| /external/XNNPACK/src/f32-qs8-vcvt/gen/ |
| D | vcvt-avx512skx-x32.c | 36 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32() local
40 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32()
43 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32()
46 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32()
63 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32() local
64 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32()
65 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32()
68 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32()
87 __m512 vx0123 = _mm512_maskz_loadu_ps(vmask, x); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32() local
88 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32()
[all …]
|
| D | vcvt-avx512skx-x64.c | 36 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64() local
42 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64()
47 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64()
52 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64()
73 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64() local
74 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64()
75 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64()
78 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64()
97 __m512 vx0123 = _mm512_maskz_loadu_ps(vmask, x); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64() local
98 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64()
[all …]
|
| D | vcvt-avx512skx-x96.c | 37 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96() local
45 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96()
52 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96()
59 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96()
88 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96() local
89 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96()
90 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96()
93 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96()
112 __m512 vx0123 = _mm512_maskz_loadu_ps(vmask, x); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96() local
113 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96()
[all …]
|
| D | vcvt-avx512skx-x128.c | 36 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128() local
46 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128()
55 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128()
64 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128()
97 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128() local
98 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128()
99 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128()
102 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128()
121 __m512 vx0123 = _mm512_maskz_loadu_ps(vmask, x); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128() local
122 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128()
[all …]
|
| /external/XNNPACK/src/f32-qu8-vcvt/gen/ |
| D | vcvt-avx512skx-x32.c | 36 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32() local
40 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32()
43 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32()
46 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32()
63 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32() local
64 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32()
