AMDGPU: Implement isFMAFasterThanFMulAndFAdd for f16

[• tstellarAMD]

LGTM.

test/CodeGen/AMDGPU/llvm.fmuladd.f16.ll

1–4

--check-prefixes=

[arsenm]

r290307

lib/Target/AMDGPU/SIISelLowering.cpp

bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
   case MVT::f32:
     // This is as fast on some subtargets. However, we always have full rate f32
     // mad available which returns the same result as the separate operations
     // which we should prefer over fma. We can't use this if we want to support
     // denormals, so only report this in these cases.
     return Subtarget->hasFP32Denormals() && Subtarget->hasFastFMAF32();
   case MVT::f64:
     return true;
+  case MVT::f16:
+    return Subtarget->has16BitInsts() && Subtarget->hasFP16Denormals();
   default:
     break;
   }
 
   return false;
 }
 
 //===----------------------------------------------------------------------===//

test/CodeGen/AMDGPU/llvm.fmuladd.f16.ll

-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
-; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+; RUN: llc -march=amdgcn -mattr=-fp16-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI -check-prefix=SI-FLUSH %s
+; RUN: llc -march=amdgcn -mcpu=fiji -mattr=-fp16-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI -check-prefix=VI-FLUSH %s
+; RUN: llc -march=amdgcn -mattr=+fp16-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI -check-prefix=SI-DENORM %s
+; RUN: llc -march=amdgcn -mcpu=fiji -mattr=+fp16-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI -check-prefix=VI-DENORM %s
 
 declare half @llvm.fmuladd.f16(half %a, half %b, half %c)
 declare <2 x half> @llvm.fmuladd.v2f16(<2 x half> %a, <2 x half> %b, <2 x half> %c)
 
 ; GCN-LABEL: {{^}}fmuladd_f16
 ; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
 ; GCN: buffer_load_ushort v[[B_F16:[0-9]+]]
 ; GCN: buffer_load_ushort v[[C_F16:[0-9]+]]
 ; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]]
 ; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]]
 ; SI:  v_cvt_f32_f16_e32 v[[C_F32:[0-9]+]], v[[C_F16]]
 ; SI:  v_mac_f32_e32 v[[C_F32]], v[[B_F32]], v[[A_F32]]
 ; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[C_F32]]
 ; SI:  buffer_store_short v[[R_F16]]
-; VI:  v_mac_f16_e32 v[[C_F16]], v[[B_F16]], v[[A_F16]]
-; VI:  buffer_store_short v[[C_F16]]
+
+; VI-FLUSH: v_mac_f16_e32 v[[C_F16]], v[[B_F16]], v[[A_F16]]
+; VI-FLUSH: buffer_store_short v[[C_F16]]
+
+; VI-DENORM: v_fma_f16 [[RESULT:v[0-9]+]], v[[A_F16]], v[[B_F16]], v[[C_F16]]
+; VI-DENORM: buffer_store_short [[RESULT]]
+
 ; GCN: s_endpgm
 define void @fmuladd_f16(
     half addrspace(1)* %r,
     half addrspace(1)* %a,
     half addrspace(1)* %b,
     half addrspace(1)* %c) {
   %a.val = load half, half addrspace(1)* %a
   %b.val = load half, half addrspace(1)* %b
   %c.val = load half, half addrspace(1)* %c
   %r.val = call half @llvm.fmuladd.f16(half %a.val, half %b.val, half %c.val)
   store half %r.val, half addrspace(1)* %r
   ret void
 }
 
 ; GCN-LABEL: {{^}}fmuladd_f16_imm_a
 ; GCN: buffer_load_ushort v[[B_F16:[0-9]+]]
 ; GCN: buffer_load_ushort v[[C_F16:[0-9]+]]
 ; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], 0x4200{{$}}
 ; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]]
 ; SI:  v_cvt_f32_f16_e32 v[[C_F32:[0-9]+]], v[[C_F16]]
 ; SI:  v_mac_f32_e32 v[[C_F32]], v[[A_F32]], v[[B_F32]]
 ; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[C_F32]]
 ; SI:  buffer_store_short v[[R_F16]]
-; VI:  v_mac_f16_e32 v[[C_F16]], 0x4200, v[[B_F16]]
-; VI:  buffer_store_short v[[C_F16]]
+
+; VI-FLUSH: v_mac_f16_e32 v[[C_F16]], 0x4200, v[[B_F16]]
+; VI-FLUSH: buffer_store_short v[[C_F16]]
+
+; VI-DENORM: v_mov_b32_e32 [[KA:v[0-9]+]], 0x4200
+; VI-DENORM: v_fma_f16 [[RESULT:v[0-9]+]], [[KA]], v[[B_F16]], v[[C_F16]]
+; VI-DENORM: buffer_store_short [[RESULT]]
+
 ; GCN: s_endpgm
 define void @fmuladd_f16_imm_a(
     half addrspace(1)* %r,
     half addrspace(1)* %b,
     half addrspace(1)* %c) {
   %b.val = load half, half addrspace(1)* %b
   %c.val = load half, half addrspace(1)* %c
   %r.val = call half @llvm.fmuladd.f16(half 3.0, half %b.val, half %c.val)
   store half %r.val, half addrspace(1)* %r
   ret void
 }
 
