intel/brw, lp: enable lower_pack_64_4x16

The compiler won't be able to emit pack_64_4x16, so we should prevent
nir_opt_algebraic to optimize to it. This fixes an infinite optimization
loop inside brw_nir_optimize:

nir_copy_prop
    16x4     %77 = @load_global (%80)
    32    %61995 = pack_32_2x16_split %77.x, %77.y
    32    %61998 = pack_32_2x16_split %77.z, %77.w
    64    %61999 = pack_64_2x32_split %61995, %61998
    64       %76 = iadd %100, %79
                   @store_global (%61999, %76)

nir_opt_algebraic
    16x4     %77 = @load_global (%80)
    32    %61995 = pack_32_2x16_split %77.x, %77.y
    32    %61998 = pack_32_2x16_split %77.z, %77.w
    16x4  %62000 = vec4 %77.x, %77.y, %77.z, %77.w
    64    %62001 = pack_64_4x16 %62000
    64       %76 = iadd %100, %79
                   @store_global (%62001, %76)

nir_lower_pack
    16x4     %77 = @load_global (%80)
    16x4  %62000 = vec4 %77.x, %77.y, %77.z, %77.w
    16    %62002 = mov %62000.y
    16    %62003 = mov %62000.x
    32    %62004 = pack_32_2x16_split %62003, %62002
    16    %62005 = mov %62000.w
    16    %62006 = mov %62000.z
    32    %62007 = pack_32_2x16_split %62006, %62005
    64    %62008 = pack_64_2x32_split %62004, %62007
    64       %76 = iadd %100, %79
                   @store_global (%62008, %76)

// brw_nir_optimize loops here

nir_copy_prop
    16x4     %77 = @load_global (%80)
    32    %62004 = pack_32_2x16_split %77.x, %77.y
    32    %62007 = pack_32_2x16_split %77.z, %77.w
    64    %62008 = pack_64_2x32_split %62004, %62007
    64       %76 = iadd %100, %79
                   @store_global (%62008, %76)

llvmpipe has a similar issue inside lp_build_opt_nir

Fixes: b1bc691b0f ("nir/algebraic: add and improve pack/unpack patterns")
Acked-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33347>
This commit is contained in:
Karol Herbst 2025-02-03 09:50:59 +01:00 committed by Marge Bot
parent 480d8bea2b
commit dad5ee1039
2 changed files with 2 additions and 0 deletions

View file

@ -446,6 +446,7 @@ static const struct nir_shader_compiler_options gallivm_nir_options = {
.lower_pack_unorm_2x16 = true,
.lower_pack_unorm_4x8 = true,
.lower_pack_half_2x16 = true,
.lower_pack_64_4x16 = true,
.lower_pack_split = true,
.lower_unpack_snorm_2x16 = true,
.lower_unpack_snorm_4x8 = true,

View file

@ -62,6 +62,7 @@ const struct nir_shader_compiler_options brw_scalar_nir_options = {
.lower_pack_snorm_4x8 = true,
.lower_pack_unorm_2x16 = true,
.lower_pack_unorm_4x8 = true,
.lower_pack_64_4x16 = true,
.lower_scmp = true,
.lower_to_scalar = true,
.lower_uadd_carry = true,