| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x4-minmax-wasmsimd-arm.c | 60 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm() local
114 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm() local
150 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm() local
|
| D | up8x4-minmax-wasmsimd-x86.c | 60 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86() local
114 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86() local
150 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86() local
|
| D | up8x4-minmax-sse.c | 60 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse() local
113 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse() local
149 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse() local
|
| D | up8x4-wasmsimd.c | 58 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd() local
109 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd() local
144 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd() local
|
| D | up8x4-minmax-neon.c | 61 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neon() local
103 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neon() local
129 float32x4_t vacc0123p0 = vld1q_f32(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon() local
|
| D | up8x4-minmax-sse-acc2.c | 60 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local
116 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local
154 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local
|
| D | up8x4-minmax-neonfma.c | 61 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma() local
103 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma() local
129 float32x4_t vacc0123p0 = vld1q_f32(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma() local
|
| D | up8x4-minmax-wasmsimd-arm-acc2.c | 60 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local
117 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local
155 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local
|
| D | up8x4-minmax-wasmsimd-x86-acc2.c | 60 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local
117 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local
155 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local
|
| D | up8x4-minmax-neonfma-acc2.c | 61 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local
106 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local
134 float32x4_t vacc0123p0 = vld1q_f32(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local
|
| D | up8x4-minmax-neon-acc2.c | 61 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local
106 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local
134 float32x4_t vacc0123p0 = vld1q_f32(w); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local
|
| D | up4x4-wasmsimd.c | 58 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_ukernel_up4x4__wasmsimd() local
94 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_ukernel_up4x4__wasmsimd() local
|
| D | up4x4-minmax-wasmsimd-arm.c | 60 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm() local
98 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm() local
|
| D | up4x4-minmax-wasmsimd-x86.c | 60 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86() local
98 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86() local
|
| D | up4x4-minmax-sse.c | 60 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up4x4__sse() local
97 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up4x4__sse() local
|
| D | up4x4-minmax-neon.c | 61 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neon() local
87 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neon() local
|
| D | up4x4-minmax-sse-acc2.c | 60 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up4x4__sse_acc2() local
99 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up4x4__sse_acc2() local
|
| D | up4x4-minmax-neonfma.c | 61 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma() local
87 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma() local
|
| D | up4x4-minmax-wasmsimd-arm-acc2.c | 60 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm_acc2() local
100 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm_acc2() local
|
| D | up4x4-minmax-wasmsimd-x86-acc2.c | 60 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86_acc2() local
100 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86_acc2() local
|
| D | up4x4-minmax-neonfma-acc2.c | 61 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2() local
89 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2() local
|
| D | up4x4-minmax-neon-acc2.c | 61 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2() local
89 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2() local
|
| D | up8x9-wasmsimd.c | 83 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
179 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
244 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
|
| D | up8x9-minmax-wasmsimd-x86-acc2.c | 85 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
187 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
255 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
|
| D | up8x9-minmax-neon-acc2.c | 86 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
166 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
214 float32x4_t vacc0123p0 = vld1q_f32(w); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
|