[ORC-RT][LoongArch] Add initial support for loongarch64 in ELFNixPlatform (#123575)
[llvm-project.git] / llvm / test / MC / AMDGPU / gfx12_asm_vop3_from_vop1_dpp8-fake16.s
blob98b997c7a2428d7bacce14959966c35254a32b76
1 // RUN: llvm-mc -triple=amdgcn -mcpu=gfx1200 -mattr=+wavefrontsize32,-real-true16 -show-encoding %s | FileCheck --check-prefix=GFX12 %s
3 v_bfrev_b32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
4 // GFX12: [0x05,0x00,0xb8,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
6 v_bfrev_b32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
7 // GFX12: [0x05,0x00,0xb8,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
9 v_bfrev_b32_e64_dpp v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0
10 // GFX12: [0xff,0x00,0xb8,0xd5,0xe9,0x00,0x00,0x00,0xff,0x00,0x00,0x00]
12 v_ceil_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
13 // GFX12: [0x05,0x00,0xdc,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
15 v_ceil_f16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
16 // GFX12: [0x05,0x00,0xdc,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
18 v_ceil_f16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
19 // GFX12: [0x05,0x00,0xdc,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
21 v_ceil_f16_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
22 // GFX12: [0xff,0x81,0xdc,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
24 v_ceil_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
25 // GFX12: [0x05,0x00,0xa2,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
27 v_ceil_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
28 // GFX12: [0x05,0x00,0xa2,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
30 v_ceil_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
31 // GFX12: [0x05,0x00,0xa2,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
33 v_ceil_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
34 // GFX12: [0xff,0x81,0xa2,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
36 v_cls_i32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
37 // GFX12: [0x05,0x00,0xbb,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
39 v_cls_i32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
40 // GFX12: [0x05,0x00,0xbb,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
42 v_cls_i32_e64_dpp v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0
43 // GFX12: [0xff,0x00,0xbb,0xd5,0xe9,0x00,0x00,0x00,0xff,0x00,0x00,0x00]
45 v_clz_i32_u32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
46 // GFX12: [0x05,0x00,0xb9,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
48 v_clz_i32_u32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
49 // GFX12: [0x05,0x00,0xb9,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
51 v_clz_i32_u32_e64_dpp v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0
52 // GFX12: [0xff,0x00,0xb9,0xd5,0xe9,0x00,0x00,0x00,0xff,0x00,0x00,0x00]
54 v_cos_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
55 // GFX12: [0x05,0x00,0xe1,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
57 v_cos_f16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
58 // GFX12: [0x05,0x00,0xe1,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
60 v_cos_f16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
61 // GFX12: [0x05,0x00,0xe1,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
63 v_cos_f16_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
64 // GFX12: [0xff,0x81,0xe1,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
66 v_cos_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
67 // GFX12: [0x05,0x00,0xb6,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
69 v_cos_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
70 // GFX12: [0x05,0x00,0xb6,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
72 v_cos_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
73 // GFX12: [0x05,0x00,0xb6,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
75 v_cos_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
76 // GFX12: [0xff,0x81,0xb6,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
78 v_ctz_i32_b32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
79 // GFX12: [0x05,0x00,0xba,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
81 v_ctz_i32_b32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
82 // GFX12: [0x05,0x00,0xba,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
84 v_ctz_i32_b32_e64_dpp v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0
85 // GFX12: [0xff,0x00,0xba,0xd5,0xe9,0x00,0x00,0x00,0xff,0x00,0x00,0x00]
87 v_cvt_f32_fp8_e64_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
88 // GFX12: encoding: [0x05,0x00,0xec,0xd5,0xe9,0x00,0x00,0x00,0x01,0x88,0xc6,0xfa]
90 v_cvt_f32_fp8_e64_dpp v1, v3 dpp8:[7,6,5,4,3,2,1,0]
