Skip to content

[NVPTX] Add folding for cvt.rn.bf16x2.f32 #116109

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Nov 14, 2024
Merged

[NVPTX] Add folding for cvt.rn.bf16x2.f32 #116109

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
45db970
[NVPTX] Add folding for cvt.rn.bf16x2.f32
AlexMaclean Nov 11, 2024
3f8f50d
address comments
AlexMaclean Nov 13, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 14 additions & 0 deletions llvm/lib/Target/NVPTX/NVPTXInstrInfo.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -739,6 +739,20 @@ let hasSideEffects = false in {
def CVT_f16x2_e5m2x2 : CVT_f16x2_fp8<"e5m2">;
}

def fpround_oneuse : PatFrag<(ops node:$a), (fpround node:$a), [{
return N->hasOneUse();
}]>;

def : Pat<(v2bf16 (build_vector (bf16 (fpround_oneuse Float32Regs:$a)),
(bf16 (fpround_oneuse Float32Regs:$b)))),
(CVT_bf16x2_f32 Float32Regs:$a, Float32Regs:$b, CvtRN)>,
Requires<[hasPTX<70>, hasSM<80>, hasBF16Math]>;

def : Pat<(v2f16 (build_vector (f16 (fpround_oneuse Float32Regs:$a)),
(f16 (fpround_oneuse Float32Regs:$b)))),
(CVT_f16x2_f32 Float32Regs:$a, Float32Regs:$b, CvtRN)>,
Requires<[hasPTX<70>, hasSM<80>, useFP16Math]>;

//-----------------------------------
// Selection instructions (selp)
//-----------------------------------
Expand Down
54 changes: 18 additions & 36 deletions llvm/test/CodeGen/NVPTX/bf16-instructions.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -204,7 +204,7 @@ define <2 x bfloat> @test_faddx2(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
;
; SM80-LABEL: test_faddx2(
; SM80: {
; SM80-NEXT: .reg .b16 %rs<7>;
; SM80-NEXT: .reg .b16 %rs<5>;
; SM80-NEXT: .reg .b32 %r<4>;
; SM80-NEXT: .reg .f32 %f<7>;
; SM80-EMPTY:
Expand All @@ -216,18 +216,16 @@ define <2 x bfloat> @test_faddx2(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
; SM80-NEXT: mov.b32 {%rs3, %rs4}, %r1;
; SM80-NEXT: cvt.f32.bf16 %f2, %rs4;
; SM80-NEXT: add.rn.f32 %f3, %f2, %f1;
; SM80-NEXT: cvt.rn.bf16.f32 %rs5, %f3;
; SM80-NEXT: cvt.f32.bf16 %f4, %rs1;
; SM80-NEXT: cvt.f32.bf16 %f5, %rs3;
; SM80-NEXT: add.rn.f32 %f6, %f5, %f4;
; SM80-NEXT: cvt.rn.bf16.f32 %rs6, %f6;
; SM80-NEXT: mov.b32 %r3, {%rs6, %rs5};
; SM80-NEXT: cvt.rn.bf16x2.f32 %r3, %f6, %f3;
; SM80-NEXT: st.param.b32 [func_retval0], %r3;
; SM80-NEXT: ret;
;
; SM80-FTZ-LABEL: test_faddx2(
