[version]
SoC_version=Ascend950PR_957bx
Short_SoC_version=Ascend950
AIC_version=AIC-C-310
CCEC_AIC_version=dav-c310-cube
CCEC_AIV_version=dav-c310-vec
CCEC_CUBE_version=dav-c310-cube
CCEC_VECTOR_version=dav-c310-vec
Compiler_aicpu_support_os=true
NpuArch=3510
Chip_type=15
[DieInfo]
io_die_num=2
[SoCInfo]
ai_core_cnt=28
cube_core_cnt=28
vector_core_cnt=56
ai_cpu_cnt=5
memory_type=0
memory_size=120259084288
l2_type=0
l2_size=117440512
l2PageNum=64
normal_stream_num=65535
normal_stream_depth=2049
huge_stream_num=0
huge_stream_depth=0
prefetch_num=8
core_type_list=CubeCore,VectorCore
cube_vector_combine=split
hardware_sync_betweencore=1
vir_type_list=7,14,28
support_bf16=1
support_inf_nan=1
[AICoreSpec]
cube_freq=1650
cube_m_size=16
cube_n_size=16
cube_k_size=16
vec_calc_size=128
l0_a_size=65536
l0_b_size=65536
l0_c_size=262144
l1_size=524288
fb0_size=4096
fb1_size=2048
fb2_size=0
fb3_size=0
bt_size=4096
smask_buffer=0
ub_size=253952
ubblock_size=32
ubbank_size=4096
ubbank_num=16
ubsubbank_num=64
ubburst_in_one_block=32
ubbank_group_num=8
support_fixpipe=1
sparsity=0
l0a_layout_zN=1
vector_reg_width=256
predicate_reg_width=32
wide_reg_width=768
padding_size=0
[AICoreMemoryRates]
ddr_rate=27
ddr_read_rate=27
ddr_write_rate=27
l2_rate=100
l2_read_rate=100
l2_write_rate=100
l1_to_l0_a_rate=256
l1_to_l0_b_rate=256
l1_to_ub_rate=128
l0_c_to_ub_rate=128
ub_to_l2_rate=128
ub_to_ddr_rate=128
ub_to_l1_rate=128
[DtypeMKN]
Default=16,16,16
DT_FLOAT8_E4M3FN=16,32,16
DT_HIFLOAT8=16,32,16
DT_HIFLOAT4=16,64,16
DT_FLOAT4_E2M1=16,64,16
DT_FLOAT4_E1M2=16,64,16
DT_INT8=16,32,16
DT_UINT8=16,32,16
DT_INT4=16,64,16
DT_INT2=16,128,16
DT_UINT2=16,128,16
DT_UINT1=16,256,16
DT_FLOAT=16,8,16
DT_INT32=16,8,16
[CPUCache]
AICPUSyncBySW=0
TSCPUSyncBySW=0
[AICoreintrinsicDtypeMap]
0=Intrinsic_vrelu|float16,float32,int32
1=Intrinsic_vexp|float16,float32
2=Intrinsic_vln|float16,float32
3=Intrinsic_vabs|float16,float32
4=Intrinsic_vrec|float16,float32
5=Intrinsic_vrsqrt|float16,float32
6=Intrinsic_vnot|int16,uint16
7=Intrinsic_vadd|float16,float32,int32,int16
8=Intrinsic_vsub|float16,float32,int32,int16
9=Intrinsic_vmul|float16,float32,int32,int16
10=Intrinsic_vmax|float16,float32,int32,int16
11=Intrinsic_vmin|float16,float32,int32,int16
12=Intrinsic_vand|int16,uint16
13=Intrinsic_vor|int16,uint16
14=Intrinsic_vadds|float16,float32,int32,int16
15=Intrinsic_vmuls|float16,float32,int32,int16
16=Intrinsic_vaxpy|f16f16f16,f32f32f32,float16,float32
17=Intrinsic_vcmpv_ne|float16,float32
18=Intrinsic_vcmpv_eq|float16,float32,int32
19=Intrinsic_vcmpv_gt|float16,float32
20=Intrinsic_vcmpv_ge|float16,float32
21=Intrinsic_vcmpv_lt|float16,float32
22=Intrinsic_vcmpv_le|float16,float32
