================
@@ -2181,27 +2183,43 @@ def : Pat<(int_nvvm_mulhi_ui i32:$a, i32:$b),
(MUL_HI_U32rr $a, $b)>;
def : Pat<(int_nvvm_mulhi_ll i64:$a, i64:$b), (MUL_HI_S64rr $a, $b)>;
def : Pat<(int_nvvm_mulhi_ull i64:$a, i64:$b), (MUL_HI_U64rr $a, $b)>;
-def INT_NVVM_MUL_RN_FTZ_F : F_MATH_2<"mul.rn.ftz.f32", B32, B32, B32,
int_nvvm_mul_rn_ftz_f>;
-def INT_NVVM_MUL_RN_F : F_MATH_2<"mul.rn.f32", B32, B32, B32,
int_nvvm_mul_rn_f>;
-def INT_NVVM_MUL_RZ_FTZ_F : F_MATH_2<"mul.rz.ftz.f32", B32, B32, B32,
int_nvvm_mul_rz_ftz_f>;
-def INT_NVVM_MUL_RZ_F : F_MATH_2<"mul.rz.f32", B32, B32, B32,
int_nvvm_mul_rz_f>;
-def INT_NVVM_MUL_RM_FTZ_F : F_MATH_2<"mul.rm.ftz.f32", B32, B32, B32,
int_nvvm_mul_rm_ftz_f>;
-def INT_NVVM_MUL_RM_F : F_MATH_2<"mul.rm.f32", B32, B32, B32,
int_nvvm_mul_rm_f>;
-def INT_NVVM_MUL_RP_FTZ_F : F_MATH_2<"mul.rp.ftz.f32", B32, B32, B32,
int_nvvm_mul_rp_ftz_f>;
-def INT_NVVM_MUL_RP_F : F_MATH_2<"mul.rp.f32", B32, B32, B32,
int_nvvm_mul_rp_f>;
-
-def INT_NVVM_MUL_RN_D : F_MATH_2<"mul.rn.f64", B64, B64, B64,
int_nvvm_mul_rn_d>;
-def INT_NVVM_MUL_RZ_D : F_MATH_2<"mul.rz.f64", B64, B64, B64,
int_nvvm_mul_rz_d>;
-def INT_NVVM_MUL_RM_D : F_MATH_2<"mul.rm.f64", B64, B64, B64,
int_nvvm_mul_rm_d>;
-def INT_NVVM_MUL_RP_D : F_MATH_2<"mul.rp.f64", B64, B64, B64,
int_nvvm_mul_rp_d>;
-
def INT_NVVM_MUL24_I : F_MATH_2<"mul24.lo.s32", B32, B32, B32,
int_nvvm_mul24_i>;
def INT_NVVM_MUL24_UI : F_MATH_2<"mul24.lo.u32", B32, B32, B32,
int_nvvm_mul24_ui>;
-def INT_NVVM_MUL_RN_SAT_F16 : F_MATH_2<"mul.rn.sat.f16", B16, B16, B16,
int_nvvm_mul_rn_sat_f16>;
-def INT_NVVM_MUL_RN_FTZ_SAT_F16 : F_MATH_2<"mul.rn.ftz.sat.f16", B16, B16,
B16, int_nvvm_mul_rn_ftz_sat_f16>;
-def INT_NVVM_MUL_RN_SAT_F16X2 : F_MATH_2<"mul.rn.sat.f16x2", B32, B32, B32,
int_nvvm_mul_rn_sat_v2f16>;
-def INT_NVVM_MUL_RN_FTZ_SAT_F16X2 : F_MATH_2<"mul.rn.ftz.sat.f16x2", B32, B32,
B32, int_nvvm_mul_rn_ftz_sat_v2f16>;
+// f16/f16x2
+foreach t = [F16RT, F16X2RT] in
+ foreach ftz = ["", "ftz"] in
+ foreach sat = ["", "sat"] in
+ def INT_NVVM_MUL_RN_ # !toupper(StrJoin<"_", [ftz, sat, t.PtxType]>.ret)
:
+ F_MATH_2_RND_TY<StrJoin<".", ["mul.rn", ftz, sat, t.PtxType]>.ret, t,
+ !cast<Intrinsic>(
+ StrJoin<"_", ["int_nvvm_fmul", ftz, sat]>.ret),
+ fp_rnd_rn_imm>;
+
+// bf16/bf16x2
+foreach t = [BF16RT, BF16X2RT] in
+ def INT_NVVM_MUL_RN_ # !toupper(t.PtxType) :
+ F_MATH_2_RND_TY<"mul.rn." # t.PtxType, t, int_nvvm_fmul, fp_rnd_rn_imm,
+ BF16ArithPreds>;
+
+// f32/f32x2/f64
+foreach ftz = ["", "ftz"] in {
+ foreach sat = ["", "sat"] in
+ def StrJoin<"_", ["INT_NVVM_MUL", !toupper(ftz), !toupper(sat), "F"]>.ret :
+ F_MATH_2_RNDOP_TY<StrJoin<".", ["mul.${rnd}", ftz, sat, "f32"]>.ret,
+ F32RT,
+ !cast<Intrinsic>(
+ StrJoin<"_", ["int_nvvm_fmul", ftz, sat]>.ret)>;
----------------
durga4github wrote:
fold into 2210
https://github.com/llvm/llvm-project/pull/224546
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
