[llvm-branch-commits] [clang] [llvm] [AMDGPU] Add support for `v_tanh_f32` on gfx1250 (PR #149360)
Shilei Tian via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Thu Jul 17 10:17:24 PDT 2025
https://github.com/shiltian created https://github.com/llvm/llvm-project/pull/149360
Co-authored-by: Mekhanoshin, Stanislav <Stanislav.Mekhanoshin at amd.com>
>From 2fa6c545f78a345feb30c1ac27e9874106b5870c Mon Sep 17 00:00:00 2001
From: Shilei Tian <i at tianshilei.me>
Date: Thu, 17 Jul 2025 13:03:14 -0400
Subject: [PATCH] [AMDGPU] Add support for `v_tanh_f32` on gfx1250
Co-authored-by: Mekhanoshin, Stanislav <Stanislav.Mekhanoshin at amd.com>
---
clang/include/clang/Basic/BuiltinsAMDGPU.def | 1 +
.../CodeGenOpenCL/builtins-amdgcn-gfx1250.cl | 19 +++++
llvm/lib/Target/AMDGPU/AMDGPU.td | 10 +++
llvm/lib/Target/AMDGPU/GCNSubtarget.h | 3 +
llvm/lib/Target/AMDGPU/VOP1Instructions.td | 4 +
llvm/lib/TargetParser/TargetParser.cpp | 1 +
llvm/test/CodeGen/AMDGPU/llvm.amdgcn.tanh.ll | 84 +++++++++++++++++++
7 files changed, 122 insertions(+)
diff --git a/clang/include/clang/Basic/BuiltinsAMDGPU.def b/clang/include/clang/Basic/BuiltinsAMDGPU.def
index 3b6ad7d90be3c..4111837d962b5 100644
--- a/clang/include/clang/Basic/BuiltinsAMDGPU.def
+++ b/clang/include/clang/Basic/BuiltinsAMDGPU.def
@@ -669,6 +669,7 @@ TARGET_BUILTIN(__builtin_amdgcn_s_monitor_sleep, "vIs", "n", "gfx1250-insts")
TARGET_BUILTIN(__builtin_amdgcn_s_wait_asynccnt, "vIUs", "n", "gfx1250-insts")
TARGET_BUILTIN(__builtin_amdgcn_s_wait_tensorcnt, "vIUs", "n", "gfx1250-insts")
+TARGET_BUILTIN(__builtin_amdgcn_tanhf, "ff", "nc", "tanh-insts")
TARGET_BUILTIN(__builtin_amdgcn_tanh_bf16, "yy", "nc", "bf16-trans-insts")
TARGET_BUILTIN(__builtin_amdgcn_rcp_bf16, "yy", "nc", "bf16-trans-insts")
TARGET_BUILTIN(__builtin_amdgcn_rsq_bf16, "yy", "nc", "bf16-trans-insts")
diff --git a/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx1250.cl b/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx1250.cl
index a1f984c129276..e120a46c6327b 100644
--- a/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx1250.cl
+++ b/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx1250.cl
@@ -42,6 +42,25 @@ void test_s_wait_tensorcnt() {
__builtin_amdgcn_s_wait_tensorcnt(0);
}
+// CHECK-LABEL: @test_tanh_f32(
+// CHECK-NEXT: entry:
+// CHECK-NEXT: [[OUT_ADDR:%.*]] = alloca ptr addrspace(1), align 8, addrspace(5)
+// CHECK-NEXT: [[A_ADDR:%.*]] = alloca float, align 4, addrspace(5)
+// CHECK-NEXT: [[OUT_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[OUT_ADDR]] to ptr
+// CHECK-NEXT: [[A_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[A_ADDR]] to ptr
+// CHECK-NEXT: store ptr addrspace(1) [[OUT:%.*]], ptr [[OUT_ADDR_ASCAST]], align 8
