Home
last modified time | relevance | path

Searched refs:vsumOPQRSTUV (Results 1 – 14 of 14) sorted by relevance

/external/XNNPACK/src/qs8-gavgpool/gen/
D7p7x-minmax-rndnu-neon-c32.c60 int16x8_t vsumOPQRSTUV = vaddl_s8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() local
69 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
77 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi3xOPQRSTUV); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
85 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi4xOPQRSTUV); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
93 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi5xOPQRSTUV); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
97 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi6xOPQRSTUV); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
105 const int32x4_t vaccOPQR = vaddw_s16(vinit_bias, vget_low_s16(vsumOPQRSTUV)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
106 const int32x4_t vaccSTUV = vaddw_s16(vinit_bias, vget_high_s16(vsumOPQRSTUV)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
172 int16x8_t vsumOPQRSTUV = vaddl_s8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() local
181 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
[all …]
D7p7x-minmax-fp32-neonv8-c32.c61 int16x8_t vsumOPQRSTUV = vaddl_s8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() local
70 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
78 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi3xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
86 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi4xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
94 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi5xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
98 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi6xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
106 const int32x4_t vaccOPQR = vaddw_s16(vinit_bias, vget_low_s16(vsumOPQRSTUV)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
107 const int32x4_t vaccSTUV = vaddw_s16(vinit_bias, vget_high_s16(vsumOPQRSTUV)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
173 int16x8_t vsumOPQRSTUV = vaddl_s8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() local
182 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
[all …]
D7p7x-minmax-fp32-neon-c32.c60 int16x8_t vsumOPQRSTUV = vaddl_s8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() local
69 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
77 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi3xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
85 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi4xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
93 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi5xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
97 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi6xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
105 const int32x4_t vaccOPQR = vaddw_s16(vinit_bias, vget_low_s16(vsumOPQRSTUV)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
106 const int32x4_t vaccSTUV = vaddw_s16(vinit_bias, vget_high_s16(vsumOPQRSTUV)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
172 int16x8_t vsumOPQRSTUV = vaddl_s8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() local
181 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
[all …]
D7x-minmax-fp32-neonv8-c32.c79 int16x8_t vsumOPQRSTUV = vaddl_s8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32() local
88 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
96 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi3xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
104 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi4xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
112 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi5xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
116 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi6xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
124 int32x4_t vaccOPQR = vaddw_s16(vinit_bias, vget_low_s16(vsumOPQRSTUV)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
125 int32x4_t vaccSTUV = vaddw_s16(vinit_bias, vget_high_s16(vsumOPQRSTUV)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
D7x-minmax-rndnu-neon-c32.c80 int16x8_t vsumOPQRSTUV = vaddl_s8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32() local
89 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32()
97 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi3xOPQRSTUV); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32()
105 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi4xOPQRSTUV); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32()
113 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi5xOPQRSTUV); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32()
117 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi6xOPQRSTUV); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32()
125 int32x4_t vaccOPQR = vaddw_s16(vinit_bias, vget_low_s16(vsumOPQRSTUV)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32()
126 int32x4_t vaccSTUV = vaddw_s16(vinit_bias, vget_high_s16(vsumOPQRSTUV)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32()
D7x-minmax-fp32-neon-c32.c79 int16x8_t vsumOPQRSTUV = vaddl_s8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c32() local
88 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
96 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi3xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
104 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi4xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
112 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi5xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
116 vsumOPQRSTUV = vaddw_s8(vsumOPQRSTUV, vi6xOPQRSTUV); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
124 int32x4_t vaccOPQR = vaddw_s16(vinit_bias, vget_low_s16(vsumOPQRSTUV)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
125 int32x4_t vaccSTUV = vaddw_s16(vinit_bias, vget_high_s16(vsumOPQRSTUV)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
/external/XNNPACK/src/qu8-gavgpool/gen/
D7p7x-minmax-fp32-neonv8-c32.c61 uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() local
70 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
78 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi3xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
86 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi4xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
94 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi5xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
98 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi6xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
106 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumOPQRSTUV))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
107 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumOPQRSTUV))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
173 uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() local
182 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
[all …]
D7p7x-minmax-fp32-neon-c32.c60 uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() local
69 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
77 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi3xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
85 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi4xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
93 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi5xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
97 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi6xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
105 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumOPQRSTUV))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
106 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumOPQRSTUV))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
172 uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() local
181 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
[all …]
D7p7x-minmax-rndnu-neon-c32.c60 uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() local
69 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
77 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi3xOPQRSTUV); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
85 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi4xOPQRSTUV); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
93 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi5xOPQRSTUV); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
97 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi6xOPQRSTUV); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
105 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumOPQRSTUV))); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
106 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumOPQRSTUV))); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
172 uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() local
181 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
[all …]
D7x-minmax-fp32-neon-c32.c79 uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c32() local
88 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
96 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi3xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
104 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi4xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
112 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi5xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
116 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi6xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
124 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumOPQRSTUV))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
125 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumOPQRSTUV))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
D7x-minmax-fp32-neonv8-c32.c79 uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32() local
88 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
96 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi3xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
104 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi4xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
112 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi5xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
116 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi6xOPQRSTUV); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
124 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumOPQRSTUV))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
125 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumOPQRSTUV))); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
D7x-minmax-rndnu-neon-c32.c80 uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32() local
89 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32()
97 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi3xOPQRSTUV); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32()
105 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi4xOPQRSTUV); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32()
113 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi5xOPQRSTUV); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32()
117 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi6xOPQRSTUV); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32()
125 … = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsumOPQRSTUV))); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32()
126 …= vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsumOPQRSTUV))); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32()
/external/XNNPACK/src/qu8-dwconv/gen/
Dup32x25-minmax-rndnu-neon-mul8.c220 uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8() local
246 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()
272 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi3xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()
298 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi4xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()
324 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi5xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()
350 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi6xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()
376 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi7xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()
402 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi8xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()
428 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi9xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()
454 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi10xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()
[all …]
Dup32x9-minmax-rndnu-neon-mul8.c140 uint16x8_t vsumOPQRSTUV = vaddl_u8(vi0xOPQRSTUV, vi1xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8() local
166 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi2xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8()
192 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi3xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8()
218 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi4xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8()
244 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi5xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8()
270 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi6xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8()
296 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi7xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8()
322 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi8xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8()
339 …pretq_s32_u32(vmlsl_u16(vreinterpretq_u32_s32(vaccOPQR), vget_low_u16(vsumOPQRSTUV), vget_low_u16(… in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8()
340 …retq_s32_u32(vmlsl_u16(vreinterpretq_u32_s32(vaccSTUV), vget_high_u16(vsumOPQRSTUV), vget_high_u16… in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8()