65 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32()
68 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32()
87 __m512 vx0123 = _mm512_maskz_loadu_ps(vmask, x); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32() local
88 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32()
[all …]
|
| D | vcvt-avx512skx-x64.c | 36 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64() local
42 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64()
47 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64()
52 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64()
73 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64() local
74 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64()
75 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64()
78 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64()
97 __m512 vx0123 = _mm512_maskz_loadu_ps(vmask, x); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64() local
98 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64()
[all …]
|
| D | vcvt-avx512skx-x96.c | 37 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96() local
45 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96()
52 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96()
59 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96()
88 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96() local
89 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96()
90 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96()
93 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96()
112 __m512 vx0123 = _mm512_maskz_loadu_ps(vmask, x); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96() local
113 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96()
[all …]
|
| D | vcvt-avx512skx-x128.c | 36 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128() local
46 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128()
55 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128()
64 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128()
97 __m512 vx0123 = _mm512_loadu_ps(x); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128() local
98 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128()
99 vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128()
102 const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128()
121 __m512 vx0123 = _mm512_maskz_loadu_ps(vmask, x); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128() local
122 vx0123 = _mm512_mul_ps(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128()
[all …]
|
| /external/XNNPACK/src/f32-vhswish/gen/ |
| D | vhswish-sse-x8.c | 33 const __m128 vx0123 = _mm_loadu_ps(x); in xnn_f32_vhswish_ukernel__sse_x8() local
37 __m128 vacc0123 = _mm_mul_ps(vx0123, vsixth); in xnn_f32_vhswish_ukernel__sse_x8()
49 vacc0123 = _mm_mul_ps(vacc0123, vx0123); in xnn_f32_vhswish_ukernel__sse_x8()
57 const __m128 vx0123 = _mm_loadu_ps(x); in xnn_f32_vhswish_ukernel__sse_x8() local
59 __m128 vacc0123 = _mm_mul_ps(vx0123, vsixth); in xnn_f32_vhswish_ukernel__sse_x8()
63 vacc0123 = _mm_mul_ps(vacc0123, vx0123); in xnn_f32_vhswish_ukernel__sse_x8()
68 const __m128 vx0123 = _mm_loadu_ps(x); in xnn_f32_vhswish_ukernel__sse_x8() local
69 __m128 vacc0123 = _mm_mul_ps(vx0123, vsixth); in xnn_f32_vhswish_ukernel__sse_x8()
73 vacc0123 = _mm_mul_ps(vacc0123, vx0123); in xnn_f32_vhswish_ukernel__sse_x8()
|
