Home
last modified time | relevance | path

Searched refs:veGHIJ (Results 1 – 25 of 72) sorted by relevance

123

/external/XNNPACK/src/f32-vsigmoid/gen/
Dvsigmoid-wasmsimd-rr2-p5-div-x20.c116 const v128_t veGHIJ = wasm_f32x4_add(vsGHIJ, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x20() local
122 const v128_t vdGHIJ = wasm_f32x4_add(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x20()
128 v128_t vfGHIJ = wasm_f32x4_div(veGHIJ, vdGHIJ); in xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x20()
Dvsigmoid-neonfma-rr1-p5-div-x20.c108 const float32x4_t veGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20() local
114 const float32x4_t vdGHIJ = vaddq_f32(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
120 float32x4_t vfGHIJ = vdivq_f32(veGHIJ, vdGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
Dvsigmoid-sse41-rr2-p5-div-x20.c116 __m128 veGHIJ = _mm_add_ps(_mm_mul_ps(vtGHIJ, vpGHIJ), vsGHIJ); in xnn_f32_vsigmoid_ukernel__sse41_rr2_p5_div_x20() local
122 __m128 vdGHIJ = _mm_add_ps(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__sse41_rr2_p5_div_x20()
128 __m128 vfGHIJ = _mm_div_ps(veGHIJ, vdGHIJ); in xnn_f32_vsigmoid_ukernel__sse41_rr2_p5_div_x20()
Dvsigmoid-neonfma-rr1-p5-div-x24.c119 const float32x4_t veGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24() local
126 const float32x4_t vdGHIJ = vaddq_f32(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24()
133 float32x4_t vfGHIJ = vdivq_f32(veGHIJ, vdGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x24()
Dvsigmoid-neonfma-rr1-p5-nr2fma-x20.c108 const float32x4_t veGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() local
114 const float32x4_t vdGHIJ = vaddq_f32(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
138 float32x4_t vfGHIJ = vmulq_f32(veGHIJ, vrGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
Dvsigmoid-sse2-rr2-p5-div-x20.c116 __m128 veGHIJ = _mm_add_ps(_mm_mul_ps(vtGHIJ, vpGHIJ), vsGHIJ); in xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x20() local
122 __m128 vdGHIJ = _mm_add_ps(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x20()
128 __m128 vfGHIJ = _mm_div_ps(veGHIJ, vdGHIJ); in xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x20()
Dvsigmoid-sse41-rr2-p5-div-x24.c128 __m128 veGHIJ = _mm_add_ps(_mm_mul_ps(vtGHIJ, vpGHIJ), vsGHIJ); in xnn_f32_vsigmoid_ukernel__sse41_rr2_p5_div_x24() local
135 __m128 vdGHIJ = _mm_add_ps(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__sse41_rr2_p5_div_x24()
142 __m128 vfGHIJ = _mm_div_ps(veGHIJ, vdGHIJ); in xnn_f32_vsigmoid_ukernel__sse41_rr2_p5_div_x24()
Dvsigmoid-neonfma-rr1-p5-nr1recps1fma-x20.c108 const float32x4_t veGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() local
114 const float32x4_t vdGHIJ = vaddq_f32(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
138 float32x4_t vfGHIJ = vmulq_f32(veGHIJ, vrGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
Dvsigmoid-neon-rr2-p5-nr2recps-x20.c115 const float32x4_t veGHIJ = vmlaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x20() local
121 const float32x4_t vdGHIJ = vaddq_f32(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
145 float32x4_t vfGHIJ = vmulq_f32(veGHIJ, vrGHIJ); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
Dvsigmoid-neonfma-rr1-p5-nr2recps-x20.c108 const float32x4_t veGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() local
114 const float32x4_t vdGHIJ = vaddq_f32(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
138 float32x4_t vfGHIJ = vmulq_f32(veGHIJ, vrGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
Dvsigmoid-wasmsimd-rr2-p5-div-x24.c128 const v128_t veGHIJ = wasm_f32x4_add(vsGHIJ, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x24() local
135 const v128_t vdGHIJ = wasm_f32x4_add(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x24()
142 v128_t vfGHIJ = wasm_f32x4_div(veGHIJ, vdGHIJ); in xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x24()
Dvsigmoid-neonfma-rr1-p5-nr2recps-x24.c119 const float32x4_t veGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() local
