| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | neon-frac-p9-p10-nr1recps-x16.c | 48 float32x4_t vn0123 = vld1q_f32(x); x += 4; in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16() local
54 vn0123 = vminq_f32(vn0123, vsigmoid_maxinput); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
55 vn0123 = vmaxq_f32(vn0123, vsigmoid_mininput); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
64 const float32x4_t vn0123_sq = vmulq_f32(vn0123, vn0123); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
90 vnum0123 = vmulq_f32(vn0123, vnum0123); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
147 float32x4_t vn0123 = vld1q_f32(x); x += 4; in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16() local
149 vn0123 = vminq_f32(vn0123, vsigmoid_maxinput); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
150 vn0123 = vmaxq_f32(vn0123, vsigmoid_mininput); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
152 const float32x4_t vn0123_sq = vmulq_f32(vn0123, vn0123); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
160 vnum0123 = vmulq_f32(vn0123, vnum0123); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
[all …]
|
| D | psimd-p5-div-x8.c | 63 psimd_f32 vn0123 = psimd_qfma_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8() local
68 const psimd_f32 vs0123 = (psimd_f32) ((psimd_u32) vn0123 << 23); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
72 vn0123 = psimd_sub_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
77 psimd_f32 vt0123 = psimd_qfma_f32(vz0123, vn0123, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
80 vt0123 = psimd_qfma_f32(vt0123, vn0123, vln2_lo); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
|
| D | sse41-p5-div-x8.c | 63 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8() local
68 const __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23)); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
72 vn0123 = _mm_sub_ps(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
77 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
80 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
|
| D | psimd-p5-div-x12.c | 65 psimd_f32 vn0123 = psimd_qfma_f32(vmagic_bias, vz0123, vminus_log2e); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12() local
71 const psimd_f32 vs0123 = (psimd_f32) ((psimd_u32) vn0123 << 23); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
76 vn0123 = psimd_sub_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
82 psimd_f32 vt0123 = psimd_qfma_f32(vz0123, vn0123, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
86 vt0123 = psimd_qfma_f32(vt0123, vn0123, vln2_lo); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
|
| D | neon-rr2-lut2048-p1-nr2recps-x8.c | 63 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e_x2048); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8() local
76 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
80 …const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_… in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
104 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
109 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_o2048_hi); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
112 vt0123 = vmlaq_f32(vt0123, vn0123, vln2_o2048_lo); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
|
| /external/XNNPACK/src/f32-sigmoid/ |
| D | neon-frac-p9-p10-nr1recps.c.in | 109 float32x4_t vn0123 = vld1q_f32(x); x += 4; variable
111 vn0123 = vminq_f32(vn0123, vsigmoid_maxinput);
112 vn0123 = vmaxq_f32(vn0123, vsigmoid_mininput);
114 const float32x4_t vn0123_sq = vmulq_f32(vn0123, vn0123);
122 vnum0123 = vmulq_f32(vn0123, vnum0123);
142 float32x4_t vn0123 = vld1q_f32(x); variable
144 vn0123 = vminq_f32(vn0123, vsigmoid_maxinput);
145 vn0123 = vmaxq_f32(vn0123, vsigmoid_mininput);
147 const float32x4_t vn0123_sq = vmulq_f32(vn0123, vn0123);
155 vnum0123 = vmulq_f32(vn0123, vnum0123);
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | psimd-p5-x4.c | 53 psimd_f32 vn0123 = psimd_qfma_f32(vmagic_bias, vx0123, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4() local
57 const psimd_f32 vs0123 = (psimd_f32) ((psimd_u32) vn0123 << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4()
60 vn0123 = psimd_sub_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4()
64 psimd_f32 vt0123 = psimd_qfma_f32(vx0123, vn0123, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4()
66 vt0123 = psimd_qfma_f32(vt0123, vn0123, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4()
|
| D | sse2-p5-x4.c | 53 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vx0123, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4() local
57 const __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4()
60 vn0123 = _mm_sub_ps(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4()
64 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vx0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4()
66 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4()
|
| D | psimd-p5-x8.c | 55 psimd_f32 vn0123 = psimd_qfma_f32(vmagic_bias, vx0123, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8() local
60 const psimd_f32 vs0123 = (psimd_f32) ((psimd_u32) vn0123 << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
64 vn0123 = psimd_sub_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
69 psimd_f32 vt0123 = psimd_qfma_f32(vx0123, vn0123, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
72 vt0123 = psimd_qfma_f32(vt0123, vn0123, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
|
| D | psimd-p5-x8-acc2.c | 56 psimd_f32 vn0123 = psimd_qfma_f32(vmagic_bias, vx0123, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2() local
61 const psimd_f32 vs0123 = (psimd_f32) ((psimd_u32) vn0123 << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
65 vn0123 = psimd_sub_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
70 psimd_f32 vt0123 = psimd_qfma_f32(vx0123, vn0123, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
73 vt0123 = psimd_qfma_f32(vt0123, vn0123, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
|
| D | neon-lut64-p2-x8-acc2.c | 61 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vx0123, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2() local
74 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
78 …const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
102 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
107 float32x4_t vt0123 = vmlaq_f32(vx0123, vn0123, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
110 vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
|
| D | neonfma-lut64-p2-x8-acc2.c | 60 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vx0123, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2() local