23=Intrinsic_vconv|f322f32r,f322f32f,f322f32c,deq,f322f32a,f322f32z,f322f16,f322f16r,f322f16f,f322f16c,f322f16a,f322f16z,f322f16o,f322s64r,f322s64f,f322s64c,f322s64a,f322s64z,f322s32r,f322s32f,f322s32c,f322s32a,f322s32z,f322s16r,f322s16f,f322s16c,f322s16a,f322s16z,f162f32,f162s32r,f162s32f,f162s32c,f162s32a,f162s32z,f162s16r,f162s16f,f162s16c,f162s16a,f162s16z,f162s8,f162s8r,f162s8f,f162s8c,f162s8a,f162s8z,f162u8,f162u8r,f162u8f,f162u8c,f162u8a,f162u8z,u82f16,s82f16,s162f16,s162f16r,s162f16f,s162f16c,s162f16a,s162f16z,s162f32,s162s8,s162u8,s322f32,s322f32r,s322f32f,s322f32c,s322f32a,s322f32z,s322f16,s322s64,s322s16,s642s32,s642f32r,s642f32f,s642f32c,s642f32a,s642f32z,f322bf16r,f322bf16f,f322bf16c,f322bf16a,f322bf16z,bf162f32,bf162s32r,bf162s32f,bf162s32c,bf162s32a,bf162s32z
24=Intrinsic_vcadd|float16,float32
25=Intrinsic_vcmax|float16,float32
26=Intrinsic_vcmin|float16,float32
27=Intrinsic_vbi|float16,float32
28=Intrinsic_vlrelu|float16,float32
29=Intrinsic_scatter_vconv|f322f16,f162f32,f162s8,f162u8,s322f16,s322f32,u82f16,s82f16,f162s32r,f162s8f,f162u8f,f162s32f,f322s32f,f162s8c,f162u8c,f162s32c,f322s32c,f162s8a,f162u8a,f162s32a,f322s32a,f162s8z,f162u8z,f162s32z,f322s32z,f322f16o
30=Intrinsic_vreduce|int16,uint16,float16,float32,int32,uint32
31=Intrinsic_vmaxs|float16,float32,int32,int16
32=Intrinsic_vcmp|float16,float32
33=Intrinsic_vsel|float16,float32
34=Intrinsic_vsqrt|float16,float32
35=Intrinsic_vdiv|float16,float32
36=Intrinsic_vmins|float16,float32,int32,int16
37=Intrinsic_vshl|int16,uint16,int32,uint32
38=Intrinsic_vshr|int16,uint16,int32,uint32
39=Intrinsic_vnchwconv|float16,float32,int32,int16,uint32,uint16,int8,uint8
40=Intrinsic_vreducev2|int16,uint16,float16,float32,int32,uint32
41=Intrinsic_data_move_out2l1|hif8,f8,u8,s8,f16,bf16,u16,s16,f32,s32,u32
42=Intrinsic_data_move_out2l0a|u8,s8,f16,u16,s16,f32,s32,u32
43=Intrinsic_data_move_out2l0b|u8,s8,f16,u16,s16,f32,s32,u32
44=Intrinsic_data_move_l12l0a|hif8,f8,u8,s8,f16,bf16,u16,s16,f32,s32,u32
45=Intrinsic_data_move_l12l0b|hif8,f8,u8,s8,f16,bf16,u16,s16,f32,s32,u32
46=Intrinsic_data_move_l12out|u8,s8,f16,u16,s16,f32,s32,u32
47=Intrinsic_data_move_l12ub|u8,s8,f16,u16,s16,f32,s32,u32
48=Intrinsic_data_move_ub2l1|u8,s8,f16,bf16,u16,s16,f32,s32,u32
49=Intrinsic_data_move_ub2out|u8,s8,f16,u16,s16,f32,s32,u32
50=Intrinsic_data_move_l12bt|u8,s8,f16,bf16,u16,s16,f32,s32,u32
51=Intrinsic_data_move_out2ub|u8,s8,f16,u16,s16,f32,s32,u32
52=Intrinsic_data_move_ub2ub|u8,s8,f16,u16,s16,f32,s32,u32
53=Intrinsic_data_move_transpose_l12l0a|u8,s8,f16,u16,s16,f32,s32,u32
54=Intrinsic_data_move_transpose_l12l0b|u8,s8,f16,u16,s16,f32,s32,u32
55=Intrinsic_set_l1|f16,u16,s16,f32,s32,u32
56=Intrinsic_set_l0a|f16,u16,s16,f32,s32,u32
57=Intrinsic_set_l0b|f16,u16,s16,f32,s32,u32
58=Intrinsic_fix_pipe_l0c2l1|f32,s32,f16
59=Intrinsic_fix_pipe_l0c2out|f32,s32,f16
60=Intrinsic_fix_pipe_l0c2ub|f32,s32,f16
61=Intrinsic_fix_pipe_l12fb|u8,s8,f16,u16,s16,f32,s32,u32
62=Intrinsic_fix_pipe_unit_list|pre_conv,pre_act,post_eltwise,post_act,post_quant,post_transform
63=Intrinsic_fix_pipe_pre_conv_func_list|requant,dequant,quant,cast
