| /external/XNNPACK/src/x32-transpose/gen/ |
| D | 4x4-reuse-multi-zip-neon.c | 56 const uint32x4_t v2_2 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride); in xnn_x32_transpose_ukernel__4x4_reuse_multi_zip_neon() local 82 const uint32x4_t v2_2 = vld1q_u32(i2); in xnn_x32_transpose_ukernel__4x4_reuse_multi_zip_neon() local
|
| D | 4x4-multi-multi-zip-neon.c | 60 const uint32x4_t v2_2 = vld1q_u32(i2); i2 = (uint32_t*) ((uintptr_t) i2 + input_offset); in xnn_x32_transpose_ukernel__4x4_multi_multi_zip_neon() local 84 const uint32x4_t v2_2 = vld1q_u32(i2); in xnn_x32_transpose_ukernel__4x4_multi_multi_zip_neon() local
|
| D | 4x4-reuse-dec-zip-neon.c | 48 const uint32x4_t v2_2 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride); in xnn_x32_transpose_ukernel__4x4_reuse_dec_zip_neon() local 85 const uint32x4_t v2_2 = vld1q_u32(i2); in xnn_x32_transpose_ukernel__4x4_reuse_dec_zip_neon() local
|
| D | 4x4-reuse-multi-sse2.c | 58 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i0); in xnn_x32_transpose_ukernel__4x4_reuse_multi_sse2() local 95 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2); in xnn_x32_transpose_ukernel__4x4_reuse_multi_sse2() local
|
| D | 4x4-multi-dec-zip-neon.c | 52 const uint32x4_t v2_2 = vld1q_u32(i2); i2 = (uint32_t*) ((uintptr_t) i2 + input_offset); in xnn_x32_transpose_ukernel__4x4_multi_dec_zip_neon() local 87 const uint32x4_t v2_2 = vld1q_u32(i2); in xnn_x32_transpose_ukernel__4x4_multi_dec_zip_neon() local
|
| D | 4x4-reuse-switch-zip-neon.c | 47 const uint32x4_t v2_2 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride); in xnn_x32_transpose_ukernel__4x4_reuse_switch_zip_neon() local 83 const uint32x4_t v2_2 = vld1q_u32(i2); in xnn_x32_transpose_ukernel__4x4_reuse_switch_zip_neon() local
|
| D | 4x4-reuse-mov-zip-neon.c | 48 const uint32x4_t v2_2 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride); in xnn_x32_transpose_ukernel__4x4_reuse_mov_zip_neon() local 88 const uint32x4_t v2_2 = vld1q_u32(i2); in xnn_x32_transpose_ukernel__4x4_reuse_mov_zip_neon() local
|
| D | 4x4-multi-switch-zip-neon.c | 51 const uint32x4_t v2_2 = vld1q_u32(i2); i2 = (uint32_t*) ((uintptr_t) i2 + input_offset); in xnn_x32_transpose_ukernel__4x4_multi_switch_zip_neon() local 85 const uint32x4_t v2_2 = vld1q_u32(i2); in xnn_x32_transpose_ukernel__4x4_multi_switch_zip_neon() local
|
| D | 4x4-multi-mov-zip-neon.c | 52 const uint32x4_t v2_2 = vld1q_u32(i2); i2 = (uint32_t*) ((uintptr_t) i2 + input_offset); in xnn_x32_transpose_ukernel__4x4_multi_mov_zip_neon() local 90 const uint32x4_t v2_2 = vld1q_u32(i2); in xnn_x32_transpose_ukernel__4x4_multi_mov_zip_neon() local
|
| D | 4x4-multi-multi-sse2.c | 62 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2); in xnn_x32_transpose_ukernel__4x4_multi_multi_sse2() local 97 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2); in xnn_x32_transpose_ukernel__4x4_multi_multi_sse2() local
|
| D | 4x4-reuse-mov-sse2.c | 50 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i0); in xnn_x32_transpose_ukernel__4x4_reuse_mov_sse2() local 97 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2); in xnn_x32_transpose_ukernel__4x4_reuse_mov_sse2() local
|
| D | 4x4-reuse-switch-sse2.c | 49 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i0); in xnn_x32_transpose_ukernel__4x4_reuse_switch_sse2() local 95 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2); in xnn_x32_transpose_ukernel__4x4_reuse_switch_sse2() local
|
