/external/XNNPACK/src/f32-dwconv/gen/ |
D | up8x3-wasmsimd.c | 53 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_ukernel_up8x3__wasmsimd() local 95 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_ukernel_up8x3__wasmsimd() local 124 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_ukernel_up8x3__wasmsimd() local
|
D | up8x3-minmax-wasmsimd-arm.c | 55 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm() local 100 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm() local 130 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm() local
|
D | up8x3-minmax-neonfma.c | 56 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma() local 91 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma() local 113 float32x4_t vacc0123p0 = vld1q_f32(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma() local
|
D | up8x3-minmax-wasmsimd-x86.c | 55 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86() local 100 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86() local 130 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86() local
|
D | up8x3-minmax-sse.c | 55 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse() local 99 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse() local 129 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse() local
|
D | up8x3-minmax-neon.c | 56 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neon() local 91 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neon() local 113 float32x4_t vacc0123p0 = vld1q_f32(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon() local
|
D | up8x3-minmax-neonfma-acc2.c | 56 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma_acc2() local 94 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma_acc2() local 118 float32x4_t vacc0123p0 = vld1q_f32(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma_acc2() local
|
D | up8x3-minmax-wasmsimd-arm-acc2.c | 55 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm_acc2() local 103 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm_acc2() local 135 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm_acc2() local
|
D | up8x3-minmax-neon-acc2.c | 56 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2() local 94 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2() local 118 float32x4_t vacc0123p0 = vld1q_f32(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2() local
|
D | up8x3-minmax-sse-acc2.c | 55 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse_acc2() local 102 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse_acc2() local 134 __m128 vacc0123p0 = _mm_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse_acc2() local
|
D | up8x3-minmax-wasmsimd-x86-acc2.c | 55 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86_acc2() local 103 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86_acc2() local 135 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86_acc2() 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-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-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-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-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-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-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-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 | up16x3-minmax-neon.c | 56 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up16x3__neon() local 117 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up16x3__neon() local 139 float32x4_t vacc0123p0 = vld1q_f32(w); in xnn_f32_dwconv_minmax_ukernel_up16x3__neon() local
|
D | up16x3-minmax-neonfma.c | 56 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up16x3__neonfma() local 117 float32x4_t vacc0123p0 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up16x3__neonfma() local 139 float32x4_t vacc0123p0 = vld1q_f32(w); in xnn_f32_dwconv_minmax_ukernel_up16x3__neonfma() 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 | up4x3-wasmsimd.c | 53 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_ukernel_up4x3__wasmsimd() local 83 v128_t vacc0123p0 = wasm_v128_load(w); in xnn_f32_dwconv_ukernel_up4x3__wasmsimd() local
|