| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | neon-frac-p9-p10-nr1recps-x16.c | 50 float32x4_t vn89AB = vld1q_f32(x); x += 4; in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16() local
58 vn89AB = vminq_f32(vn89AB, vsigmoid_maxinput); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
59 vn89AB = vmaxq_f32(vn89AB, vsigmoid_mininput); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
66 const float32x4_t vn89AB_sq = vmulq_f32(vn89AB, vn89AB); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
92 vnum89AB = vmulq_f32(vn89AB, vnum89AB); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
|
| D | psimd-p5-div-x12.c | 67 psimd_f32 vn89AB = psimd_qfma_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12() local
73 const psimd_f32 vs89AB = (psimd_f32) ((psimd_u32) vn89AB << 23); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
78 vn89AB = psimd_sub_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
84 psimd_f32 vt89AB = psimd_qfma_f32(vz89AB, vn89AB, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
88 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vln2_lo); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
|
| D | psimd-p5-div-x16.c | 69 psimd_f32 vn89AB = psimd_qfma_f32(vmagic_bias, vz89AB, vminus_log2e); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16() local
76 const psimd_f32 vs89AB = (psimd_f32) ((psimd_u32) vn89AB << 23); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
82 vn89AB = psimd_sub_f32(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
89 psimd_f32 vt89AB = psimd_qfma_f32(vz89AB, vn89AB, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
94 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vln2_lo); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
|
| D | sse41-p5-div-x12.c | 67 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12() local
73 const __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23)); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
78 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
84 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
88 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x12()
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | psimd-p5-x12.c | 59 psimd_f32 vn89AB = psimd_qfma_f32(vmagic_bias, vx89AB, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12() local
65 const psimd_f32 vs89AB = (psimd_f32) ((psimd_u32) vn89AB << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
70 vn89AB = psimd_sub_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
76 psimd_f32 vt89AB = psimd_qfma_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
80 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
|
| D | psimd-p5-x12-acc2.c | 60 psimd_f32 vn89AB = psimd_qfma_f32(vmagic_bias, vx89AB, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2() local
66 const psimd_f32 vs89AB = (psimd_f32) ((psimd_u32) vn89AB << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
71 vn89AB = psimd_sub_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
77 psimd_f32 vt89AB = psimd_qfma_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
81 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
|
| D | psimd-p5-x12-acc3.c | 61 psimd_f32 vn89AB = psimd_qfma_f32(vmagic_bias, vx89AB, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3() local
67 const psimd_f32 vs89AB = (psimd_f32) ((psimd_u32) vn89AB << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
72 vn89AB = psimd_sub_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
78 psimd_f32 vt89AB = psimd_qfma_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
82 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
|
| D | sse2-p5-x12.c | 59 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vx89AB, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12() local
65 const __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
70 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
76 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vx89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
80 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
|
| D | neon-p5-x12.c | 63 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vx89AB, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12() local
69 … const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
74 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
80 float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
84 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
|
| D | neonfma-p5-x12.c | 62 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vx89AB, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() local
68 … const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
73 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
79 float32x4_t vt89AB = vfmaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
83 vt89AB = vfmaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
|
| D | neonfma-p5-x12-acc2.c | 63 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vx89AB, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2() local
69 … const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
74 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
80 float32x4_t vt89AB = vfmaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
84 vt89AB = vfmaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
|
| D | neon-p5-x12-acc2.c | 64 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vx89AB, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2() local
70 … const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
75 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
81 float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
85 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
|
| D | neonfma-lut64-p2-x12-acc3.c | 65 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vx89AB, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc3() local
79 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc3()
88 …const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc3()
117 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc3()
123 float32x4_t vt89AB = vfmaq_f32(vx89AB, vn89AB, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc3()
127 vt89AB = vfmaq_f32(vt89AB, vn89AB, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc3()
|
| D | neon-lut64-p2-x12-acc3.c | 66 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vx89AB, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc3() local
80 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc3()
89 …const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc3()
118 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc3()
124 float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc3()
128 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc3()
|
| D | neon-lut64-p2-x12-acc2.c | 65 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vx89AB, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc2() local
79 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc2()
88 …const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc2()
117 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc2()
123 float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc2()
127 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc2()
|
| D | neon-lut64-p2-x12.c | 64 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vx89AB, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12() local
78 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12()
87 …const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12()
116 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12()
122 float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12()
126 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12()
|
| D | neonfma-lut64-p2-x12-acc2.c | 64 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vx89AB, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc2() local
78 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc2()
87 …const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc2()
116 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc2()
122 float32x4_t vt89AB = vfmaq_f32(vx89AB, vn89AB, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc2()
126 vt89AB = vfmaq_f32(vt89AB, vn89AB, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc2()
|
| D | neonfma-lut64-p2-x12.c | 63 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vx89AB, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12() local
77 …const int32x4_t ve89AB = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn89AB), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12()
86 …const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12()
115 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12()
121 float32x4_t vt89AB = vfmaq_f32(vx89AB, vn89AB, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12()
125 vt89AB = vfmaq_f32(vt89AB, vn89AB, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12()
|
| D | psimd-p5-x16-acc2.c | 62 psimd_f32 vn89AB = psimd_qfma_f32(vmagic_bias, vx89AB, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2() local
69 const psimd_f32 vs89AB = (psimd_f32) ((psimd_u32) vn89AB << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
75 vn89AB = psimd_sub_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
82 psimd_f32 vt89AB = psimd_qfma_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
87 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
|
| D | sse2-p5-x12-acc3.c | 61 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vx89AB, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3() local
67 const __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
72 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
78 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vx89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
82 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
|
| D | psimd-p5-x16.c | 61 psimd_f32 vn89AB = psimd_qfma_f32(vmagic_bias, vx89AB, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16() local
68 const psimd_f32 vs89AB = (psimd_f32) ((psimd_u32) vn89AB << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
74 vn89AB = psimd_sub_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
81 psimd_f32 vt89AB = psimd_qfma_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
86 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
|
| D | psimd-p5-x16-acc4.c | 64 psimd_f32 vn89AB = psimd_qfma_f32(vmagic_bias, vx89AB, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4() local
71 const psimd_f32 vs89AB = (psimd_f32) ((psimd_u32) vn89AB << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
77 vn89AB = psimd_sub_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
84 psimd_f32 vt89AB = psimd_qfma_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
89 vt89AB = psimd_qfma_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
|
| D | neonfma-p5-x12-acc3.c | 64 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vx89AB, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3() local
70 … const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
75 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
81 float32x4_t vt89AB = vfmaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
85 vt89AB = vfmaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
|
| D | neon-p5-x12-acc3.c | 65 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vx89AB, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3() local
71 … const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
76 vn89AB = vsubq_f32(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
82 float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
86 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
|
| D | sse2-p5-x12-acc2.c | 60 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vx89AB, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2() local
66 const __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2()
71 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2()
77 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vx89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2()
81 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2()
|