Home
last modified time | relevance | path

Searched refs:vtCDEF (Results 1 – 25 of 135) sorted by relevance

123456

/external/XNNPACK/src/f32-velu/gen/
Dvelu-wasmsimd-x86-rr2-p6-x16.c78 v128_t vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_hi), vzCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16() local
87 vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
96 vtCDEF = wasm_v128_andnot(vtCDEF, vsatmCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
97 v128_t vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtCDEF), vc5); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
102 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc4); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
107 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc3); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
112 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc2); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
117 vpCDEF = wasm_f32x4_mul(vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
125 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
131 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
Dvelu-wasmsimd-arm-rr2-p6-x16.c75 v128_t vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_hi), vzCDEF); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16() local
80 vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
85 v128_t vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtCDEF), vc5); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
90 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc4); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
95 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc3); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
100 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc2); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
105 vpCDEF = wasm_f32x4_mul(vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
113 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
119 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
Dvelu-sse41-rr2-p6-x16.c75 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16() local
80 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
85 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc6, vtCDEF), vc5); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
90 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc4); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
95 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
100 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
105 vpCDEF = _mm_mul_ps(vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
113 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
119 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vtCDEF); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
Dvelu-neon-rr2-p6-x16.c73 float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_velu_ukernel__neon_rr2_p6_x16() local
78 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
83 float32x4_t vpCDEF = vmlaq_f32(vc5, vc6, vtCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
88 vpCDEF = vmlaq_f32(vc4, vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
93 vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
98 vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
103 vpCDEF = vmulq_f32(vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
111 vtCDEF = vmulq_f32(vtCDEF, vsCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
117 vpCDEF = vmlaq_f32(vtCDEF, vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
Dvelu-wasmsimd-x86-rr2-p6-x20.c83 v128_t vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_hi), vzCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20() local
94 vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
105 vtCDEF = wasm_v128_andnot(vtCDEF, vsatmCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
106 v128_t vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtCDEF), vc5); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
113 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc4); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
119 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc3); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
125 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc2); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
131 vpCDEF = wasm_f32x4_mul(vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
140 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
148 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
Dvelu-sse2-rr2-p6-x16.c75 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16() local
80 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
85 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc6, vtCDEF), vc5); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
90 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc4); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
95 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
100 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
105 vpCDEF = _mm_mul_ps(vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
113 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
119 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vtCDEF); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
Dvelu-wasmsimd-arm-rr2-p6-x20.c80 v128_t vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_hi), vzCDEF); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20() local
86 vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
92 v128_t vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtCDEF), vc5); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
98 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc4); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
104 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc3); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
110 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc2); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
116 vpCDEF = wasm_f32x4_mul(vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
125 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
133 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
Dvelu-sse41-rr2-p6-x20.c80 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20() local
86 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
92 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc6, vtCDEF), vc5); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
98 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc4); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
104 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
110 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