; SM80-FTZ: {
; SM80-FTZ-NEXT: .reg .b16 %rs<7>;
; SM80-FTZ-NEXT: .reg .b16 %rs<5>;
; SM80-FTZ-NEXT: .reg .b32 %r<4>;
; SM80-FTZ-NEXT: .reg .f32 %f<7>;
; SM80-FTZ-EMPTY:
Expand All @@ -239,12 +237,10 @@ define <2 x bfloat> @test_faddx2(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
; SM80-FTZ-NEXT: mov.b32 {%rs3, %rs4}, %r1;
; SM80-FTZ-NEXT: cvt.ftz.f32.bf16 %f2, %rs4;
; SM80-FTZ-NEXT: add.rn.ftz.f32 %f3, %f2, %f1;
; SM80-FTZ-NEXT: cvt.rn.bf16.f32 %rs5, %f3;
; SM80-FTZ-NEXT: cvt.ftz.f32.bf16 %f4, %rs1;
; SM80-FTZ-NEXT: cvt.ftz.f32.bf16 %f5, %rs3;
; SM80-FTZ-NEXT: add.rn.ftz.f32 %f6, %f5, %f4;
; SM80-FTZ-NEXT: cvt.rn.bf16.f32 %rs6, %f6;
; SM80-FTZ-NEXT: mov.b32 %r3, {%rs6, %rs5};
; SM80-FTZ-NEXT: cvt.rn.bf16x2.f32 %r3, %f6, %f3;
; SM80-FTZ-NEXT: st.param.b32 [func_retval0], %r3;
; SM80-FTZ-NEXT: ret;
;
Expand Down Expand Up @@ -311,7 +307,7 @@ define <2 x bfloat> @test_fsubx2(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
;
; SM80-LABEL: test_fsubx2(
; SM80: {
; SM80-NEXT: .reg .b16 %rs<7>;
; SM80-NEXT: .reg .b16 %rs<5>;
; SM80-NEXT: .reg .b32 %r<4>;
; SM80-NEXT: .reg .f32 %f<7>;
; SM80-EMPTY:
Expand All @@ -323,18 +319,16 @@ define <2 x bfloat> @test_fsubx2(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
; SM80-NEXT: mov.b32 {%rs3, %rs4}, %r1;
; SM80-NEXT: cvt.f32.bf16 %f2, %rs4;
; SM80-NEXT: sub.rn.f32 %f3, %f2, %f1;
; SM80-NEXT: cvt.rn.bf16.f32 %rs5, %f3;
; SM80-NEXT: cvt.f32.bf16 %f4, %rs1;
; SM80-NEXT: cvt.f32.bf16 %f5, %rs3;
; SM80-NEXT: sub.rn.f32 %f6, %f5, %f4;
; SM80-NEXT: cvt.rn.bf16.f32 %rs6, %f6;
; SM80-NEXT: mov.b32 %r3, {%rs6, %rs5};
; SM80-NEXT: cvt.rn.bf16x2.f32 %r3, %f6, %f3;
; SM80-NEXT: st.param.b32 [func_retval0], %r3;
; SM80-NEXT: ret;
;
; SM80-FTZ-LABEL: test_fsubx2(
; SM80-FTZ: {
; SM80-FTZ-NEXT: .reg .b16 %rs<7>;
; SM80-FTZ-NEXT: .reg .b16 %rs<5>;
; SM80-FTZ-NEXT: .reg .b32 %r<4>;
; SM80-FTZ-NEXT: .reg .f32 %f<7>;
; SM80-FTZ-EMPTY:
Expand All @@ -346,12 +340,10 @@ define <2 x bfloat> @test_fsubx2(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
; SM80-FTZ-NEXT: mov.b32 {%rs3, %rs4}, %r1;
; SM80-FTZ-NEXT: cvt.ftz.f32.bf16 %f2, %rs4;
; SM80-FTZ-NEXT: sub.rn.ftz.f32 %f3, %f2, %f1;
; SM80-FTZ-NEXT: cvt.rn.bf16.f32 %rs5, %f3;
; SM80-FTZ-NEXT: cvt.ftz.f32.bf16 %f4, %rs1;
; SM80-FTZ-NEXT: cvt.ftz.f32.bf16 %f5, %rs3;