73 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
77 …const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
101 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
106 float32x4_t vt0123 = vfmaq_f32(vx0123, vn0123, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
109 vt0123 = vfmaq_f32(vt0123, vn0123, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
|
| D | neon-lut64-p2-x8.c | 60 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vx0123, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8() local
73 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
77 …const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
101 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
106 float32x4_t vt0123 = vmlaq_f32(vx0123, vn0123, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
109 vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
|
| D | neonfma-lut64-p2-x8.c | 59 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vx0123, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8() local
72 …const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
76 …const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
100 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
105 float32x4_t vt0123 = vfmaq_f32(vx0123, vn0123, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
108 vt0123 = vfmaq_f32(vt0123, vn0123, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
|
| D | neon-p5-x8-acc2.c | 60 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vx0123, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2() local
65 … const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
69 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
74 float32x4_t vt0123 = vmlaq_f32(vx0123, vn0123, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
77 vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
|
| D | sse2-p5-x8.c | 55 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vx0123, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8() local
60 const __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
64 vn0123 = _mm_sub_ps(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
69 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vx0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
72 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
|
| D | neonfma-p5-x8-acc2.c | 59 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vx0123, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2() local
64 … const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
68 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
73 float32x4_t vt0123 = vfmaq_f32(vx0123, vn0123, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
76 vt0123 = vfmaq_f32(vt0123, vn0123, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
|
| D | neon-p5-x8.c | 59 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vx0123, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8() local
64 … const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
68 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
73 float32x4_t vt0123 = vmlaq_f32(vx0123, vn0123, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
76 vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
|
| D | sse2-p5-x8-acc2.c | 56 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vx0123, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2() local
61 const __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
65 vn0123 = _mm_sub_ps(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
70 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vx0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
73 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
|
| D | neonfma-p5-x8.c | 58 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vx0123, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8() local
63 … const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
67 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
72 float32x4_t vt0123 = vfmaq_f32(vx0123, vn0123, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
75 vt0123 = vfmaq_f32(vt0123, vn0123, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
|
| D | psimd-p5-x12.c | 57 psimd_f32 vn0123 = psimd_qfma_f32(vmagic_bias, vx0123, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12() local
63 const psimd_f32 vs0123 = (psimd_f32) ((psimd_u32) vn0123 << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
68 vn0123 = psimd_sub_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
74 psimd_f32 vt0123 = psimd_qfma_f32(vx0123, vn0123, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
78 vt0123 = psimd_qfma_f32(vt0123, vn0123, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
|
| D | psimd-p5-x12-acc2.c | 58 psimd_f32 vn0123 = psimd_qfma_f32(vmagic_bias, vx0123, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2() local
64 const psimd_f32 vs0123 = (psimd_f32) ((psimd_u32) vn0123 << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
69 vn0123 = psimd_sub_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
75 psimd_f32 vt0123 = psimd_qfma_f32(vx0123, vn0123, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
79 vt0123 = psimd_qfma_f32(vt0123, vn0123, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
|
| D | psimd-p5-x12-acc3.c | 59 psimd_f32 vn0123 = psimd_qfma_f32(vmagic_bias, vx0123, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3() local
65 const psimd_f32 vs0123 = (psimd_f32) ((psimd_u32) vn0123 << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
70 vn0123 = psimd_sub_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
76 psimd_f32 vt0123 = psimd_qfma_f32(vx0123, vn0123, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
80 vt0123 = psimd_qfma_f32(vt0123, vn0123, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
|
| D | sse2-p5-x12.c | 57 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vx0123, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12() local
63 const __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
68 vn0123 = _mm_sub_ps(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
74 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vx0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
78 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
|
| D | neon-p5-x12.c | 61 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vx0123, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12() local
67 … const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
72 vn0123 = vsubq_f32(vn0123, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
78 float32x4_t vt0123 = vmlaq_f32(vx0123, vn0123, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
82 vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
|