[Mesa-dev] [v2 4/9] i965: Estimate batch space per shader stage

[Pohjolainen, Topi]

On Tue, Jan 31, 2017 at 11:12:31AM -0800, Jason Ekstrand wrote:
>    On Tue, Jan 31, 2017 at 10:38 AM, Jason Ekstrand
>    <[1]jason at jlekstrand.net> wrote:
> 
>    On Tue, Jan 31, 2017 at 8:15 AM, Topi Pohjolainen
>    <[2]topi.pohjolainen at gmail.com> wrote:
> 
>      Current estimate doesn't consider space needed for surface states
>      and it only calculates for one shader stage. Each stage can have
>      its own sampler and surface state configuration.
>      While this is only matter of runtime dynamics we don't seem to hit
>      it currently. However, this becomes visible with blorp tex uploads
>      (HSW with piglit test max-samplers). One runs out of space while
>      batch wrapping isn't allowed.
> 
>    Also, what happens when this case is triggered?  Do we assert fail?

Assert on debug build, I didn't try release.

>    GPU Hang?  Just nicely handle it by flushing the batch?
>    --Jason
> 
>      v2: Rebase on top of current upstream
>      Signed-off-by: Topi Pohjolainen <[3]topi.pohjolainen at intel.com>
>      CC: Kenneth Graunke <[4]kenneth at whitecape.org>
>      CC: Jason Ekstrand <[5]jason at jlekstrand.net>
>      ---
>       src/mesa/drivers/dri/i965/brw_draw.c | 49
>      +++++++++++++++++++++++++++++++++---
>       1 file changed, 46 insertions(+), 3 deletions(-)
>      diff --git a/src/mesa/drivers/dri/i965/brw_draw.c
>      b/src/mesa/drivers/dri/i965/brw_draw.c
>      index 0db7311..83a9f33 100644
>      --- a/src/mesa/drivers/dri/i965/brw_draw.c
>      +++ b/src/mesa/drivers/dri/i965/brw_draw.c
>      @@ -395,6 +395,51 @@ brw_postdraw_set_buffers_need_resolve(struct
>      brw_context *brw)
> 
>        }
>     }
>    +static unsigned
>    +brw_get_num_active_samplers(const struct gl_context *ctx,
>    +                            const struct gl_program *prog)
>    +{
>    +   const unsigned last = util_last_bit(prog->SamplersUsed);
>    +   unsigned count = 0;
>    +
>    +   for (unsigned s = 0; s < last; s++) {
>    +      if (prog->SamplersUsed & (1 << s)) {
>    +         const unsigned unit = prog->SamplerUnits[s];
>    +         if (ctx->Texture.Unit[unit]._Current)
>    +            ++count;
>    +      }
>    +   }
>    +
>    +   return count;
>    +}
>    +
>    +static unsigned
>    +brw_estimate_batch_space_for_textures(const struct brw_context *brw)
>    +{
>    +   const struct gl_context *ctx = &brw->ctx;
>    +   unsigned total = 0;
>    +
>    +   for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
> 
>      +      const struct gl_program *prog =
>      ctx->_Shader->CurrentProgram[i];
>      +
>      +      if (prog == NULL)
>      +         continue;
>      +
>      +      const unsigned num_samplers = brw_get_num_active_samplers(ct
>      x, prog);
>      +      const unsigned sampler_needs_per_tex_unit =
>      +         16 /* sampler_state_size */ +
>      +         sizeof(struct gen5_sampler_default_color);
>      +      const unsigned surface_state_needs_per_tex_unit =
>      +         ALIGN(brw->isl_dev.ss.size, brw->isl_dev.ss.align) +
>      +         4 /* binding table pointer */;
>      +      const unsigned total_per_tex_unit =
>      sampler_needs_per_tex_unit +
>      +
>      surface_state_needs_per_tex_unit;
>      +      total += (num_samplers * total_per_tex_unit);
> 
>    This isn't exactly correct.  While it's true that a binding table entry
>    only consumes 4 bytes, binding table sizes have to be rounded up to 64B
>    so, if you have a number of samplers that is not a multiple of 16, this
>    will underestimate.

Good catch, thanks!

> 
>      +   }
>      +
>      +   return total;
>      +}
>      +
>       /* May fail if out of video memory for texture or vbo upload, or on
>        * fallback conditions.
>        */
>      @@ -477,11 +522,9 @@ brw_try_draw_prims(struct gl_context *ctx,
> 
>        for (i = 0; i < nr_prims; i++) {
>           int estimated_max_prim_size;
>    -      const int sampler_state_size = 16;
>           estimated_max_prim_size = 512; /* batchbuffer commands */
>    -      estimated_max_prim_size += BRW_MAX_TEX_UNIT *
>    -         (sampler_state_size + sizeof(struct
>    gen5_sampler_default_color));
>    +      estimated_max_prim_size += brw_estimate_batch_space_for_t
>    extures(brw);
>           estimated_max_prim_size += 1024; /* gen6 VS push constants */
>           estimated_max_prim_size += 1024; /* gen6 WM push constants */
>           estimated_max_prim_size += 512; /* misc. pad */
>    --
>    2.5.5
> 
> References
> 
>    1. mailto:jason at jlekstrand.net
>    2. mailto:topi.pohjolainen at gmail.com
>    3. mailto:topi.pohjolainen at intel.com
>    4. mailto:kenneth at whitecape.org
>    5. mailto:jason at jlekstrand.net