; SM80-FTZ-NEXT: sub.rn.ftz.f32 %f6, %f5, %f4;
; SM80-FTZ-NEXT: cvt.rn.bf16.f32 %rs6, %f6;
; SM80-FTZ-NEXT: mov.b32 %r3, {%rs6, %rs5};
; SM80-FTZ-NEXT: cvt.rn.bf16x2.f32 %r3, %f6, %f3;
; SM80-FTZ-NEXT: st.param.b32 [func_retval0], %r3;
; SM80-FTZ-NEXT: ret;
;
Expand Down Expand Up @@ -418,7 +410,7 @@ define <2 x bfloat> @test_fmulx2(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
;
; SM80-LABEL: test_fmulx2(
; SM80: {
; SM80-NEXT: .reg .b16 %rs<7>;
; SM80-NEXT: .reg .b16 %rs<5>;
; SM80-NEXT: .reg .b32 %r<4>;
; SM80-NEXT: .reg .f32 %f<7>;
; SM80-EMPTY:
Expand All @@ -430,18 +422,16 @@ define <2 x bfloat> @test_fmulx2(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
; SM80-NEXT: mov.b32 {%rs3, %rs4}, %r1;
; SM80-NEXT: cvt.f32.bf16 %f2, %rs4;
; SM80-NEXT: mul.rn.f32 %f3, %f2, %f1;
; SM80-NEXT: cvt.rn.bf16.f32 %rs5, %f3;
; SM80-NEXT: cvt.f32.bf16 %f4, %rs1;
; SM80-NEXT: cvt.f32.bf16 %f5, %rs3;
; SM80-NEXT: mul.rn.f32 %f6, %f5, %f4;
; SM80-NEXT: cvt.rn.bf16.f32 %rs6, %f6;
; SM80-NEXT: mov.b32 %r3, {%rs6, %rs5};
; SM80-NEXT: cvt.rn.bf16x2.f32 %r3, %f6, %f3;
; SM80-NEXT: st.param.b32 [func_retval0], %r3;
; SM80-NEXT: ret;
;
; SM80-FTZ-LABEL: test_fmulx2(
; SM80-FTZ: {
; SM80-FTZ-NEXT: .reg .b16 %rs<7>;
; SM80-FTZ-NEXT: .reg .b16 %rs<5>;
; SM80-FTZ-NEXT: .reg .b32 %r<4>;
; SM80-FTZ-NEXT: .reg .f32 %f<7>;
; SM80-FTZ-EMPTY:
Expand All @@ -453,12 +443,10 @@ define <2 x bfloat> @test_fmulx2(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
; SM80-FTZ-NEXT: mov.b32 {%rs3, %rs4}, %r1;
; SM80-FTZ-NEXT: cvt.ftz.f32.bf16 %f2, %rs4;
; SM80-FTZ-NEXT: mul.rn.ftz.f32 %f3, %f2, %f1;
; SM80-FTZ-NEXT: cvt.rn.bf16.f32 %rs5, %f3;
; SM80-FTZ-NEXT: cvt.ftz.f32.bf16 %f4, %rs1;
; SM80-FTZ-NEXT: cvt.ftz.f32.bf16 %f5, %rs3;
; SM80-FTZ-NEXT: mul.rn.ftz.f32 %f6, %f5, %f4;
; SM80-FTZ-NEXT: cvt.rn.bf16.f32 %rs6, %f6;
; SM80-FTZ-NEXT: mov.b32 %r3, {%rs6, %rs5};
; SM80-FTZ-NEXT: cvt.rn.bf16x2.f32 %r3, %f6, %f3;
; SM80-FTZ-NEXT: st.param.b32 [func_retval0], %r3;
; SM80-FTZ-NEXT: ret;
;
Expand Down Expand Up @@ -525,7 +513,7 @@ define <2 x bfloat> @test_fdiv(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
;
; SM80-LABEL: test_fdiv(
; SM80: {
; SM80-NEXT: .reg .b16 %rs<7>;
; SM80-NEXT: .reg .b16 %rs<5>;
; SM80-NEXT: .reg .b32 %r<4>;
; SM80-NEXT: .reg .f32 %f<7>;
; SM80-EMPTY:
Expand All @@ -537,18 +525,16 @@ define <2 x bfloat> @test_fdiv(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
; SM80-NEXT: mov.b32 {%rs3, %rs4}, %r1;
; SM80-NEXT: cvt.f32.bf16 %f2, %rs4;
; SM80-NEXT: div.rn.f32 %f3, %f2, %f1;
; SM80-NEXT: cvt.rn.bf16.f32 %rs5, %f3;
; SM80-NEXT: cvt.f32.bf16 %f4, %rs1;
; SM80-NEXT: cvt.f32.bf16 %f5, %rs3;
; SM80-NEXT: div.rn.f32 %f6, %f5, %f4;
; SM80-NEXT: cvt.rn.bf16.f32 %rs6, %f6;
; SM80-NEXT: mov.b32 %r3, {%rs6, %rs5};
; SM80-NEXT: cvt.rn.bf16x2.f32 %r3, %f6, %f3;
; SM80-NEXT: st.param.b32 [func_retval0], %r3;
; SM80-NEXT: ret;
;
; SM80-FTZ-LABEL: test_fdiv(
; SM80-FTZ: {
; SM80-FTZ-NEXT: .reg .b16 %rs<7>;
; SM80-FTZ-NEXT: .reg .b16 %rs<5>;
; SM80-FTZ-NEXT: .reg .b32 %r<4>;
; SM80-FTZ-NEXT: .reg .f32 %f<7>;
; SM80-FTZ-EMPTY:
Expand All @@ -560,18 +546,16 @@ define <2 x bfloat> @test_fdiv(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
; SM80-FTZ-NEXT: mov.b32 {%rs3, %rs4}, %r1;
; SM80-FTZ-NEXT: cvt.ftz.f32.bf16 %f2, %rs4;
; SM80-FTZ-NEXT: div.rn.ftz.f32 %f3, %f2, %f1;
; SM80-FTZ-NEXT: cvt.rn.bf16.f32 %rs5, %f3;
; SM80-FTZ-NEXT: cvt.ftz.f32.bf16 %f4, %rs1;
; SM80-FTZ-NEXT: cvt.ftz.f32.bf16 %f5, %rs3;
; SM80-FTZ-NEXT: div.rn.ftz.f32 %f6, %f5, %f4;
; SM80-FTZ-NEXT: cvt.rn.bf16.f32 %rs6, %f6;
; SM80-FTZ-NEXT: mov.b32 %r3, {%rs6, %rs5};
; SM80-FTZ-NEXT: cvt.rn.bf16x2.f32 %r3, %f6, %f3;
; SM80-FTZ-NEXT: st.param.b32 [func_retval0], %r3;
; SM80-FTZ-NEXT: ret;
;
; SM90-LABEL: test_fdiv(
; SM90: {
; SM90-NEXT: .reg .b16 %rs<7>;
; SM90-NEXT: .reg .b16 %rs<5>;
; SM90-NEXT: .reg .b32 %r<4>;
; SM90-NEXT: .reg .f32 %f<7>;
; SM90-EMPTY:
Expand All @@ -583,12 +567,10 @@ define <2 x bfloat> @test_fdiv(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
; SM90-NEXT: mov.b32 {%rs3, %rs4}, %r1;
; SM90-NEXT: cvt.f32.bf16 %f2, %rs4;
; SM90-NEXT: div.rn.f32 %f3, %f2, %f1;
; SM90-NEXT: cvt.rn.bf16.f32 %rs5, %f3;
; SM90-NEXT: cvt.f32.bf16 %f4, %rs1;
; SM90-NEXT: cvt.f32.bf16 %f5, %rs3;
; SM90-NEXT: div.rn.f32 %f6, %f5, %f4;
; SM90-NEXT: cvt.rn.bf16.f32 %rs6, %f6;
; SM90-NEXT: mov.b32 %r3, {%rs6, %rs5};
; SM90-NEXT: cvt.rn.bf16x2.f32 %r3, %f6, %f3;
; SM90-NEXT: st.param.b32 [func_retval0], %r3;
; SM90-NEXT: ret;
%r = fdiv <2 x bfloat> %a, %b
Expand Down
8 changes: 2 additions & 6 deletions llvm/test/CodeGen/NVPTX/bf16x2-instructions-approx.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,9 +13,7 @@ declare <2 x bfloat> @llvm.cos.f16(<2 x bfloat> %a) #0
