1; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tahiti -denormal-fp-math=preserve-sign -denormal-fp-math-f32=preserve-sign -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=SI -check-prefix=SI-FLUSH %s
2; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=fiji -denormal-fp-math=preserve-sign -denormal-fp-math-f32=preserve-sign -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=VI -check-prefix=VI-FLUSH %s
3; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tahiti -denormal-fp-math=ieee -denormal-fp-math-f32=preserve-sign -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=SI -check-prefix=SI-DENORM %s
4; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=fiji -denormal-fp-math=ieee -denormal-fp-math-f32=preserve-sign -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=VI -check-prefix=VI-DENORM %s
5; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx1010 -denormal-fp-math=preserve-sign -denormal-fp-math-f32=preserve-sign -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=GFX10 -check-prefix=GFX10-FLUSH %s
6; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx1010 -denormal-fp-math=ieee -denormal-fp-math-f32=preserve-sign -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=GFX10 -check-prefix=GFX10-DENORM %s
7
8declare half @llvm.fmuladd.f16(half %a, half %b, half %c)
9declare <2 x half> @llvm.fmuladd.v2f16(<2 x half> %a, <2 x half> %b, <2 x half> %c)
10
11; GCN-LABEL: {{^}}fmuladd_f16
12; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
13; GCN: buffer_load_ushort v[[B_F16:[0-9]+]]
14; GCN: buffer_load_ushort v[[C_F16:[0-9]+]]
15; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]]
16; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]]
17; SI:  v_cvt_f32_f16_e32 v[[C_F32:[0-9]+]], v[[C_F16]]
18; SI:  v_mac_f32_e32 v[[C_F32]], v[[A_F32]], v[[B_F32]]
19; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[C_F32]]
20; SI:  buffer_store_short v[[R_F16]]
21
22; VI-FLUSH: v_mac_f16_e32 v[[C_F16]], v[[A_F16]], v[[B_F16]]
23; VI-FLUSH: buffer_store_short v[[C_F16]]
24
25; VI-DENORM: v_fma_f16 [[RESULT:v[0-9]+]], v[[A_F16]], v[[B_F16]], v[[C_F16]]
26; VI-DENORM: buffer_store_short [[RESULT]]
27
28; GFX10-FLUSH: v_mul_f16_e32 [[MUL:v[0-9]+]], v[[A_F16]], v[[B_F16]]
29; GFX10-FLUSH: v_add_f16_e32 [[ADD:v[0-9]+]], [[MUL]], v[[C_F16]]
30; GFX10-FLUSH: buffer_store_short [[ADD]]
31
32; GFX10-DENORM: v_fmac_f16_e32 v[[C_F16]], v[[A_F16]], v[[B_F16]]
33; GFX10-DENORM: buffer_store_short v[[C_F16]],
34
35; GCN: s_endpgm
36define amdgpu_kernel void @fmuladd_f16(
37    half addrspace(1)* %r,
38    half addrspace(1)* %a,
39    half addrspace(1)* %b,
40    half addrspace(1)* %c) {
41  %a.val = load half, half addrspace(1)* %a
42  %b.val = load half, half addrspace(1)* %b
43  %c.val = load half, half addrspace(1)* %c
44  %r.val = call half @llvm.fmuladd.f16(half %a.val, half %b.val, half %c.val)
45  store half %r.val, half addrspace(1)* %r
46  ret void
47}
48
49; GCN-LABEL: {{^}}fmuladd_f16_imm_a
50; GCN: buffer_load_ushort v[[B_F16:[0-9]+]]
51; GCN: buffer_load_ushort v[[C_F16:[0-9]+]]
52; SI:  v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]]
