[Mesa-dev] [PATCH 06/18] i965/cs: Add support for the SEND message that terminates a CS thread.

[Matt Turner]

On Tue, Apr 14, 2015 at 11:35 AM, Jordan Justen
<jordan.l.justen at intel.com> wrote:
> On 2015-04-13 23:41:15, Matt Turner wrote:
>> On Sat, Mar 14, 2015 at 9:54 PM, Jordan Justen
>> <jordan.l.justen at intel.com> wrote:
>> > From: Paul Berry <stereotype441 at gmail.com>
>> >
>> > This message must be sent to the "thread spawner" unit.
>> >
>> > Reviewed-by: Jordan Justen <jordan.l.justen at intel.com>
>> > ---
>> >  src/mesa/drivers/dri/i965/brw_eu.h      |  4 +++
>> >  src/mesa/drivers/dri/i965/brw_eu_emit.c | 46 +++++++++++++++++++++++++++++++++
>> >  2 files changed, 50 insertions(+)
>> >
>> > diff --git a/src/mesa/drivers/dri/i965/brw_eu.h b/src/mesa/drivers/dri/i965/brw_eu.h
>> > index d9ad5bd..708c216 100644
>> > --- a/src/mesa/drivers/dri/i965/brw_eu.h
>> > +++ b/src/mesa/drivers/dri/i965/brw_eu.h
>> > @@ -349,6 +349,10 @@ brw_jump_scale(const struct brw_context *brw)
>> >     return 1;
>> >  }
>> >
>> > +void brw_cs_terminate(struct brw_compile *p,
>> > +                      unsigned msg_reg_nr,
>> > +                      struct brw_reg src0);
>> > +
>> >  /* If/else/endif.  Works by manipulating the execution flags on each
>> >   * channel.
>> >   */
>> > diff --git a/src/mesa/drivers/dri/i965/brw_eu_emit.c b/src/mesa/drivers/dri/i965/brw_eu_emit.c
>> > index 43e5783..9c166a3 100644
>> > --- a/src/mesa/drivers/dri/i965/brw_eu_emit.c
>> > +++ b/src/mesa/drivers/dri/i965/brw_eu_emit.c
>> > @@ -2895,3 +2895,49 @@ void brw_shader_time_add(struct brw_compile *p,
>> >                                       0 /* response length */,
>> >                                       false /* header present */);
>> >  }
>> > +
>> > +
>> > +static void
>> > +brw_set_cs_terminate_message(struct brw_context *brw,
>> > +                             struct brw_compile *p,
>> > +                             struct brw_inst *insn)
>> > +{
>> > +   /* Terminate a compute shader by sending a message to the thread spawner.
>> > +    * Note that even though the thread has a URB resource associated with it,
>> > +    * we set the "do not dereference URB" bit, because the URB resource is
>> > +    * managed by the fixed-function unit, so it will free it automatically.
>> > +    */
>> > +   brw_set_message_descriptor(p, insn, BRW_SFID_THREAD_SPAWNER,
>> > +                              1 /* msg_length */,
>> > +                              0 /* response_length */,
>> > +                              false /* header_present */,
>> > +                              true /* end_of_thread */);
>> > +   brw_inst_set_ts_opcode(brw, insn, 0); /* Dereference resource */
>> > +   brw_inst_set_ts_request_type(brw, insn, 0); /* Root thread */
>> > +   brw_inst_set_ts_resource_select(brw, insn, 1); /* Do not dereference URB */
>> > +}
>> > +
>> > +
>> > +/**
>> > + * Emit the SEND message to terminate a compute shader.
>> > + */
>> > +void
>> > +brw_cs_terminate(struct brw_compile *p,
>> > +                 unsigned msg_reg_nr,
>> > +                 struct brw_reg src0)
>> > +{
>> > +   struct brw_context *brw = p->brw;
>> > +   struct brw_inst *insn;
>> > +
>> > +   gen6_resolve_implied_move(p, &src0, msg_reg_nr);
>> > +   insn = next_insn(p, BRW_OPCODE_SEND);
>> > +   brw_set_dest(p, insn, brw_null_reg());
>> > +   brw_set_src0(p, insn, src0);
>> > +   brw_set_src1(p, insn, brw_imm_d(0));
>> > +
>> > +   if (brw->gen < 6)
>> > +      brw_inst_set_cond_modifier(brw, insn, msg_reg_nr);
>>
>> Presumably gen <= 6 can't do compute shaders, so what's going on here
>> and in the call to gen6_resolve_implied_move above?
>
> Ken mentioned that I needed to drop the "i965/fs: Don't assign high
> regs for CS termination send" patch. This caused me to rework these
> patches a bit. In the end, I dropped this patch, and moved this code
> to be in the visitor & generator:
>
> http://cgit.freedesktop.org/~jljusten/mesa/commit/?h=cs-27&id=fc0ccf68
> http://cgit.freedesktop.org/~jljusten/mesa/commit/?h=cs-27&id=d12ec0e8
>
> Does this look better?

Yep, those look good.

Thanks!