64=Intrinsic_fix_pipe_pre_conv_requant|s322s8,s322u8,s322s4
65=Intrinsic_fix_pipe_pre_conv_dequant|s322f16
66=Intrinsic_fix_pipe_pre_conv_quant|f322s8,f322u8,f322s4
67=Intrinsic_fix_pipe_pre_conv_cast|f322f16,f322bf16
68=Intrinsic_fix_pipe_pre_act_func_list|normal_relu,scalar_relu,vector_relu,clip_relu
69=Intrinsic_fix_pipe_pre_act_normal_relu|f32,f16,bf16,s8,u8,s4,s32
70=Intrinsic_fix_pipe_pre_act_scalar_relu|f16,s8,u8,s4
71=Intrinsic_fix_pipe_pre_act_vector_relu|f16,s8,u8,s4
72=Intrinsic_fix_pipe_pre_act_clip_relu|f16,s8,u8,s4
73=Intrinsic_mmad|s8,f162f32,s4,bf162f32,f322f32,h322f32,s32s8s8,f16f16f16,f32f16f16,s32u8u2,f16f16u2,s32s4s4,e4m3e4m3,e4m3e5m2,e5m2e4m3,e5m2e5m2
74=Intrinsic_mmad_sp|s8
75=Intrinsic_vcgadd|float16,float32
76=Intrinsic_vconcat|float16,float32
77=Intrinsic_sqrt|int64,float32
78=Intrinsic_abs|int64,float32
79=Intrinsic_bcnt0|uint64
80=Intrinsic_bcnt1|uint64
81=Intrinsic_max|int64,float32
82=Intrinsic_min|int64,float32
83=Intrinsic_vcopy|int16,uint16,float16,float32,int32,uint32
84=Intrinsic_vbitsort32|float16,float32
85=Intrinsic_vmrgsort4v2|float16,float32
86=Intrinsic_vgatherb|u16u16,s16s16,f16f16,u32u32,s32s32,f32f32
87=Intrinsic_vmadd|float16,float32
88=Intrinsic_vcmpvs_ne|float16,float32
89=Intrinsic_vcmpvs_eq|int32,float16,float32
90=Intrinsic_vcmpvs_gt|float16,float32
91=Intrinsic_vcmpvs_ge|float16,float32
92=Intrinsic_vcmpvs_lt|float16,float32
93=Intrinsic_vcmpvs_le|float16,float32
94=Intrinsic_vcbd|u82u16,u82u32,s82s16,s82s32,u162u8,u162u32,s162u8,s162u32,s162s32,u322u8,u322u16,u322s16,s322u8,s322u16,s322s16
95=Intrinsic_vcgmax|float16,float32
96=Intrinsic_clz|uint64
97=Intrinsic_vmla|f16f16f16,f32f32f32,float16,float32
98=Intrinsic_vcgmin|float16,float32
99=Intrinsic_vcpadd|float16,float32
100=Intrinsic_vtranspose|uint16,int16,float16
101=Intrinsic_vscatter|int16,uint16,float16,float32,int32,uint32
102=Intrinsic_vgather|int16,uint16,float16,float32,int32,uint32
103=Intrinsic_vbrcb|u16u16,s16s16,f16f16,u32u32,s32s32,f32f32
104=Intrinsic_data_move_pad|uint8,int8,uint16,int16,float16,int32,float32,uint32,bfloat16
105=Intrinsic_data_move_out2l1_nd2nz|hif8,f8,u8,s8,f16,bf16,u16,s16,f32,s32,u32
106=Intrinsic_data_move_out2l1_dn2nz|hif8,f8,u8,s8,f16,bf16,u16,s16,f32,s32,u32
107=Intrinsic_fix_pipe_post_transform_nz2nd|f32,f16,bf16,s8,u8,s4,s32
108=Intrinsic_fix_pipe_support_multiple_output|true
[VectorCoreSpec]
vec_freq=1650
vec_calc_size=128
smask_buffer=0
ub_size=253952
ubblock_size=32
ubbank_size=4096
ubbank_num=16
ubsubbank_num=64
ubburst_in_one_block=32
ubbank_group_num=8
vector_reg_width=256
predicate_reg_width=32
wide_reg_width=768
[VectorCoreMemoryRates]
ddr_rate=14
ddr_read_rate=14
ddr_write_rate=14
l2_rate=100
l2_read_rate=100
l2_write_rate=100
ub_to_l2_rate=42
ub_to_ddr_rate=40
[VectorCoreintrinsicDtypeMap]
0=Intrinsic_vrelu|float16,float32,int32
1=Intrinsic_vexp|float16,float32
2=Intrinsic_vln|float16,float32
3=Intrinsic_vabs|float16,float32