+// CHECK-NEXT: store float [[A:%.*]], ptr [[A_ADDR_ASCAST]], align 4
+// CHECK-NEXT: [[TMP0:%.*]] = load float, ptr [[A_ADDR_ASCAST]], align 4
+// CHECK-NEXT: [[TMP1:%.*]] = call float @llvm.amdgcn.tanh.f32(float [[TMP0]])
+// CHECK-NEXT: [[TMP2:%.*]] = load ptr addrspace(1), ptr [[OUT_ADDR_ASCAST]], align 8
+// CHECK-NEXT: store float [[TMP1]], ptr addrspace(1) [[TMP2]], align 4
+// CHECK-NEXT: ret void
+//
+void test_tanh_f32(global float* out, float a)
+{
+ *out = __builtin_amdgcn_tanhf(a);
+}
+
// CHECK-LABEL: @test_tanh_bf16(
// CHECK-NEXT: entry:
// CHECK-NEXT: [[OUT_ADDR:%.*]] = alloca ptr addrspace(1), align 8, addrspace(5)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
index faf59c1541fc0..0e0e83b7a6b54 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPU.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -1118,6 +1118,12 @@ def FeatureBitOp3Insts : SubtargetFeature<"bitop3-insts",
"Has v_bitop3_b32/v_bitop3_b16 instructions"
>;
+def FeatureTanhInsts : SubtargetFeature<"tanh-insts",
+ "HasTanhInsts",
+ "true",
+ "Has v_tanh_f32/f16 instructions"
+>;
+
def FeatureTransposeLoadF4F6Insts : SubtargetFeature<"transpose-load-f4f6-insts",
"HasTransposeLoadF4F6Insts",
"true",
@@ -1979,6 +1985,7 @@ def FeatureISAVersion12_50 : FeatureSet<
FeatureScalarDwordx3Loads,
FeatureDPPSrc1SGPR,
FeatureBitOp3Insts,
+ FeatureTanhInsts,
FeatureTransposeLoadF4F6Insts,
FeatureBF16TransInsts,
FeatureBF16ConversionInsts,
@@ -2703,6 +2710,9 @@ def HasPseudoScalarTrans : Predicate<"Subtarget->hasPseudoScalarTrans()">,
def HasBitOp3Insts : Predicate<"Subtarget->hasBitOp3Insts()">,
AssemblerPredicate<(all_of FeatureBitOp3Insts)>;
+def HasTanhInsts : Predicate<"Subtarget->hasTanhInsts()">,
+ AssemblerPredicate<(all_of FeatureTanhInsts)>;
+
def HasTransposeLoadF4F6Insts : Predicate<"Subtarget->hasTransposeLoadF4F6Insts()">,
AssemblerPredicate<(all_of FeatureTransposeLoadF4F6Insts)>;
diff --git a/llvm/lib/Target/AMDGPU/GCNSubtarget.h b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
index 67c6daaa24c2a..268162bcada47 100644
--- a/llvm/lib/Target/AMDGPU/GCNSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
@@ -234,6 +234,7 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
bool HasRestrictedSOffset = false;
bool Has64BitLiterals = false;
bool HasBitOp3Insts = false;
+ bool HasTanhInsts = false;
bool HasTransposeLoadF4F6Insts = false;
bool HasPrngInst = false;
bool HasBVHDualAndBVH8Insts = false;
@@ -1380,6 +1381,8 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
return HasMinimum3Maximum3F16;
}
+ bool hasTanhInsts() const { return HasTanhInsts; }
+
bool hasAddPC64Inst() const { return GFX1250Insts; }
bool hasMinimum3Maximum3PKF16() const {
diff --git a/llvm/lib/Target/AMDGPU/VOP1Instructions.td b/llvm/lib/Target/AMDGPU/VOP1Instructions.td
