| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | psimd-p5-x8.c | 56 psimd_f32 vn4567 = psimd_qfma_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8() local
61 const psimd_f32 vs4567 = (psimd_f32) ((psimd_u32) vn4567 << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
65 vn4567 = psimd_sub_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
70 psimd_f32 vt4567 = psimd_qfma_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
73 vt4567 = psimd_qfma_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
|
| D | psimd-p5-x8-acc2.c | 57 psimd_f32 vn4567 = psimd_qfma_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2() local
62 const psimd_f32 vs4567 = (psimd_f32) ((psimd_u32) vn4567 << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
66 vn4567 = psimd_sub_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
71 psimd_f32 vt4567 = psimd_qfma_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
74 vt4567 = psimd_qfma_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
|
| D | neon-lut64-p2-x8-acc2.c | 62 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2() local
75 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
81 …const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
103 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
108 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
111 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
|
| D | neonfma-lut64-p2-x8-acc2.c | 61 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vx4567, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2() local
74 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
80 …const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
102 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
107 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
110 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
|
| D | neon-lut64-p2-x8.c | 61 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8() local
74 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
80 …const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
102 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
107 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
110 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
|
| D | neonfma-lut64-p2-x8.c | 60 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vx4567, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8() local
73 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
79 …const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
101 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
106 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
109 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
|
| D | neon-p5-x8-acc2.c | 61 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2() local
66 … const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
70 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
75 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
78 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
|
| D | sse2-p5-x8.c | 56 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vx4567, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8() local
61 const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
65 vn4567 = _mm_sub_ps(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
70 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vx4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
73 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
|
| D | neonfma-p5-x8-acc2.c | 60 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2() local
65 … const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
69 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
74 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
77 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
|
| D | neon-p5-x8.c | 60 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8() local
65 … const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
69 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
74 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
77 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
|
| D | sse2-p5-x8-acc2.c | 57 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vx4567, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2() local
62 const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
66 vn4567 = _mm_sub_ps(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
71 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vx4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
74 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
|
| D | neonfma-p5-x8.c | 59 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8() local
64 … const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
68 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
73 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
76 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
|
| D | psimd-p5-x12.c | 58 psimd_f32 vn4567 = psimd_qfma_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12() local
64 const psimd_f32 vs4567 = (psimd_f32) ((psimd_u32) vn4567 << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
69 vn4567 = psimd_sub_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
75 psimd_f32 vt4567 = psimd_qfma_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
79 vt4567 = psimd_qfma_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
|
| D | psimd-p5-x12-acc2.c | 59 psimd_f32 vn4567 = psimd_qfma_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2() local
65 const psimd_f32 vs4567 = (psimd_f32) ((psimd_u32) vn4567 << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
70 vn4567 = psimd_sub_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
76 psimd_f32 vt4567 = psimd_qfma_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
80 vt4567 = psimd_qfma_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
|
| D | psimd-p5-x12-acc3.c | 60 psimd_f32 vn4567 = psimd_qfma_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3() local
66 const psimd_f32 vs4567 = (psimd_f32) ((psimd_u32) vn4567 << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
71 vn4567 = psimd_sub_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
77 psimd_f32 vt4567 = psimd_qfma_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
81 vt4567 = psimd_qfma_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
|
| D | sse2-p5-x12.c | 58 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vx4567, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12() local
64 const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
69 vn4567 = _mm_sub_ps(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
75 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vx4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
79 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
|
| D | neon-p5-x12.c | 62 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12() local
68 … const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
73 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
79 float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
83 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
|
| D | neonfma-p5-x12.c | 61 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() local
67 … const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
72 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
78 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
82 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
|
| D | neonfma-p5-x12-acc2.c | 62 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vx4567, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2() local
68 … const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
73 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
79 float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
83 vt4567 = vfmaq_f32(vt4567, vn4567, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
|
| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | neon-frac-p9-p10-nr1recps-x16.c | 49 float32x4_t vn4567 = vld1q_f32(x); x += 4; in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16() local
56 vn4567 = vminq_f32(vn4567, vsigmoid_maxinput); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
57 vn4567 = vmaxq_f32(vn4567, vsigmoid_mininput); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
65 const float32x4_t vn4567_sq = vmulq_f32(vn4567, vn4567); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
91 vnum4567 = vmulq_f32(vn4567, vnum4567); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
|
| D | psimd-p5-div-x8.c | 64 psimd_f32 vn4567 = psimd_qfma_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8() local
69 const psimd_f32 vs4567 = (psimd_f32) ((psimd_u32) vn4567 << 23); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
73 vn4567 = psimd_sub_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
78 psimd_f32 vt4567 = psimd_qfma_f32(vz4567, vn4567, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
81 vt4567 = psimd_qfma_f32(vt4567, vn4567, vln2_lo); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
|
| D | sse41-p5-div-x8.c | 64 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8() local
69 const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23)); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
73 vn4567 = _mm_sub_ps(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
78 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
81 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
|
| D | psimd-p5-div-x12.c | 66 psimd_f32 vn4567 = psimd_qfma_f32(vmagic_bias, vz4567, vminus_log2e); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12() local
72 const psimd_f32 vs4567 = (psimd_f32) ((psimd_u32) vn4567 << 23); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
77 vn4567 = psimd_sub_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
83 psimd_f32 vt4567 = psimd_qfma_f32(vz4567, vn4567, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
87 vt4567 = psimd_qfma_f32(vt4567, vn4567, vln2_lo); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
|
| D | neon-rr2-lut2048-p1-nr2recps-x8.c | 64 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8() local
77 …const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
81 …const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_… in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
105 vn4567 = vsubq_f32(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
110 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_o2048_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
113 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_o2048_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
|
| D | sse2-p5-div-x8.c | 64 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8() local
69 const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23)); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
73 vn4567 = _mm_sub_ps(vn4567, vmagic_bias); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
78 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
81 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
|