| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | psimd-p5-x8.c | 70 psimd_f32 vt4567 = psimd_qfma_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8() local
73 vt4567 = psimd_qfma_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
77 psimd_f32 vp4567 = psimd_qfma_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
80 vp4567 = psimd_qfma_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
83 vp4567 = psimd_qfma_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
86 vp4567 = psimd_qfma_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
93 vt4567 = psimd_mul_f32(vt4567, vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
96 psimd_f32 vf4567 = psimd_qfma_f32(vs4567, vt4567, vp4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
|
| D | psimd-p5-x8-acc2.c | 71 psimd_f32 vt4567 = psimd_qfma_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2() local
74 vt4567 = psimd_qfma_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
78 psimd_f32 vp4567 = psimd_qfma_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
81 vp4567 = psimd_qfma_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
84 vp4567 = psimd_qfma_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
87 vp4567 = psimd_qfma_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
94 vt4567 = psimd_mul_f32(vt4567, vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
97 psimd_f32 vf4567 = psimd_qfma_f32(vs4567, vt4567, vp4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
|
| D | neon-p5-x8-acc2.c | 75 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2() local
78 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
82 float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
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()
98 vt4567 = vmulq_f32(vt4567, vs4567); 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()
|
| D | sse2-p5-x8.c | 70 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vx4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8() local
73 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
77 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
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()
93 vt4567 = _mm_mul_ps(vt4567, vs4567); 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 | neonfma-p5-x8-acc2.c | 74 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2() local
77 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
81 float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
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()
97 vt4567 = vmulq_f32(vt4567, vs4567); 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 | neon-p5-x8.c | 74 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8() local
77 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
81 float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
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()
97 vt4567 = vmulq_f32(vt4567, vs4567); 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-acc2.c | 71 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vx4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2() local
74 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
78 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
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()
94 vt4567 = _mm_mul_ps(vt4567, vs4567); 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.c | 73 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8() local
76 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
80 float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
83 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
86 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
89 vp4567 = vfmaq_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
96 vt4567 = vmulq_f32(vt4567, vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
99 float32x4_t vf4567 = vfmaq_f32(vs4567, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
|
| D | psimd-p5-x12.c | 75 psimd_f32 vt4567 = psimd_qfma_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12() local
79 vt4567 = psimd_qfma_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
84 psimd_f32 vp4567 = psimd_qfma_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
88 vp4567 = psimd_qfma_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
92 vp4567 = psimd_qfma_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
96 vp4567 = psimd_qfma_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
104 vt4567 = psimd_mul_f32(vt4567, vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
108 psimd_f32 vf4567 = psimd_qfma_f32(vs4567, vt4567, vp4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
|
| D | psimd-p5-x12-acc2.c | 76 psimd_f32 vt4567 = psimd_qfma_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2() local
80 vt4567 = psimd_qfma_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
85 psimd_f32 vp4567 = psimd_qfma_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
89 vp4567 = psimd_qfma_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
93 vp4567 = psimd_qfma_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
97 vp4567 = psimd_qfma_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
105 vt4567 = psimd_mul_f32(vt4567, vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
109 psimd_f32 vf4567 = psimd_qfma_f32(vs4567, vt4567, vp4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
|
| D | psimd-p5-x12-acc3.c | 77 psimd_f32 vt4567 = psimd_qfma_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3() local
81 vt4567 = psimd_qfma_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
86 psimd_f32 vp4567 = psimd_qfma_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
90 vp4567 = psimd_qfma_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
94 vp4567 = psimd_qfma_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
98 vp4567 = psimd_qfma_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
106 vt4567 = psimd_mul_f32(vt4567, vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
110 psimd_f32 vf4567 = psimd_qfma_f32(vs4567, vt4567, vp4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
|
| D | sse2-p5-x12.c | 75 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vx4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12() local
79 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
84 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
88 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
92 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
96 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
104 vt4567 = _mm_mul_ps(vt4567, vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
108 __m128 vf4567 = _mm_add_ps(_mm_mul_ps(vt4567, vp4567), vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
|
| D | neon-p5-x12.c | 79 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12() local
83 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
88 float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
92 vp4567 = vmlaq_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
96 vp4567 = vmlaq_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
100 vp4567 = vmlaq_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
108 vt4567 = vmulq_f32(vt4567, vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
112 float32x4_t vf4567 = vmlaq_f32(vs4567, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
|
| D | neonfma-p5-x12.c | 78 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() local
82 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
87 float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
91 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
95 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
99 vp4567 = vfmaq_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
107 vt4567 = vmulq_f32(vt4567, vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
111 float32x4_t vf4567 = vfmaq_f32(vs4567, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
|