; CHECK-DAG: cvt.f32.bf16 [[AF1:%f[0-9]+]], [[A1]];
; CHECK-DAG: sin.approx.f32 [[RF0:%f[0-9]+]], [[AF0]];
; CHECK-DAG: sin.approx.f32 [[RF1:%f[0-9]+]], [[AF1]];
; CHECK-DAG: cvt.rn.bf16.f32 [[R0:%rs[0-9]+]], [[RF0]];
; CHECK-DAG: cvt.rn.bf16.f32 [[R1:%rs[0-9]+]], [[RF1]];
; CHECK: mov.b32 [[R:%r[0-9]+]], {[[R0]], [[R1]]}
; CHECK: cvt.rn.bf16x2.f32 [[R:%r[0-9]+]], [[RF0]], [[RF1]]
; CHECK: st.param.b32 [func_retval0], [[R]];
; CHECK: ret;
define <2 x bfloat> @test_sin(<2 x bfloat> %a) #0 #1 {
Expand All @@ -30,9 +28,7 @@ define <2 x bfloat> @test_sin(<2 x bfloat> %a) #0 #1 {
; CHECK-DAG: cvt.f32.bf16 [[AF1:%f[0-9]+]], [[A1]];
; CHECK-DAG: cos.approx.f32 [[RF0:%f[0-9]+]], [[AF0]];
; CHECK-DAG: cos.approx.f32 [[RF1:%f[0-9]+]], [[AF1]];
; CHECK-DAG: cvt.rn.bf16.f32 [[R0:%rs[0-9]+]], [[RF0]];
; CHECK-DAG: cvt.rn.bf16.f32 [[R1:%rs[0-9]+]], [[RF1]];
; CHECK: mov.b32 [[R:%r[0-9]+]], {[[R0]], [[R1]]}
; CHECK: cvt.rn.bf16x2.f32 [[R:%r[0-9]+]], [[RF0]], [[RF1]]
; CHECK: st.param.b32 [func_retval0], [[R]];
; CHECK: ret;
define <2 x bfloat> @test_cos(<2 x bfloat> %a) #0 #1 {
Expand Down
36 changes: 10 additions & 26 deletions llvm/test/CodeGen/NVPTX/bf16x2-instructions.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,9 +26,7 @@ define <2 x bfloat> @test_ret_const() #0 {
; SM80-DAG: cvt.f32.bf16 [[FA1:%f[0-9]+]], [[A1]]
; SM80-DAG: add.rn.f32 [[FR0:%f[0-9]+]], [[FA0]], 0f3F800000;
; SM80-DAG: add.rn.f32 [[FR1:%f[0-9]+]], [[FA1]], 0f40000000;
; SM80-DAG: cvt.rn.bf16.f32 [[R0:%rs[0-9]+]], [[FR0]]
; SM80-DAG: cvt.rn.bf16.f32 [[R1:%rs[0-9]+]], [[FR1]]
; SM80-DAG: mov.b32 [[R:%r[0-9]+]], {[[R0]], [[R1]]}
; SM80-DAG: cvt.rn.bf16x2.f32 [[R:%r[0-9]+]], [[FR0]], [[FR1]];
;
; CHECK-NEXT: st.param.b32 [func_retval0], [[R]];
; CHECK-NEXT: ret;
Expand Down Expand Up @@ -68,9 +66,7 @@ define bfloat @test_fadd_imm_1(bfloat %a) #0 {
; SM80-DAG: cvt.f32.bf16 [[FB1:%f[0-9]+]], [[B1]];
; SM80-DAG: sub.rn.f32 [[FR0:%f[0-9]+]], [[FA0]], [[FB0]];
; SM80-DAG: sub.rn.f32 [[FR1:%f[0-9]+]], [[FA1]], [[FB1]];
; SM80-DAG: cvt.rn.bf16.f32 [[R0:%rs[0-9]+]], [[FR0]];
; SM80-DAG: cvt.rn.bf16.f32 [[R1:%rs[0-9]+]], [[FR1]];
; SM80: mov.b32 [[R:%r[0-9]+]], {[[R0]], [[R1]]};
; SM80-DAG: cvt.rn.bf16x2.f32 [[R:%r[0-9]+]], [[FR0]], [[FR1]];