91 // GFX12: encoding: [0x01,0x00,0xec,0xd5,0xe9,0x00,0x00,0x00,0x03,0x77,0x39,0x05]
93 v_cvt_f32_bf8_e64_dpp v5, v1 dpp8:[0,1,2,3,4,5,6,7]
94 // GFX12: encoding: [0x05,0x00,0xed,0xd5,0xe9,0x00,0x00,0x00,0x01,0x88,0xc6,0xfa]
96 v_cvt_f32_bf8_e64_dpp v1, v3 dpp8:[7,6,5,4,3,2,1,0]
97 // GFX12: encoding: [0x01,0x00,0xed,0xd5,0xe9,0x00,0x00,0x00,0x03,0x77,0x39,0x05]
99 v_cvt_f16_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
100 // GFX12: [0x05,0x00,0x8a,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
102 v_cvt_f16_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
103 // GFX12: [0x05,0x00,0x8a,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
105 v_cvt_f16_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
106 // GFX12: [0x05,0x00,0x8a,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
108 v_cvt_f16_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
109 // GFX12: [0xff,0x81,0x8a,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
111 v_cvt_f16_i16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
112 // GFX12: [0x05,0x00,0xd1,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
114 v_cvt_f16_i16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
115 // GFX12: [0x05,0x00,0xd1,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
117 v_cvt_f16_i16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
118 // GFX12: [0x05,0x00,0xd1,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
120 v_cvt_f16_i16_e64_dpp v255, v255 clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
121 // GFX12: [0xff,0x80,0xd1,0xd5,0xe9,0x00,0x00,0x18,0xff,0x00,0x00,0x00]
123 v_cvt_f16_u16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
124 // GFX12: [0x05,0x00,0xd0,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
126 v_cvt_f16_u16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
127 // GFX12: [0x05,0x00,0xd0,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
129 v_cvt_f16_u16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
130 // GFX12: [0x05,0x00,0xd0,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
132 v_cvt_f16_u16_e64_dpp v255, v255 clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
133 // GFX12: [0xff,0x80,0xd0,0xd5,0xe9,0x00,0x00,0x18,0xff,0x00,0x00,0x00]
135 v_cvt_f32_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
136 // GFX12: [0x05,0x00,0x8b,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
138 v_cvt_f32_f16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
139 // GFX12: [0x05,0x00,0x8b,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
141 v_cvt_f32_f16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
142 // GFX12: [0x05,0x00,0x8b,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
144 v_cvt_f32_f16_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
145 // GFX12: [0xff,0x81,0x8b,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
147 v_cvt_f32_fp8 v1, v2 dpp8:[7,6,5,4,3,2,1,0]
148 // GFX12: v_cvt_f32_fp8_dpp v1, v2 dpp8:[7,6,5,4,3,2,1,0] ; encoding: [0xe9,0xd8,0x02,0x7e,0x02,0x77,0x39,0x05]
150 v_cvt_f32_fp8 v1, v2 byte_sel:0 dpp8:[7,6,5,4,3,2,1,0]
151 // GFX12: v_cvt_f32_fp8_e64_dpp v1, v2 dpp8:[7,6,5,4,3,2,1,0] ; encoding: [0x01,0x00,0xec,0xd5,0xe9,0x00,0x00,0x00,0x02,0x77,0x39,0x05]
153 v_cvt_f32_fp8 v1, v2 byte_sel:1 dpp8:[7,6,5,4,3,2,1,0]
154 // GFX12: v_cvt_f32_fp8_e64_dpp v1, v2 byte_sel:1 dpp8:[7,6,5,4,3,2,1,0] ; encoding: [0x01,0x10,0xec,0xd5,0xe9,0x00,0x00,0x00,0x02,0x77,0x39,0x05]
156 v_cvt_f32_fp8 v1, v2 byte_sel:2 dpp8:[7,6,5,4,3,2,1,0]
157 // GFX12: v_cvt_f32_fp8_e64_dpp v1, v2 byte_sel:2 dpp8:[7,6,5,4,3,2,1,0] ; encoding: [0x01,0x08,0xec,0xd5,0xe9,0x00,0x00,0x00,0x02,0x77,0x39,0x05]
159 v_cvt_f32_fp8 v1, v2 byte_sel:3 dpp8:[7,6,5,4,3,2,1,0]
160 // GFX12: v_cvt_f32_fp8_e64_dpp v1, v2 byte_sel:3 dpp8:[7,6,5,4,3,2,1,0] ; encoding: [0x01,0x18,0xec,0xd5,0xe9,0x00,0x00,0x00,0x02,0x77,0x39,0x05]
162 v_cvt_f32_bf8 v1, v2 dpp8:[7,6,5,4,3,2,1,0]
163 // GFX12: v_cvt_f32_bf8_dpp v1, v2 dpp8:[7,6,5,4,3,2,1,0] ; encoding: [0xe9,0xda,0x02,0x7e,0x02,0x77,0x39,0x05]
165 v_cvt_f32_bf8 v1, v2 byte_sel:0 dpp8:[7,6,5,4,3,2,1,0]
166 // GFX12: v_cvt_f32_bf8_e64_dpp v1, v2 dpp8:[7,6,5,4,3,2,1,0] ; encoding: [0x01,0x00,0xed,0xd5,0xe9,0x00,0x00,0x00,0x02,0x77,0x39,0x05]
168 v_cvt_f32_bf8 v1, v2 byte_sel:1 dpp8:[7,6,5,4,3,2,1,0]
169 // GFX12: v_cvt_f32_bf8_e64_dpp v1, v2 byte_sel:1 dpp8:[7,6,5,4,3,2,1,0] ; encoding: [0x01,0x10,0xed,0xd5,0xe9,0x00,0x00,0x00,0x02,0x77,0x39,0x05]
171 v_cvt_f32_bf8 v1, v2 byte_sel:2 dpp8:[7,6,5,4,3,2,1,0]
172 // GFX12: v_cvt_f32_bf8_e64_dpp v1, v2 byte_sel:2 dpp8:[7,6,5,4,3,2,1,0] ; encoding: [0x01,0x08,0xed,0xd5,0xe9,0x00,0x00,0x00,0x02,0x77,0x39,0x05]
174 v_cvt_f32_bf8 v1, v2 byte_sel:3 dpp8:[7,6,5,4,3,2,1,0]
175 // GFX12: v_cvt_f32_bf8_e64_dpp v1, v2 byte_sel:3 dpp8:[7,6,5,4,3,2,1,0] ; encoding: [0x01,0x18,0xed,0xd5,0xe9,0x00,0x00,0x00,0x02,0x77,0x39,0x05]