116 vpCDEF = _mm_mul_ps(vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
125 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
133 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vtCDEF); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
Dvelu-neon-rr2-p6-x20.c78 float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_velu_ukernel__neon_rr2_p6_x20() local
84 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
90 float32x4_t vpCDEF = vmlaq_f32(vc5, vc6, vtCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
96 vpCDEF = vmlaq_f32(vc4, vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
102 vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
108 vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
114 vpCDEF = vmulq_f32(vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
123 vtCDEF = vmulq_f32(vtCDEF, vsCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
131 vpCDEF = vmlaq_f32(vtCDEF, vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
Dvelu-wasmsimd-x86-rr2-p6-x24.c88 v128_t vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_hi), vzCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24() local
101 vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
114 vtCDEF = wasm_v128_andnot(vtCDEF, vsatmCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
115 v128_t vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtCDEF), vc5); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
124 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc4); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
131 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc3); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
138 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc2); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
145 vpCDEF = wasm_f32x4_mul(vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
155 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
165 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vtCDEF); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
Dvelu-neonfma-rr1-p6-x16.c72 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vminus_ln2); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16() local
77 float32x4_t vpCDEF = vfmaq_f32(vc5, vc6, vtCDEF); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
82 vpCDEF = vfmaq_f32(vc4, vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
87 vpCDEF = vfmaq_f32(vc3, vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
92 vpCDEF = vfmaq_f32(vc2, vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
97 vpCDEF = vmulq_f32(vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
105 vtCDEF = vmulq_f32(vtCDEF, vsCDEF); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
111 vpCDEF = vfmaq_f32(vtCDEF, vpCDEF, vtCDEF); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Dneon-p5-x16.c86 float32x4_t vtCDEF = vmlaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16() local
91 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
97 float32x4_t vpCDEF = vmlaq_f32(vc4, vc5, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
102 vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
107 vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
112 vpCDEF = vmlaq_f32(vc1, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
121 vtCDEF = vmulq_f32(vtCDEF, vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
126 float32x4_t vfCDEF = vmlaq_f32(vsCDEF, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
Dwasmsimd-p5-x16.c82 v128_t vtCDEF = wasm_f32x4_add(vxCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_hi)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16() local
87 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_lo)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16()
93 v128_t vpCDEF = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16()
98 vpCDEF = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpCDEF, vtCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16()
103 vpCDEF = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpCDEF, vtCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16()
108 vpCDEF = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpCDEF, vtCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16()
117 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16()
122 v128_t vfCDEF = wasm_f32x4_add(vsCDEF, wasm_f32x4_mul(vtCDEF, vpCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16()
Dneonfma-p5-x16.c85 float32x4_t vtCDEF = vfmaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16() local
90 vtCDEF = vfmaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
96 float32x4_t vpCDEF = vfmaq_f32(vc4, vc5, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
101 vpCDEF = vfmaq_f32(vc3, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
106 vpCDEF = vfmaq_f32(vc2, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
111 vpCDEF = vfmaq_f32(vc1, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
120 vtCDEF = vmulq_f32(vtCDEF, vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
125 float32x4_t vfCDEF = vfmaq_f32(vsCDEF, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
Dsse2-p5-x16-acc4.c85 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vxCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4() local
90 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4()
96 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc5, vtCDEF), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4()
101 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4()
106 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4()
111 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4()
120 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4()
125 __m128 vfCDEF = _mm_add_ps(_mm_mul_ps(vtCDEF, vpCDEF), vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4()
Dneonfma-p5-x16-acc2.c86 float32x4_t vtCDEF = vfmaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2() local
91 vtCDEF = vfmaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
97 float32x4_t vpCDEF = vfmaq_f32(vc4, vc5, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
102 vpCDEF = vfmaq_f32(vc3, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
107 vpCDEF = vfmaq_f32(vc2, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
112 vpCDEF = vfmaq_f32(vc1, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
121 vtCDEF = vmulq_f32(vtCDEF, vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
126 float32x4_t vfCDEF = vfmaq_f32(vsCDEF, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
Dneon-p5-x16-acc2.c87 float32x4_t vtCDEF = vmlaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2() local
92 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
98 float32x4_t vpCDEF = vmlaq_f32(vc4, vc5, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
