/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
D | psimd-p5-x4.c | 69 psimd_f32 vp0123 = psimd_qfma_f32(vc4, vc5, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4() local 71 vp0123 = psimd_qfma_f32(vc3, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4() 73 vp0123 = psimd_qfma_f32(vc2, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4() 75 vp0123 = psimd_qfma_f32(vc1, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4() 83 psimd_f32 vf0123 = psimd_qfma_f32(vs0123, vt0123, vp0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x4()
|
D | sse2-p5-x4.c | 69 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc5, vt0123), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4() local 71 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4() 73 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4() 75 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4() 83 __m128 vf0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4()
|
D | psimd-p5-x8.c | 76 psimd_f32 vp0123 = psimd_qfma_f32(vc4, vc5, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8() local 79 vp0123 = psimd_qfma_f32(vc3, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8() 82 vp0123 = psimd_qfma_f32(vc2, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8() 85 vp0123 = psimd_qfma_f32(vc1, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8() 95 psimd_f32 vf0123 = psimd_qfma_f32(vs0123, vt0123, vp0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
|
D | psimd-p5-x8-acc2.c | 77 psimd_f32 vp0123 = psimd_qfma_f32(vc4, vc5, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2() local 80 vp0123 = psimd_qfma_f32(vc3, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2() 83 vp0123 = psimd_qfma_f32(vc2, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2() 86 vp0123 = psimd_qfma_f32(vc1, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2() 96 psimd_f32 vf0123 = psimd_qfma_f32(vs0123, vt0123, vp0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
|
D | neon-p5-x8-acc2.c | 81 float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2() local 84 vp0123 = vmlaq_f32(vc3, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2() 87 vp0123 = vmlaq_f32(vc2, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2() 90 vp0123 = vmlaq_f32(vc1, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2() 100 float32x4_t vf0123 = vmlaq_f32(vs0123, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
|
D | sse2-p5-x8.c | 76 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc5, vt0123), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8() local 79 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8() 82 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8() 85 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8() 95 __m128 vf0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
|
D | neonfma-p5-x8-acc2.c | 80 float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2() local 83 vp0123 = vfmaq_f32(vc3, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2() 86 vp0123 = vfmaq_f32(vc2, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2() 89 vp0123 = vfmaq_f32(vc1, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2() 99 float32x4_t vf0123 = vfmaq_f32(vs0123, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
|
D | neon-p5-x8.c | 80 float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8() local 83 vp0123 = vmlaq_f32(vc3, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8() 86 vp0123 = vmlaq_f32(vc2, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8() 89 vp0123 = vmlaq_f32(vc1, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8() 99 float32x4_t vf0123 = vmlaq_f32(vs0123, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
|
D | sse2-p5-x8-acc2.c | 77 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc5, vt0123), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2() local 80 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2() 83 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2() 86 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2() 96 __m128 vf0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
|
D | neonfma-p5-x8.c | 79 float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8() local 82 vp0123 = vfmaq_f32(vc3, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8() 85 vp0123 = vfmaq_f32(vc2, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8() 88 vp0123 = vfmaq_f32(vc1, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8() 98 float32x4_t vf0123 = vfmaq_f32(vs0123, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
|
D | psimd-p5-x12.c | 83 psimd_f32 vp0123 = psimd_qfma_f32(vc4, vc5, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12() local 87 vp0123 = psimd_qfma_f32(vc3, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12() 91 vp0123 = psimd_qfma_f32(vc2, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12() 95 vp0123 = psimd_qfma_f32(vc1, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12() 107 psimd_f32 vf0123 = psimd_qfma_f32(vs0123, vt0123, vp0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
|
D | psimd-p5-x12-acc2.c | 84 psimd_f32 vp0123 = psimd_qfma_f32(vc4, vc5, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2() local 88 vp0123 = psimd_qfma_f32(vc3, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2() 92 vp0123 = psimd_qfma_f32(vc2, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2() 96 vp0123 = psimd_qfma_f32(vc1, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2() 108 psimd_f32 vf0123 = psimd_qfma_f32(vs0123, vt0123, vp0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
|
D | psimd-p5-x12-acc3.c | 85 psimd_f32 vp0123 = psimd_qfma_f32(vc4, vc5, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3() local 89 vp0123 = psimd_qfma_f32(vc3, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3() 93 vp0123 = psimd_qfma_f32(vc2, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3() 97 vp0123 = psimd_qfma_f32(vc1, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3() 109 psimd_f32 vf0123 = psimd_qfma_f32(vs0123, vt0123, vp0123); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
