Home
last modified time | relevance | path

Searched refs:vpGHIJ (Results 1 – 25 of 84) sorted by relevance

1234

/external/XNNPACK/src/f32-velu/gen/
Dvelu-neonfma-rr1-p6-x20.c84 float32x4_t vpGHIJ = vfmaq_f32(vc5, vc6, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20() local
90 vpGHIJ = vfmaq_f32(vc4, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
96 vpGHIJ = vfmaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
102 vpGHIJ = vfmaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
108 vpGHIJ = vmulq_f32(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
125 vpGHIJ = vfmaq_f32(vtGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
131 const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
Dvelu-wasmsimd-arm-rr2-p6-x20.c93 v128_t vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20() local
99 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
105 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
111 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
117 vpGHIJ = wasm_f32x4_mul(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
134 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
140 const v128_t veGHIJ = wasm_f32x4_mul(wasm_f32x4_add(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20()
Dvelu-sse41-rr2-p6-x20.c93 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20() local
99 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
105 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
111 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
117 vpGHIJ = _mm_mul_ps(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
134 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
140 const __m128 veGHIJ = _mm_mul_ps(_mm_add_ps(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__sse41_rr2_p6_x20()
Dvelu-neon-rr2-p6-x20.c91 float32x4_t vpGHIJ = vmlaq_f32(vc5, vc6, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x20() local
97 vpGHIJ = vmlaq_f32(vc4, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
103 vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
109 vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
115 vpGHIJ = vmulq_f32(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
132 vpGHIJ = vmlaq_f32(vtGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
138 const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
Dvelu-neon-rr2-p6-x24.c98 float32x4_t vpGHIJ = vmlaq_f32(vc5, vc6, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x24() local
105 vpGHIJ = vmlaq_f32(vc4, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
112 vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
119 vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
126 vpGHIJ = vmulq_f32(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
146 vpGHIJ = vmlaq_f32(vtGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
153 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.c108 v128_t vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20() local
114 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
120 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
126 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
132 vpGHIJ = wasm_f32x4_mul(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
149 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
155 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.c93 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20() local
99 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20()
105 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20()
111 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20()
117 vpGHIJ = _mm_mul_ps(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20()
134 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20()
140 const __m128 veGHIJ = _mm_mul_ps(_mm_add_ps(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__sse2_rr2_p6_x20()
Dvelu-neonfma-rr1-p6-x24.c90 float32x4_t vpGHIJ = vfmaq_f32(vc5, vc6, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24() local
97 vpGHIJ = vfmaq_f32(vc4, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
104 vpGHIJ = vfmaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
111 vpGHIJ = vfmaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
118 vpGHIJ = vmulq_f32(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
138 vpGHIJ = vfmaq_f32(vtGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
145 const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
Dvelu-sse41-rr2-p6-x24.c100 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__sse41_rr2_p6_x24() local
107 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__sse41_rr2_p6_x24()
114 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__sse41_rr2_p6_x24()
121 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__sse41_rr2_p6_x24()
128 vpGHIJ = _mm_mul_ps(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__sse41_rr2_p6_x24()
148 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__sse41_rr2_p6_x24()
155 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.c100 v128_t vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24() local
107 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
114 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
121 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
128 vpGHIJ = wasm_f32x4_mul(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
148 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
155 const v128_t veGHIJ = wasm_f32x4_mul(wasm_f32x4_add(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x24()
Dvelu-wasmsimd-x86-rr2-p6-x24.c117 v128_t vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24() local
125 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
132 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
139 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
146 vpGHIJ = wasm_f32x4_mul(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
166 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
173 const v128_t veGHIJ = wasm_f32x4_mul(wasm_f32x4_add(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
Dvelu-sse2-rr2-p6-x24.c100 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc6, vtGHIJ), vc5); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24() local
107 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc4); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24()
114 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24()