103 vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
108 vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
113 vpCDEF = vmlaq_f32(vc1, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
122 vtCDEF = vmulq_f32(vtCDEF, vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
127 float32x4_t vfCDEF = vmlaq_f32(vsCDEF, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
Dwasmsimd-p5-x16-acc2.c83 v128_t vtCDEF = wasm_f32x4_add(vxCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_hi)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc2() local
88 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_lo)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc2()
94 v128_t vpCDEF = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc2()
99 vpCDEF = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpCDEF, vtCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc2()
104 vpCDEF = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpCDEF, vtCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc2()
109 vpCDEF = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpCDEF, vtCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc2()
118 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc2()
123 v128_t vfCDEF = wasm_f32x4_add(vsCDEF, wasm_f32x4_mul(vtCDEF, vpCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc2()
Dsse2-p5-x16-acc2.c83 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vxCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2() local
88 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2()
94 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc5, vtCDEF), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2()
99 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2()
104 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2()
109 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2()
118 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2()
123 __m128 vfCDEF = _mm_add_ps(_mm_mul_ps(vtCDEF, vpCDEF), vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2()
Dwasmsimd-p5-x16-acc4.c85 v128_t vtCDEF = wasm_f32x4_add(vxCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_hi)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc4() local
90 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vminus_ln2_lo)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc4()
96 v128_t vpCDEF = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc4()
101 vpCDEF = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpCDEF, vtCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc4()
106 vpCDEF = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpCDEF, vtCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc4()
111 vpCDEF = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpCDEF, vtCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc4()
120 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc4()
125 v128_t vfCDEF = wasm_f32x4_add(vsCDEF, wasm_f32x4_mul(vtCDEF, vpCDEF)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc4()
Dneonfma-p5-x16-acc4.c88 float32x4_t vtCDEF = vfmaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4() local
93 vtCDEF = vfmaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
99 float32x4_t vpCDEF = vfmaq_f32(vc4, vc5, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
104 vpCDEF = vfmaq_f32(vc3, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
109 vpCDEF = vfmaq_f32(vc2, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
114 vpCDEF = vfmaq_f32(vc1, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
123 vtCDEF = vmulq_f32(vtCDEF, vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
128 float32x4_t vfCDEF = vfmaq_f32(vsCDEF, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
Dneon-p5-x16-acc4.c89 float32x4_t vtCDEF = vmlaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4() local
94 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
100 float32x4_t vpCDEF = vmlaq_f32(vc4, vc5, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
105 vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
110 vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
115 vpCDEF = vmlaq_f32(vc1, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
124 vtCDEF = vmulq_f32(vtCDEF, vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
129 float32x4_t vfCDEF = vmlaq_f32(vsCDEF, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
Dsse2-p5-x16.c82 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vxCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16() local
87 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
93 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc5, vtCDEF), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
98 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
103 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
108 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
117 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
122 __m128 vfCDEF = _mm_add_ps(_mm_mul_ps(vtCDEF, vpCDEF), vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
Dneonfma-p5-x20.c90 float32x4_t vtCDEF = vfmaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20() local
96 vtCDEF = vfmaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
103 float32x4_t vpCDEF = vfmaq_f32(vc4, vc5, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
109 vpCDEF = vfmaq_f32(vc3, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
115 vpCDEF = vfmaq_f32(vc2, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
121 vpCDEF = vfmaq_f32(vc1, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
131 vtCDEF = vmulq_f32(vtCDEF, vsCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
137 float32x4_t vfCDEF = vfmaq_f32(vsCDEF, vpCDEF, vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
/external/XNNPACK/src/f32-sigmoid/gen/
Dwasmsimd-p5-div-x16.c68 v128_t vtCDEF = wasm_f32x4_add(vzCDEF, wasm_f32x4_mul(vnCDEF, vln2_hi)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16() local
73 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vln2_lo)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
78 v128_t vpCDEF = wasm_f32x4_add(vc4, wasm_f32x4_mul(vtCDEF, vc5)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
83 vpCDEF = wasm_f32x4_add(vc3, wasm_f32x4_mul(vtCDEF, vpCDEF)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
88 vpCDEF = wasm_f32x4_add(vc2, wasm_f32x4_mul(vtCDEF, vpCDEF)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
93 vpCDEF = wasm_f32x4_add(vc1, wasm_f32x4_mul(vtCDEF, vpCDEF)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
98 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()
103 const v128_t veCDEF = wasm_f32x4_add(vsCDEF, wasm_f32x4_mul(vtCDEF, vpCDEF)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16()

123456