126 const float32x4_t vdGHIJ = vaddq_f32(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
154 float32x4_t vfGHIJ = vmulq_f32(veGHIJ, vrGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
Dvsigmoid-neon-rr2-p5-nr2recps-x24.c127 const float32x4_t veGHIJ = vmlaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x24() local
134 const float32x4_t vdGHIJ = vaddq_f32(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
162 float32x4_t vfGHIJ = vmulq_f32(veGHIJ, vrGHIJ); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
Dvsigmoid-neonfma-rr1-p5-nr1recps1fma-x24.c119 const float32x4_t veGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() local
126 const float32x4_t vdGHIJ = vaddq_f32(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
154 float32x4_t vfGHIJ = vmulq_f32(veGHIJ, vrGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
Dvsigmoid-sse2-rr2-p5-div-x24.c128 __m128 veGHIJ = _mm_add_ps(_mm_mul_ps(vtGHIJ, vpGHIJ), vsGHIJ); in xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x24() local
135 __m128 vdGHIJ = _mm_add_ps(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x24()
142 __m128 vfGHIJ = _mm_div_ps(veGHIJ, vdGHIJ); in xnn_f32_vsigmoid_ukernel__sse2_rr2_p5_div_x24()
Dvsigmoid-neonfma-rr1-p5-nr2fma-x24.c119 const float32x4_t veGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() local
126 const float32x4_t vdGHIJ = vaddq_f32(veGHIJ, vone); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
154 float32x4_t vfGHIJ = vmulq_f32(veGHIJ, vrGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
/external/XNNPACK/src/f32-velu/gen/
Dvelu-sse41-rr2-p6-x20.c139 const __m128 veGHIJ = _mm_mul_ps(_mm_add_ps(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20() local
151 const __m128 vyGHIJ = _mm_blendv_ps(vxGHIJ, veGHIJ, vxGHIJ); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
Dvelu-neonfma-rr1-p6-x20.c130 const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20() local
147 const float32x4_t vyGHIJ = vbslq_f32(vmGHIJ, veGHIJ, vxGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
Dvelu-neon-rr2-p6-x20.c137 const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__neon_rr2_p6_x20() local
154 const float32x4_t vyGHIJ = vbslq_f32(vmGHIJ, veGHIJ, vxGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
Dvelu-neonfma-rr1-p6-x24.c144 const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24() local
164 const float32x4_t vyGHIJ = vbslq_f32(vmGHIJ, veGHIJ, vxGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
Dvelu-wasmsimd-arm-rr2-p6-x20.c139 const v128_t veGHIJ = wasm_f32x4_mul(wasm_f32x4_add(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20() local
156 const v128_t vyGHIJ = wasm_v128_bitselect(veGHIJ, vxGHIJ, vsignmGHIJ); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
Dvelu-neon-rr2-p6-x24.c152 const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__neon_rr2_p6_x24() local
172 const float32x4_t vyGHIJ = vbslq_f32(vmGHIJ, veGHIJ, vxGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
Dvelu-wasmsimd-x86-rr2-p6-x20.c154 const v128_t veGHIJ = wasm_f32x4_mul(wasm_f32x4_add(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20() local
171 const v128_t vyGHIJ = wasm_v128_bitselect(veGHIJ, vxGHIJ, vsignmGHIJ); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
Dvelu-sse2-rr2-p6-x20.c139 const __m128 veGHIJ = _mm_mul_ps(_mm_add_ps(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20() local
156 const __m128 vyGHIJ = _mm_or_ps(_mm_and_ps(veGHIJ, vmGHIJ), _mm_andnot_ps(vmGHIJ, vxGHIJ)); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20()
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Dneon-rr2-lut64-p2-x20.c63 …const int32x4_t veGHIJ = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vnGHIJ), vmovq_n_s32(INT32_C(… in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_lut64_p2_x20() local
112 …const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlGHIJ), veGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_lut64_p2_x20()

123