index ff89b8badeed0..8c35fea8259f4 100644
--- a/llvm/lib/Target/AMDGPU/VOP1Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP1Instructions.td
@@ -366,6 +366,9 @@ defm V_SQRT_F64 : VOP1Inst <"v_sqrt_f64", VOP_F64_F64, int_amdgcn_sqrt>;
let TRANS = 1, SchedRW = [WriteTrans32] in {
defm V_SIN_F32 : VOP1Inst <"v_sin_f32", VOP_F32_F32, AMDGPUsin>;
defm V_COS_F32 : VOP1Inst <"v_cos_f32", VOP_F32_F32, AMDGPUcos>;
+
+let SubtargetPredicate = HasTanhInsts in
+defm V_TANH_F32 : VOP1Inst <"v_tanh_f32", VOP_F32_F32, int_amdgcn_tanh>;
} // End TRANS = 1, SchedRW = [WriteTrans32]
defm V_NOT_B32 : VOP1Inst <"v_not_b32", VOP_I32_I32>;
@@ -1138,6 +1141,7 @@ defm V_CVT_F32_F16 : VOP1_Real_FULL_t16_and_fake16_gfx11_gfx12<0x00b>;
defm V_MOV_B64 : VOP1_Real_FULL <GFX1250Gen, 0x1d>;
+defm V_TANH_F32 : VOP1_Real_FULL<GFX1250Gen, 0x01e>;
defm V_TANH_BF16 : VOP1_Real_FULL_t16_and_fake16_gfx1250<0x04a>;
defm V_CVT_F32_BF16 : VOP1_Real_FULL_t16_and_fake16_gfx1250<0x072, "v_cvt_f32_bf16", "V_CVT_F32_BF16_gfx1250">;
defm V_CVT_PK_F16_FP8 : VOP1_Real_FULL_t16_and_fake16_gfx1250<0x075>;
diff --git a/llvm/lib/TargetParser/TargetParser.cpp b/llvm/lib/TargetParser/TargetParser.cpp
index d7e206ef8cd4f..4ca7444a73b35 100644
--- a/llvm/lib/TargetParser/TargetParser.cpp
+++ b/llvm/lib/TargetParser/TargetParser.cpp
@@ -443,6 +443,7 @@ void AMDGPU::fillAMDGPUFeatureMap(StringRef GPU, const Triple &T,
Features["gfx1250-insts"] = true;
Features["bitop3-insts"] = true;
Features["prng-inst"] = true;
+ Features["tanh-insts"] = true;
Features["transpose-load-f4f6-insts"] = true;
Features["bf16-trans-insts"] = true;
Features["fp8-conversion-insts"] = true;
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.tanh.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.tanh.ll
index 344c0112e4a54..91a2a0b651132 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.tanh.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.tanh.ll
@@ -7,8 +7,92 @@
; FIXME: t16 doesn't work at the moment because the store of s16 under t16 mode fails to select.
; FIXME: GlobalISel does not work with bf16
+declare float @llvm.amdgcn.tanh.f32(float) #0
declare bfloat @llvm.amdgcn.tanh.bf16(bfloat) #0
+define amdgpu_kernel void @tanh_f32(ptr addrspace(1) %out, float %src) #1 {
+; SDAG-REAL16-LABEL: tanh_f32:
+; SDAG-REAL16: ; %bb.0:
+; SDAG-REAL16-NEXT: s_load_b96 s[0:2], s[4:5], 0x0
+; SDAG-REAL16-NEXT: v_mov_b32_e32 v1, 0
+; SDAG-REAL16-NEXT: s_wait_kmcnt 0x0
+; SDAG-REAL16-NEXT: v_tanh_f32_e32 v0, s2
+; SDAG-REAL16-NEXT: global_store_b32 v1, v0, s[0:1]
+; SDAG-REAL16-NEXT: s_endpgm
+;
+; SDAG-FAKE16-LABEL: tanh_f32:
+; SDAG-FAKE16: ; %bb.0:
+; SDAG-FAKE16-NEXT: s_load_b96 s[0:2], s[4:5], 0x0
