[Mesa-dev] [PATCH 22/24] i965/gen7: Handle atomic instructions from the VEC4 back-end.

[Paul Berry]

On 15 September 2013 00:19, Francisco Jerez <currojerez at riseup.net> wrote:

> This can deal with all the 15 32-bit untyped atomic operations the
> hardware supports, but only INC and PREDEC are going to be exposed
> through the API for now.
> ---
>  src/mesa/drivers/dri/i965/brw_vec4.h           |  7 +++
>  src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp | 64
> ++++++++++++++++++++++++++
>  2 files changed, 71 insertions(+)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_vec4.h
> b/src/mesa/drivers/dri/i965/brw_vec4.h
> index 37e1da0..7678925 100644
> --- a/src/mesa/drivers/dri/i965/brw_vec4.h
> +++ b/src/mesa/drivers/dri/i965/brw_vec4.h
> @@ -499,6 +499,13 @@ public:
>     void emit_shader_time_write(enum shader_time_shader_type type,
>                                 src_reg value);
>
> +   void emit_untyped_atomic(unsigned atomic_op, unsigned surf_index,
> +                            unsigned offset, dst_reg dst, src_reg src0,
> +                            src_reg src1);
> +
> +   void emit_untyped_surface_read(unsigned surf_index, unsigned offset,
> +                                  dst_reg dst);
> +
>     src_reg get_scratch_offset(vec4_instruction *inst,
>                               src_reg *reladdr, int reg_offset);
>     src_reg get_pull_constant_offset(vec4_instruction *inst,
> diff --git a/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
> b/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
> index a19686b..c3d4506 100644
> --- a/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
> @@ -2464,8 +2464,72 @@ vec4_visitor::visit(ir_end_primitive *)
>  }
>
>  void
> +vec4_visitor::emit_untyped_atomic(unsigned atomic_op, unsigned surf_index,
> +                                  unsigned offset, dst_reg dst,
> +                                  src_reg src0, src_reg src1)
> +{
> +   unsigned mlen = 0;
> +
> +   /* Set the atomic operation offset. */
> +   emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen, 0), WRITEMASK_X),
> +            src_reg(offset)));
> +   mlen++;
> +
> +   /* Set the atomic operation arguments. */
> +   if (src0.file != BAD_FILE) {
> +      emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen, 0), WRITEMASK_X),
> src0));
> +      mlen++;
> +   }
> +
> +   if (src1.file != BAD_FILE) {
> +      emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen, 0), WRITEMASK_X),
> src1));
> +      mlen++;
> +   }
> +
> +   /* Emit the instruction. */
> +   vec4_instruction *inst = emit(SHADER_OPCODE_UNTYPED_ATOMIC, dst,
> +                                 src_reg(atomic_op), src_reg(surf_index));
>

It's probably worth putting in a comment here to note that pre-Haswell,
this uses the SIMD8 atomic message (since there is no SIMD4x2 atomic
message).  But that's ok because the dst register is a float, so the
instruction will be emitted with WRITEMASK_X, meaning that the operation
will be appropriately masked down to operate on just 2 vertices.


> +   inst->base_mrf = 0;
> +   inst->mlen = mlen;
> +}
> +
> +void
> +vec4_visitor::emit_untyped_surface_read(unsigned surf_index,
> +                                        unsigned offset, dst_reg dst)
> +{
> +   /* Set the surface read offset. */
> +   emit(MOV(brw_writemask(brw_uvec_mrf(8, 0, 0), WRITEMASK_X),
> +            src_reg(offset)));
> +
> +   /* Emit the instruction. */
> +   vec4_instruction *inst = emit(SHADER_OPCODE_UNTYPED_SURFACE_READ,
> +                                 dst, src_reg(surf_index));
>

A similar coment about masking would be nice here.


> +   inst->base_mrf = 0;
> +   inst->mlen = 1;
> +}
> +
> +void
>  vec4_visitor::visit(ir_atomic *ir)
>  {
> +   ir_variable *loc = ir->location->variable_referenced();
> +   unsigned surf_index = SURF_INDEX_VEC4_ABO(loc->atomic.buffer_index);
> +
> +   result = src_reg(this, ir->type);
> +
> +   switch (ir->op) {
> +   case ir_atomic_read:
> +      emit_untyped_surface_read(surf_index, loc->atomic.offset,
> +                                dst_reg(result));
> +      break;
> +   case ir_atomic_inc:
> +      emit_untyped_atomic(BRW_AOP_INC, surf_index, loc->atomic.offset,
> +                          dst_reg(result), src_reg(), src_reg());
> +      break;
> +   case ir_atomic_dec:
> +      emit_untyped_atomic(BRW_AOP_PREDEC, surf_index, loc->atomic.offset,
> +                          dst_reg(result), src_reg(), src_reg());
>

My questions about fs_reg() in the last patch apply here too.


> +      break;
> +   }
>  }
>
>  void
> --
> 1.8.3.4
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20130926/9c6a4000/attachment.html>