| D | vhswish-sse-x4.c | 33 const __m128 vx0123 = _mm_loadu_ps(x); in xnn_f32_vhswish_ukernel__sse_x4() local
36 __m128 vacc0123 = _mm_mul_ps(vx0123, vsixth); in xnn_f32_vhswish_ukernel__sse_x4()
44 vacc0123 = _mm_mul_ps(vacc0123, vx0123); in xnn_f32_vhswish_ukernel__sse_x4()
50 const __m128 vx0123 = _mm_loadu_ps(x); in xnn_f32_vhswish_ukernel__sse_x4() local
51 __m128 vacc0123 = _mm_mul_ps(vx0123, vsixth); in xnn_f32_vhswish_ukernel__sse_x4()
55 vacc0123 = _mm_mul_ps(vacc0123, vx0123); in xnn_f32_vhswish_ukernel__sse_x4()
|
| D | vhswish-wasmsimd-x8.c | 33 v128_t vx0123 = wasm_v128_load(x); in xnn_f32_vhswish_ukernel__wasmsimd_x8() local
37 v128_t vacc0123 = wasm_f32x4_add(vx0123, vthree); in xnn_f32_vhswish_ukernel__wasmsimd_x8()
38 vx0123 = wasm_f32x4_mul(vx0123, vsixth); in xnn_f32_vhswish_ukernel__wasmsimd_x8()
48 vacc0123 = wasm_f32x4_mul(vacc0123, vx0123); in xnn_f32_vhswish_ukernel__wasmsimd_x8()
|
| /external/XNNPACK/src/qs8-vcvt/gen/ |
| D | vcvt-armsimd32-x8.c | 30 const int8x4_t vx0123 = (int8x4_t) unaligned_indexed_load_u32(x, 0); in xnn_qs8_vcvt_ukernel__armsimd32_x8() local
34 const int16x2_t vx02 = __sxtab16(vminus_input_zero_point, vx0123); in xnn_qs8_vcvt_ukernel__armsimd32_x8()
35 const int16x2_t vx13 = __sxtab16(vminus_input_zero_point, __ror(vx0123, 8)); in xnn_qs8_vcvt_ukernel__armsimd32_x8()
68 const int8x4_t vx0123 = (int8x4_t) unaligned_load_u32(x); in xnn_qs8_vcvt_ukernel__armsimd32_x8() local
71 const int16x2_t vx02 = __sxtab16(vminus_input_zero_point, vx0123); in xnn_qs8_vcvt_ukernel__armsimd32_x8()
72 const int16x2_t vx13 = __sxtab16(vminus_input_zero_point, __ror(vx0123, 8)); in xnn_qs8_vcvt_ukernel__armsimd32_x8()
91 const int8x4_t vx0123 = (int8x4_t) unaligned_load_u32(x); in xnn_qs8_vcvt_ukernel__armsimd32_x8() local
93 const int16x2_t vx02 = __sxtab16(vminus_input_zero_point, vx0123); in xnn_qs8_vcvt_ukernel__armsimd32_x8()
94 const int16x2_t vx13 = __sxtab16(vminus_input_zero_point, __ror(vx0123, 8)); in xnn_qs8_vcvt_ukernel__armsimd32_x8()
|
| D | vcvt-armsimd32-x4.c | 30 const int8x4_t vx0123 = (int8x4_t) unaligned_load_u32(x); in xnn_qs8_vcvt_ukernel__armsimd32_x4() local
33 const int16x2_t vx02 = __sxtab16(vminus_input_zero_point, vx0123); in xnn_qs8_vcvt_ukernel__armsimd32_x4()
34 const int16x2_t vx13 = __sxtab16(vminus_input_zero_point, __ror(vx0123, 8)); in xnn_qs8_vcvt_ukernel__armsimd32_x4()
53 const int8x4_t vx0123 = (int8x4_t) unaligned_load_u32(x); in xnn_qs8_vcvt_ukernel__armsimd32_x4() local
55 const int16x2_t vx02 = __sxtab16(vminus_input_zero_point, vx0123); in xnn_qs8_vcvt_ukernel__armsimd32_x4()
56 const int16x2_t vx13 = __sxtab16(vminus_input_zero_point, __ror(vx0123, 8)); in xnn_qs8_vcvt_ukernel__armsimd32_x4()
|
| /external/XNNPACK/src/qu8-vcvt/gen/ |
| D | vcvt-armsimd32-x8.c | 30 const uint8x4_t vx0123 = (uint8x4_t) unaligned_indexed_load_u32(x, 0); in xnn_qu8_vcvt_ukernel__armsimd32_x8() local
34 const uint16x2_t vx02 = __uxtab16(vminus_input_zero_point, vx0123); in xnn_qu8_vcvt_ukernel__armsimd32_x8()
35 const uint16x2_t vx13 = __uxtab16(vminus_input_zero_point, __ror(vx0123, 8)); in xnn_qu8_vcvt_ukernel__armsimd32_x8()
68 const uint8x4_t vx0123 = (uint8x4_t) unaligned_load_u32(x); in xnn_qu8_vcvt_ukernel__armsimd32_x8() local
71 const uint16x2_t vx02 = __uxtab16(vminus_input_zero_point, vx0123); in xnn_qu8_vcvt_ukernel__armsimd32_x8()
72 const uint16x2_t vx13 = __uxtab16(vminus_input_zero_point, __ror(vx0123, 8)); in xnn_qu8_vcvt_ukernel__armsimd32_x8()