177 v_cvt_f32_i32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
178 // GFX12: [0x05,0x00,0x85,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
180 v_cvt_f32_i32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
181 // GFX12: [0x05,0x00,0x85,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
183 v_cvt_f32_i32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
184 // GFX12: [0x05,0x00,0x85,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
186 v_cvt_f32_i32_e64_dpp v255, v255 clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
187 // GFX12: [0xff,0x80,0x85,0xd5,0xe9,0x00,0x00,0x18,0xff,0x00,0x00,0x00]
189 v_cvt_f32_u32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
190 // GFX12: [0x05,0x00,0x86,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
192 v_cvt_f32_u32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
193 // GFX12: [0x05,0x00,0x86,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
195 v_cvt_f32_u32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
196 // GFX12: [0x05,0x00,0x86,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
198 v_cvt_f32_u32_e64_dpp v255, v255 clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
199 // GFX12: [0xff,0x80,0x86,0xd5,0xe9,0x00,0x00,0x18,0xff,0x00,0x00,0x00]
201 v_cvt_f32_ubyte0_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
202 // GFX12: [0x05,0x00,0x91,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
204 v_cvt_f32_ubyte0_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
205 // GFX12: [0x05,0x00,0x91,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
207 v_cvt_f32_ubyte0_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
208 // GFX12: [0x05,0x00,0x91,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
210 v_cvt_f32_ubyte0_e64_dpp v255, v255 clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
211 // GFX12: [0xff,0x80,0x91,0xd5,0xe9,0x00,0x00,0x18,0xff,0x00,0x00,0x00]
213 v_cvt_f32_ubyte1_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
214 // GFX12: [0x05,0x00,0x92,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
216 v_cvt_f32_ubyte1_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
217 // GFX12: [0x05,0x00,0x92,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
219 v_cvt_f32_ubyte1_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
220 // GFX12: [0x05,0x00,0x92,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
222 v_cvt_f32_ubyte1_e64_dpp v255, v255 clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
223 // GFX12: [0xff,0x80,0x92,0xd5,0xe9,0x00,0x00,0x18,0xff,0x00,0x00,0x00]
225 v_cvt_f32_ubyte2_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
226 // GFX12: [0x05,0x00,0x93,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
228 v_cvt_f32_ubyte2_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
229 // GFX12: [0x05,0x00,0x93,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
231 v_cvt_f32_ubyte2_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
232 // GFX12: [0x05,0x00,0x93,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
234 v_cvt_f32_ubyte2_e64_dpp v255, v255 clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
235 // GFX12: [0xff,0x80,0x93,0xd5,0xe9,0x00,0x00,0x18,0xff,0x00,0x00,0x00]
237 v_cvt_f32_ubyte3_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
238 // GFX12: [0x05,0x00,0x94,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
240 v_cvt_f32_ubyte3_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
241 // GFX12: [0x05,0x00,0x94,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
243 v_cvt_f32_ubyte3_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
244 // GFX12: [0x05,0x00,0x94,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
246 v_cvt_f32_ubyte3_e64_dpp v255, v255 clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
247 // GFX12: [0xff,0x80,0x94,0xd5,0xe9,0x00,0x00,0x18,0xff,0x00,0x00,0x00]
249 v_cvt_floor_i32_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
250 // GFX12: [0x05,0x00,0x8d,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
252 v_cvt_floor_i32_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
253 // GFX12: [0x05,0x00,0x8d,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
255 v_cvt_floor_i32_f32_e64_dpp v255, -|v255| dpp8:[0,0,0,0,0,0,0,0] fi:0
256 // GFX12: [0xff,0x01,0x8d,0xd5,0xe9,0x00,0x00,0x20,0xff,0x00,0x00,0x00]
258 v_cvt_flr_i32_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
259 // GFX12: [0x05,0x00,0x8d,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
261 v_cvt_flr_i32_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
262 // GFX12: [0x05,0x00,0x8d,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
264 v_cvt_flr_i32_f32_e64_dpp v255, -|v255| dpp8:[0,0,0,0,0,0,0,0] fi:0
