From 1b8ef49f48ae3634e4903422a9d9c11864c03cb1 Mon Sep 17 00:00:00 2001 From: Jason Ekstrand Date: Thu, 12 Oct 2017 16:17:03 -0700 Subject: [PATCH] intel/fs: Use a pair of 1-wide MOVs instead of SEL for any/all For some reason, the any/all predicates don't work properly with SIMD32. In particular, it appears that a SEL with a QtrCtrl of 2H doesn't read the correct subset of the flag register and you end up getting garbage in the second half. Work around this by using a pair of 1-wide MOVs and scattering the result. This fixes the any/all instructions for SIMD32. Reviewed-by: Matt Turner Cc: mesa-stable@lists.freedesktop.org --- src/intel/compiler/brw_fs_nir.cpp | 42 ++++++++++++++++++++++++------- 1 file changed, 33 insertions(+), 9 deletions(-) diff --git a/src/intel/compiler/brw_fs_nir.cpp b/src/intel/compiler/brw_fs_nir.cpp index 3143bc6eea1..f433e3b6c3d 100644 --- a/src/intel/compiler/brw_fs_nir.cpp +++ b/src/intel/compiler/brw_fs_nir.cpp @@ -4208,12 +4208,20 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr } bld.CMP(bld.null_reg_d(), get_nir_src(instr->src[0]), brw_imm_d(0), BRW_CONDITIONAL_NZ); - dest.type = BRW_REGISTER_TYPE_D; - bld.MOV(dest, brw_imm_d(-1)); + /* For some reason, the any/all predicates don't work properly with + * SIMD32. In particular, it appears that a SEL with a QtrCtrl of 2H + * doesn't read the correct subset of the flag register and you end up + * getting garbage in the second half. Work around this by using a pair + * of 1-wide MOVs and scattering the result. + */ + fs_reg res1 = ubld.vgrf(BRW_REGISTER_TYPE_D); + ubld.MOV(res1, brw_imm_d(0)); set_predicate(dispatch_width == 8 ? BRW_PREDICATE_ALIGN1_ANY8H : dispatch_width == 16 ? BRW_PREDICATE_ALIGN1_ANY16H : BRW_PREDICATE_ALIGN1_ANY32H, - bld.SEL(dest, dest, brw_imm_d(0))); + ubld.MOV(res1, brw_imm_d(-1))); + + bld.MOV(retype(dest, BRW_REGISTER_TYPE_D), component(res1, 0)); break; } case nir_intrinsic_vote_all: { @@ -4232,12 +4240,20 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr } bld.CMP(bld.null_reg_d(), get_nir_src(instr->src[0]), brw_imm_d(0), BRW_CONDITIONAL_NZ); - dest.type = BRW_REGISTER_TYPE_D; - bld.MOV(dest, brw_imm_d(-1)); + /* For some reason, the any/all predicates don't work properly with + * SIMD32. In particular, it appears that a SEL with a QtrCtrl of 2H + * doesn't read the correct subset of the flag register and you end up + * getting garbage in the second half. Work around this by using a pair + * of 1-wide MOVs and scattering the result. + */ + fs_reg res1 = ubld.vgrf(BRW_REGISTER_TYPE_D); + ubld.MOV(res1, brw_imm_d(0)); set_predicate(dispatch_width == 8 ? BRW_PREDICATE_ALIGN1_ALL8H : dispatch_width == 16 ? BRW_PREDICATE_ALIGN1_ALL16H : BRW_PREDICATE_ALIGN1_ALL32H, - bld.SEL(dest, dest, brw_imm_d(0))); + ubld.MOV(res1, brw_imm_d(-1))); + + bld.MOV(retype(dest, BRW_REGISTER_TYPE_D), component(res1, 0)); break; } case nir_intrinsic_vote_eq: { @@ -4258,12 +4274,20 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr } bld.CMP(bld.null_reg_d(), value, uniformized, BRW_CONDITIONAL_Z); - dest.type = BRW_REGISTER_TYPE_D; - bld.MOV(dest, brw_imm_d(-1)); + /* For some reason, the any/all predicates don't work properly with + * SIMD32. In particular, it appears that a SEL with a QtrCtrl of 2H + * doesn't read the correct subset of the flag register and you end up + * getting garbage in the second half. Work around this by using a pair + * of 1-wide MOVs and scattering the result. + */ + fs_reg res1 = ubld.vgrf(BRW_REGISTER_TYPE_D); + ubld.MOV(res1, brw_imm_d(0)); set_predicate(dispatch_width == 8 ? BRW_PREDICATE_ALIGN1_ALL8H : dispatch_width == 16 ? BRW_PREDICATE_ALIGN1_ALL16H : BRW_PREDICATE_ALIGN1_ALL32H, - bld.SEL(dest, dest, brw_imm_d(0))); + ubld.MOV(res1, brw_imm_d(-1))); + + bld.MOV(retype(dest, BRW_REGISTER_TYPE_D), component(res1, 0)); break; } -- 2.30.2