| D | neonfma-p5-x12-acc2.c | 79 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2() local
83 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
88 float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
92 vp4567 = vfmaq_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
96 vp4567 = vfmaq_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
100 vp4567 = vfmaq_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
108 vt4567 = vmulq_f32(vt4567, vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
112 float32x4_t vf4567 = vfmaq_f32(vs4567, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
|
| D | neon-p5-x12-acc2.c | 80 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2() local
84 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
89 float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
93 vp4567 = vmlaq_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
97 vp4567 = vmlaq_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
101 vp4567 = vmlaq_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
109 vt4567 = vmulq_f32(vt4567, vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
113 float32x4_t vf4567 = vmlaq_f32(vs4567, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
|
| D | psimd-p5-x16-acc2.c | 81 psimd_f32 vt4567 = psimd_qfma_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2() local
86 vt4567 = psimd_qfma_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
92 psimd_f32 vp4567 = psimd_qfma_f32(vc4, vc5, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
97 vp4567 = psimd_qfma_f32(vc3, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
102 vp4567 = psimd_qfma_f32(vc2, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
107 vp4567 = psimd_qfma_f32(vc1, vp4567, vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
116 vt4567 = psimd_mul_f32(vt4567, vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
121 psimd_f32 vf4567 = psimd_qfma_f32(vs4567, vt4567, vp4567); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
|
| D | sse2-p5-x12-acc3.c | 77 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vx4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3() local
81 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
86 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
90 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
94 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
98 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
106 vt4567 = _mm_mul_ps(vt4567, vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
110 __m128 vf4567 = _mm_add_ps(_mm_mul_ps(vt4567, vp4567), vs4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
|
| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | psimd-p5-div-x8.c | 78 psimd_f32 vt4567 = psimd_qfma_f32(vz4567, vn4567, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8() local
81 vt4567 = psimd_qfma_f32(vt4567, vn4567, vln2_lo); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
86 psimd_f32 vp4567 = psimd_qfma_f32(vc4, vt4567, vc5); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
89 vp4567 = psimd_qfma_f32(vc3, vt4567, vp4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
92 vp4567 = psimd_qfma_f32(vc2, vt4567, vp4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
95 vp4567 = psimd_qfma_f32(vc1, vt4567, vp4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
102 vt4567 = psimd_mul_f32(vt4567, vs4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
105 const psimd_f32 ve4567 = psimd_qfma_f32(vs4567, vt4567, vp4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
|
| D | sse41-p5-div-x8.c | 78 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8() local
81 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
85 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
88 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
91 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
94 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc1); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
101 vt4567 = _mm_mul_ps(vt4567, vs4567); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
104 __m128 ve4567 = _mm_add_ps(_mm_mul_ps(vt4567, vp4567), vs4567); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
|
| D | psimd-p5-div-x12.c | 83 psimd_f32 vt4567 = psimd_qfma_f32(vz4567, vn4567, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12() local
87 vt4567 = psimd_qfma_f32(vt4567, vn4567, vln2_lo); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
93 psimd_f32 vp4567 = psimd_qfma_f32(vc4, vt4567, vc5); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
97 vp4567 = psimd_qfma_f32(vc3, vt4567, vp4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
101 vp4567 = psimd_qfma_f32(vc2, vt4567, vp4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
105 vp4567 = psimd_qfma_f32(vc1, vt4567, vp4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
113 vt4567 = psimd_mul_f32(vt4567, vs4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
117 const psimd_f32 ve4567 = psimd_qfma_f32(vs4567, vt4567, vp4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
|
| D | sse2-p5-div-x8.c | 78 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8() local
81 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
85 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
88 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
91 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
94 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc1); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
101 vt4567 = _mm_mul_ps(vt4567, vs4567); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
104 __m128 ve4567 = _mm_add_ps(_mm_mul_ps(vt4567, vp4567), vs4567); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
|
| D | neon-rr2-p5-nr2recps-x8.c | 77 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8() local
80 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
84 float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
87 vp4567 = vmlaq_f32(vc3, vp4567, vt4567); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
90 vp4567 = vmlaq_f32(vc2, vp4567, vt4567); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
93 vp4567 = vmlaq_f32(vc1, vp4567, vt4567); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
100 vt4567 = vmulq_f32(vt4567, vs4567); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
103 float32x4_t ve4567 = vmlaq_f32(vs4567, vp4567, vt4567); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
|
| D | psimd-p5-div-x16.c | 88 psimd_f32 vt4567 = psimd_qfma_f32(vz4567, vn4567, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16() local
93 vt4567 = psimd_qfma_f32(vt4567, vn4567, vln2_lo); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
100 psimd_f32 vp4567 = psimd_qfma_f32(vc4, vt4567, vc5); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
105 vp4567 = psimd_qfma_f32(vc3, vt4567, vp4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
110 vp4567 = psimd_qfma_f32(vc2, vt4567, vp4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
115 vp4567 = psimd_qfma_f32(vc1, vt4567, vp4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
124 vt4567 = psimd_mul_f32(vt4567, vs4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
129 const psimd_f32 ve4567 = psimd_qfma_f32(vs4567, vt4567, vp4567); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
|
| D | sse41-p5-div-x12.c | 83 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12() local
87 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
92 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
96 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
100 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
104 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc1); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
112 vt4567 = _mm_mul_ps(vt4567, vs4567); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
116 __m128 ve4567 = _mm_add_ps(_mm_mul_ps(vt4567, vp4567), vs4567); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
|