| D | 4x4-multi-switch-sse2.c | 53 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2); in xnn_x32_transpose_ukernel__4x4_multi_switch_sse2() local 97 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2); in xnn_x32_transpose_ukernel__4x4_multi_switch_sse2() local
|
| D | 4x4-multi-mov-sse2.c | 54 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2); in xnn_x32_transpose_ukernel__4x4_multi_mov_sse2() local 99 const __m128i v2_2 = _mm_loadu_si128((const __m128i*) i2); in xnn_x32_transpose_ukernel__4x4_multi_mov_sse2() local
|
| /external/XNNPACK/src/x16-transpose/gen/ |
| D | 8x8-reuse-dec-zip-neon.c | 57 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6); in xnn_x16_transpose_ukernel__8x8_reuse_dec_zip_neon() local 138 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6); in xnn_x16_transpose_ukernel__8x8_reuse_dec_zip_neon() local
|
| D | 8x8-reuse-switch-zip-neon.c | 56 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6); in xnn_x16_transpose_ukernel__8x8_reuse_switch_zip_neon() local 128 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6); in xnn_x16_transpose_ukernel__8x8_reuse_switch_zip_neon() local
|
| D | 8x8-multi-dec-zip-neon.c | 65 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6); in xnn_x16_transpose_ukernel__8x8_multi_dec_zip_neon() local 140 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6); in xnn_x16_transpose_ukernel__8x8_multi_dec_zip_neon() local
|
| D | 8x8-reuse-multi-zip-neon.c | 81 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6); in xnn_x16_transpose_ukernel__8x8_reuse_multi_zip_neon() local 139 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6); in xnn_x16_transpose_ukernel__8x8_reuse_multi_zip_neon() local
|
| D | 8x8-reuse-mov-zip-neon.c | 57 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6); in xnn_x16_transpose_ukernel__8x8_reuse_mov_zip_neon() local 145 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6); in xnn_x16_transpose_ukernel__8x8_reuse_mov_zip_neon() local
|
| D | 8x8-multi-switch-zip-neon.c | 64 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6); in xnn_x16_transpose_ukernel__8x8_multi_switch_zip_neon() local 130 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6); in xnn_x16_transpose_ukernel__8x8_multi_switch_zip_neon() local
|
| D | 8x8-multi-mov-zip-neon.c | 65 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6); in xnn_x16_transpose_ukernel__8x8_multi_mov_zip_neon() local 147 const uint16x8x2_t v2_2 = vzipq_u16(v3_2, v3_6); in xnn_x16_transpose_ukernel__8x8_multi_mov_zip_neon() local
|
| D | 8x8-multi-switch-sse2.c | 72 const __m128i v2_2 = _mm_unpacklo_epi16(v3_2, v3_3); in xnn_x16_transpose_ukernel__8x8_multi_switch_sse2() local 158 const __m128i v2_2 = _mm_unpacklo_epi16(v3_2, v3_3); in xnn_x16_transpose_ukernel__8x8_multi_switch_sse2() local
|
| D | 8x8-reuse-switch-sse2.c | 64 const __m128i v2_2 = _mm_unpacklo_epi16(v3_2, v3_3); in xnn_x16_transpose_ukernel__8x8_reuse_switch_sse2() local 156 const __m128i v2_2 = _mm_unpacklo_epi16(v3_2, v3_3); in xnn_x16_transpose_ukernel__8x8_reuse_switch_sse2() local
|
| D | 8x8-reuse-multi-sse2.c | 89 const __m128i v2_2 = _mm_unpacklo_epi16(v3_2, v3_3); in xnn_x16_transpose_ukernel__8x8_reuse_multi_sse2() local 168 const __m128i v2_2 = _mm_unpacklo_epi16(v3_2, v3_3); in xnn_x16_transpose_ukernel__8x8_reuse_multi_sse2() local
|
| D | 8x8-reuse-mov-sse2.c | 65 const __m128i v2_2 = _mm_unpacklo_epi16(v3_2, v3_3); in xnn_x16_transpose_ukernel__8x8_reuse_mov_sse2() local 166 const __m128i v2_2 = _mm_unpacklo_epi16(v3_2, v3_3); in xnn_x16_transpose_ukernel__8x8_reuse_mov_sse2() local
|