265 // GFX12: [0xff,0x01,0x8d,0xd5,0xe9,0x00,0x00,0x20,0xff,0x00,0x00,0x00]
267 v_cvt_i16_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
268 // GFX12: [0x05,0x00,0xd3,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
270 v_cvt_i16_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
271 // GFX12: [0x05,0x00,0xd3,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
273 v_cvt_i16_f16_e64_dpp v255, -|v255| clamp dpp8:[0,0,0,0,0,0,0,0] fi:0
274 // GFX12: [0xff,0x81,0xd3,0xd5,0xe9,0x00,0x00,0x20,0xff,0x00,0x00,0x00]
276 v_cvt_i32_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
277 // GFX12: [0x05,0x00,0x88,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
279 v_cvt_i32_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
280 // GFX12: [0x05,0x00,0x88,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
282 v_cvt_i32_f32_e64_dpp v255, -|v255| clamp dpp8:[0,0,0,0,0,0,0,0] fi:0
283 // GFX12: [0xff,0x81,0x88,0xd5,0xe9,0x00,0x00,0x20,0xff,0x00,0x00,0x00]
285 v_cvt_i32_i16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
286 // GFX12: [0x05,0x00,0xea,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
288 v_cvt_i32_i16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
289 // GFX12: [0x05,0x00,0xea,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
291 v_cvt_i32_i16_e64_dpp v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0
292 // GFX12: [0xff,0x00,0xea,0xd5,0xe9,0x00,0x00,0x00,0xff,0x00,0x00,0x00]
294 v_cvt_nearest_i32_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
295 // GFX12: [0x05,0x00,0x8c,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
297 v_cvt_nearest_i32_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
298 // GFX12: [0x05,0x00,0x8c,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
300 v_cvt_nearest_i32_f32_e64_dpp v255, -|v255| dpp8:[0,0,0,0,0,0,0,0] fi:0
301 // GFX12: [0xff,0x01,0x8c,0xd5,0xe9,0x00,0x00,0x20,0xff,0x00,0x00,0x00]
303 v_cvt_norm_i16_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
304 // GFX12: [0x05,0x00,0xe3,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
306 v_cvt_norm_i16_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
307 // GFX12: [0x05,0x00,0xe3,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
309 v_cvt_norm_i16_f16_e64_dpp v255, -|v255| dpp8:[0,0,0,0,0,0,0,0] fi:0
310 // GFX12: [0xff,0x01,0xe3,0xd5,0xe9,0x00,0x00,0x20,0xff,0x00,0x00,0x00]
312 v_cvt_norm_u16_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
313 // GFX12: [0x05,0x00,0xe4,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
315 v_cvt_norm_u16_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
316 // GFX12: [0x05,0x00,0xe4,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
318 v_cvt_norm_u16_f16_e64_dpp v255, -|v255| dpp8:[0,0,0,0,0,0,0,0] fi:0
319 // GFX12: [0xff,0x01,0xe4,0xd5,0xe9,0x00,0x00,0x20,0xff,0x00,0x00,0x00]
321 v_cvt_off_f32_i4_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
322 // GFX12: [0x05,0x00,0x8e,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
324 v_cvt_off_f32_i4_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
325 // GFX12: [0x05,0x00,0x8e,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
327 v_cvt_off_f32_i4_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
328 // GFX12: [0x05,0x00,0x8e,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
330 v_cvt_off_f32_i4_e64_dpp v255, v255 clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
331 // GFX12: [0xff,0x80,0x8e,0xd5,0xe9,0x00,0x00,0x18,0xff,0x00,0x00,0x00]
333 v_cvt_rpi_i32_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
334 // GFX12: [0x05,0x00,0x8c,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
336 v_cvt_rpi_i32_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
337 // GFX12: [0x05,0x00,0x8c,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
339 v_cvt_rpi_i32_f32_e64_dpp v255, -|v255| dpp8:[0,0,0,0,0,0,0,0] fi:0
340 // GFX12: [0xff,0x01,0x8c,0xd5,0xe9,0x00,0x00,0x20,0xff,0x00,0x00,0x00]
342 v_cvt_u16_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
343 // GFX12: [0x05,0x00,0xd2,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
345 v_cvt_u16_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
346 // GFX12: [0x05,0x00,0xd2,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
348 v_cvt_u16_f16_e64_dpp v255, -|v255| clamp dpp8:[0,0,0,0,0,0,0,0] fi:0
349 // GFX12: [0xff,0x81,0xd2,0xd5,0xe9,0x00,0x00,0x20,0xff,0x00,0x00,0x00]
351 v_cvt_u32_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
352 // GFX12: [0x05,0x00,0x87,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
354 v_cvt_u32_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
355 // GFX12: [0x05,0x00,0x87,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
357 v_cvt_u32_f32_e64_dpp v255, -|v255| clamp dpp8:[0,0,0,0,0,0,0,0] fi:0
