> +      if (brw->gen == 6) {
> +         /* [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache
> +          * Flush Enable = 1, a PIPE_CONTROL with any non-zero
> +          * post-sync-op is required.
> +          */
> +         brw_emit_post_sync_nonzero_flush(brw);
> +      }
> +
> +      brw_emit_pipe_control_flush(brw,
> +                                  PIPE_CONTROL_DEPTH_CACHE_FLUSH |
> +                                  PIPE_CONTROL_RENDER_TARGET_FLUSH |
> +                                  PIPE_CONTROL_CS_STALL);
> +
> +      brw_emit_pipe_control_flush(brw,
> +                                  PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
> +                                  PIPE_CONTROL_CONST_CACHE_INVALIDATE);

What's the purpose of the const cache invalidation in the last pipe
control command?

> +   } else {
> +      brw_emit_mi_flush(brw);
> +   }
>  }
>  
>  void
> -- 
> 2.7.4
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 212 bytes
Desc: not available
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20160629/841a386e/attachment.sig>