[Mesa-dev] [PATCH] i965/nir/vec4: fill the type of the dst and src when loading an uniform

Jason Ekstrand jason at jlekstrand.net
Tue Sep 15 16:39:43 PDT 2015


On Tue, Sep 15, 2015 at 3:23 PM, Alejandro PiƱeiro <apinheiro at igalia.com> wrote:
> Until know, nir_intrinsic_load_uniform was using a default type. This
          "now"
> caused some type mismatch between movs and alu operations, so copy
> propagation optimization was not applied to remove unneeded movs if
> negate modifier was involved. This was first detected on minus
> (negate+add) operations with uniforms involved.
>
> This patch is equivalent to commit 4de86, but that one only affected
> inputs, which information, included the type, was already being stored.
> So this patch includes saving the type on nir_setup_uniforms.

Can't we just fix copy propagation?  If the copy operation is a dumb
MOV where there's no source/dest modifiers and the source and
destination types are the same, then we can copy-propagate regardless
of whether the type of the MOV matches the type of the thing we're
copying into.  Because NIR is typeless, we're liable to get type
mismatches on MOVs and that's just kind of part of life now.  If I
recall correctly, we did that in the FS backend and it worked out
pretty well.

--Jason

> Shader DB results (taking into account only vec4):
>
> total instructions in shared programs: 1754416 -> 1742146 (-0.70%)
> instructions in affected programs:     677491 -> 665221 (-1.81%)
> helped:                                5381
> HURT:                                  0
> ---
>  src/mesa/drivers/dri/i965/brw_vec4.h           | 1 +
>  src/mesa/drivers/dri/i965/brw_vec4_nir.cpp     | 4 ++++
>  src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp | 1 +
>  3 files changed, 6 insertions(+)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_vec4.h b/src/mesa/drivers/dri/i965/brw_vec4.h
> index 01c6e84..f6d2c82 100644
> --- a/src/mesa/drivers/dri/i965/brw_vec4.h
> +++ b/src/mesa/drivers/dri/i965/brw_vec4.h
> @@ -163,6 +163,7 @@ public:
>     int *uniform_vector_size;
>     int uniform_array_size; /*< Size of uniform_[vector_]size arrays */
>     int uniforms;
> +   brw_reg_type *uniform_type;
>
>     src_reg shader_start_time;
>
> diff --git a/src/mesa/drivers/dri/i965/brw_vec4_nir.cpp b/src/mesa/drivers/dri/i965/brw_vec4_nir.cpp
> index 175d92b..0b624fa 100644
> --- a/src/mesa/drivers/dri/i965/brw_vec4_nir.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_vec4_nir.cpp
> @@ -143,6 +143,7 @@ vec4_visitor::nir_setup_uniforms(nir_shader *shader)
>
>           assert(uniforms < uniform_array_size);
>           this->uniform_size[uniforms] = type_size_vec4(var->type);
> +         this->uniform_type[uniforms] = brw_type_for_base_type(var->type);
>
>           if (strncmp(var->name, "gl_", 3) == 0)
>              nir_setup_builtin_uniform(var);
> @@ -553,6 +554,9 @@ vec4_visitor::nir_emit_intrinsic(nir_intrinsic_instr *instr)
>        src = src_reg(dst_reg(UNIFORM, instr->const_index[0]));
>        src.reg_offset = instr->const_index[1];
>
> +      assert(instr->const_index[0] < uniform_array_size);
> +      dest.type = src.type = this->uniform_type[instr->const_index[0]];
> +
>        if (has_indirect) {
>           src_reg tmp = get_nir_src(instr->src[0], BRW_REGISTER_TYPE_D, 1);
>           src.reladdr = new(mem_ctx) src_reg(tmp);
> diff --git a/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp b/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
> index 0465770..8e00612 100644
> --- a/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
> @@ -3747,6 +3747,7 @@ vec4_visitor::vec4_visitor(const struct brw_compiler *compiler,
>
>     this->uniform_size = rzalloc_array(mem_ctx, int, this->uniform_array_size);
>     this->uniform_vector_size = rzalloc_array(mem_ctx, int, this->uniform_array_size);
> +   this->uniform_type = rzalloc_array(mem_ctx, brw_reg_type, this->uniform_array_size);
>  }
>
>  vec4_visitor::~vec4_visitor()
> --
> 2.1.4
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev


More information about the mesa-dev mailing list