358 // GFX12: [0xff,0x81,0x87,0xd5,0xe9,0x00,0x00,0x20,0xff,0x00,0x00,0x00]
360 v_cvt_u32_u16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
361 // GFX12: [0x05,0x00,0xeb,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
363 v_cvt_u32_u16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
364 // GFX12: [0x05,0x00,0xeb,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
366 v_cvt_u32_u16_e64_dpp v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0
367 // GFX12: [0xff,0x00,0xeb,0xd5,0xe9,0x00,0x00,0x00,0xff,0x00,0x00,0x00]
369 v_exp_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
370 // GFX12: [0x05,0x00,0xd8,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
372 v_exp_f16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
373 // GFX12: [0x05,0x00,0xd8,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
375 v_exp_f16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
376 // GFX12: [0x05,0x00,0xd8,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
378 v_exp_f16_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
379 // GFX12: [0xff,0x81,0xd8,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
381 v_exp_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
382 // GFX12: [0x05,0x00,0xa5,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
384 v_exp_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
385 // GFX12: [0x05,0x00,0xa5,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
387 v_exp_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
388 // GFX12: [0x05,0x00,0xa5,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
390 v_exp_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
391 // GFX12: [0xff,0x81,0xa5,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
393 v_ffbh_i32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
394 // GFX12: [0x05,0x00,0xbb,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
396 v_ffbh_i32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
397 // GFX12: [0x05,0x00,0xbb,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
399 v_ffbh_i32_e64_dpp v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0
400 // GFX12: [0xff,0x00,0xbb,0xd5,0xe9,0x00,0x00,0x00,0xff,0x00,0x00,0x00]
402 v_ffbh_u32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
403 // GFX12: [0x05,0x00,0xb9,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
405 v_ffbh_u32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
406 // GFX12: [0x05,0x00,0xb9,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
408 v_ffbh_u32_e64_dpp v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0
409 // GFX12: [0xff,0x00,0xb9,0xd5,0xe9,0x00,0x00,0x00,0xff,0x00,0x00,0x00]
411 v_ffbl_b32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
412 // GFX12: [0x05,0x00,0xba,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
414 v_ffbl_b32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
415 // GFX12: [0x05,0x00,0xba,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
417 v_ffbl_b32_e64_dpp v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0
418 // GFX12: [0xff,0x00,0xba,0xd5,0xe9,0x00,0x00,0x00,0xff,0x00,0x00,0x00]
420 v_floor_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
421 // GFX12: [0x05,0x00,0xdb,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
423 v_floor_f16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
424 // GFX12: [0x05,0x00,0xdb,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
426 v_floor_f16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
427 // GFX12: [0x05,0x00,0xdb,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
429 v_floor_f16_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
430 // GFX12: [0xff,0x81,0xdb,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
432 v_floor_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
433 // GFX12: [0x05,0x00,0xa4,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
435 v_floor_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
436 // GFX12: [0x05,0x00,0xa4,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
438 v_floor_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
439 // GFX12: [0x05,0x00,0xa4,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
441 v_floor_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
442 // GFX12: [0xff,0x81,0xa4,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
444 v_fract_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
445 // GFX12: [0x05,0x00,0xdf,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
447 v_fract_f16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
448 // GFX12: [0x05,0x00,0xdf,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
450 v_fract_f16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
451 // GFX12: [0x05,0x00,0xdf,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
