Home
last modified time | relevance | path

Searched refs:vs89AB (Results 1 – 25 of 187) sorted by relevance

12345678

/external/XNNPACK/src/f32-velu/gen/
Dvelu-wasmsimd-x86-rr2-p6-x12.c60 v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12() local
78 vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
107 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
108 vs89AB = wasm_f32x4_sub(vs89AB, vone); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
116 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
Dvelu-wasmsimd-x86-rr2-p6-x16.c63 v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16() local
85 vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
122 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
123 vs89AB = wasm_f32x4_sub(vs89AB, vone); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
134 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
Dvelu-wasmsimd-x86-rr2-lut16-p3-x12.c92 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x12() local
106 vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x12()
123 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x12()
124 vs89AB = wasm_f32x4_sub(vs89AB, vone); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x12()
132 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x12()
Dvelu-wasmsimd-x86-rr2-p6-x20.c66 v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20() local
92 vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
137 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
138 vs89AB = wasm_f32x4_sub(vs89AB, vone); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
152 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
Dvelu-wasmsimd-x86-rr2-lut16-p3-x16.c104 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x16() local
122 vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x16()
144 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x16()
145 vs89AB = wasm_f32x4_sub(vs89AB, vone); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x16()
156 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x16()
Dvelu-neon-rr2-p6-x12.c61 float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23)); in xnn_f32_velu_ukernel__neon_rr2_p6_x12() local
96 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
97 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
105 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
Dvelu-neonfma-rr1-p6-x12.c60 float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12() local
91 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
92 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
100 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
Dvelu-wasmsimd-arm-rr2-p6-x12.c60 v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12() local
98 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12()
99 vs89AB = wasm_f32x4_sub(vs89AB, vone); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12()
107 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12()
Dvelu-sse41-rr2-p6-x12.c60 __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23)); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12() local
98 vt89AB = _mm_mul_ps(vt89AB, vs89AB); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
99 vs89AB = _mm_sub_ps(vs89AB, vone); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
107 const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
Dvelu-sse2-rr2-p6-x12.c60 __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23)); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12() local
98 vt89AB = _mm_mul_ps(vt89AB, vs89AB); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
99 vs89AB = _mm_sub_ps(vs89AB, vone); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
107 const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
Dvelu-neonfma-rr1-p6-x16.c63 float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16() local
102 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
103 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
114 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
Dvelu-wasmsimd-x86-rr2-p6-x24.c69 v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24() local
99 vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
152 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
153 vs89AB = wasm_f32x4_sub(vs89AB, vone); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
170 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
Dvelu-wasmsimd-x86-rr2-lut16-p3-x20.c116 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x20() local
138 vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x20()
165 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x20()
166 vs89AB = wasm_f32x4_sub(vs89AB, vone); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x20()
180 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x20()
Dvelu-neon-rr2-p6-x16.c64 float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23)); in xnn_f32_velu_ukernel__neon_rr2_p6_x16() local
108 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
109 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
120 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
Dvelu-sse41-rr2-p6-x16.c63 __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23)); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16() local
110 vt89AB = _mm_mul_ps(vt89AB, vs89AB); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
111 vs89AB = _mm_sub_ps(vs89AB, vone); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
122 const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
Dvelu-wasmsimd-arm-rr2-lut16-p3-x12.c92 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x12() local
114 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x12()
115 vs89AB = wasm_f32x4_sub(vs89AB, vone); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x12()
123 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x12()
Dvelu-wasmsimd-arm-rr2-p6-x16.c63 v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16() local
110 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
111 vs89AB = wasm_f32x4_sub(vs89AB, vone); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
122 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
Dvelu-wasmsimd-x86-rr2-lut16-p3-x24.c128 v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() local
154 vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24()
186 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24()
187 vs89AB = wasm_f32x4_sub(vs89AB, vone); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24()
204 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24()
Dvelu-sse41-rr2-p6-x20.c66 __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23)); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20() local
122 vt89AB = _mm_mul_ps(vt89AB, vs89AB); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
123 vs89AB = _mm_sub_ps(vs89AB, vone); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
137 const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
Dvelu-neonfma-rr1-p6-x20.c66 float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20() local
113 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
114 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
128 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
Dvelu-neonfma-rr1-lut16-p3-x12.c90 float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x12() local
108 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x12()
109 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x12()
117 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x12()
Dvelu-sse2-rr2-p6-x16.c63 __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23)); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16() local
110 vt89AB = _mm_mul_ps(vt89AB, vs89AB); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
111 vs89AB = _mm_sub_ps(vs89AB, vone); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
122 const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
Dvelu-neon-rr2-p6-x20.c67 float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23)); in xnn_f32_velu_ukernel__neon_rr2_p6_x20() local
120 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
121 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
135 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
Dvelu-neon-rr2-lut16-p3-x12.c91 float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vl89AB, ven89AB)); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x12() local
113 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x12()
114 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x12()
122 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x12()
Dvelu-neonfma-rr1-p6-x24.c69 float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24() local
124 vt89AB = vmulq_f32(vt89AB, vs89AB); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
125 vs89AB = vsubq_f32(vs89AB, vone); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
142 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()

12345678