121 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24()
128 vpGHIJ = _mm_mul_ps(vpGHIJ, vtGHIJ); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24()
148 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vtGHIJ); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24()
155 const __m128 veGHIJ = _mm_mul_ps(_mm_add_ps(vpGHIJ, vsGHIJ), valpha); in xnn_f32_velu_ukernel__sse2_rr2_p6_x24()
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Dneonfma-p5-x20.c104 float32x4_t vpGHIJ = vfmaq_f32(vc4, vc5, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20() local
110 vpGHIJ = vfmaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
116 vpGHIJ = vfmaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
122 vpGHIJ = vfmaq_f32(vc1, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
138 float32x4_t vfGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
Dwasmsimd-p5-x20-acc5.c105 v128_t vpGHIJ = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20_acc5() local
111 vpGHIJ = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20_acc5()
117 vpGHIJ = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20_acc5()
123 vpGHIJ = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20_acc5()
139 v128_t vfGHIJ = wasm_f32x4_add(vsGHIJ, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20_acc5()
Dneon-p5-x20-acc5.c109 float32x4_t vpGHIJ = vmlaq_f32(vc4, vc5, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5() local
115 vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
121 vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
127 vpGHIJ = vmlaq_f32(vc1, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
143 float32x4_t vfGHIJ = vmlaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
Dneonfma-p5-x20-acc5.c108 float32x4_t vpGHIJ = vfmaq_f32(vc4, vc5, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5() local
114 vpGHIJ = vfmaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
120 vpGHIJ = vfmaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
126 vpGHIJ = vfmaq_f32(vc1, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
142 float32x4_t vfGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
Dsse2-p5-x20.c101 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc5, vtGHIJ), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20() local
107 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20()
113 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20()
119 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20()
135 __m128 vfGHIJ = _mm_add_ps(_mm_mul_ps(vtGHIJ, vpGHIJ), vsGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20()
Dneon-p5-x20.c105 float32x4_t vpGHIJ = vmlaq_f32(vc4, vc5, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20() local
111 vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
117 vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
123 vpGHIJ = vmlaq_f32(vc1, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
139 float32x4_t vfGHIJ = vmlaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
Dneon-p5-x20-acc2.c106 float32x4_t vpGHIJ = vmlaq_f32(vc4, vc5, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2() local
112 vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
118 vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
124 vpGHIJ = vmlaq_f32(vc1, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
140 float32x4_t vfGHIJ = vmlaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
Dsse2-p5-x20-acc2.c102 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc5, vtGHIJ), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc2() local
108 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc2()
114 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc2()
120 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc2()
136 __m128 vfGHIJ = _mm_add_ps(_mm_mul_ps(vtGHIJ, vpGHIJ), vsGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc2()
Dneonfma-p5-x20-acc2.c105 float32x4_t vpGHIJ = vfmaq_f32(vc4, vc5, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2() local
111 vpGHIJ = vfmaq_f32(vc3, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
117 vpGHIJ = vfmaq_f32(vc2, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
123 vpGHIJ = vfmaq_f32(vc1, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
139 float32x4_t vfGHIJ = vfmaq_f32(vsGHIJ, vpGHIJ, vtGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
Dwasmsimd-p5-x20-acc2.c102 v128_t vpGHIJ = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20_acc2() local
108 vpGHIJ = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20_acc2()
114 vpGHIJ = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20_acc2()
120 vpGHIJ = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20_acc2()
136 v128_t vfGHIJ = wasm_f32x4_add(vsGHIJ, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20_acc2()
Dwasmsimd-p5-x20.c101 v128_t vpGHIJ = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20() local
107 vpGHIJ = wasm_f32x4_add(vc3, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20()
113 vpGHIJ = wasm_f32x4_add(vc2, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20()
119 vpGHIJ = wasm_f32x4_add(vc1, wasm_f32x4_mul(vpGHIJ, vtGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20()
135 v128_t vfGHIJ = wasm_f32x4_add(vsGHIJ, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20()
Dsse2-p5-x20-acc5.c105 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc5, vtGHIJ), vc4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc5() local
111 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc5()
117 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc5()
123 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc1); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc5()
139 __m128 vfGHIJ = _mm_add_ps(_mm_mul_ps(vtGHIJ, vpGHIJ), vsGHIJ); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc5()
/external/XNNPACK/src/f32-sigmoid/gen/
Dwasmsimd-p5-div-x20.c86 v128_t vpGHIJ = wasm_f32x4_add(vc4, wasm_f32x4_mul(vtGHIJ, vc5)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20() local
92 vpGHIJ = wasm_f32x4_add(vc3, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
98 vpGHIJ = wasm_f32x4_add(vc2, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
104 vpGHIJ = wasm_f32x4_add(vc1, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()
116 const v128_t veGHIJ = wasm_f32x4_add(vsGHIJ, wasm_f32x4_mul(vtGHIJ, vpGHIJ)); in xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20()

1234