| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-neonfma-rr1-p6-x8.c | 62 float32x4_t vp4567 = vfmaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x8() local
65 vp4567 = vfmaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x8()
68 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x8()
71 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x8()
74 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x8()
82 vp4567 = vfmaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x8()
85 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x8()
|
| D | velu-neon-rr2-p6-x8.c | 66 float32x4_t vp4567 = vmlaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x8() local
69 vp4567 = vmlaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
72 vp4567 = vmlaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
75 vp4567 = vmlaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
78 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
86 vp4567 = vmlaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
89 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
|
| D | velu-sse41-rr2-p6-x8.c | 68 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8() local
71 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
74 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
77 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
80 vp4567 = _mm_mul_ps(vp4567, vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
88 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
91 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
|
| D | velu-wasmsimd-x86-rr2-p6-x8.c | 74 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8() local
77 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8()
80 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8()
83 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8()
86 vp4567 = wasm_f32x4_mul(vp4567, vt4567); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8()
94 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8()
97 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8()
|
| D | velu-wasmsimd-arm-rr2-p6-x8.c | 68 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x8() local
71 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x8()
74 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x8()
77 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x8()
80 vp4567 = wasm_f32x4_mul(vp4567, vt4567); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x8()
88 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x8()
91 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x8()
|
| D | velu-sse2-rr2-p6-x8.c | 68 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__sse2_rr2_p6_x8() local
71 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__sse2_rr2_p6_x8()
74 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__sse2_rr2_p6_x8()
77 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__sse2_rr2_p6_x8()
80 vp4567 = _mm_mul_ps(vp4567, vt4567); in xnn_f32_velu_ukernel__sse2_rr2_p6_x8()
88 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__sse2_rr2_p6_x8()
91 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__sse2_rr2_p6_x8()
|
| D | velu-neon-rr2-p6-x12.c | 73 float32x4_t vp4567 = vmlaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x12() local
77 vp4567 = vmlaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
81 vp4567 = vmlaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
85 vp4567 = vmlaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
89 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
100 vp4567 = vmlaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
104 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
|
| D | velu-neonfma-rr1-p6-x12.c | 68 float32x4_t vp4567 = vfmaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12() local
72 vp4567 = vfmaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
76 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
80 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
84 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
95 vp4567 = vfmaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
99 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
|
| D | velu-wasmsimd-arm-rr2-p6-x12.c | 75 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12() local
79 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12()
83 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12()
87 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12()
91 vp4567 = wasm_f32x4_mul(vp4567, vt4567); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12()
102 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12()
106 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12()
|
| D | velu-sse41-rr2-p6-x12.c | 75 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12() local
79 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
83 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
87 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
91 vp4567 = _mm_mul_ps(vp4567, vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
102 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
106 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
|
| D | velu-wasmsimd-x86-rr2-p6-x12.c | 83 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12() local
88 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
92 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
96 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
100 vp4567 = wasm_f32x4_mul(vp4567, vt4567); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
111 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
115 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
|
| D | velu-sse2-rr2-p6-x12.c | 75 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12() local
79 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
83 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
87 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
91 vp4567 = _mm_mul_ps(vp4567, vt4567); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
102 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
106 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
|
| D | velu-neonfma-rr1-p6-x16.c | 74 float32x4_t vp4567 = vfmaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16() local
79 vp4567 = vfmaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
84 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
89 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
94 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
108 vp4567 = vfmaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
113 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
|
| D | velu-neon-rr2-p6-x16.c | 80 float32x4_t vp4567 = vmlaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x16() local
85 vp4567 = vmlaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
90 vp4567 = vmlaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
95 vp4567 = vmlaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
100 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
114 vp4567 = vmlaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
119 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
|
| D | velu-sse41-rr2-p6-x16.c | 82 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16() local
87 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
92 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
97 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
102 vp4567 = _mm_mul_ps(vp4567, vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
116 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
121 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
|
| D | velu-wasmsimd-arm-rr2-p6-x16.c | 82 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16() local
87 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
92 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
97 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
102 vp4567 = wasm_f32x4_mul(vp4567, vt4567); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
116 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
121 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
|
| D | velu-wasmsimd-x86-rr2-p6-x16.c | 92 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16() local
99 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
104 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
109 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
114 vp4567 = wasm_f32x4_mul(vp4567, vt4567); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
128 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
133 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
|
| D | velu-sse41-rr2-p6-x20.c | 89 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20() local
95 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
101 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
107 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
113 vp4567 = _mm_mul_ps(vp4567, vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
130 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
136 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
|
| D | velu-neonfma-rr1-p6-x20.c | 80 float32x4_t vp4567 = vfmaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20() local
86 vp4567 = vfmaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
92 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
98 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
104 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
121 vp4567 = vfmaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
127 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
|
| D | velu-sse2-rr2-p6-x16.c | 82 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16() local
87 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
92 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
97 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
102 vp4567 = _mm_mul_ps(vp4567, vt4567); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
116 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
121 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
|
| D | velu-neon-rr2-p6-x20.c | 87 float32x4_t vp4567 = vmlaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x20() local
93 vp4567 = vmlaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
99 vp4567 = vmlaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
105 vp4567 = vmlaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
111 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
128 vp4567 = vmlaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
134 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | neonfma-rr1-p5-x8.c | 60 float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x8() local
63 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x8()
66 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x8()
69 vp4567 = vfmaq_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x8()
75 float32x4_t vf4567 = vfmaq_f32(vs4567, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x8()
|
| D | neonfma-rr1-p5-x8-acc2.c | 61 float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x8_acc2() local
64 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x8_acc2()
67 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x8_acc2()
70 vp4567 = vfmaq_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x8_acc2()
76 float32x4_t vf4567 = vfmaq_f32(vs4567, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x8_acc2()
|
| D | neon-rr2-p5-x8-acc2.c | 65 float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x8_acc2() local
68 vp4567 = vmlaq_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x8_acc2()
71 vp4567 = vmlaq_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x8_acc2()
74 vp4567 = vmlaq_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x8_acc2()
80 float32x4_t vf4567 = vmlaq_f32(vs4567, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x8_acc2()
|
| /external/XNNPACK/src/f32-vsigmoid/gen/ |
| D | vsigmoid-neonfma-rr1-p5-div-x8.c | 57 float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x8() local
60 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x8()
63 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x8()
66 vp4567 = vfmaq_f32(vc1, vp4567, vt4567); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x8()
72 const float32x4_t ve4567 = vfmaq_f32(vs4567, vp4567, vt4567); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x8()
|