| /external/XNNPACK/src/x8-transposec/gen/ |
| D | 8x8-reuse-dec-zip-neon.c | 49 const uint8x8_t v3_3 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); in xnn_x8_transposec_ukernel__8x8_reuse_dec_zip_neon() local 118 const uint8x8_t v3_3 = vld1_u8(i3); in xnn_x8_transposec_ukernel__8x8_reuse_dec_zip_neon() local
|
| D | 8x8-multi-switch-zip-neon.c | 56 const uint8x8_t v3_3 = vld1_u8(i3); i3 = (uint8_t*) ((uintptr_t) i3 + input_offset); in xnn_x8_transposec_ukernel__8x8_multi_switch_zip_neon() local 113 const uint8x8_t v3_3 = vld1_u8(i3); in xnn_x8_transposec_ukernel__8x8_multi_switch_zip_neon() local
|
| D | 8x8-reuse-multi-zip-neon.c | 73 const uint8x8_t v3_3 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); in xnn_x8_transposec_ukernel__8x8_reuse_multi_zip_neon() local 119 const uint8x8_t v3_3 = vld1_u8(i3); in xnn_x8_transposec_ukernel__8x8_reuse_multi_zip_neon() local
|
| D | 8x8-multi-dec-zip-neon.c | 57 const uint8x8_t v3_3 = vld1_u8(i3); i3 = (uint8_t*) ((uintptr_t) i3 + input_offset); in xnn_x8_transposec_ukernel__8x8_multi_dec_zip_neon() local 123 const uint8x8_t v3_3 = vld1_u8(i3); in xnn_x8_transposec_ukernel__8x8_multi_dec_zip_neon() local
|
| D | 8x8-reuse-mov-zip-neon.c | 49 const uint8x8_t v3_3 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); in xnn_x8_transposec_ukernel__8x8_reuse_mov_zip_neon() local 125 const uint8x8_t v3_3 = vld1_u8(i3); in xnn_x8_transposec_ukernel__8x8_reuse_mov_zip_neon() local
|
| D | 8x8-reuse-switch-zip-neon.c | 48 const uint8x8_t v3_3 = vld1_u8(i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); in xnn_x8_transposec_ukernel__8x8_reuse_switch_zip_neon() local 108 const uint8x8_t v3_3 = vld1_u8(i3); in xnn_x8_transposec_ukernel__8x8_reuse_switch_zip_neon() local
|
| D | 8x8-multi-mov-zip-neon.c | 57 const uint8x8_t v3_3 = vld1_u8(i3); i3 = (uint8_t*) ((uintptr_t) i3 + input_offset); in xnn_x8_transposec_ukernel__8x8_multi_mov_zip_neon() local 130 const uint8x8_t v3_3 = vld1_u8(i3); in xnn_x8_transposec_ukernel__8x8_multi_mov_zip_neon() local
|
| D | 16x16-reuse-switch-zip-neon.c | 65 const uint8x16x2_t v3_3 = vzipq_u8(v4_3, v4_11); in xnn_x8_transposec_ukernel__16x16_reuse_switch_zip_neon() local 213 const uint8x16x2_t v3_3 = vzipq_u8(v4_3, v4_11); in xnn_x8_transposec_ukernel__16x16_reuse_switch_zip_neon() local
|
| /external/XNNPACK/src/x16-transposec/gen/ |
| D | 8x8-reuse-dec-zip-neon.c | 49 const uint16x8_t v3_3 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); in xnn_x16_transposec_ukernel__8x8_reuse_dec_zip_neon() local 118 const uint16x8_t v3_3 = vld1q_u16(i3); in xnn_x16_transposec_ukernel__8x8_reuse_dec_zip_neon() local
|
| D | 8x8-multi-switch-zip-neon.c | 56 const uint16x8_t v3_3 = vld1q_u16(i3); i3 = (uint16_t*) ((uintptr_t) i3 + input_offset); in xnn_x16_transposec_ukernel__8x8_multi_switch_zip_neon() local 113 const uint16x8_t v3_3 = vld1q_u16(i3); in xnn_x16_transposec_ukernel__8x8_multi_switch_zip_neon() local
|
| D | 8x8-reuse-multi-zip-neon.c | 73 const uint16x8_t v3_3 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); in xnn_x16_transposec_ukernel__8x8_reuse_multi_zip_neon() local 119 const uint16x8_t v3_3 = vld1q_u16(i3); in xnn_x16_transposec_ukernel__8x8_reuse_multi_zip_neon() local
|
| D | 8x8-multi-dec-zip-neon.c | 57 const uint16x8_t v3_3 = vld1q_u16(i3); i3 = (uint16_t*) ((uintptr_t) i3 + input_offset); in xnn_x16_transposec_ukernel__8x8_multi_dec_zip_neon() local 123 const uint16x8_t v3_3 = vld1q_u16(i3); in xnn_x16_transposec_ukernel__8x8_multi_dec_zip_neon() local
