1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
2 ; RUN: llc -mtriple=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX6 %s
3 ; RUN: llc -mtriple=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX8 %s
4 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11 %s
5 ; RUN: llc -mtriple=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -check-prefixes=CYPRESS %s
6 ; RUN: llc -mtriple=r600 -mcpu=cayman -verify-machineinstrs < %s | FileCheck -check-prefixes=CAYMAN %s
8 declare float @llvm.convert.from.fp16.f32(i16) nounwind readnone
10 define amdgpu_kernel void @test_convert_fp16_to_fp32(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) nounwind {
11 ; GFX6-LABEL: test_convert_fp16_to_fp32:
13 ; GFX6-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
14 ; GFX6-NEXT: s_mov_b32 s7, 0xf000
15 ; GFX6-NEXT: s_mov_b32 s6, -1
16 ; GFX6-NEXT: s_mov_b32 s10, s6
17 ; GFX6-NEXT: s_mov_b32 s11, s7
18 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
19 ; GFX6-NEXT: s_mov_b32 s8, s2
20 ; GFX6-NEXT: s_mov_b32 s9, s3
21 ; GFX6-NEXT: buffer_load_ushort v0, off, s[8:11], 0
22 ; GFX6-NEXT: s_mov_b32 s4, s0
23 ; GFX6-NEXT: s_mov_b32 s5, s1
24 ; GFX6-NEXT: s_waitcnt vmcnt(0)
25 ; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
26 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
29 ; GFX8-LABEL: test_convert_fp16_to_fp32:
31 ; GFX8-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
32 ; GFX8-NEXT: s_mov_b32 s7, 0xf000
33 ; GFX8-NEXT: s_mov_b32 s6, -1
34 ; GFX8-NEXT: s_mov_b32 s10, s6
35 ; GFX8-NEXT: s_mov_b32 s11, s7
36 ; GFX8-NEXT: s_waitcnt lgkmcnt(0)
37 ; GFX8-NEXT: s_mov_b32 s8, s2
38 ; GFX8-NEXT: s_mov_b32 s9, s3
39 ; GFX8-NEXT: buffer_load_ushort v0, off, s[8:11], 0
40 ; GFX8-NEXT: s_mov_b32 s4, s0
41 ; GFX8-NEXT: s_mov_b32 s5, s1
42 ; GFX8-NEXT: s_waitcnt vmcnt(0)
43 ; GFX8-NEXT: v_cvt_f32_f16_e32 v0, v0
44 ; GFX8-NEXT: buffer_store_dword v0, off, s[4:7], 0
47 ; GFX11-LABEL: test_convert_fp16_to_fp32:
49 ; GFX11-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
50 ; GFX11-NEXT: s_mov_b32 s6, -1
51 ; GFX11-NEXT: s_mov_b32 s7, 0x31016000
52 ; GFX11-NEXT: s_mov_b32 s10, s6
53 ; GFX11-NEXT: s_mov_b32 s11, s7
54 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
55 ; GFX11-NEXT: s_mov_b32 s8, s2
56 ; GFX11-NEXT: s_mov_b32 s9, s3
57 ; GFX11-NEXT: s_mov_b32 s4, s0
58 ; GFX11-NEXT: buffer_load_u16 v0, off, s[8:11], 0
59 ; GFX11-NEXT: s_mov_b32 s5, s1
60 ; GFX11-NEXT: s_waitcnt vmcnt(0)
61 ; GFX11-NEXT: v_cvt_f32_f16_e32 v0, v0
62 ; GFX11-NEXT: buffer_store_b32 v0, off, s[4:7], 0
64 ; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
65 ; GFX11-NEXT: s_endpgm
67 ; CYPRESS-LABEL: test_convert_fp16_to_fp32:
69 ; CYPRESS-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
70 ; CYPRESS-NEXT: TEX 0 @6
71 ; CYPRESS-NEXT: ALU 2, @9, KC0[CB0:0-32], KC1[]
72 ; CYPRESS-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
73 ; CYPRESS-NEXT: CF_END
75 ; CYPRESS-NEXT: Fetch clause starting at 6:
76 ; CYPRESS-NEXT: VTX_READ_16 T0.X, T0.X, 0, #1
77 ; CYPRESS-NEXT: ALU clause starting at 8:
78 ; CYPRESS-NEXT: MOV * T0.X, KC0[2].Z,
79 ; CYPRESS-NEXT: ALU clause starting at 9:
80 ; CYPRESS-NEXT: FLT16_TO_FLT32 T0.X, T0.X,
81 ; CYPRESS-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
82 ; CYPRESS-NEXT: 2(2.802597e-45), 0(0.000000e+00)
84 ; CAYMAN-LABEL: test_convert_fp16_to_fp32:
86 ; CAYMAN-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
87 ; CAYMAN-NEXT: TEX 0 @6
88 ; CAYMAN-NEXT: ALU 2, @9, KC0[CB0:0-32], KC1[]
89 ; CAYMAN-NEXT: MEM_RAT_CACHELESS STORE_DWORD T0.X, T1.X
92 ; CAYMAN-NEXT: Fetch clause starting at 6:
93 ; CAYMAN-NEXT: VTX_READ_16 T0.X, T0.X, 0, #1
94 ; CAYMAN-NEXT: ALU clause starting at 8:
95 ; CAYMAN-NEXT: MOV * T0.X, KC0[2].Z,
96 ; CAYMAN-NEXT: ALU clause starting at 9:
97 ; CAYMAN-NEXT: FLT16_TO_FLT32 * T0.X, T0.X,
98 ; CAYMAN-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
99 ; CAYMAN-NEXT: 2(2.802597e-45), 0(0.000000e+00)
100 %val = load i16, ptr addrspace(1) %in, align 2
101 %cvt = call float @llvm.convert.from.fp16.f32(i16 %val) nounwind readnone
102 store float %cvt, ptr addrspace(1) %out, align 4