453 v_fract_f16_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
454 // GFX12: [0xff,0x81,0xdf,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
456 v_fract_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
457 // GFX12: [0x05,0x00,0xa0,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
459 v_fract_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
460 // GFX12: [0x05,0x00,0xa0,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
462 v_fract_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
463 // GFX12: [0x05,0x00,0xa0,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
465 v_fract_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
466 // GFX12: [0xff,0x81,0xa0,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
468 v_frexp_exp_i16_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
469 // GFX12: [0x05,0x00,0xda,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
471 v_frexp_exp_i16_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
472 // GFX12: [0x05,0x00,0xda,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
474 v_frexp_exp_i16_f16_e64_dpp v255, -|v255| dpp8:[0,0,0,0,0,0,0,0] fi:0
475 // GFX12: [0xff,0x01,0xda,0xd5,0xe9,0x00,0x00,0x20,0xff,0x00,0x00,0x00]
477 v_frexp_exp_i32_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
478 // GFX12: [0x05,0x00,0xbf,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
480 v_frexp_exp_i32_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
481 // GFX12: [0x05,0x00,0xbf,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
483 v_frexp_exp_i32_f32_e64_dpp v255, -|v255| dpp8:[0,0,0,0,0,0,0,0] fi:0
484 // GFX12: [0xff,0x01,0xbf,0xd5,0xe9,0x00,0x00,0x20,0xff,0x00,0x00,0x00]
486 v_frexp_mant_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
487 // GFX12: [0x05,0x00,0xd9,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
489 v_frexp_mant_f16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
490 // GFX12: [0x05,0x00,0xd9,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
492 v_frexp_mant_f16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
493 // GFX12: [0x05,0x00,0xd9,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
495 v_frexp_mant_f16_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
496 // GFX12: [0xff,0x81,0xd9,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
498 v_frexp_mant_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
499 // GFX12: [0x05,0x00,0xc0,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
501 v_frexp_mant_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
502 // GFX12: [0x05,0x00,0xc0,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
504 v_frexp_mant_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
505 // GFX12: [0x05,0x00,0xc0,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
507 v_frexp_mant_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
508 // GFX12: [0xff,0x81,0xc0,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
510 v_log_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
511 // GFX12: [0x05,0x00,0xd7,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
513 v_log_f16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
514 // GFX12: [0x05,0x00,0xd7,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
516 v_log_f16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
517 // GFX12: [0x05,0x00,0xd7,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
519 v_log_f16_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
520 // GFX12: [0xff,0x81,0xd7,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
522 v_log_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
523 // GFX12: [0x05,0x00,0xa7,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
525 v_log_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
526 // GFX12: [0x05,0x00,0xa7,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
528 v_log_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
529 // GFX12: [0x05,0x00,0xa7,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
531 v_log_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
532 // GFX12: [0xff,0x81,0xa7,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
534 v_mov_b32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
535 // GFX12: [0x05,0x00,0x81,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
537 v_mov_b32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
538 // GFX12: [0x05,0x00,0x81,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
540 v_mov_b32_e64_dpp v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0
541 // GFX12: [0xff,0x00,0x81,0xd5,0xe9,0x00,0x00,0x00,0xff,0x00,0x00,0x00]
543 v_movrels_b32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
544 // GFX12: [0x05,0x00,0xc3,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
