Home
last modified time | relevance | path

Searched refs:vpGHIJ (Results 1 – 25 of 85) sorted by relevance

1234

/external/XNNPACK/src/f32-velu/gen/
Dvelu-sse41-rr2-p6-x20.c92 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20() local
98 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
104 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
110 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
116 vpGHIJ = _mm_mul_ps(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
133 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
139 const __m128 veGHIJ = _mm_mul_ps(_mm_add_ps(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
Dvelu-neonfma-rr1-p6-x20.c83 float32x4_t vpGHIJ = vfmaq_f32(vc5, vc6, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20() local
89 vpGHIJ = vfmaq_f32(vc4, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
95 vpGHIJ = vfmaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
101 vpGHIJ = vfmaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
107 vpGHIJ = vmulq_f32(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
124 vpGHIJ = vfmaq_f32(vtGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
130 const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
Dvelu-neon-rr2-p6-x20.c90 float32x4_t vpGHIJ = vmlaq_f32(vc5, vc6, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x20() local
96 vpGHIJ = vmlaq_f32(vc4, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
102 vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
108 vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
114 vpGHIJ = vmulq_f32(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
131 vpGHIJ = vmlaq_f32(vtGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
137 const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
Dvelu-neonfma-rr1-p6-x24.c89 float32x4_t vpGHIJ = vfmaq_f32(vc5, vc6, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24() local
96 vpGHIJ = vfmaq_f32(vc4, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
103 vpGHIJ = vfmaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
110 vpGHIJ = vfmaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
117 vpGHIJ = vmulq_f32(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
137 vpGHIJ = vfmaq_f32(vtGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
144 const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
Dvelu-wasmsimd-arm-rr2-p6-x20.c92 v128_t vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20() local
98 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
104 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
110 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
116 vpGHIJ = wasm_f32x4_mul(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
133 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
139 const v128_t veGHIJ = wasm_f32x4_mul(wasm_f32x4_add(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
Dvelu-neon-rr2-p6-x24.c97 float32x4_t vpGHIJ = vmlaq_f32(vc5, vc6, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x24() local
104 vpGHIJ = vmlaq_f32(vc4, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
111 vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
118 vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
125 vpGHIJ = vmulq_f32(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
145 vpGHIJ = vmlaq_f32(vtGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
152 const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
Dvelu-wasmsimd-x86-rr2-p6-x20.c107 v128_t vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20() local
113 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
119 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
125 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
131 vpGHIJ = wasm_f32x4_mul(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
148 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
154 const v128_t veGHIJ = wasm_f32x4_mul(wasm_f32x4_add(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
Dvelu-sse2-rr2-p6-x20.c92 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20() local
98 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20()
104 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20()
110 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20()
116 vpGHIJ = _mm_mul_ps(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20()
133 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20()
139 const __m128 veGHIJ = _mm_mul_ps(_mm_add_ps(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20()
Dvelu-sse41-rr2-p6-x24.c99 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__sse41_rr2_p6_x24() local
106 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__sse41_rr2_p6_x24()
113 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__sse41_rr2_p6_x24()
120 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__sse41_rr2_p6_x24()
127 vpGHIJ = _mm_mul_ps(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__sse41_rr2_p6_x24()
147 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__sse41_rr2_p6_x24()
154 const __m128 veGHIJ = _mm_mul_ps(_mm_add_ps(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__sse41_rr2_p6_x24()
Dvelu-wasmsimd-arm-rr2-p6-x24.c99 v128_t vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24() local
106 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
113 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
120 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
127 vpGHIJ = wasm_f32x4_mul(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
147 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
154 const v128_t veGHIJ = wasm_f32x4_mul(wasm_f32x4_add(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
Dvelu-sse2-rr2-p6-x24.c99 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24() local
106 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24()
113 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24()
120 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24()
127 vpGHIJ = _mm_mul_ps(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24()
147 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24()
154 const __m128 veGHIJ = _mm_mul_ps(_mm_add_ps(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24()
Dvelu-wasmsimd-x86-rr2-p6-x24.c116 v128_t vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24() local
124 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