4=Intrinsic_vrec|float16,float32
5=Intrinsic_vrsqrt|float16,float32
6=Intrinsic_vnot|int16,uint16
7=Intrinsic_vadd|float16,float32,int32,int16
8=Intrinsic_vsub|float16,float32,int32,int16
9=Intrinsic_vmul|float16,float32,int32,int16
10=Intrinsic_vmax|float16,float32,int32,int16
11=Intrinsic_vmin|float16,float32,int32,int16
12=Intrinsic_vand|int16,uint16
13=Intrinsic_vor|int16,uint16
14=Intrinsic_vadds|float16,float32,int32,int16
15=Intrinsic_vmuls|float16,float32,int32,int16
16=Intrinsic_vaxpy|f16f16f16,f32f32f32,float16,float32
17=Intrinsic_vcmpv_ne|float16,float32
18=Intrinsic_vcmpv_eq|float16,float32,int32
19=Intrinsic_vcmpv_gt|float16,float32
20=Intrinsic_vcmpv_ge|float16,float32
21=Intrinsic_vcmpv_lt|float16,float32
22=Intrinsic_vcmpv_le|float16,float32
23=Intrinsic_vconv|f322f32r,f322f32f,f322f32c,deq,f322f32a,f322f32z,f322f16,f322f16r,f322f16f,f322f16c,f322f16a,f322f16z,f322f16o,f322s64r,f322s64f,f322s64c,f322s64a,f322s64z,f322s32r,f322s32f,f322s32c,f322s32a,f322s32z,f322s16r,f322s16f,f322s16c,f322s16a,f322s16z,f162f32,f162s32r,f162s32f,f162s32c,f162s32a,f162s32z,f162s16r,f162s16f,f162s16c,f162s16a,f162s16z,f162s8,f162s8r,f162s8f,f162s8c,f162s8a,f162s8z,f162u8,f162u8r,f162u8f,f162u8c,f162u8a,f162u8z,u82f16,s82f16,s162f16,s162f16r,s162f16f,s162f16c,s162f16a,s162f16z,s162f32,s162s8,s162u8,s322f32,s322f32r,s322f32f,s322f32c,s322f32a,s322f32z,s322f16,s322s64,s322s16,s642s32,s642f32r,s642f32f,s642f32c,s642f32a,s642f32z,f322bf16r,f322bf16f,f322bf16c,f322bf16a,f322bf16z,bf162f32,bf162s32r,bf162s32f,bf162s32c,bf162s32a,bf162s32z
24=Intrinsic_vcadd|float16,float32
25=Intrinsic_vcmax|float16,float32
26=Intrinsic_vcmin|float16,float32
27=Intrinsic_vbi|float16,float32
28=Intrinsic_vlrelu|float16,float32
29=Intrinsic_scatter_vconv|f322f16,f162f32,f162s8,f162u8,s322f16,s322f32,u82f16,s82f16,f162s32r,f162s8f,f162u8f,f162s32f,f322s32f,f162s8c,f162u8c,f162s32c,f322s32c,f162s8a,f162u8a,f162s32a,f322s32a,f162s8z,f162u8z,f162s32z,f322s32z,f322f16o
30=Intrinsic_vreduce|int16,uint16,float16,float32,int32,uint32
31=Intrinsic_vmaxs|float16,float32,int32,int16
32=Intrinsic_vcmp|float16,float32
33=Intrinsic_vsel|float16,float32
34=Intrinsic_vsqrt|float16,float32
35=Intrinsic_vdiv|float16,float32
36=Intrinsic_vmins|float16,float32,int32,int16
37=Intrinsic_vshl|int16,uint16,int32,uint32
38=Intrinsic_vshr|int16,uint16,int32,uint32
39=Intrinsic_vnchwconv|float16,float32,int32,int16,uint32,uint16,int8,uint8
40=Intrinsic_vreducev2|int16,uint16,float16,float32,int32,uint32
41=Intrinsic_vcpadd|float16,float32
42=Intrinsic_vgatherb|u16u16,s16s16,f16f16,u32u32,s32s32,f32f32
43=Intrinsic_vbrcb|u16u16,s16s16,f16f16,u32u32,s32s32,f32f32
44=Intrinsic_sqrt|int64,float32
45=Intrinsic_abs|int64,float32
46=Intrinsic_bcnt0|uint64
47=Intrinsic_bcnt1|uint64
48=Intrinsic_clz|uint64
49=Intrinsic_max|int64,float32
50=Intrinsic_min|int64,float32
51=Intrinsic_vcgadd|float16,float32