; CHECK: st.param.b32 [func_retval0], [[R]];
; CHECK: ret;
Expand All @@ -93,9 +89,7 @@ define <2 x bfloat> @test_fsubx2(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
; SM80-DAG: cvt.f32.bf16 [[FB1:%f[0-9]+]], [[B1]];
; SM80-DAG: mul.rn.f32 [[FR0:%f[0-9]+]], [[FA0]], [[FB0]];
; SM80-DAG: mul.rn.f32 [[FR1:%f[0-9]+]], [[FA1]], [[FB1]];
; SM80-DAG: cvt.rn.bf16.f32 [[R0:%rs[0-9]+]], [[FR0]];
; SM80-DAG: cvt.rn.bf16.f32 [[R1:%rs[0-9]+]], [[FR1]];
; SM80: mov.b32 [[R:%r[0-9]+]], {[[R0]], [[R1]]};
; SM80-DAG: cvt.rn.bf16x2.f32 [[R:%r[0-9]+]], [[FR0]], [[FR1]];

; CHECK: st.param.b32 [func_retval0], [[R]];
; CHECK: ret;
Expand All @@ -116,9 +110,7 @@ define <2 x bfloat> @test_fmulx2(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
; CHECK-DAG: cvt.f32.bf16 [[FB1:%f[0-9]+]], [[B1]];
; CHECK-DAG: div.rn.f32 [[FR0:%f[0-9]+]], [[FA0]], [[FB0]];
; CHECK-DAG: div.rn.f32 [[FR1:%f[0-9]+]], [[FA1]], [[FB1]];
; CHECK-DAG: cvt.rn.bf16.f32 [[R0:%rs[0-9]+]], [[FR0]];
; CHECK-DAG: cvt.rn.bf16.f32 [[R1:%rs[0-9]+]], [[FR1]];
; CHECK-NEXT: mov.b32 [[R:%r[0-9]+]], {[[R0]], [[R1]]}
; CHECK: cvt.rn.bf16x2.f32 [[R:%r[0-9]+]], [[FR0]], [[FR1]];
; CHECK-NEXT: st.param.b32 [func_retval0], [[R]];
; CHECK-NEXT: ret;

Expand Down Expand Up @@ -287,9 +279,7 @@ define <2 x bfloat> @test_select_cc_bf16_f32(<2 x bfloat> %a, <2 x bfloat> %b,

; CHECK-LABEL: test_fptrunc_2xfloat(
; CHECK: ld.param.v2.f32 {[[A0:%f[0-9]+]], [[A1:%f[0-9]+]]}, [test_fptrunc_2xfloat_param_0];
; CHECK-DAG: cvt.rn.bf16.f32 [[R0:%rs[0-9]+]], [[A0]];
; CHECK-DAG: cvt.rn.bf16.f32 [[R1:%rs[0-9]+]], [[A1]];
; CHECK: mov.b32 [[R:%r[0-9]+]], {[[R0]], [[R1]]}
; CHECK: cvt.rn.bf16x2.f32 [[R:%r[0-9]+]], [[A0]], [[A1]];
; CHECK: st.param.b32 [func_retval0], [[R]];
; CHECK: ret;
define <2 x bfloat> @test_fptrunc_2xfloat(<2 x float> %a) #0 {
Expand Down Expand Up @@ -359,9 +349,7 @@ declare <2 x bfloat> @llvm.fmuladd.f16(<2 x bfloat> %a, <2 x bfloat> %b, <2 x bf
; CHECK-DAG: cvt.f32.bf16 [[AF1:%f[0-9]+]], [[A1]];
; CHECK-DAG: sqrt.rn.f32 [[RF0:%f[0-9]+]], [[AF0]];
; CHECK-DAG: sqrt.rn.f32 [[RF1:%f[0-9]+]], [[AF1]];
; CHECK-DAG: cvt.rn.bf16.f32 [[R0:%rs[0-9]+]], [[RF0]];
; CHECK-DAG: cvt.rn.bf16.f32 [[R1:%rs[0-9]+]], [[RF1]];
; CHECK: mov.b32 [[R:%r[0-9]+]], {[[R0]], [[R1]]}