131 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
138 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
145 vpGHIJ = wasm_f32x4_mul(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
165 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
172 const v128_t veGHIJ = wasm_f32x4_mul(wasm_f32x4_add(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Dneonfma-rr1-p5-x20-acc5.c85 float32x4_t vpGHIJ = vfmaq_f32(vc4, vc5, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5() local
91 vpGHIJ = vfmaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5()
97 vpGHIJ = vfmaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5()
103 vpGHIJ = vfmaq_f32(vc1, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5()
115 float32x4_t vfGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc5()
Dneonfma-rr1-p5-x20.c81 float32x4_t vpGHIJ = vfmaq_f32(vc4, vc5, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20() local
87 vpGHIJ = vfmaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20()
93 vpGHIJ = vfmaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20()
99 vpGHIJ = vfmaq_f32(vc1, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20()
111 float32x4_t vfGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20()
Dneon-rr2-p5-x20.c88 float32x4_t vpGHIJ = vmlaq_f32(vc4, vc5, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20() local
94 vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20()
100 vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20()
106 vpGHIJ = vmlaq_f32(vc1, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20()
118 float32x4_t vfGHIJ = vmlaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20()
Dneonfma-rr1-p5-x20-acc2.c82 float32x4_t vpGHIJ = vfmaq_f32(vc4, vc5, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2() local
88 vpGHIJ = vfmaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2()
94 vpGHIJ = vfmaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2()
100 vpGHIJ = vfmaq_f32(vc1, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2()
112 float32x4_t vfGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x20_acc2()
Dwasmsimd-rr2-p5-x20-acc2.c99 v128_t vpGHIJ = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x20_acc2() local
105 vpGHIJ = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x20_acc2()
111 vpGHIJ = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x20_acc2()
117 vpGHIJ = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x20_acc2()
133 v128_t vfGHIJ = wasm_f32x4_add(vsGHIJ, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x20_acc2()
Dsse2-rr2-p5-x20-acc2.c99 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc5, vtGHIJ), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20_acc2() local
105 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20_acc2()
111 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20_acc2()
117 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20_acc2()
133 __m128 vfGHIJ = _mm_add_ps(_mm_mul_ps(vtGHIJ, vpGHIJ), vsGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20_acc2()
Dwasmsimd-rr2-p5-x20-acc5.c102 v128_t vpGHIJ = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x20_acc5() local
108 vpGHIJ = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x20_acc5()
114 vpGHIJ = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x20_acc5()
120 vpGHIJ = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x20_acc5()
136 v128_t vfGHIJ = wasm_f32x4_add(vsGHIJ, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x20_acc5()
Dsse2-rr2-p5-x20-acc5.c102 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc5, vtGHIJ), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20_acc5() local
108 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20_acc5()
114 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20_acc5()
120 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20_acc5()
136 __m128 vfGHIJ = _mm_add_ps(_mm_mul_ps(vtGHIJ, vpGHIJ), vsGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20_acc5()
Dneon-rr2-p5-x20-acc5.c92 float32x4_t vpGHIJ = vmlaq_f32(vc4, vc5, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5() local
98 vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5()
104 vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5()
110 vpGHIJ = vmlaq_f32(vc1, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5()
122 float32x4_t vfGHIJ = vmlaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc5()
Dsse2-rr2-p5-x20.c98 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc5, vtGHIJ), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20() local
104 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20()
110 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20()
116 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20()
132 __m128 vfGHIJ = _mm_add_ps(_mm_mul_ps(vtGHIJ, vpGHIJ), vsGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20()
Dneon-rr2-p5-x20-acc2.c89 float32x4_t vpGHIJ = vmlaq_f32(vc4, vc5, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2() local
95 vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2()
101 vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2()
107 vpGHIJ = vmlaq_f32(vc1, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2()
119 float32x4_t vfGHIJ = vmlaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x20_acc2()
/external/XNNPACK/src/f32-vsigmoid/gen/
Dvsigmoid-wasmsimd-rr2-p5-div-x20.c86 v128_t vpGHIJ = wasm_f32x4_add(vc4, wasm_f32x4_mul(vtGHIJ, vc5)); in xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x20() local
92 vpGHIJ = wasm_f32x4_add(vc3, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x20()
98 vpGHIJ = wasm_f32x4_add(vc2, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x20()
104 vpGHIJ = wasm_f32x4_add(vc1, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x20()
116 const v128_t veGHIJ = wasm_f32x4_add(vsGHIJ, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x20()
Dvsigmoid-neonfma-rr1-p5-div-x20.c78 float32x4_t vpGHIJ = vfmaq_f32(vc4, vc5, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20() local
84 vpGHIJ = vfmaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
90 vpGHIJ = vfmaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
96 vpGHIJ = vfmaq_f32(vc1, vpGHIJ, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()
108 const float32x4_t veGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_div_x20()

1234