| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-neonfma-rr1-p6-x8.c | 63 float32x4_t vp4567 = vfmaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x8() local
66 vp4567 = vfmaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x8()
69 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x8()
72 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x8()
75 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x8()
83 vp4567 = vfmaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x8()
86 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 | 67 float32x4_t vp4567 = vmlaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x8() local
70 vp4567 = vmlaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
73 vp4567 = vmlaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
76 vp4567 = vmlaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
79 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
87 vp4567 = vmlaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x8()
90 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 | 69 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8() local
72 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
75 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
78 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
81 vp4567 = _mm_mul_ps(vp4567, vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
89 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
92 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__sse41_rr2_p6_x8()
|
| D | velu-wasmsimd-arm-rr2-p6-x8.c | 69 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x8() local
72 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x8()
75 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x8()
78 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x8()
81 vp4567 = wasm_f32x4_mul(vp4567, vt4567); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x8()
89 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x8()
92 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-neonfma-rr1-p6-x12.c | 69 float32x4_t vp4567 = vfmaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12() local
73 vp4567 = vfmaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
77 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
81 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
85 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
96 vp4567 = vfmaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
100 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x12()
|
| D | velu-sse2-rr2-p6-x8.c | 69 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__sse2_rr2_p6_x8() local
72 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__sse2_rr2_p6_x8()
75 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__sse2_rr2_p6_x8()
78 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__sse2_rr2_p6_x8()
81 vp4567 = _mm_mul_ps(vp4567, vt4567); in xnn_f32_velu_ukernel__sse2_rr2_p6_x8()
89 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__sse2_rr2_p6_x8()
92 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__sse2_rr2_p6_x8()
|
| D | velu-wasmsimd-x86-rr2-p6-x8.c | 75 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8() local
78 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8()
81 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8()
84 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8()
87 vp4567 = wasm_f32x4_mul(vp4567, vt4567); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8()
95 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8()
98 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-neon-rr2-p6-x12.c | 74 float32x4_t vp4567 = vmlaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x12() local
78 vp4567 = vmlaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
82 vp4567 = vmlaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
86 vp4567 = vmlaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
90 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
101 vp4567 = vmlaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
105 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__neon_rr2_p6_x12()
|
| D | velu-wasmsimd-arm-rr2-p6-x12.c | 76 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12() local
80 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12()
84 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12()
88 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12()
92 vp4567 = wasm_f32x4_mul(vp4567, vt4567); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12()
103 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x12()
107 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-sse2-rr2-p6-x12.c | 76 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12() local
80 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
84 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
88 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
92 vp4567 = _mm_mul_ps(vp4567, vt4567); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
103 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
107 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__sse2_rr2_p6_x12()
|
| D | velu-sse41-rr2-p6-x12.c | 76 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12() local
80 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
84 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
88 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
92 vp4567 = _mm_mul_ps(vp4567, vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
103 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
107 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__sse41_rr2_p6_x12()
|
| D | velu-neonfma-rr1-p6-x16.c | 75 float32x4_t vp4567 = vfmaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16() local
80 vp4567 = vfmaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
85 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
90 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
95 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
109 vp4567 = vfmaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
114 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x16()
|
| D | velu-wasmsimd-arm-rr2-p6-x16.c | 83 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16() local
88 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
93 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
98 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
103 vp4567 = wasm_f32x4_mul(vp4567, vt4567); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
117 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16()
122 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-x12.c | 84 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12() local
89 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
93 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
97 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
101 vp4567 = wasm_f32x4_mul(vp4567, vt4567); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
112 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
116 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-sse41-rr2-p6-x16.c | 83 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16() local
88 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
93 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
98 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
103 vp4567 = _mm_mul_ps(vp4567, vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
117 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
122 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__sse41_rr2_p6_x16()
|
| D | velu-neon-rr2-p6-x16.c | 81 float32x4_t vp4567 = vmlaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x16() local
86 vp4567 = vmlaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
91 vp4567 = vmlaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
96 vp4567 = vmlaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
101 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
115 vp4567 = vmlaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
120 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__neon_rr2_p6_x16()
|
| D | velu-neonfma-rr1-p6-x20.c | 81 float32x4_t vp4567 = vfmaq_f32(vc5, vc6, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20() local
87 vp4567 = vfmaq_f32(vc4, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
93 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
99 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
105 vp4567 = vmulq_f32(vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
122 vp4567 = vfmaq_f32(vt4567, vp4567, vt4567); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
128 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 | 83 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16() local
88 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
93 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
98 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
103 vp4567 = _mm_mul_ps(vp4567, vt4567); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
117 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
122 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha); in xnn_f32_velu_ukernel__sse2_rr2_p6_x16()
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | sse2-p5-x8-acc2.c | 78 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2() local
81 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
84 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
87 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
97 __m128 vf4567 = _mm_add_ps(_mm_mul_ps(vt4567, vp4567), vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
|
| D | neonfma-p5-x8-acc2.c | 81 float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2() local
84 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
87 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
90 vp4567 = vfmaq_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
100 float32x4_t vf4567 = vfmaq_f32(vs4567, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
|
| D | wasmsimd-p5-x8.c | 77 v128_t vp4567 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt4567)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8() local
80 vp4567 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp4567, vt4567)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8()
83 vp4567 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp4567, vt4567)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8()
86 vp4567 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp4567, vt4567)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8()
96 v128_t vf4567 = wasm_f32x4_add(vs4567, wasm_f32x4_mul(vt4567, vp4567)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8()
|
| D | neon-p5-x8.c | 81 float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8() local
84 vp4567 = vmlaq_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
87 vp4567 = vmlaq_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
90 vp4567 = vmlaq_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
100 float32x4_t vf4567 = vmlaq_f32(vs4567, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
|
| D | sse2-p5-x8.c | 77 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8() local
80 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
83 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
86 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
96 __m128 vf4567 = _mm_add_ps(_mm_mul_ps(vt4567, vp4567), vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
|
| D | neon-p5-x8-acc2.c | 82 float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2() local
85 vp4567 = vmlaq_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
88 vp4567 = vmlaq_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
91 vp4567 = vmlaq_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
101 float32x4_t vf4567 = vmlaq_f32(vs4567, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
|
| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | neonfma-rr1-p5-div-x8.c | 57 float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8() local
60 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8()
63 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8()
66 vp4567 = vfmaq_f32(vc1, vp4567, vt4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8()
72 const float32x4_t ve4567 = vfmaq_f32(vs4567, vp4567, vt4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8()
|