; CHECK-DAG: cvt.rn.bf16x2.f32 [[R:%r[0-9]+]], [[RF0]], [[RF1]];
; CHECK: st.param.b32 [func_retval0], [[R]];
; CHECK: ret;
define <2 x bfloat> @test_sqrt(<2 x bfloat> %a) #0 {
Expand Down Expand Up @@ -436,9 +424,7 @@ define <2 x bfloat> @test_maxnum(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
; SM80-DAG: cvt.f32.bf16 [[FA1:%f[0-9]+]], [[A1]];
; SM80-DAG: cvt.rmi.f32.f32 [[RF0:%f[0-9]+]], [[FA0]];
; SM80-DAG: cvt.rmi.f32.f32 [[RF1:%f[0-9]+]], [[FA1]];
; SM80-DAG: cvt.rn.bf16.f32 [[R0:%rs[0-9]+]], [[RF0]];
; SM80-DAG: cvt.rn.bf16.f32 [[R1:%rs[0-9]+]], [[RF1]];
; CHECK: mov.b32 [[R:%r[0-9]+]], {[[R0]], [[R1]]}
; SM80: cvt.rn.bf16x2.f32 [[R:%r[0-9]+]], [[RF0]], [[RF1]];
; CHECK: st.param.b32 [func_retval0], [[R]];
; CHECK: ret;
define <2 x bfloat> @test_floor(<2 x bfloat> %a) #0 {
Expand All @@ -455,9 +441,7 @@ define <2 x bfloat> @test_floor(<2 x bfloat> %a) #0 {
; SM80-DAG: cvt.f32.bf16 [[FA1:%f[0-9]+]], [[A1]];
; SM80-DAG: cvt.rpi.f32.f32 [[RF0:%f[0-9]+]], [[FA0]];
; SM80-DAG: cvt.rpi.f32.f32 [[RF1:%f[0-9]+]], [[FA1]];
; SM80-DAG: cvt.rn.bf16.f32 [[R0:%rs[0-9]+]], [[RF0]];
; SM80-DAG: cvt.rn.bf16.f32 [[R1:%rs[0-9]+]], [[RF1]];
; CHECK: mov.b32 [[R:%r[0-9]+]], {[[R0]], [[R1]]}
; SM80: cvt.rn.bf16x2.f32 [[R:%r[0-9]+]], [[RF0]], [[RF1]];
; CHECK: st.param.b32 [func_retval0], [[R]];
; CHECK: ret;
define <2 x bfloat> @test_ceil(<2 x bfloat> %a) #0 {
Expand All @@ -470,7 +454,7 @@ define <2 x bfloat> @test_ceil(<2 x bfloat> %a) #0 {
; CHECK-DAG: mov.b32 {[[A0:%rs[0-9]+]], [[A1:%rs[0-9]+]]}, [[A]];
; SM90: cvt.rzi.bf16.bf16 [[R1:%rs[0-9]+]], [[A1]];
; SM90: cvt.rzi.bf16.bf16 [[R0:%rs[0-9]+]], [[A0]];
; CHECK: mov.b32 [[R:%r[0-9]+]], {[[R0]], [[R1]]}
; SM90: mov.b32 [[R:%r[0-9]+]], {[[R0]], [[R1]]}
; CHECK: st.param.b32 [func_retval0], [[R]];
; CHECK: ret;
define <2 x bfloat> @test_trunc(<2 x bfloat> %a) #0 {
Expand All @@ -483,7 +467,7 @@ define <2 x bfloat> @test_trunc(<2 x bfloat> %a) #0 {
; CHECK-DAG: mov.b32 {[[A0:%rs[0-9]+]], [[A1:%rs[0-9]+]]}, [[A]];
; SM90: cvt.rni.bf16.bf16 [[R1:%rs[0-9]+]], [[A1]];
; SM90: cvt.rni.bf16.bf16 [[R0:%rs[0-9]+]], [[A0]];
; CHECK: mov.b32 [[R:%r[0-9]+]], {[[R0]], [[R1]]}
; SM90: mov.b32 [[R:%r[0-9]+]], {[[R0]], [[R1]]}
; CHECK: st.param.b32 [func_retval0], [[R]];
; CHECK: ret;
define <2 x bfloat> @test_rint(<2 x bfloat> %a) #0 {
Expand Down
Loading
Loading