52=Intrinsic_vectorized_psts|uint8,int8,uint16,int16,uint32,int32,float16,float32
53=Intrinsic_vgather|int16,uint16,float16,float32,int32,uint32
98=Intrinsic_vectorized_vadds|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
99=Intrinsic_vectorized_vmuls|u8u8,s8s8,u16u16,s16s16,s32s32,f16f16,f32f32,u8s24,s8s24,u16s48,s16s48,u16s64,s16s64
100=Intrinsic_vectorized_vmul|u8u8,s8s8,u16u16,s16s16,s32s32,f16f16,f32f32,u8s24,s8s24,u16s48,s16s48,u16s64,s16s64
101=Intrinsic_vectorized_sub|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32,s8s24,u8s24,s16s48,u16s48
102=Intrinsic_vectorized_vexp|f16f16,f32f32
103=Intrinsic_vectorized_vadd|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32,s8s24,u8s24,s16s48,u16s48
104=Intrinsic_vectorized_vdiv|u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
105=Intrinsic_vectorized_vmaxs|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
106=Intrinsic_vectorized_vmins|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
107=Intrinsic_vectorized_vcadd|s8s16,s16s32,s32s32,u8u16,u16u32,u32u32,f16f16,f32f32
108=Intrinsic_vectorized_vcmax|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
109=Intrinsic_vectorized_vcmin|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
110=Intrinsic_vectorized_vxor|bb,u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
111=Intrinsic_vectorized_vand|bb,u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
112=Intrinsic_vectorized_vor|bb,u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
113=Intrinsic_vectorized_vsel|bb,u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
114=Intrinsic_vectorized_vdup|uint8,int8,uint16,int16,uint32,int32,float16,float32
115=Intrinsic_vectorized_vdupm|uint8,int8,uint16,int16,uint32,int32,float16,float32
116=Intrinsic_vectorized_vdups|uint8,int8,uint16,int16,uint32,int32,float16,float32
117=Intrinsic_vectorized_vdupi|uint8,int8,uint16,int16,float16
118=Intrinsic_vectorized_vcmp_eq|uint8,int8,uint16,int16,uint32,int32,float16,float32
119=Intrinsic_vectorized_vcmp_ne|uint8,int8,uint16,int16,uint32,int32,float16,float32
120=Intrinsic_vectorized_vcmp_gt|uint8,int8,uint16,int16,uint32,int32,float16,float32
121=Intrinsic_vectorized_vcmp_ge|uint8,int8,uint16,int16,uint32,int32,float16,float32
122=Intrinsic_vectorized_vcmp_lt|uint8,int8,uint16,int16,uint32,int32,float16,float32
123=Intrinsic_vectorized_vcmp_le|uint8,int8,uint16,int16,uint32,int32,float16,float32
124=Intrinsic_vectorized_vcmps_eq|uint8,int8,uint16,int16,uint32,int32,float16,float32
125=Intrinsic_vectorized_vcmps_ne|uint8,int8,uint16,int16,uint32,int32,float16,float32
126=Intrinsic_vectorized_vcmps_gt|uint8,int8,uint16,int16,uint32,int32,float16,float32
127=Intrinsic_vectorized_vcmps_ge|uint8,int8,uint16,int16,uint32,int32,float16,float32
128=Intrinsic_vectorized_vcmps_lt|uint8,int8,uint16,int16,uint32,int32,float16,float32
129=Intrinsic_vectorized_vcmps_le|uint8,int8,uint16,int16,uint32,int32,float16,float32
