| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | neon-frac-p9-p10-nr1recps-x16.c | 51 float32x4_t vnCDEF = vld1q_f32(x); x += 4; in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16() local
60 vnCDEF = vminq_f32(vnCDEF, vsigmoid_maxinput); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
61 vnCDEF = vmaxq_f32(vnCDEF, vsigmoid_mininput); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
67 const float32x4_t vnCDEF_sq = vmulq_f32(vnCDEF, vnCDEF); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
93 vnumCDEF = vmulq_f32(vnCDEF, vnumCDEF); in xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16()
|
| D | psimd-p5-div-x16.c | 70 psimd_f32 vnCDEF = psimd_qfma_f32(vmagic_bias, vzCDEF, vminus_log2e); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16() local
77 const psimd_f32 vsCDEF = (psimd_f32) ((psimd_u32) vnCDEF << 23); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
83 vnCDEF = psimd_sub_f32(vnCDEF, vmagic_bias); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
90 psimd_f32 vtCDEF = psimd_qfma_f32(vzCDEF, vnCDEF, vln2_hi); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
95 vtCDEF = psimd_qfma_f32(vtCDEF, vnCDEF, vln2_lo); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x16()
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | psimd-p5-x16-acc2.c | 63 psimd_f32 vnCDEF = psimd_qfma_f32(vmagic_bias, vxCDEF, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2() local
70 const psimd_f32 vsCDEF = (psimd_f32) ((psimd_u32) vnCDEF << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
76 vnCDEF = psimd_sub_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
83 psimd_f32 vtCDEF = psimd_qfma_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
88 vtCDEF = psimd_qfma_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
|
| D | psimd-p5-x16.c | 62 psimd_f32 vnCDEF = psimd_qfma_f32(vmagic_bias, vxCDEF, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16() local
69 const psimd_f32 vsCDEF = (psimd_f32) ((psimd_u32) vnCDEF << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
75 vnCDEF = psimd_sub_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
82 psimd_f32 vtCDEF = psimd_qfma_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
87 vtCDEF = psimd_qfma_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
|
| D | psimd-p5-x16-acc4.c | 65 psimd_f32 vnCDEF = psimd_qfma_f32(vmagic_bias, vxCDEF, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4() local
72 const psimd_f32 vsCDEF = (psimd_f32) ((psimd_u32) vnCDEF << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
78 vnCDEF = psimd_sub_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
85 psimd_f32 vtCDEF = psimd_qfma_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
90 vtCDEF = psimd_qfma_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
|
| D | neonfma-p5-x16-acc4.c | 68 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vxCDEF, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4() local
75 … const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
81 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
88 float32x4_t vtCDEF = vfmaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
93 vtCDEF = vfmaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
|
| D | psimd-p5-x20-acc2.c | 65 psimd_f32 vnCDEF = psimd_qfma_f32(vmagic_bias, vxCDEF, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2() local
73 const psimd_f32 vsCDEF = (psimd_f32) ((psimd_u32) vnCDEF << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
80 vnCDEF = psimd_sub_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
88 psimd_f32 vtCDEF = psimd_qfma_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
94 vtCDEF = psimd_qfma_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
|
| D | neon-p5-x16-acc2.c | 67 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vxCDEF, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2() local
74 … const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
80 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
87 float32x4_t vtCDEF = vmlaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
92 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
|
| D | sse2-p5-x16.c | 62 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vxCDEF, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16() local
69 const __m128 vsCDEF = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnCDEF), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
75 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
82 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vxCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
87 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
|
| D | neon-p5-x16.c | 66 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vxCDEF, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16() local
73 … const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
79 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
86 float32x4_t vtCDEF = vmlaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
91 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
|
| D | neonfma-p5-x16.c | 65 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vxCDEF, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16() local
72 … const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
78 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
85 float32x4_t vtCDEF = vfmaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
90 vtCDEF = vfmaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
|
| D | sse2-p5-x16-acc4.c | 65 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vxCDEF, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4() local
72 const __m128 vsCDEF = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnCDEF), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4()
78 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4()
85 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vxCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4()
90 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4()
|
| D | neonfma-p5-x16-acc2.c | 66 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vxCDEF, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2() local
73 … const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
79 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
86 float32x4_t vtCDEF = vfmaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
91 vtCDEF = vfmaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
|
| D | sse2-p5-x16-acc2.c | 63 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vxCDEF, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2() local
70 const __m128 vsCDEF = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnCDEF), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2()
76 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2()
83 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vxCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2()
88 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2()
|
| D | neon-p5-x16-acc4.c | 69 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vxCDEF, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4() local
76 … const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
82 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
89 float32x4_t vtCDEF = vmlaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
94 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
|
| D | psimd-p5-x20.c | 64 psimd_f32 vnCDEF = psimd_qfma_f32(vmagic_bias, vxCDEF, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20() local
72 const psimd_f32 vsCDEF = (psimd_f32) ((psimd_u32) vnCDEF << 23); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20()
79 vnCDEF = psimd_sub_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20()
87 psimd_f32 vtCDEF = psimd_qfma_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20()
93 vtCDEF = psimd_qfma_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20()
|
| D | neonfma-lut64-p2-x16-acc4.c | 69 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vxCDEF, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4() local
84 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
96 …const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
132 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
139 float32x4_t vtCDEF = vfmaq_f32(vxCDEF, vnCDEF, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
144 vtCDEF = vfmaq_f32(vtCDEF, vnCDEF, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4()
|
| D | neon-lut64-p2-x16-acc2.c | 68 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vxCDEF, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2() local
83 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
95 …const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
131 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
138 float32x4_t vtCDEF = vmlaq_f32(vxCDEF, vnCDEF, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
143 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2()
|
| D | neon-lut64-p2-x16.c | 67 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vxCDEF, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16() local
82 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
94 …const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
130 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
137 float32x4_t vtCDEF = vmlaq_f32(vxCDEF, vnCDEF, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
142 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16()
|
| D | neonfma-lut64-p2-x16-acc2.c | 67 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vxCDEF, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2() local
82 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
94 …const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
130 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
137 float32x4_t vtCDEF = vfmaq_f32(vxCDEF, vnCDEF, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
142 vtCDEF = vfmaq_f32(vtCDEF, vnCDEF, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2()
|
| D | neon-lut64-p2-x16-acc4.c | 70 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vxCDEF, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4() local
85 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
97 …const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
133 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
140 float32x4_t vtCDEF = vmlaq_f32(vxCDEF, vnCDEF, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
145 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4()
|
| D | neonfma-lut64-p2-x16.c | 66 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vxCDEF, vlog2e_x64); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16() local
81 …const int32x4_t veCDEF = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnCDEF), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
93 …const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
129 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
136 float32x4_t vtCDEF = vfmaq_f32(vxCDEF, vnCDEF, vminus_ln2_o64_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
141 vtCDEF = vfmaq_f32(vtCDEF, vnCDEF, vminus_ln2_o64_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16()
|
| D | neonfma-p5-x20-acc2.c | 68 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vxCDEF, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2() local
76 … const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
83 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
91 float32x4_t vtCDEF = vfmaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
97 vtCDEF = vfmaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
|
| D | sse2-p5-x20.c | 64 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vxCDEF, vlog2e), vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20() local
72 const __m128 vsCDEF = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnCDEF), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20()
79 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20()
87 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vxCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20()
93 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20()
|
| D | neonfma-p5-x20-acc5.c | 71 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vxCDEF, vlog2e); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5() local
79 … const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
86 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
94 float32x4_t vtCDEF = vfmaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
100 vtCDEF = vfmaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
|