| /external/XNNPACK/src/qs8-vaddc/gen/ |
| D | minmax-neon-ld128-x16.c | 43 …const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), vget_low_s8(va_zero_point)… in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x16() local
46 const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), va_zero_point); in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x16() local
50 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x16()
51 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x16()
76 const int16x8_t vxa01234567 = vsubl_s8(va01234567, vget_low_s8(va_zero_point)); in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x16() local
78 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x16() local
81 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x16()
82 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x16()
|
| D | minmax-neon-ld64-x8.c | 38 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8() local
40 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8()
41 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8()
60 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8() local
62 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8()
63 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8()
|
| D | minmax-neon-ld128-x32.c | 44 …const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), vget_low_s8(va_zero_point)… in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x32() local
49 const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), va_zero_point); in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x32() local
55 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x32()
56 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x32()
95 const int16x8_t vxa01234567 = vsubl_s8(va01234567, vget_low_s8(va_zero_point)); in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x32() local
97 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x32() local
100 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x32()
101 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qs8_vaddc_minmax_ukernel__neon_ld128_x32()
|
| /external/XNNPACK/src/qu8-vaddc/gen/ |
| D | minmax-neon-ld128-x16.c | 43 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), vget… in xnn_qu8_vaddc_minmax_ukernel__neon_ld128_x16() local
46 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), va_z… in xnn_qu8_vaddc_minmax_ukernel__neon_ld128_x16() local
50 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qu8_vaddc_minmax_ukernel__neon_ld128_x16()
51 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qu8_vaddc_minmax_ukernel__neon_ld128_x16()
76 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, vget_low_u8(va_zero_point… in xnn_qu8_vaddc_minmax_ukernel__neon_ld128_x16() local
78 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); in xnn_qu8_vaddc_minmax_ukernel__neon_ld128_x16() local
81 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qu8_vaddc_minmax_ukernel__neon_ld128_x16()
82 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qu8_vaddc_minmax_ukernel__neon_ld128_x16()
|
| D | minmax-neon-ld64-x8.c | 38 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8() local
40 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8()
41 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8()
60 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8() local
62 int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8()
63 int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8()
|
| /external/XNNPACK/src/qu8-vmulc/gen/ |
| D | minmax-fp32-neonv8-ld128-x16.c | 42 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), vget… in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local
45 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), va_z… in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local
49 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16()
50 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16()
97 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, vget_low_u8(va_zero_point… in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local
99 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local
102 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16()
103 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16()
|
| D | minmax-fp32-neon-ld128-x16.c | 42 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), vget… in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld128_x16() local
45 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), va_z… in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld128_x16() local
49 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld128_x16()
50 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld128_x16()
100 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, vget_low_u8(va_zero_point… in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld128_x16() local
102 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld128_x16() local
105 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld128_x16()
106 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld128_x16()
|
| D | minmax-rndnu-neon-ld128-x16.c | 43 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), vget… in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local
46 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), va_z… in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local
50 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16()
51 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16()
98 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, vget_low_u8(va_zero_point… in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local
100 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local
103 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16()
104 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16()
|
| D | minmax-fp32-sse41-mul16-ld64-x8.c | 40 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
42 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
43 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
74 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
76 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
77 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
|
| D | minmax-fp32-neonv8-ld64-x8.c | 37 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
39 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
40 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
75 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
77 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
78 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
|
| /external/XNNPACK/src/qs8-vmulc/gen/ |
| D | minmax-fp32-neonv8-ld128-x16.c | 42 …const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), vget_low_s8(va_zero_point)… in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local
45 const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), va_zero_point); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local
49 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16()
50 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16()
97 const int16x8_t vxa01234567 = vsubl_s8(va01234567, vget_low_s8(va_zero_point)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local
99 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local
102 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16()
103 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16()
|
| D | minmax-rndnu-neon-ld128-x16.c | 43 …const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), vget_low_s8(va_zero_point)… in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local
46 const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), va_zero_point); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local
50 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16()
51 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16()
98 const int16x8_t vxa01234567 = vsubl_s8(va01234567, vget_low_s8(va_zero_point)); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local
100 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local
103 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16()
104 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16()
|
| D | minmax-fp32-neon-ld128-x16.c | 42 …const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), vget_low_s8(va_zero_point)… in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld128_x16() local
45 const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), va_zero_point); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld128_x16() local
49 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld128_x16()
50 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld128_x16()
100 const int16x8_t vxa01234567 = vsubl_s8(va01234567, vget_low_s8(va_zero_point)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld128_x16() local
102 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld128_x16() local
105 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld128_x16()
106 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld128_x16()
|
| D | minmax-fp32-sse41-mul16-ld64-x8.c | 40 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
42 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
43 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
74 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
76 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
77 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