546 v_movrels_b32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
547 // GFX12: [0x05,0x00,0xc3,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
549 v_movrels_b32_e64_dpp v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0
550 // GFX12: [0xff,0x00,0xc3,0xd5,0xe9,0x00,0x00,0x00,0xff,0x00,0x00,0x00]
552 v_not_b16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
553 // GFX12: [0x05,0x00,0xe9,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
555 v_not_b16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
556 // GFX12: [0x05,0x00,0xe9,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
558 v_not_b16_e64_dpp v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0
559 // GFX12: [0xff,0x00,0xe9,0xd5,0xe9,0x00,0x00,0x00,0xff,0x00,0x00,0x00]
561 v_not_b32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
562 // GFX12: [0x05,0x00,0xb7,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
564 v_not_b32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
565 // GFX12: [0x05,0x00,0xb7,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
567 v_not_b32_e64_dpp v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0
568 // GFX12: [0xff,0x00,0xb7,0xd5,0xe9,0x00,0x00,0x00,0xff,0x00,0x00,0x00]
570 v_rcp_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
571 // GFX12: [0x05,0x00,0xd4,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
573 v_rcp_f16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
574 // GFX12: [0x05,0x00,0xd4,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
576 v_rcp_f16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
577 // GFX12: [0x05,0x00,0xd4,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
579 v_rcp_f16_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
580 // GFX12: [0xff,0x81,0xd4,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
582 v_rcp_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
583 // GFX12: [0x05,0x00,0xaa,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
585 v_rcp_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
586 // GFX12: [0x05,0x00,0xaa,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
588 v_rcp_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
589 // GFX12: [0x05,0x00,0xaa,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
591 v_rcp_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
592 // GFX12: [0xff,0x81,0xaa,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
594 v_rcp_iflag_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
595 // GFX12: [0x05,0x00,0xab,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
597 v_rcp_iflag_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
598 // GFX12: [0x05,0x00,0xab,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
600 v_rcp_iflag_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
601 // GFX12: [0x05,0x00,0xab,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
603 v_rcp_iflag_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
604 // GFX12: [0xff,0x81,0xab,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
606 v_rndne_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
607 // GFX12: [0x05,0x00,0xde,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
609 v_rndne_f16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
610 // GFX12: [0x05,0x00,0xde,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
612 v_rndne_f16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
613 // GFX12: [0x05,0x00,0xde,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
615 v_rndne_f16_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
616 // GFX12: [0xff,0x81,0xde,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
618 v_rndne_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
619 // GFX12: [0x05,0x00,0xa3,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
621 v_rndne_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
622 // GFX12: [0x05,0x00,0xa3,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
624 v_rndne_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
625 // GFX12: [0x05,0x00,0xa3,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
627 v_rndne_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
628 // GFX12: [0xff,0x81,0xa3,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
630 v_rsq_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
631 // GFX12: [0x05,0x00,0xd6,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
633 v_rsq_f16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
634 // GFX12: [0x05,0x00,0xd6,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
636 v_rsq_f16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
637 // GFX12: [0x05,0x00,0xd6,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
639 v_rsq_f16_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