130=Intrinsic_vectorized_vscatter|s8u16,u8u16,s16u16,u16u16,s32u32,u32u32,f16u16,f32u32
131=Intrinsic_vectorized_vmov|bb,u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
132=Intrinsic_vectorized_vsqrt|f16f16,f32f32
133=Intrinsic_vectorized_vshls|u8u8,u16u16,s16s16,u32u32,s32s32
134=Intrinsic_vectorized_vrnds|s16s16,s32s32
135=Intrinsic_vectorized_vlrelu|f16f16,f32f32
136=Intrinsic_vectorized_vsadds|s16s16
137=Intrinsic_vectorized_vcbmax|uint8,int8,uint16,int16,uint32,int32,float16,float32
138=Intrinsic_vectorized_vcbmin|uint8,int8,uint16,int16,uint32,int32,float16,float32
139=Intrinsic_vectorized_vmula|u8u8,s8s8,u16u16,s16s16,s32s32,f16f16,f32f32
140=Intrinsic_vectorized_vmadd|f16f16,f32f32
142=Intrinsic_vectorized_vrnd|s16s16,s32s32
143=Intrinsic_vectorized_vavg|u8u8,s8s8,u16u16,s16s16
144=Intrinsic_vectorized_vadd3|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32
146=Intrinsic_vectorized_vadif|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32
147=Intrinsic_vectorized_vsad|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32
149=Intrinsic_vectorized_vsadd|s16s16
150=Intrinsic_vectorized_vssub|s16s16
154=Intrinsic_vectorized_vrsqrt|f16f16,f32f32
155=Intrinsic_vectorized_vnot|bb,u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
156=Intrinsic_vectorized_vneg|s8s8,s16s16,s32s32,f16f16,f32f32
157=Intrinsic_vectorized_vln|f16f16,f32f32
158=Intrinsic_vectorized_vcls|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32
160=Intrinsic_vectorized_vintlv|uint8,int8,uint16,int16,uint32,int32,float16,float32
161=Intrinsic_vectorized_vdintlv|uint8,int8,uint16,int16,uint32,int32,float16,float32
162=Intrinsic_vectorized_vselr|u8u8u8,s8s8u8,u16u16u16,s16s16u16,f16f16u16
163=Intrinsic_vectorized_vpack|bb,u16u8,s16u8,u32u16,s32u16
164=Intrinsic_vectorized_vunpack|bb,u8u16,s8s16,u16u32,s16s32
165=Intrinsic_vectorized_vsqz|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
166=Intrinsic_vectorized_vbr|uint8,int8,uint16,int16,uint32,int32,float16,float32
167=Intrinsic_vectorized_vci|int8,int16,int32,float16,float32
168=Intrinsic_vectorized_vgatherb|int8,uint8,int16,uint16,int32,uint32,float16,float32
169=Intrinsic_vectorized_vgather2|s16u16s8,u16u16u8,s16u16s16,u16u16u16,s32u32s32,u32u32u32,f16u16f16,f32u32f32
172=Intrinsic_vectorized_vmax|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
173=Intrinsic_vectorized_vsub|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32,s8s24,u8s24,s16s48,u16s48
174=Intrinsic_vectorized_vbcnt|u8s8,s8s8,u16s16,s16s16,u32s32,s32s32
175=Intrinsic_vectorized_pld|uint8,int8,uint16,int16,uint32,int32,float16,float32
176=Intrinsic_vectorized_vaxpy|f16f16,f32f32
177=Intrinsic_vectorized_vmin|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
178=Intrinsic_vectorized_vrec|f16f16,f32f32
179=Intrinsic_vectorized_vlog|float16,float32
180=Intrinsic_vectorized_vabs|s8s8,s16s16,s32s32,f16f16,f32f32