|
| D | 8x8-reuse-switch-zip-neon.c | 48 const uint16x8_t v3_3 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); in xnn_x16_transposec_ukernel__8x8_reuse_switch_zip_neon() local 108 const uint16x8_t v3_3 = vld1q_u16(i3); in xnn_x16_transposec_ukernel__8x8_reuse_switch_zip_neon() local
|
| D | 8x8-reuse-mov-zip-neon.c | 49 const uint16x8_t v3_3 = vld1q_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); in xnn_x16_transposec_ukernel__8x8_reuse_mov_zip_neon() local 125 const uint16x8_t v3_3 = vld1q_u16(i3); in xnn_x16_transposec_ukernel__8x8_reuse_mov_zip_neon() local
|
| D | 8x8-multi-mov-zip-neon.c | 57 const uint16x8_t v3_3 = vld1q_u16(i3); i3 = (uint16_t*) ((uintptr_t) i3 + input_offset); in xnn_x16_transposec_ukernel__8x8_multi_mov_zip_neon() local 130 const uint16x8_t v3_3 = vld1q_u16(i3); in xnn_x16_transposec_ukernel__8x8_multi_mov_zip_neon() local
|
| D | 8x8-reuse-multi-wasmsimd.c | 76 const v128_t v3_3 = wasm_v128_load(i0); in xnn_x16_transposec_ukernel__8x8_reuse_multi_wasmsimd() local 146 const v128_t v3_3 = wasm_v128_load(i3); in xnn_x16_transposec_ukernel__8x8_reuse_multi_wasmsimd() local
|
| D | 8x8-reuse-switch-sse2.c | 53 const __m128i v3_3 = _mm_loadu_si128((const __m128i*) i0); in xnn_x16_transposec_ukernel__8x8_reuse_switch_sse2() local 138 const __m128i v3_3 = _mm_loadu_si128((const __m128i*) i3); in xnn_x16_transposec_ukernel__8x8_reuse_switch_sse2() local
|
| D | 8x8-multi-switch-sse2.c | 61 const __m128i v3_3 = _mm_loadu_si128((const __m128i*) i3); in xnn_x16_transposec_ukernel__8x8_multi_switch_sse2() local 143 const __m128i v3_3 = _mm_loadu_si128((const __m128i*) i3); in xnn_x16_transposec_ukernel__8x8_multi_switch_sse2() local
|
| D | 8x8-reuse-mov-sse2.c | 54 const __m128i v3_3 = _mm_loadu_si128((const __m128i*) i0); in xnn_x16_transposec_ukernel__8x8_reuse_mov_sse2() local 148 const __m128i v3_3 = _mm_loadu_si128((const __m128i*) i3); in xnn_x16_transposec_ukernel__8x8_reuse_mov_sse2() local
|
| D | 8x8-reuse-switch-wasmsimd.c | 51 const v128_t v3_3 = wasm_v128_load(i0); in xnn_x16_transposec_ukernel__8x8_reuse_switch_wasmsimd() local 134 const v128_t v3_3 = wasm_v128_load(i3); in xnn_x16_transposec_ukernel__8x8_reuse_switch_wasmsimd() local
|
| D | 8x8-reuse-multi-sse2.c | 78 const __m128i v3_3 = _mm_loadu_si128((const __m128i*) i0); in xnn_x16_transposec_ukernel__8x8_reuse_multi_sse2() local 150 const __m128i v3_3 = _mm_loadu_si128((const __m128i*) i3); in xnn_x16_transposec_ukernel__8x8_reuse_multi_sse2() local
|
| D | 8x8-multi-switch-wasmsimd.c | 59 const v128_t v3_3 = wasm_v128_load(i3); in xnn_x16_transposec_ukernel__8x8_multi_switch_wasmsimd() local 139 const v128_t v3_3 = wasm_v128_load(i3); in xnn_x16_transposec_ukernel__8x8_multi_switch_wasmsimd() local
|
| D | 8x8-reuse-mov-wasmsimd.c | 52 const v128_t v3_3 = wasm_v128_load(i0); in xnn_x16_transposec_ukernel__8x8_reuse_mov_wasmsimd() local 144 const v128_t v3_3 = wasm_v128_load(i3); in xnn_x16_transposec_ukernel__8x8_reuse_mov_wasmsimd() local
|
| D | 8x8-multi-mov-sse2.c | 62 const __m128i v3_3 = _mm_loadu_si128((const __m128i*) i3); in xnn_x16_transposec_ukernel__8x8_multi_mov_sse2() local 153 const __m128i v3_3 = _mm_loadu_si128((const __m128i*) i3); in xnn_x16_transposec_ukernel__8x8_multi_mov_sse2() local
|
| D | 8x8-multi-mov-wasmsimd.c | 60 const v128_t v3_3 = wasm_v128_load(i3); in xnn_x16_transposec_ukernel__8x8_multi_mov_wasmsimd() local 149 const v128_t v3_3 = wasm_v128_load(i3); in xnn_x16_transposec_ukernel__8x8_multi_mov_wasmsimd() local
|