[AMDGPU] Add the code generation support for llvm.[sin/cos].bf16

[shiltian]

This is a partial support because some other instructions have not been upstreamed yet.

[shiltian]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• [AMDGPU] Add the code generation support for llvm.[sin/cos].bf16 #149631 👈 (View in Graphite)
• [MI400][SDAG] Lower unsafe bf16 divisions #149628
• [NFC][AMDGPU] Add an IR test for v_cvt_f16_bf8 #149627
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Shilei Tian (shiltian)

Changes

This is a partial support because some other instructions have not been upstreamed yet.

---

Full diff: https://github.com/llvm/llvm-project/pull/149631.diff

3 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+1-1)
• (added) llvm/test/CodeGen/AMDGPU/llvm.cos.bf16.ll (+38)
• (added) llvm/test/CodeGen/AMDGPU/llvm.sin.bf16.ll (+38)

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 79487dcec3525..181db6291b361 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -620,7 +620,7 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
 
     // BF16 - VOP1 Actions.
     if (Subtarget->hasBF16TransInsts())
-      setOperationAction(ISD::FDIV, MVT::bf16, Custom);
+      setOperationAction({ISD::FCOS, ISD::FSIN, ISD::FDIV}, MVT::bf16, Custom);
 
     setOperationAction({ISD::FP_TO_SINT, ISD::FP_TO_UINT}, MVT::f16, Promote);
     setOperationAction({ISD::FP_TO_SINT, ISD::FP_TO_UINT}, MVT::bf16, Promote);
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.cos.bf16.ll b/llvm/test/CodeGen/AMDGPU/llvm.cos.bf16.ll
new file mode 100644
index 0000000000000..ced96ee98e0ad
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/llvm.cos.bf16.ll
@@ -0,0 +1,38 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 %s -o - | FileCheck -check-prefixes=GCN %s
+; xUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 -verify-machineinstrs %s -o - | FileCheck -check-prefix=GCN %s
+
+; FIXME: GlobalISel does not work with bf16
+
+declare bfloat @llvm.cos.bf16(bfloat) #0
+
+define amdgpu_kernel void @cos_bf16_constant_4(ptr addrspace(1) %out) #1 {
+; GCN-LABEL: cos_bf16_constant_4:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
+; GCN-NEXT:    v_cos_bf16_e32 v0, 0x3f23
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    s_wait_kmcnt 0x0
+; GCN-NEXT:    global_store_b16 v1, v0, s[0:1]
+; GCN-NEXT:    s_endpgm
+  %cos = call bfloat @llvm.cos.bf16(bfloat 4.0) #0
+  store bfloat %cos, ptr addrspace(1) %out, align 2
+  ret void
+}
+
+define amdgpu_kernel void @cos_bf16_constant_100(ptr addrspace(1) %out) #1 {
+; GCN-LABEL: cos_bf16_constant_100:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
+; GCN-NEXT:    v_cos_bf16_e32 v0, 0x417f
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    s_wait_kmcnt 0x0
+; GCN-NEXT:    global_store_b16 v1, v0, s[0:1]
+; GCN-NEXT:    s_endpgm
+  %cos = call bfloat @llvm.cos.bf16(bfloat 100.0) #0
+  store bfloat %cos, ptr addrspace(1) %out, align 2
+  ret void
+}
+
+attributes #0 = { nounwind readnone }
+attributes #1 = { nounwind }
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.sin.bf16.ll b/llvm/test/CodeGen/AMDGPU/llvm.sin.bf16.ll
new file mode 100644
index 0000000000000..7a355a36b15bf
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/llvm.sin.bf16.ll
@@ -0,0 +1,38 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 %s -o - | FileCheck -check-prefixes=GCN %s
+; xUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 -verify-machineinstrs %s -o - | FileCheck -check-prefix=GCN %s
+
+; FIXME: GlobalISel does not work with bf16
+
+declare bfloat @llvm.sin.bf16(bfloat) #0
+
+define amdgpu_kernel void @sin_bf16_constant_4(ptr addrspace(1) %out) #1 {
+; GCN-LABEL: sin_bf16_constant_4:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
+; GCN-NEXT:    v_sin_bf16_e32 v0, 0x3f23
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    s_wait_kmcnt 0x0
+; GCN-NEXT:    global_store_b16 v1, v0, s[0:1]
+; GCN-NEXT:    s_endpgm
+  %sin = call bfloat @llvm.sin.bf16(bfloat 4.0) #0
+  store bfloat %sin, ptr addrspace(1) %out, align 2
+  ret void
+}
+
+define amdgpu_kernel void @sin_bf16_constant_100(ptr addrspace(1) %out) #1 {
+; GCN-LABEL: sin_bf16_constant_100:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
+; GCN-NEXT:    v_sin_bf16_e32 v0, 0x417f
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    s_wait_kmcnt 0x0
+; GCN-NEXT:    global_store_b16 v1, v0, s[0:1]
+; GCN-NEXT:    s_endpgm
+  %sin = call bfloat @llvm.sin.bf16(bfloat 100.0) #0
+  store bfloat %sin, ptr addrspace(1) %out, align 2
+  ret void
+}
+
+attributes #0 = { nounwind readnone }
+attributes #1 = { nounwind }