181=Intrinsic_vectorized_plt|uint8,int8,uint16,int16,float16,int32,float32,uint32
182=Intrinsic_data_move_gather|uint8,int8,uint16,int16,float16,int32,float32,uint32,uint64,int64
183=Intrinsic_vector_filter1d|u8s16,s8s16,u16s16,s16s16
184=Intrinsic_vectorized_vabsdif|u8u8,s8s8,u16u16,s16s16,s32s32,f16f16,f32f32
185=Intrinsic_vectorized_vshl|u8u8,u16u16,s16s16,u32u32,s32s32
186=Intrinsic_vectorized_vshr|u8u8,u16u16,s16s16,u32u32,s32s32,s24s8,s24u8,s48s16,s48u16,s64s32,s24s16,s24u16,s48s32
188=Intrinsic_vectorized_vrelu|s32s32,f16f16,f32f32
189=Intrinsic_vectorized_pge|uint8,int8,uint16,int16,float16,int32,float32,uint32
190=Intrinsic_vectorized_pset|uint8,int8,uint16,int16,float16,int32,float32,uint32
191=Intrinsic_vectorized_pslide|uint8,int8,uint16,int16,float16,int32,float32,uint32
192=Intrinsic_data_move_pad|uint8,int8,uint16,int16,float16,int32,float32,uint32,bfloat16
193=Intrinsic_vectorized_vsld|uint8,int8,uint16,int16,float16,int32,float32,uint32
194=Intrinsic_vectorized_cast|f16f32,f32f16,u8f16,s8f16,f16u8,f16s8,f16s16,f16s32,s16f16,s16f32,f32s16,f32s32,s32f32,u8u16,u8u32,s8s16,s8s32,u16u8,u16u32,s16u8,s16u32,s16s32,u32u8,u32u16,u32s16,s32u8,s32u16,s32s16
196=Intrinsic_vectorized_vshrs|u8u8,u16u16,s16s16,u32u32,s32s32,s24s8,s24u8,s48s16,s48u16,s64s32,s24s16,s24u16,s48s32
197=Intrinsic_vectorized_pst|uint8,int8,uint16,int16,uint32,int32,float16,float32
198=Intrinsic_vectorized_dhist|u8u32
199=Intrinsic_vectorized_dhistv2|u8u16
200=Intrinsic_vectorized_chist|u8u32
201=Intrinsic_vectorized_chistv2|u8u16
202=Intrinsic_vectorized_vfilter|s8s48,u8s48,s16s48,u16s48
203=Intrinsic_vectorized_vslide|u8u8,s8s8,u16u16,s16s16,u32u32,s32s32,f16f16,f32f32
204=Intrinsic_vectorized_fmax|u8u8
205=Intrinsic_vectorized_fmin|u8u8
206=Intrinsic_vectorized_vusqz|bs8,bs16,bs32,bu8,bu16,bu32
207=Intrinsic_vectorized_vtrc|f16f16,f32f32
208=Intrinsic_fix_pipe_l0c2out|f32,s32,f16
209=Intrinsic_vcopy|int16,uint16,float16,float32,int32,uint32
210=Intrinsic_vbitsort32|float16,float32
211=Intrinsic_vmrgsort4v2|float16,float32
212=Intrinsic_vmadd|float16,float32
213=Intrinsic_vcmpvs_ne|float16,float32
214=Intrinsic_vcmpvs_eq|int32,float16,float32
215=Intrinsic_vcmpvs_gt|float16,float32
216=Intrinsic_vcmpvs_ge|float16,float32
217=Intrinsic_vcmpvs_lt|float16,float32
218=Intrinsic_vcmpvs_le|float16,float32
219=Intrinsic_vcbd|u82u16,u82u32,s82s16,s82s32,u162u8,u162u32,s162u8,s162u32,s162s32,u322u8,u322u16,u322s16,s322u8,s322u16,s322s16
220=Intrinsic_vcgmax|float16,float32
221=Intrinsic_vmla|f16f16f16,f32f32f32,float16,float32
222=Intrinsic_vcgmin|float16,float32
223=Intrinsic_vtranspose|uint16,int16,float16
224=Intrinsic_vscatter|int16,uint16,float16,float32,int32,uint32
225=Intrinsic_vci|float16,float32,int16,int32
226=Intrinsic_vbrcb|u16u16,s16s16,f16f16,u32u32,s32s32,f32f32
227=Intrinsic_vectorized_plds|uint8,int8,uint16,int16,uint32,int32,float16,float32
228=Intrinsic_vbitsort|float16