|
| D | minmax-fp32-neonv8-ld64-x8.c | 37 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
39 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
40 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
75 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
77 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
78 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
|
| /external/XNNPACK/src/qs8-vmul/gen/ |
| D | minmax-fp32-neonv8-ld128-x16.c | 42 …const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), vget_low_s8(va_zero_point)… in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16() local
47 const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), va_zero_point); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16() local
53 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16()
54 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16()
102 const int16x8_t vxa01234567 = vsubl_s8(va01234567, vget_low_s8(va_zero_point)); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16() local
105 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16() local
109 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16()
110 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16()
|
| D | minmax-rndnu-neon-ld128-x16.c | 43 …const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), vget_low_s8(va_zero_point)… in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld128_x16() local
48 const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), va_zero_point); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld128_x16() local
54 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567)); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld128_x16()
55 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567)); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld128_x16()
103 const int16x8_t vxa01234567 = vsubl_s8(va01234567, vget_low_s8(va_zero_point)); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld128_x16() local
106 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld128_x16() local
110 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567)); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld128_x16()
111 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567)); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld128_x16()
|
| D | minmax-fp32-neon-ld128-x16.c | 42 …const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), vget_low_s8(va_zero_point)… in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld128_x16() local
47 const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), va_zero_point); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld128_x16() local
53 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld128_x16()
54 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld128_x16()
105 const int16x8_t vxa01234567 = vsubl_s8(va01234567, vget_low_s8(va_zero_point)); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld128_x16() local
108 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld128_x16() local
112 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld128_x16()
113 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld128_x16()
|
| D | minmax-fp32-wasmsimd-mul32-ld64-x8.c | 39 const v128_t vxa01234567 = wasm_i16x8_sub(va01234567, va_zero_point); in xnn_qs8_vmul_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8() local
42 …v128_t vacc0123 = wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa01234567), wasm_i32x4_extend_low_i… in xnn_qs8_vmul_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8()
43 …v128_t vacc4567 = wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa01234567), wasm_i32x4_extend_high… in xnn_qs8_vmul_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8()
74 const v128_t vxa01234567 = wasm_i16x8_sub(va01234567, va_zero_point); in xnn_qs8_vmul_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8() local
77 …v128_t vacc0123 = wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa01234567), wasm_i32x4_extend_low_i… in xnn_qs8_vmul_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8()
78 …v128_t vacc4567 = wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa01234567), wasm_i32x4_extend_high… in xnn_qs8_vmul_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8()
|
| D | minmax-fp32-neonv8-ld64-x8.c | 36 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local
39 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
40 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
76 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local
79 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
80 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
|
| /external/XNNPACK/src/qu8-vadd/gen/ |
| D | minmax-neon-ld128-x16.c | 43 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), vget… in xnn_qu8_vadd_minmax_ukernel__neon_ld128_x16() local
48 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), va_z… in xnn_qu8_vadd_minmax_ukernel__neon_ld128_x16() local
54 int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qu8_vadd_minmax_ukernel__neon_ld128_x16()
55 int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qu8_vadd_minmax_ukernel__neon_ld128_x16()
86 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, vget_low_u8(va_zero_point… in xnn_qu8_vadd_minmax_ukernel__neon_ld128_x16() local
89 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); in xnn_qu8_vadd_minmax_ukernel__neon_ld128_x16() local
93 int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qu8_vadd_minmax_ukernel__neon_ld128_x16()
94 int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qu8_vadd_minmax_ukernel__neon_ld128_x16()
|
| D | minmax-neon-ld64-x8.c | 37 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8() local
40 int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8()
41 int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8()
64 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8() local
67 int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8()
68 int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qu8_vadd_minmax_ukernel__neon_ld64_x8()
|
| /external/XNNPACK/src/qs8-vadd/gen/ |
| D | minmax-neon-ld128-x16.c | 43 …const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), vget_low_s8(va_zero_point)… in xnn_qs8_vadd_minmax_ukernel__neon_ld128_x16() local
48 const int16x8_t vxa01234567 = vsubl_s8(vget_low_s8(va0123456789ABCDEF), va_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld128_x16() local
54 int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qs8_vadd_minmax_ukernel__neon_ld128_x16()
55 int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qs8_vadd_minmax_ukernel__neon_ld128_x16()
86 const int16x8_t vxa01234567 = vsubl_s8(va01234567, vget_low_s8(va_zero_point)); in xnn_qs8_vadd_minmax_ukernel__neon_ld128_x16() local
89 const int16x8_t vxa01234567 = vsubl_s8(va01234567, va_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld128_x16() local
93 int32x4_t vacc0123 = vmulq_s32(vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier); in xnn_qs8_vadd_minmax_ukernel__neon_ld128_x16()
94 int32x4_t vacc4567 = vmulq_s32(vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier); in xnn_qs8_vadd_minmax_ukernel__neon_ld128_x16()
|
| /external/XNNPACK/src/qu8-vmul/gen/ |
| D | minmax-fp32-neonv8-ld128-x16.c | 42 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), vget… in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16() local
47 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), va_z… in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16() local
53 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567)); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16()
54 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567)); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16()
102 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, vget_low_u8(va_zero_point… in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16() local
105 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16() local
109 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567)); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16()
110 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567)); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld128_x16()
|
| D | minmax-rndnu-neon-ld128-x16.c | 43 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), vget… in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld128_x16() local
48 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(va0123456789ABCDEF), va_z… in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld128_x16() local
54 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567)); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld128_x16()
55 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567)); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld128_x16()
103 …const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, vget_low_u8(va_zero_point… in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld128_x16() local
106 const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point)); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld128_x16() local
110 int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb01234567)); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld128_x16()
111 int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb01234567)); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld128_x16()
|