91 const uint8x4_t vx0123 = (uint8x4_t) unaligned_load_u32(x); in xnn_qu8_vcvt_ukernel__armsimd32_x8() local
93 const uint16x2_t vx02 = __uxtab16(vminus_input_zero_point, vx0123); in xnn_qu8_vcvt_ukernel__armsimd32_x8()
94 const uint16x2_t vx13 = __uxtab16(vminus_input_zero_point, __ror(vx0123, 8)); in xnn_qu8_vcvt_ukernel__armsimd32_x8()
|
| D | vcvt-armsimd32-x4.c | 30 const uint8x4_t vx0123 = (uint8x4_t) unaligned_load_u32(x); in xnn_qu8_vcvt_ukernel__armsimd32_x4() local
33 const uint16x2_t vx02 = __uxtab16(vminus_input_zero_point, vx0123); in xnn_qu8_vcvt_ukernel__armsimd32_x4()
34 const uint16x2_t vx13 = __uxtab16(vminus_input_zero_point, __ror(vx0123, 8)); in xnn_qu8_vcvt_ukernel__armsimd32_x4()
53 const uint8x4_t vx0123 = (uint8x4_t) unaligned_load_u32(x); in xnn_qu8_vcvt_ukernel__armsimd32_x4() local
55 const uint16x2_t vx02 = __uxtab16(vminus_input_zero_point, vx0123); in xnn_qu8_vcvt_ukernel__armsimd32_x4()
56 const uint16x2_t vx13 = __uxtab16(vminus_input_zero_point, __ror(vx0123, 8)); in xnn_qu8_vcvt_ukernel__armsimd32_x4()
|
| /external/XNNPACK/src/f32-vlrelu/gen/ |
| D | vlrelu-sse-x4.c | 30 __m128 vx0123 = _mm_loadu_ps(x); in xnn_f32_vlrelu_ukernel__sse_x4() local
33 __m128 vacc0123 = _mm_max_ps(_mm_setzero_ps(), vx0123); in xnn_f32_vlrelu_ukernel__sse_x4()
34 vx0123 = _mm_min_ps(vx0123, vzero); in xnn_f32_vlrelu_ukernel__sse_x4()
36 vacc0123 = _mm_add_ps(vacc0123, _mm_mul_ps(vx0123, vslope)); in xnn_f32_vlrelu_ukernel__sse_x4()
|
| D | vlrelu-wasmsimd-minmax-x8.c | 30 v128_t vx0123 = wasm_v128_load(x); in xnn_f32_vlrelu_ukernel__wasmsimd_minmax_x8() local
34 v128_t vacc0123 = wasm_i32x4_max(vx0123, vzero); in xnn_f32_vlrelu_ukernel__wasmsimd_minmax_x8()
35 vx0123 = wasm_i32x4_min(vx0123, vzero); in xnn_f32_vlrelu_ukernel__wasmsimd_minmax_x8()
39 vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vx0123, vslope)); in xnn_f32_vlrelu_ukernel__wasmsimd_minmax_x8()
|
| D | vlrelu-sse-x8.c | 30 __m128 vx0123 = _mm_loadu_ps(x); in xnn_f32_vlrelu_ukernel__sse_x8() local
34 __m128 vacc0123 = _mm_max_ps(_mm_setzero_ps(), vx0123); in xnn_f32_vlrelu_ukernel__sse_x8()
35 vx0123 = _mm_min_ps(vx0123, vzero); in xnn_f32_vlrelu_ukernel__sse_x8()
39 vacc0123 = _mm_add_ps(vacc0123, _mm_mul_ps(vx0123, vslope)); in xnn_f32_vlrelu_ukernel__sse_x8()
|
| D | vlrelu-sse41-x4.c | 29 const __m128 vx0123 = _mm_loadu_ps(x); in xnn_f32_vlrelu_ukernel__sse41_x4() local
32 __m128 vacc0123 = _mm_mul_ps(vx0123, vslope); in xnn_f32_vlrelu_ukernel__sse41_x4()
34 vacc0123 = _mm_blendv_ps(vx0123, vacc0123, vx0123); in xnn_f32_vlrelu_ukernel__sse41_x4()
|
| /external/XNNPACK/src/qs8-vlrelu/gen/ |
| D | vlrelu-armsimd32-x8.c | 31 const int8x4_t vx0123 = (int8x4_t) unaligned_indexed_load_u32(x, 0); in xnn_qs8_vlrelu_ukernel__armsimd32_x8() local
35 int16x2_t vx02 = __sxtb16(vx0123); in xnn_qs8_vlrelu_ukernel__armsimd32_x8()
36 int16x2_t vx13 = __sxtb16(__ror(vx0123, 8)); in xnn_qs8_vlrelu_ukernel__armsimd32_x8()
78 const int8x4_t vx0123 = (int8x4_t) unaligned_load_u32(x); in xnn_qs8_vlrelu_ukernel__armsimd32_x8() local
81 int16x2_t vx02 = __sxtb16(vx0123); in xnn_qs8_vlrelu_ukernel__armsimd32_x8()
82 int16x2_t vx13 = __sxtb16(__ror(vx0123, 8)); in xnn_qs8_vlrelu_ukernel__armsimd32_x8()
106 const int8x4_t vx0123 = (int8x4_t) unaligned_load_u32(x); in xnn_qs8_vlrelu_ukernel__armsimd32_x8() local