+; SDAG-FAKE16-NEXT: v_mov_b32_e32 v1, 0
+; SDAG-FAKE16-NEXT: s_wait_kmcnt 0x0
+; SDAG-FAKE16-NEXT: v_tanh_f32_e32 v0, s2
+; SDAG-FAKE16-NEXT: global_store_b32 v1, v0, s[0:1]
+; SDAG-FAKE16-NEXT: s_endpgm
+ %tanh = call float @llvm.amdgcn.tanh.f32(float %src) #0
+ store float %tanh, ptr addrspace(1) %out, align 4
+ ret void
+}
+
+; TODO: Really these should be constant folded
+define amdgpu_kernel void @tanh_f32_constant_4.0(ptr addrspace(1) %out) #1 {
+; SDAG-REAL16-LABEL: tanh_f32_constant_4.0:
+; SDAG-REAL16: ; %bb.0:
+; SDAG-REAL16-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
+; SDAG-REAL16-NEXT: v_tanh_f32_e32 v0, 4.0
+; SDAG-REAL16-NEXT: v_mov_b32_e32 v1, 0
+; SDAG-REAL16-NEXT: s_wait_kmcnt 0x0
+; SDAG-REAL16-NEXT: global_store_b32 v1, v0, s[0:1]
+; SDAG-REAL16-NEXT: s_endpgm
+;
+; SDAG-FAKE16-LABEL: tanh_f32_constant_4.0:
+; SDAG-FAKE16: ; %bb.0:
+; SDAG-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
+; SDAG-FAKE16-NEXT: v_tanh_f32_e32 v0, 4.0
+; SDAG-FAKE16-NEXT: v_mov_b32_e32 v1, 0
+; SDAG-FAKE16-NEXT: s_wait_kmcnt 0x0
+; SDAG-FAKE16-NEXT: global_store_b32 v1, v0, s[0:1]
+; SDAG-FAKE16-NEXT: s_endpgm
+ %tanh = call float @llvm.amdgcn.tanh.f32(float 4.0) #0
+ store float %tanh, ptr addrspace(1) %out, align 4
+ ret void
+}
+
+define amdgpu_kernel void @tanh_f32_constant_100.0(ptr addrspace(1) %out) #1 {
+; SDAG-REAL16-LABEL: tanh_f32_constant_100.0:
+; SDAG-REAL16: ; %bb.0:
+; SDAG-REAL16-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
+; SDAG-REAL16-NEXT: v_tanh_f32_e32 v0, 0x42c80000
+; SDAG-REAL16-NEXT: v_mov_b32_e32 v1, 0
+; SDAG-REAL16-NEXT: s_wait_kmcnt 0x0
+; SDAG-REAL16-NEXT: global_store_b32 v1, v0, s[0:1]
+; SDAG-REAL16-NEXT: s_endpgm
+;
+; SDAG-FAKE16-LABEL: tanh_f32_constant_100.0:
+; SDAG-FAKE16: ; %bb.0:
+; SDAG-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
+; SDAG-FAKE16-NEXT: v_tanh_f32_e32 v0, 0x42c80000
+; SDAG-FAKE16-NEXT: v_mov_b32_e32 v1, 0
+; SDAG-FAKE16-NEXT: s_wait_kmcnt 0x0
+; SDAG-FAKE16-NEXT: global_store_b32 v1, v0, s[0:1]
+; SDAG-FAKE16-NEXT: s_endpgm
+ %tanh = call float @llvm.amdgcn.tanh.f32(float 100.0) #0
+ store float %tanh, ptr addrspace(1) %out, align 4
+ ret void
+}
+
+define amdgpu_kernel void @tanh_undef_f32(ptr addrspace(1) %out) #1 {
+; SDAG-REAL16-LABEL: tanh_undef_f32:
+; SDAG-REAL16: ; %bb.0:
+; SDAG-REAL16-NEXT: s_endpgm
+;
+; SDAG-FAKE16-LABEL: tanh_undef_f32:
+; SDAG-FAKE16: ; %bb.0:
+; SDAG-FAKE16-NEXT: s_endpgm
+ %tanh = call float @llvm.amdgcn.tanh.f32(float undef)
+ store float %tanh, ptr addrspace(1) %out, align 4
+ ret void
+}
+
define amdgpu_kernel void @tanh_bf16(ptr addrspace(1) %out, bfloat %src) #1 {
; SDAG-REAL16-LABEL: tanh_bf16:
; SDAG-REAL16: ; %bb.0:
More information about the llvm-branch-commits
mailing list