This is an archive of the discontinued LLVM Phabricator instance.

Differential D119157

[NVPTX] Add ex2 f16 support
ClosedPublic

Authored by npmiller on Feb 7 2022, 9:45 AM.

Details

Reviewers
tra
csigg
bkramer
Commits
rG510fd283fda2: [NVPTX] Add ex2.approx.f16/f16x2 support
rG69a8350c232a: [NVPTX] Add ex2.approx.f16/f16x2 support
Summary

This patch adds builtins and intrinsics for the f16 and f16x2 variants of the ex2 instruction.

These two variants were added in PTX7.0, and are supported by sm_75 and above.

Note that this isn't wired with the exp2 llvm intrinsic because the ex2 instruction is only available in its approx variant.

Running ptxas on the assembly generated by the test f16-ex2.ll works as expected.

Depends on D118977

Diff Detail

Event Timeline

npmiller created this revision.Feb 7 2022, 9:45 AM
Herald added subscribers: asavonic, hiraditya, jholewinski. · View Herald TranscriptFeb 7 2022, 9:45 AM
npmiller requested review of this revision.Feb 7 2022, 9:45 AM
Herald added projects: Restricted Project, Restricted Project. · View Herald TranscriptFeb 7 2022, 9:45 AM
Herald added subscribers: llvm-commits, cfe-commits, jdoerfert. · View Herald Transcript
tra added a subscriber: tra.Feb 7 2022, 11:49 AM

Is the patch is ready for review?

Harbormaster completed remote builds in B148025: Diff 406507.Feb 7 2022, 12:19 PM
npmiller added a comment.Feb 8 2022, 1:40 AM

Yes, it should be good to be reviewed.

npmiller added reviewers: tra, csigg, bkramer.Feb 8 2022, 1:51 AM
jchlanda mentioned this in D118977: [NVPTX] Add more FMA intriniscs/builtins.Feb 17 2022, 4:31 AM
tra accepted this revision.Feb 22 2022, 2:20 PM
This revision is now accepted and ready to land.Feb 22 2022, 2:20 PM
This revision was landed with ongoing or failed builds.Feb 23 2022, 1:57 PM
Closed by commit rG69a8350c232a: [NVPTX] Add ex2.approx.f16/f16x2 support (authored by npmiller, committed by tra). · Explain Why
This revision was automatically updated to reflect the committed changes.
tra added a commit: rG69a8350c232a: [NVPTX] Add ex2.approx.f16/f16x2 support.
tra added a commit: rG510fd283fda2: [NVPTX] Add ex2.approx.f16/f16x2 support.Mar 1 2022, 11:08 AM

Revision Contents

PathSize
llvm/
include/
llvm/
IR/
4 lines
lib/
Target/
NVPTX/
4 lines
test/
CodeGen/
NVPTX/
20 lines

Diff 410932

llvm/include/llvm/IR/IntrinsicsNVVM.td

Show First 20 Lines • Show All 822 Lines • ▼ Show 20 Lines
// //
def int_nvvm_ex2_approx_ftz_f : GCCBuiltin<"__nvvm_ex2_approx_ftz_f">, def int_nvvm_ex2_approx_ftz_f : GCCBuiltin<"__nvvm_ex2_approx_ftz_f">,
DefaultAttrsIntrinsic<[llvm_float_ty], [llvm_float_ty], [IntrNoMem]>; DefaultAttrsIntrinsic<[llvm_float_ty], [llvm_float_ty], [IntrNoMem]>;
def int_nvvm_ex2_approx_f : GCCBuiltin<"__nvvm_ex2_approx_f">, def int_nvvm_ex2_approx_f : GCCBuiltin<"__nvvm_ex2_approx_f">,
DefaultAttrsIntrinsic<[llvm_float_ty], [llvm_float_ty], [IntrNoMem]>; DefaultAttrsIntrinsic<[llvm_float_ty], [llvm_float_ty], [IntrNoMem]>;
def int_nvvm_ex2_approx_d : GCCBuiltin<"__nvvm_ex2_approx_d">, def int_nvvm_ex2_approx_d : GCCBuiltin<"__nvvm_ex2_approx_d">,
DefaultAttrsIntrinsic<[llvm_double_ty], [llvm_double_ty], [IntrNoMem]>; DefaultAttrsIntrinsic<[llvm_double_ty], [llvm_double_ty], [IntrNoMem]>;
def int_nvvm_ex2_approx_f16 : GCCBuiltin<"__nvvm_ex2_approx_f16">,
DefaultAttrsIntrinsic<[llvm_half_ty], [llvm_half_ty], [IntrNoMem]>;
def int_nvvm_ex2_approx_f16x2 : GCCBuiltin<"__nvvm_ex2_approx_f16x2">,
DefaultAttrsIntrinsic<[llvm_v2f16_ty], [llvm_v2f16_ty], [IntrNoMem]>;
def int_nvvm_lg2_approx_ftz_f : GCCBuiltin<"__nvvm_lg2_approx_ftz_f">, def int_nvvm_lg2_approx_ftz_f : GCCBuiltin<"__nvvm_lg2_approx_ftz_f">,
DefaultAttrsIntrinsic<[llvm_float_ty], [llvm_float_ty], [IntrNoMem]>; DefaultAttrsIntrinsic<[llvm_float_ty], [llvm_float_ty], [IntrNoMem]>;
def int_nvvm_lg2_approx_f : GCCBuiltin<"__nvvm_lg2_approx_f">, def int_nvvm_lg2_approx_f : GCCBuiltin<"__nvvm_lg2_approx_f">,
DefaultAttrsIntrinsic<[llvm_float_ty], [llvm_float_ty], [IntrNoMem]>; DefaultAttrsIntrinsic<[llvm_float_ty], [llvm_float_ty], [IntrNoMem]>;
def int_nvvm_lg2_approx_d : GCCBuiltin<"__nvvm_lg2_approx_d">, def int_nvvm_lg2_approx_d : GCCBuiltin<"__nvvm_lg2_approx_d">,
DefaultAttrsIntrinsic<[llvm_double_ty], [llvm_double_ty], [IntrNoMem]>; DefaultAttrsIntrinsic<[llvm_double_ty], [llvm_double_ty], [IntrNoMem]>;
▲ Show 20 Lines • Show All 3,821 Lines • Show Last 20 Lines