640 // GFX12: [0xff,0x81,0xd6,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
642 v_rsq_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
643 // GFX12: [0x05,0x00,0xae,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
645 v_rsq_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
646 // GFX12: [0x05,0x00,0xae,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
648 v_rsq_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
649 // GFX12: [0x05,0x00,0xae,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
651 v_rsq_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
652 // GFX12: [0xff,0x81,0xae,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
654 v_sat_pk_u8_i16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
655 // GFX12: [0x05,0x00,0xe2,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
657 v_sat_pk_u8_i16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0] fi:1
658 // GFX12: [0x05,0x00,0xe2,0xd5,0xea,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
660 v_sat_pk_u8_i16_e64_dpp v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0
661 // GFX12: [0xff,0x00,0xe2,0xd5,0xe9,0x00,0x00,0x00,0xff,0x00,0x00,0x00]
663 v_sin_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
664 // GFX12: [0x05,0x00,0xe0,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
666 v_sin_f16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
667 // GFX12: [0x05,0x00,0xe0,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
669 v_sin_f16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
670 // GFX12: [0x05,0x00,0xe0,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
672 v_sin_f16_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
673 // GFX12: [0xff,0x81,0xe0,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
675 v_sin_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
676 // GFX12: [0x05,0x00,0xb5,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
678 v_sin_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
679 // GFX12: [0x05,0x00,0xb5,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
681 v_sin_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
682 // GFX12: [0x05,0x00,0xb5,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
684 v_sin_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
685 // GFX12: [0xff,0x81,0xb5,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
687 v_sqrt_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
688 // GFX12: [0x05,0x00,0xd5,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
690 v_sqrt_f16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
691 // GFX12: [0x05,0x00,0xd5,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
693 v_sqrt_f16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
694 // GFX12: [0x05,0x00,0xd5,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
696 v_sqrt_f16_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
697 // GFX12: [0xff,0x81,0xd5,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
699 v_sqrt_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
700 // GFX12: [0x05,0x00,0xb3,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
702 v_sqrt_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
703 // GFX12: [0x05,0x00,0xb3,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
705 v_sqrt_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
706 // GFX12: [0x05,0x00,0xb3,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
708 v_sqrt_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
709 // GFX12: [0xff,0x81,0xb3,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
711 v_trunc_f16_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
712 // GFX12: [0x05,0x00,0xdd,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
714 v_trunc_f16_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
715 // GFX12: [0x05,0x00,0xdd,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
717 v_trunc_f16_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
718 // GFX12: [0x05,0x00,0xdd,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
720 v_trunc_f16_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
721 // GFX12: [0xff,0x81,0xdd,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]
723 v_trunc_f32_e64_dpp v5, v1 dpp8:[7,6,5,4,3,2,1,0]
724 // GFX12: [0x05,0x00,0xa1,0xd5,0xe9,0x00,0x00,0x00,0x01,0x77,0x39,0x05]
726 v_trunc_f32_e64_dpp v5, v1 mul:2 dpp8:[7,6,5,4,3,2,1,0]
727 // GFX12: [0x05,0x00,0xa1,0xd5,0xe9,0x00,0x00,0x08,0x01,0x77,0x39,0x05]
729 v_trunc_f32_e64_dpp v5, v1 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1
730 // GFX12: [0x05,0x00,0xa1,0xd5,0xea,0x00,0x00,0x10,0x01,0x77,0x39,0x05]
732 v_trunc_f32_e64_dpp v255, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0
733 // GFX12: [0xff,0x81,0xa1,0xd5,0xe9,0x00,0x00,0x38,0xff,0x00,0x00,0x00]