@@ -94,6 +94,18 @@ v_ceil_f16 v5.l, src_scc
94
94
v_ceil_f16 v127.l, 0xfe0b
95
95
// GFX12: v_ceil_f16_e32 v127.l, 0xfe0b ; encoding: [0xff,0xb8,0xfe,0x7e,0x0b,0xfe,0x00,0x00]
96
96
97
+ v_ceil_f16 v5.l, v1.h
98
+ // GFX12: v_ceil_f16_e32 v5.l, v1.h ; encoding: [0x81,0xb9,0x0a,0x7e]
99
+
100
+ v_ceil_f16 v5.l, v127.h
101
+ // GFX12: v_ceil_f16_e32 v5.l, v127.h ; encoding: [0xff,0xb9,0x0a,0x7e]
102
+
103
+ v_ceil_f16 v5.h, src_scc
104
+ // GFX12: v_ceil_f16_e32 v5.h, src_scc ; encoding: [0xfd,0xb8,0x0a,0x7f]
105
+
106
+ v_ceil_f16 v127.h, 0xfe0b
107
+ // GFX12: v_ceil_f16_e32 v127.h, 0xfe0b ; encoding: [0xff,0xb8,0xfe,0x7f,0x0b,0xfe,0x00,0x00]
108
+
97
109
v_ceil_f32 v5, v1
98
110
// GFX12: v_ceil_f32_e32 v5, v1 ; encoding: [0x01,0x45,0x0a,0x7e]
99
111
@@ -1886,6 +1898,18 @@ v_exp_f16 v5.l, src_scc
1886
1898
v_exp_f16 v127.l, 0xfe0b
1887
1899
// GFX12: v_exp_f16_e32 v127.l, 0xfe0b ; encoding: [0xff,0xb0,0xfe,0x7e,0x0b,0xfe,0x00,0x00]
1888
1900
1901
+ v_exp_f16 v5.l, v1.h
1902
+ // GFX12: v_exp_f16_e32 v5.l, v1.h ; encoding: [0x81,0xb1,0x0a,0x7e]
1903
+
1904
+ v_exp_f16 v5.l, v127.h
1905
+ // GFX12: v_exp_f16_e32 v5.l, v127.h ; encoding: [0xff,0xb1,0x0a,0x7e]
1906
+
1907
+ v_exp_f16 v5.h, src_scc
1908
+ // GFX12: v_exp_f16_e32 v5.h, src_scc ; encoding: [0xfd,0xb0,0x0a,0x7f]
1909
+
1910
+ v_exp_f16 v127.h, 0xfe0b
1911
+ // GFX12: v_exp_f16_e32 v127.h, 0xfe0b ; encoding: [0xff,0xb0,0xfe,0x7f,0x0b,0xfe,0x00,0x00]
1912
+
1889
1913
v_exp_f32 v5, v1
1890
1914
// GFX12: v_exp_f32_e32 v5, v1 ; encoding: [0x01,0x4b,0x0a,0x7e]
1891
1915
@@ -2111,6 +2135,18 @@ v_floor_f16 v5.l, src_scc
2111
2135
v_floor_f16 v127.l, 0xfe0b
2112
2136
// GFX12: v_floor_f16_e32 v127.l, 0xfe0b ; encoding: [0xff,0xb6,0xfe,0x7e,0x0b,0xfe,0x00,0x00]
2113
2137
2138
+ v_floor_f16 v5.l, v1.h
2139
+ // GFX12: v_floor_f16_e32 v5.l, v1.h ; encoding: [0x81,0xb7,0x0a,0x7e]
2140
+
2141
+ v_floor_f16 v5.l, v127.h
2142
+ // GFX12: v_floor_f16_e32 v5.l, v127.h ; encoding: [0xff,0xb7,0x0a,0x7e]
2143
+
2144
+ v_floor_f16 v5.h, src_scc
2145
+ // GFX12: v_floor_f16_e32 v5.h, src_scc ; encoding: [0xfd,0xb6,0x0a,0x7f]
2146
+
2147
+ v_floor_f16 v127.h, 0xfe0b
2148
+ // GFX12: v_floor_f16_e32 v127.h, 0xfe0b ; encoding: [0xff,0xb6,0xfe,0x7f,0x0b,0xfe,0x00,0x00]
2149
+
2114
2150
v_floor_f32 v5, v1
2115
2151
// GFX12: v_floor_f32_e32 v5, v1 ; encoding: [0x01,0x49,0x0a,0x7e]
2116
2152
@@ -2651,6 +2687,18 @@ v_log_f16 v5.l, src_scc
2651
2687
v_log_f16 v127.l, 0xfe0b
2652
2688
// GFX12: v_log_f16_e32 v127.l, 0xfe0b ; encoding: [0xff,0xae,0xfe,0x7e,0x0b,0xfe,0x00,0x00]
2653
2689
2690
+ v_log_f16 v5.l, v1.h
2691
+ // GFX12: v_log_f16_e32 v5.l, v1.h ; encoding: [0x81,0xaf,0x0a,0x7e]
2692
+
2693
+ v_log_f16 v5.l, v127.h
2694
+ // GFX12: v_log_f16_e32 v5.l, v127.h ; encoding: [0xff,0xaf,0x0a,0x7e]