108 int16x2_t vx02 = __sxtb16(vx0123); in xnn_qs8_vlrelu_ukernel__armsimd32_x8()
109 int16x2_t vx13 = __sxtb16(__ror(vx0123, 8)); in xnn_qs8_vlrelu_ukernel__armsimd32_x8()
|
| D | vlrelu-armsimd32-x4.c | 31 const int8x4_t vx0123 = (int8x4_t) unaligned_load_u32(x); in xnn_qs8_vlrelu_ukernel__armsimd32_x4() local
34 int16x2_t vx02 = __sxtb16(vx0123); in xnn_qs8_vlrelu_ukernel__armsimd32_x4()
35 int16x2_t vx13 = __sxtb16(__ror(vx0123, 8)); in xnn_qs8_vlrelu_ukernel__armsimd32_x4()
59 const int8x4_t vx0123 = (int8x4_t) unaligned_load_u32(x); in xnn_qs8_vlrelu_ukernel__armsimd32_x4() local
61 int16x2_t vx02 = __sxtb16(vx0123); in xnn_qs8_vlrelu_ukernel__armsimd32_x4()
62 int16x2_t vx13 = __sxtb16(__ror(vx0123, 8)); in xnn_qs8_vlrelu_ukernel__armsimd32_x4()
|
| /external/XNNPACK/src/qu8-vlrelu/gen/ |
| D | vlrelu-armsimd32-x8.c | 31 const uint8x4_t vx0123 = (uint8x4_t) unaligned_indexed_load_u32(x, 0); in xnn_qu8_vlrelu_ukernel__armsimd32_x8() local
35 uint16x2_t vx02 = __uxtb16(vx0123); in xnn_qu8_vlrelu_ukernel__armsimd32_x8()
36 uint16x2_t vx13 = __uxtb16(__ror(vx0123, 8)); in xnn_qu8_vlrelu_ukernel__armsimd32_x8()
78 const uint8x4_t vx0123 = (uint8x4_t) unaligned_load_u32(x); in xnn_qu8_vlrelu_ukernel__armsimd32_x8() local
81 uint16x2_t vx02 = __uxtb16(vx0123); in xnn_qu8_vlrelu_ukernel__armsimd32_x8()
82 uint16x2_t vx13 = __uxtb16(__ror(vx0123, 8)); in xnn_qu8_vlrelu_ukernel__armsimd32_x8()
106 const uint8x4_t vx0123 = (uint8x4_t) unaligned_load_u32(x); in xnn_qu8_vlrelu_ukernel__armsimd32_x8() local
108 uint16x2_t vx02 = __uxtb16(vx0123); in xnn_qu8_vlrelu_ukernel__armsimd32_x8()
109 uint16x2_t vx13 = __uxtb16(__ror(vx0123, 8)); in xnn_qu8_vlrelu_ukernel__armsimd32_x8()
|
| D | vlrelu-armsimd32-x4.c | 31 const uint8x4_t vx0123 = (uint8x4_t) unaligned_load_u32(x); in xnn_qu8_vlrelu_ukernel__armsimd32_x4() local
34 uint16x2_t vx02 = __uxtb16(vx0123); in xnn_qu8_vlrelu_ukernel__armsimd32_x4()
35 uint16x2_t vx13 = __uxtb16(__ror(vx0123, 8)); in xnn_qu8_vlrelu_ukernel__armsimd32_x4()
59 const uint8x4_t vx0123 = (uint8x4_t) unaligned_load_u32(x); in xnn_qu8_vlrelu_ukernel__armsimd32_x4() local
61 uint16x2_t vx02 = __uxtb16(vx0123); in xnn_qu8_vlrelu_ukernel__armsimd32_x4()
62 uint16x2_t vx13 = __uxtb16(__ror(vx0123, 8)); in xnn_qu8_vlrelu_ukernel__armsimd32_x4()
|
| /external/XNNPACK/src/f32-vrnd/gen/ |
| D | vrndd-neon-x4.c | 31 const float32x4_t vx0123 = vld1q_f32(x); x += 4; in xnn_f32_vrndd_ukernel__neon_x4() local
33 const int32x4_t vintx0123 = vcvtq_s32_f32(vx0123); in xnn_f32_vrndd_ukernel__neon_x4()
35 uint32x4_t vrndmask0123 = vcaltq_f32(vx0123, vintegral_threshold); in xnn_f32_vrndd_ukernel__neon_x4()
41 const float32x4_t vrndx0123 = vbslq_f32(vrndmask0123, vprerndx0123, vx0123); in xnn_f32_vrndd_ukernel__neon_x4()
43 const uint32x4_t vadjmask0123 = vcgtq_f32(vrndx0123, vx0123); in xnn_f32_vrndd_ukernel__neon_x4()
|
| D | vrndu-neon-x4.c | 31 const float32x4_t vx0123 = vld1q_f32(x); x += 4; in xnn_f32_vrndu_ukernel__neon_x4() local
33 const int32x4_t vintx0123 = vcvtq_s32_f32(vx0123); in xnn_f32_vrndu_ukernel__neon_x4()
35 uint32x4_t vrndmask0123 = vcaltq_f32(vx0123, vintegral_threshold); in xnn_f32_vrndu_ukernel__neon_x4()
41 const float32x4_t vrndx0123 = vbslq_f32(vrndmask0123, vprerndx0123, vx0123); in xnn_f32_vrndu_ukernel__neon_x4()
43 uint32x4_t vadjmask0123 = vcgeq_f32(vrndx0123, vx0123); in xnn_f32_vrndu_ukernel__neon_x4()
|