53; SI:  v_cvt_f32_f16_e32 v[[C_F32:[0-9]+]], v[[C_F16]]
54; SI:  v_mac_f32_e32 v[[C_F32]], 0x40400000, v[[B_F32]]
55; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[C_F32]]
56; SI:  buffer_store_short v[[R_F16]]
57
58; VI-FLUSH: v_mac_f16_e32 v[[C_F16]], 0x4200, v[[B_F16]]
59; VI-FLUSH: buffer_store_short v[[C_F16]]
60
61; VI-DENORM: s_movk_i32 [[KA:s[0-9]+]], 0x4200
62; VI-DENORM: v_fma_f16 [[RESULT:v[0-9]+]], v[[B_F16]], [[KA]], v[[C_F16]]
63; VI-DENORM: buffer_store_short [[RESULT]]
64
65; GFX10-FLUSH: v_mul_f16_e32 [[MUL:v[0-9]+]], 0x4200, v[[B_F16]]
66; GFX10-FLUSH: v_add_f16_e32 [[ADD:v[0-9]+]], [[MUL]], v[[C_F16]]
67; GFX10-FLUSH: buffer_store_short [[ADD]]
68
69; GFX10-DENORM: v_fmac_f16_e32 v[[C_F16]], 0x4200, v[[B_F16]]
70; GFX10-DENORM: buffer_store_short v[[C_F16]],
71
72; GCN: s_endpgm
73define amdgpu_kernel void @fmuladd_f16_imm_a(
74    half addrspace(1)* %r,
75    half addrspace(1)* %b,
76    half addrspace(1)* %c) {
77  %b.val = load volatile half, half addrspace(1)* %b
78  %c.val = load volatile half, half addrspace(1)* %c
79  %r.val = call half @llvm.fmuladd.f16(half 3.0, half %b.val, half %c.val)
80  store half %r.val, half addrspace(1)* %r
81  ret void
82}
83
84; GCN-LABEL: {{^}}fmuladd_f16_imm_b
85; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]
86; GCN: buffer_load_ushort v[[C_F16:[0-9]+]]
87; SI:  v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]]
88; SI:  v_cvt_f32_f16_e32 v[[C_F32:[0-9]+]], v[[C_F16]]
89; SI:  v_mac_f32_e32 v[[C_F32]], 0x40400000, v[[A_F32]]
90; SI:  v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[C_F32]]
91; SI:  buffer_store_short v[[R_F16]]
92
93; VI-FLUSH: v_mac_f16_e32 v[[C_F16]], 0x4200, v[[A_F16]]
94; VI-FLUSH: buffer_store_short v[[C_F16]]
95
96; VI-DENORM: s_movk_i32 [[KA:s[0-9]+]], 0x4200
97; VI-DENORM: v_fma_f16 [[RESULT:v[0-9]+]], v[[A_F16]], [[KA]], v[[C_F16]]
98; VI-DENORM: buffer_store_short [[RESULT]]
99
100; GFX10-FLUSH: v_mul_f16_e32 [[MUL:v[0-9]+]], 0x4200, v[[A_F16]]
101; GFX10-FLUSH: v_add_f16_e32 [[ADD:v[0-9]+]], [[MUL]], v[[C_F16]]
102; GFX10-FLUSH: buffer_store_short [[ADD]]
103
104; GFX10-DENORM: v_fmac_f16_e32 v[[C_F16]], 0x4200, v[[A_F16]]
105; GFX10-DENORM: buffer_store_short v[[C_F16]],
106
107; GCN: s_endpgm
108define amdgpu_kernel void @fmuladd_f16_imm_b(
109    half addrspace(1)* %r,
110    half addrspace(1)* %a,
111    half addrspace(1)* %c) {
112  %a.val = load volatile half, half addrspace(1)* %a
113  %c.val = load volatile half, half addrspace(1)* %c
114  %r.val = call half @llvm.fmuladd.f16(half %a.val, half 3.0, half %c.val)
115  store half %r.val, half addrspace(1)* %r
116  ret void
117}
118
119; GCN-LABEL: {{^}}fmuladd_v2f16
120; SI: buffer_load_dword v[[A_V2_F16:[0-9]+]]
121; SI: buffer_load_dword v[[B_V2_F16:[0-9]+]]
122; SI: buffer_load_dword v[[C_V2_F16:[0-9]+]]
123
124; VI-FLUSH: buffer_load_dword v[[A_V2_F16:[0-9]+]]
125; VI-FLUSH: buffer_load_dword v[[C_V2_F16:[0-9]+]]
126; VI-FLUSH: buffer_load_dword v[[B_V2_F16:[0-9]+]]
127
128; VI-DENORM: buffer_load_dword v[[A_V2_F16:[0-9]+]]
129; VI-DENORM: buffer_load_dword v[[B_V2_F16:[0-9]+]]
130; VI-DENORM: buffer_load_dword v[[C_V2_F16:[0-9]+]]
131