2695
+
2696
+ v_log_f16 v5.h, src_scc
2697
+ // GFX12: v_log_f16_e32 v5.h, src_scc ; encoding: [0xfd,0xae,0x0a,0x7f]
2698
+
2699
+ v_log_f16 v127.h, 0xfe0b
2700
+ // GFX12: v_log_f16_e32 v127.h, 0xfe0b ; encoding: [0xff,0xae,0xfe,0x7f,0x0b,0xfe,0x00,0x00]
2701
+
2654
2702
v_log_f32 v5, v1
2655
2703
// GFX12: v_log_f32_e32 v5, v1 ; encoding: [0x01,0x4f,0x0a,0x7e]
2656
2704
@@ -2964,6 +3012,18 @@ v_rcp_f16 v5.l, src_scc
2964
3012
v_rcp_f16 v127.l, 0xfe0b
2965
3013
// GFX12: v_rcp_f16_e32 v127.l, 0xfe0b ; encoding: [0xff,0xa8,0xfe,0x7e,0x0b,0xfe,0x00,0x00]
2966
3014
3015
+ v_rcp_f16 v5.l, v1.h
3016
+ // GFX12: v_rcp_f16_e32 v5.l, v1.h ; encoding: [0x81,0xa9,0x0a,0x7e]
3017
+
3018
+ v_rcp_f16 v5.l, v127.h
3019
+ // GFX12: v_rcp_f16_e32 v5.l, v127.h ; encoding: [0xff,0xa9,0x0a,0x7e]
3020
+
3021
+ v_rcp_f16 v5.h, src_scc
3022
+ // GFX12: v_rcp_f16_e32 v5.h, src_scc ; encoding: [0xfd,0xa8,0x0a,0x7f]
3023
+
3024
+ v_rcp_f16 v127.h, 0xfe0b
3025
+ // GFX12: v_rcp_f16_e32 v127.h, 0xfe0b ; encoding: [0xff,0xa8,0xfe,0x7f,0x0b,0xfe,0x00,0x00]
3026
+
2967
3027
v_rcp_f32 v5, v1
2968
3028
// GFX12: v_rcp_f32_e32 v5, v1 ; encoding: [0x01,0x55,0x0a,0x7e]
2969
3029
@@ -3291,6 +3351,18 @@ v_rsq_f16 v5.l, src_scc
3291
3351
v_rsq_f16 v127.l, 0xfe0b
3292
3352
// GFX12: v_rsq_f16_e32 v127.l, 0xfe0b ; encoding: [0xff,0xac,0xfe,0x7e,0x0b,0xfe,0x00,0x00]
3293
3353
3354
+ v_rsq_f16 v5.l, v1.h
3355
+ // GFX12: v_rsq_f16_e32 v5.l, v1.h ; encoding: [0x81,0xad,0x0a,0x7e]
3356
+
3357
+ v_rsq_f16 v5.l, v127.h
3358
+ // GFX12: v_rsq_f16_e32 v5.l, v127.h ; encoding: [0xff,0xad,0x0a,0x7e]
3359
+
3360
+ v_rsq_f16 v5.h, src_scc
3361
+ // GFX12: v_rsq_f16_e32 v5.h, src_scc ; encoding: [0xfd,0xac,0x0a,0x7f]
3362
+
3363
+ v_rsq_f16 v127.h, 0xfe0b
3364
+ // GFX12: v_rsq_f16_e32 v127.h, 0xfe0b ; encoding: [0xff,0xac,0xfe,0x7f,0x0b,0xfe,0x00,0x00]
3365
+
3294
3366
v_rsq_f32 v5, v1
3295
3367
// GFX12: v_rsq_f32_e32 v5, v1 ; encoding: [0x01,0x5d,0x0a,0x7e]
3296
3368
@@ -3585,6 +3657,18 @@ v_sqrt_f16 v5.l, src_scc
3585
3657
v_sqrt_f16 v127.l, 0xfe0b
3586
3658
// GFX12: v_sqrt_f16_e32 v127.l, 0xfe0b ; encoding: [0xff,0xaa,0xfe,0x7e,0x0b,0xfe,0x00,0x00]
3587
3659
3660
+ v_sqrt_f16 v5.l, v1.h
3661
+ // GFX12: v_sqrt_f16_e32 v5.l, v1.h ; encoding: [0x81,0xab,0x0a,0x7e]
3662
+
3663
+ v_sqrt_f16 v5.l, v127.h
3664
+ // GFX12: v_sqrt_f16_e32 v5.l, v127.h ; encoding: [0xff,0xab,0x0a,0x7e]
3665
+
3666
+ v_sqrt_f16 v5.h, src_scc
3667
+ // GFX12: v_sqrt_f16_e32 v5.h, src_scc ; encoding: [0xfd,0xaa,0x0a,0x7f]
3668
+
3669
+ v_sqrt_f16 v127.h, 0xfe0b
3670
+ // GFX12: v_sqrt_f16_e32 v127.h, 0xfe0b ; encoding: [0xff,0xaa,0xfe,0x7f,0x0b,0xfe,0x00,0x00]
3671
+
3588
3672
v_sqrt_f32 v5, v1
3589
3673
// GFX12: v_sqrt_f32_e32 v5, v1 ; encoding: [0x01,0x67,0x0a,0x7e]
3590
3674
0 commit comments