| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x3-wasmrelaxedsimd-fma.c | 70 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_ukernel_up8x3__wasmrelaxedsimd_fma() local
72 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x3__wasmrelaxedsimd_fma()
106 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_ukernel_up8x3__wasmrelaxedsimd_fma() local
107 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x3__wasmrelaxedsimd_fma()
131 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_ukernel_up8x3__wasmrelaxedsimd_fma() local
132 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x3__wasmrelaxedsimd_fma()
|
| D | up8x3-wasmsimd.c | 70 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_ukernel_up8x3__wasmsimd() local
72 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_ukernel_up8x3__wasmsimd()
106 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_ukernel_up8x3__wasmsimd() local
107 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_ukernel_up8x3__wasmsimd()
131 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_ukernel_up8x3__wasmsimd() local
132 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_ukernel_up8x3__wasmsimd()
|
| D | up8x3-minmax-wasmrelaxedsimd-fma.c | 72 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma() local
74 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma()
111 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma() local
112 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma()
137 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma() local
138 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma()
|
| D | up8x3-minmax-wasmsimd-x86-acc2.c | 72 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86_acc2() local
74 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86_acc2()
114 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86_acc2() local
115 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86_acc2()
142 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86_acc2() local
143 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86_acc2()
|
| D | up8x3-minmax-wasmrelaxedsimd-fma-acc2.c | 72 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma_acc2() local
74 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma_acc2()
114 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma_acc2() local
115 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma_acc2()
142 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma_acc2() local
143 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma_acc2()
|
| D | up8x3-minmax-sse.c | 72 const __m128 vk1x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse() local
74 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse()
110 const __m128 vk1x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse() local
111 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse()
136 const __m128 vk1x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse() local
137 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse()
|
| D | up8x3-minmax-neon.c | 69 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neon() local
71 vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon()
99 const float32x4_t vk1x0123 = vld1q_f32(w + 12); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon() local
100 vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon()
121 const float32x4_t vk1x0123 = vld1q_f32(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon() local
122 vacc0123p0 = vmlaq_f32(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon()
|
| D | up8x3-minmax-sse-acc2.c | 72 const __m128 vk1x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse_acc2() local
74 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse_acc2()
113 const __m128 vk1x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse_acc2() local
114 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse_acc2()
141 const __m128 vk1x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse_acc2() local
142 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse_acc2()
|
| D | up8x3-minmax-neonfma.c | 69 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma() local
71 vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma()
99 const float32x4_t vk1x0123 = vld1q_f32(w + 12); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma() local
100 vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma()
121 const float32x4_t vk1x0123 = vld1q_f32(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma() local
122 vacc0123p0 = vfmaq_f32(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma()
|
| D | up8x3-minmax-wasmsimd-x86.c | 72 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86() local
74 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86()
111 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86() local
112 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86()
137 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86() local
138 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86()
|
| D | up8x3-minmax-neonfma-acc2.c | 69 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma_acc2() local
71 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma_acc2()
102 const float32x4_t vk1x0123 = vld1q_f32(w + 12); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma_acc2() local
103 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma_acc2()
126 const float32x4_t vk1x0123 = vld1q_f32(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma_acc2() local
127 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma_acc2()
|
| D | up8x3-minmax-wasmsimd-arm-acc2.c | 72 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm_acc2() local
74 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm_acc2()
114 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm_acc2() local
115 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm_acc2()
142 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm_acc2() local
143 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm_acc2()
|
| D | up8x3-minmax-wasmrelaxedsimd.c | 72 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd() local
74 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd()
111 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd() local
112 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd()
137 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd() local
138 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd()
|
| D | up8x3-minmax-neon-acc2.c | 69 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2() local
71 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2()
102 const float32x4_t vk1x0123 = vld1q_f32(w + 12); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2() local
103 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2()
126 const float32x4_t vk1x0123 = vld1q_f32(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2() local
127 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2()
|
| D | up8x3-minmax-wasmrelaxedsimd-acc2.c | 72 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_acc2() local
74 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_acc2()
114 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_acc2() local
115 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_acc2()
142 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_acc2() local
143 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_acc2()
|
| D | up8x4-wasmrelaxedsimd-fma.c | 75 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_ukernel_up8x4__wasmrelaxedsimd_fma() local
77 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x4__wasmrelaxedsimd_fma()
120 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_ukernel_up8x4__wasmrelaxedsimd_fma() local
121 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x4__wasmrelaxedsimd_fma()
151 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_ukernel_up8x4__wasmrelaxedsimd_fma() local
152 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x4__wasmrelaxedsimd_fma()
|
| D | up8x3-minmax-wasmsimd-arm.c | 72 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm() local
74 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm()
111 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm() local
112 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm()
137 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm() local
138 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm()
|
| D | up8x4-minmax-wasmrelaxedsimd-fma.c | 77 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_fma() local
79 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_fma()
125 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_fma() local
126 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_fma()
157 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_fma() local
158 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_fma()
|
| D | up8x4-wasmsimd.c | 75 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd() local
77 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd()
120 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd() local
121 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd()
151 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd() local
152 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd()
|
| D | up4x3-wasmsimd.c | 65 const v128_t vk1x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_ukernel_up4x3__wasmsimd() local
66 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_ukernel_up4x3__wasmsimd()
90 const v128_t vk1x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_ukernel_up4x3__wasmsimd() local
91 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_ukernel_up4x3__wasmsimd()
|
| D | up4x3-wasmrelaxedsimd-fma.c | 65 const v128_t vk1x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_ukernel_up4x3__wasmrelaxedsimd_fma() local
66 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x3__wasmrelaxedsimd_fma()
90 const v128_t vk1x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_ukernel_up4x3__wasmrelaxedsimd_fma() local
91 vacc0123p0 = __builtin_wasm_fma_f32x4(vacc0123p0, vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x3__wasmrelaxedsimd_fma()
|
| D | up8x4-minmax-wasmrelaxedsimd-acc2.c | 77 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_acc2() local
79 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_acc2()
128 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_acc2() local
129 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_acc2()
162 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_acc2() local
163 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_acc2()
|
| D | up8x4-minmax-wasmsimd-x86-acc2.c | 77 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local
79 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2()
128 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local
129 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2()
162 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local
163 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2()
|
| D | up8x4-minmax-wasmsimd-x86.c | 77 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86() local
79 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86()
125 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86() local
126 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86()
157 const v128_t vk1x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86() local
158 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi1x0123, vk1x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86()
|
| D | up8x4-minmax-neonfma-acc2.c | 74 const float32x4_t vk1x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local
76 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2()
114 const float32x4_t vk1x0123 = vld1q_f32(w + 12); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local
115 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2()
142 const float32x4_t vk1x0123 = vld1q_f32(w + 16); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local
143 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2()
|