|
D | sse2-p5-x12.c | 83 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc5, vt0123), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12() local 87 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12() 91 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12() 95 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12() 107 __m128 vf0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
|
D | neon-p5-x12.c | 87 float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12() local 91 vp0123 = vmlaq_f32(vc3, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12() 95 vp0123 = vmlaq_f32(vc2, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12() 99 vp0123 = vmlaq_f32(vc1, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12() 111 float32x4_t vf0123 = vmlaq_f32(vs0123, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
|
D | neonfma-p5-x12.c | 86 float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() local 90 vp0123 = vfmaq_f32(vc3, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() 94 vp0123 = vfmaq_f32(vc2, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() 98 vp0123 = vfmaq_f32(vc1, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() 110 float32x4_t vf0123 = vfmaq_f32(vs0123, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
|
D | neonfma-p5-x12-acc2.c | 87 float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2() local 91 vp0123 = vfmaq_f32(vc3, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2() 95 vp0123 = vfmaq_f32(vc2, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2() 99 vp0123 = vfmaq_f32(vc1, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2() 111 float32x4_t vf0123 = vfmaq_f32(vs0123, vp0123, vt0123); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
|
/external/XNNPACK/src/f32-sigmoid/gen/ |
D | psimd-p5-div-x8.c | 85 psimd_f32 vp0123 = psimd_qfma_f32(vc4, vt0123, vc5); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8() local 88 vp0123 = psimd_qfma_f32(vc3, vt0123, vp0123); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8() 91 vp0123 = psimd_qfma_f32(vc2, vt0123, vp0123); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8() 94 vp0123 = psimd_qfma_f32(vc1, vt0123, vp0123); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8() 104 const psimd_f32 ve0123 = psimd_qfma_f32(vs0123, vt0123, vp0123); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x8()
|
D | neonfma-rr1-p5-div-x8.c | 77 float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8() local 80 vp0123 = vfmaq_f32(vc3, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8() 83 vp0123 = vfmaq_f32(vc2, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8() 86 vp0123 = vfmaq_f32(vc1, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8() 96 float32x4_t ve0123 = vfmaq_f32(vs0123, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8()
|
D | sse41-p5-div-x8.c | 84 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc5, vt0123), vc4); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8() local 87 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8() 90 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8() 93 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc1); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8() 103 __m128 ve0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123); in xnn_f32_sigmoid_ukernel__sse41_p5_div_x8()
|
D | psimd-p5-div-x12.c | 92 psimd_f32 vp0123 = psimd_qfma_f32(vc4, vt0123, vc5); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12() local 96 vp0123 = psimd_qfma_f32(vc3, vt0123, vp0123); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12() 100 vp0123 = psimd_qfma_f32(vc2, vt0123, vp0123); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12() 104 vp0123 = psimd_qfma_f32(vc1, vt0123, vp0123); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12() 116 const psimd_f32 ve0123 = psimd_qfma_f32(vs0123, vt0123, vp0123); in xnn_f32_sigmoid_ukernel__psimd_p5_div_x12()
|
D | neonfma-rr1-p5-nr2recps-x8.c | 77 float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8() local 80 vp0123 = vfmaq_f32(vc3, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8() 83 vp0123 = vfmaq_f32(vc2, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8() 86 vp0123 = vfmaq_f32(vc1, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8() 96 float32x4_t ve0123 = vfmaq_f32(vs0123, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8()
|
D | sse2-p5-div-x8.c | 84 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc5, vt0123), vc4); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8() local 87 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8() 90 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8() 93 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc1); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8() 103 __m128 ve0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x8()
|
D | neonfma-rr1-p5-nr1recps1fma-x8.c | 77 float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8() local 80 vp0123 = vfmaq_f32(vc3, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8() 83 vp0123 = vfmaq_f32(vc2, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8() 86 vp0123 = vfmaq_f32(vc1, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8() 96 float32x4_t ve0123 = vfmaq_f32(vs0123, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8()
|
D | neonfma-rr1-p5-div-x12.c | 83 float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12() local 87 vp0123 = vfmaq_f32(vc3, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12() 91 vp0123 = vfmaq_f32(vc2, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12() 95 vp0123 = vfmaq_f32(vc1, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12() 107 float32x4_t ve0123 = vfmaq_f32(vs0123, vp0123, vt0123); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12()
|