 ; GCN-LABEL: {{^}}fmuladd_f16_imm_b
 ; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
 ; GCN: buffer_load_ushort v[[C_F16:[0-9]+]]
 ; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], 0x4200{{$}}
 ; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]]
 ; SI:  v_cvt_f32_f16_e32 v[[C_F32:[0-9]+]], v[[C_F16]]
 ; SI:  v_mac_f32_e32 v[[C_F32]], v[[B_F32]], v[[A_F32]]
 ; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[C_F32]]
 ; SI:  buffer_store_short v[[R_F16]]
-; VI:  v_mac_f16_e32 v[[C_F16]], 0x4200, v[[A_F16]]
-; VI:  buffer_store_short v[[C_F16]]
+
+; VI-FLUSH: v_mac_f16_e32 v[[C_F16]], 0x4200, v[[A_F16]]
+; VI-FLUSH: buffer_store_short v[[C_F16]]
+
+; VI-DENORM: v_mov_b32_e32 [[KA:v[0-9]+]], 0x4200
+; VI-DENORM: v_fma_f16 [[RESULT:v[0-9]+]], [[KA]], v[[A_F16]], v[[C_F16]]
+; VI-DENORM buffer_store_short [[RESULT]]
+
+
 ; GCN: s_endpgm
 define void @fmuladd_f16_imm_b(
     half addrspace(1)* %r,
     half addrspace(1)* %a,
     half addrspace(1)* %c) {
   %a.val = load half, half addrspace(1)* %a
   %c.val = load half, half addrspace(1)* %c
   %r.val = call half @llvm.fmuladd.f16(half %a.val, half 3.0, half %c.val)

 ; SI:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
 ; SI:  v_cvt_f32_f16_e32 v[[C_F32_1:[0-9]+]], v[[C_F16_1]]
 ; SI:  v_mac_f32_e32 v[[C_F32_0]], v[[B_F32_0]], v[[A_F32_0]]
 ; SI:  v_cvt_f16_f32_e32 v[[R_F16_0:[0-9]+]], v[[C_F32_0]]
 ; SI:  v_mac_f32_e32 v[[C_F32_1]], v[[B_F32_1]], v[[A_F32_1]]
 ; SI:  v_cvt_f16_f32_e32 v[[R_F16_1:[0-9]+]], v[[C_F32_1]]
 ; SI:  v_and_b32_e32 v[[R_F16_LO:[0-9]+]], 0xffff, v[[R_F16_0]]
 ; SI:  v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[R_F16_1]]
-; VI:  v_mac_f16_e32 v[[C_V2_F16]], v[[B_V2_F16]], v[[A_V2_F16]]
-; VI:  v_mac_f16_e32 v[[C_F16_1]], v[[B_F16_1]], v[[A_F16_1]]
-; VI:  v_and_b32_e32 v[[R_F16_LO:[0-9]+]], 0xffff, v[[C_V2_F16]]
-; VI:  v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[C_F16_1]]
+
+
+; FIXME: and should be unnecessary
+; VI-FLUSH: v_mac_f16_e32 v[[C_V2_F16]], v[[B_V2_F16]], v[[A_V2_F16]]
+; VI-FLUSH: v_mac_f16_e32 v[[C_F16_1]], v[[B_F16_1]], v[[A_F16_1]]
+; VI-FLUSH: v_and_b32_e32 v[[R_F16_LO:[0-9]+]], 0xffff, v[[C_V2_F16]]
+; VI-FLUSH: v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[C_F16_1]]
+
+; VI-DENORM-DAG: v_fma_f16 v[[RES0:[0-9]+]], v[[A_V2_F16]], v[[B_V2_F16]], v[[C_V2_F16]]
+; VI-DENORM-DAG: v_fma_f16 v[[RES1:[0-9]+]], v[[A_F16_1]], v[[B_F16_1]], v[[C_F16_1]]
+; VI-DENORM: v_and_b32_e32 v[[R_F16_LO:[0-9]+]], 0xffff, v[[RES0]]
+; VI-DENORM: v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[RES1]]
+
 ; GCN: v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_HI]], v[[R_F16_LO]]
 ; GCN: buffer_store_dword v[[R_V2_F16]]
 ; GCN: s_endpgm
 define void @fmuladd_v2f16(
     <2 x half> addrspace(1)* %r,
     <2 x half> addrspace(1)* %a,
     <2 x half> addrspace(1)* %b,
     <2 x half> addrspace(1)* %c) {
   %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
   %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
   %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
   %r.val = call <2 x half> @llvm.fmuladd.v2f16(<2 x half> %a.val, <2 x half> %b.val, <2 x half> %c.val)
   store <2 x half> %r.val, <2 x half> addrspace(1)* %r
   ret void
 }