llvm/lib/Target/NVPTX/NVPTXIntrinsics.td

  • This file is larger than 256 KB, so syntax highlighting is disabled by default.
Show First 20 Lines • Show All 901 Lines • ▼ Show 20 Lines
// //
def INT_NVVM_EX2_APPROX_FTZ_F : F_MATH_1<"ex2.approx.ftz.f32 \t$dst, $src0;", def INT_NVVM_EX2_APPROX_FTZ_F : F_MATH_1<"ex2.approx.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_ex2_approx_ftz_f>; Float32Regs, Float32Regs, int_nvvm_ex2_approx_ftz_f>;
def INT_NVVM_EX2_APPROX_F : F_MATH_1<"ex2.approx.f32 \t$dst, $src0;", def INT_NVVM_EX2_APPROX_F : F_MATH_1<"ex2.approx.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_ex2_approx_f>; Float32Regs, Float32Regs, int_nvvm_ex2_approx_f>;
def INT_NVVM_EX2_APPROX_D : F_MATH_1<"ex2.approx.f64 \t$dst, $src0;", def INT_NVVM_EX2_APPROX_D : F_MATH_1<"ex2.approx.f64 \t$dst, $src0;",
Float64Regs, Float64Regs, int_nvvm_ex2_approx_d>; Float64Regs, Float64Regs, int_nvvm_ex2_approx_d>;
def INT_NVVM_EX2_APPROX_F16 : F_MATH_1<"ex2.approx.f16 \t$dst, $src0;",
Float16Regs, Float16Regs, int_nvvm_ex2_approx_f16, [hasPTX70, hasSM75]>;
def INT_NVVM_EX2_APPROX_F16X2 : F_MATH_1<"ex2.approx.f16x2 \t$dst, $src0;",
Float16x2Regs, Float16x2Regs, int_nvvm_ex2_approx_f16x2, [hasPTX70, hasSM75]>;
def INT_NVVM_LG2_APPROX_FTZ_F : F_MATH_1<"lg2.approx.ftz.f32 \t$dst, $src0;", def INT_NVVM_LG2_APPROX_FTZ_F : F_MATH_1<"lg2.approx.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_lg2_approx_ftz_f>; Float32Regs, Float32Regs, int_nvvm_lg2_approx_ftz_f>;
def INT_NVVM_LG2_APPROX_F : F_MATH_1<"lg2.approx.f32 \t$dst, $src0;", def INT_NVVM_LG2_APPROX_F : F_MATH_1<"lg2.approx.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_lg2_approx_f>; Float32Regs, Float32Regs, int_nvvm_lg2_approx_f>;
def INT_NVVM_LG2_APPROX_D : F_MATH_1<"lg2.approx.f64 \t$dst, $src0;", def INT_NVVM_LG2_APPROX_D : F_MATH_1<"lg2.approx.f64 \t$dst, $src0;",
Float64Regs, Float64Regs, int_nvvm_lg2_approx_d>; Float64Regs, Float64Regs, int_nvvm_lg2_approx_d>;
▲ Show 20 Lines • Show All 5,773 Lines • Show Last 20 Lines

llvm/test/CodeGen/NVPTX/f16-ex2.ll

  • This file was added.
; RUN: llc < %s -march=nvptx64 -mcpu=sm_75 -mattr=+ptx70 | FileCheck %s
declare half @llvm.nvvm.ex2.approx.f16(half)
declare <2 x half> @llvm.nvvm.ex2.approx.f16x2(<2 x half>)
; CHECK-LABEL: exp2_half
define half @exp2_half(half %0) {
; CHECK-NOT: call
; CHECK: ex2.approx.f16
%res = call half @llvm.nvvm.ex2.approx.f16(half %0);
ret half %res
}
; CHECK-LABEL: exp2_2xhalf
define <2 x half> @exp2_2xhalf(<2 x half> %0) {
; CHECK-NOT: call
; CHECK: ex2.approx.f16x2
%res = call <2 x half> @llvm.nvvm.ex2.approx.f16x2(<2 x half> %0);
ret <2 x half> %res
}