132; GFX10: buffer_load_dword v[[A_V2_F16:[0-9]+]]
133; GFX10: buffer_load_dword v[[B_V2_F16:[0-9]+]]
134; GFX10: buffer_load_dword v[[C_V2_F16:[0-9]+]]
135
136; SI: v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]]
137; SI: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
138; SI: v_lshrrev_b32_e32 v[[C_F16_1:[0-9]+]], 16, v[[C_V2_F16]]
139; SI: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
140
141; SI-DAG: v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]]
142; SI-DAG: v_cvt_f32_f16_e32 v[[C_F32_0:[0-9]+]], v[[C_V2_F16]]
143
144; SI-DAG:  v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]]
145; SI-DAG:  v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]]
146; SI-DAG:  v_cvt_f32_f16_e32 v[[C_F32_1:[0-9]+]], v[[C_F16_1]]
147; SI-DAG:  v_mac_f32_e32 v[[C_F32_0]], v[[A_F32_0]], v[[B_F32_0]]
148; SI-DAG:  v_mac_f32_e32 v[[C_F32_1]], v[[A_F32_1]], v[[B_F32_1]]
149; SI-DAG:  v_cvt_f16_f32_e32 v[[R_F16_1:[0-9]+]], v[[C_F32_1]]
150; SI-DAG:  v_cvt_f16_f32_e32 v[[R_F16_LO:[0-9]+]], v[[C_F32_0]]
151; SI-DAG:  v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[R_F16_1]]
152; SI:  v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_LO]], v[[R_F16_HI]]
153
154; VI-FLUSH:     v_lshrrev_b32_e32 v[[C_F16_1:[0-9]+]], 16, v[[C_V2_F16]]
155; VI-FLUSH-DAG: v_mac_f16_sdwa v[[C_F16_1]], v[[A_V2_F16]], v[[B_V2_F16]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
156; VI-FLUSH-DAG: v_mac_f16_e32 v[[C_V2_F16]], v[[A_V2_F16]], v[[B_V2_F16]]
157; VI-FLUSH-DAG: v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[C_F16_1]]
158; VI-FLUSH-NOT: v_and_b32
159; VI-FLUSH:     v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[C_V2_F16]], v[[R_F16_HI]]
160
161; VI-DENORM-DAG: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]]
162; VI-DENORM-DAG: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]]
163; VI-DENORM-DAG: v_lshrrev_b32_e32 v[[C_F16_1:[0-9]+]], 16, v[[C_V2_F16]]
164; VI-DENORM-DAG: v_fma_f16 v[[RES0:[0-9]+]], v[[A_V2_F16]], v[[B_V2_F16]], v[[C_V2_F16]]
165; VI-DENORM-DAG: v_fma_f16 v[[RES1:[0-9]+]], v[[A_F16_1]], v[[B_F16_1]], v[[C_F16_1]]
166; VI-DENORM-DAG: v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[RES1]]
167; VI-DENORM-NOT: v_and_b32
168; VI-DENORM: v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[RES0]], v[[R_F16_HI]]
169
170; GFX10-FLUSH: v_pk_mul_f16 [[MUL:v[0-9]+]], v[[A_V2_F16]], v[[B_V2_F16]]
171; GFX10-FLUSH: v_pk_add_f16 v[[R_V2_F16:[0-9]+]], [[MUL]], v[[C_V2_F16]]
172
173; GFX10-DENORM: v_pk_fma_f16 v[[R_V2_F16:[0-9]+]], v[[A_V2_F16]], v[[B_V2_F16]], v[[C_V2_F16]]
174
175; GCN: buffer_store_dword v[[R_V2_F16]]
176define amdgpu_kernel void @fmuladd_v2f16(
177    <2 x half> addrspace(1)* %r,
178    <2 x half> addrspace(1)* %a,
179    <2 x half> addrspace(1)* %b,
180    <2 x half> addrspace(1)* %c) {
181  %a.val = load <2 x half>, <2 x half> addrspace(1)* %a
182  %b.val = load <2 x half>, <2 x half> addrspace(1)* %b
183  %c.val = load <2 x half>, <2 x half> addrspace(1)* %c
184  %r.val = call <2 x half> @llvm.fmuladd.v2f16(<2 x half> %a.val, <2 x half> %b.val, <2 x half> %c.val)
185  store <2 x half> %r.val, <2 x half> addrspace(1)* %r
186  ret void
187}
188