/aosp_15_r20/external/XNNPACK/src/x32-transposec/gen/ |
H A D | 4x4-reuse-multi-zip-neon.c | 55 const uint32x4_t v2_1 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride); in xnn_x32_transposec_ukernel__4x4_reuse_multi_zip_neon() local 77 const uint32x4_t v2_1 = vld1q_u32(i1); in xnn_x32_transposec_ukernel__4x4_reuse_multi_zip_neon() local
|
H A D | 4x4-multi-multi-zip-neon.c | 59 const uint32x4_t v2_1 = vld1q_u32(i1); i1 = (uint32_t*) ((uintptr_t) i1 + input_offset); in xnn_x32_transposec_ukernel__4x4_multi_multi_zip_neon() local 80 const uint32x4_t v2_1 = vld1q_u32(i1); in xnn_x32_transposec_ukernel__4x4_multi_multi_zip_neon() local
|
H A D | 4x4-reuse-dec-zip-neon.c | 47 const uint32x4_t v2_1 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride); in xnn_x32_transposec_ukernel__4x4_reuse_dec_zip_neon() local 80 const uint32x4_t v2_1 = vld1q_u32(i1); in xnn_x32_transposec_ukernel__4x4_reuse_dec_zip_neon() local
|
H A D | 4x4-reuse-multi-wasmsimd.c | 56 const v128_t v2_1 = wasm_v128_load(i0); in xnn_x32_transposec_ukernel__4x4_reuse_multi_wasmsimd() local 88 const v128_t v2_1 = wasm_v128_load(i1); in xnn_x32_transposec_ukernel__4x4_reuse_multi_wasmsimd() local
|
H A D | 4x4-reuse-mov-zip-neon.c | 47 const uint32x4_t v2_1 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride); in xnn_x32_transposec_ukernel__4x4_reuse_mov_zip_neon() local 83 const uint32x4_t v2_1 = vld1q_u32(i1); in xnn_x32_transposec_ukernel__4x4_reuse_mov_zip_neon() local
|
H A D | 4x4-multi-dec-zip-neon.c | 51 const uint32x4_t v2_1 = vld1q_u32(i1); i1 = (uint32_t*) ((uintptr_t) i1 + input_offset); in xnn_x32_transposec_ukernel__4x4_multi_dec_zip_neon() local 83 const uint32x4_t v2_1 = vld1q_u32(i1); in xnn_x32_transposec_ukernel__4x4_multi_dec_zip_neon() local
|
H A D | 4x4-reuse-multi-sse2.c | 58 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i0); in xnn_x32_transposec_ukernel__4x4_reuse_multi_sse2() local 92 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1); in xnn_x32_transposec_ukernel__4x4_reuse_multi_sse2() local
|
H A D | 4x4-reuse-switch-zip-neon.c | 46 const uint32x4_t v2_1 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride); in xnn_x32_transposec_ukernel__4x4_reuse_switch_zip_neon() local 78 const uint32x4_t v2_1 = vld1q_u32(i1); in xnn_x32_transposec_ukernel__4x4_reuse_switch_zip_neon() local
|
H A D | 4x4-multi-multi-wasmsimd.c | 60 const v128_t v2_1 = wasm_v128_load(i1); in xnn_x32_transposec_ukernel__4x4_multi_multi_wasmsimd() local 91 const v128_t v2_1 = wasm_v128_load(i1); in xnn_x32_transposec_ukernel__4x4_multi_multi_wasmsimd() local
|
H A D | 4x4-multi-mov-zip-neon.c | 51 const uint32x4_t v2_1 = vld1q_u32(i1); i1 = (uint32_t*) ((uintptr_t) i1 + input_offset); in xnn_x32_transposec_ukernel__4x4_multi_mov_zip_neon() local 86 const uint32x4_t v2_1 = vld1q_u32(i1); in xnn_x32_transposec_ukernel__4x4_multi_mov_zip_neon() local
|
H A D | 4x4-multi-multi-sse2.c | 62 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1); in xnn_x32_transposec_ukernel__4x4_multi_multi_sse2() local 95 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1); in xnn_x32_transposec_ukernel__4x4_multi_multi_sse2() local
|
H A D | 4x4-multi-switch-zip-neon.c | 50 const uint32x4_t v2_1 = vld1q_u32(i1); i1 = (uint32_t*) ((uintptr_t) i1 + input_offset); in xnn_x32_transposec_ukernel__4x4_multi_switch_zip_neon() local 81 const uint32x4_t v2_1 = vld1q_u32(i1); in xnn_x32_transposec_ukernel__4x4_multi_switch_zip_neon() local
|
H A D | 4x4-reuse-switch-sse2.c | 49 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i0); in xnn_x32_transposec_ukernel__4x4_reuse_switch_sse2() local 92 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1); in xnn_x32_transposec_ukernel__4x4_reuse_switch_sse2() local
|
H A D | 4x4-reuse-mov-sse2.c | 50 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i0); in xnn_x32_transposec_ukernel__4x4_reuse_mov_sse2() local 94 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1); in xnn_x32_transposec_ukernel__4x4_reuse_mov_sse2() local
|
H A D | 4x4-reuse-mov-wasmsimd.c | 48 const v128_t v2_1 = wasm_v128_load(i0); in xnn_x32_transposec_ukernel__4x4_reuse_mov_wasmsimd() local 90 const v128_t v2_1 = wasm_v128_load(i1); in xnn_x32_transposec_ukernel__4x4_reuse_mov_wasmsimd() local
|
H A D | 4x4-reuse-switch-wasmsimd.c | 47 const v128_t v2_1 = wasm_v128_load(i0); in xnn_x32_transposec_ukernel__4x4_reuse_switch_wasmsimd() local 88 const v128_t v2_1 = wasm_v128_load(i1); in xnn_x32_transposec_ukernel__4x4_reuse_switch_wasmsimd() local
|
H A D | 4x4-multi-switch-sse2.c | 53 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1); in xnn_x32_transposec_ukernel__4x4_multi_switch_sse2() local 95 const __m128i v2_1 = _mm_loadu_si128((const __m128i*) i1); in xnn_x32_transposec_ukernel__4x4_multi_switch_sse2() local
|
/aosp_15_r20/external/XNNPACK/src/x16-transposec/gen/ |
H A D | 4x4-reuse-multi-zip-neon.c | 55 const uint16x4_t v2_1 = vld1_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); in xnn_x16_transposec_ukernel__4x4_reuse_multi_zip_neon() local 77 const uint16x4_t v2_1 = vld1_u16(i1); in xnn_x16_transposec_ukernel__4x4_reuse_multi_zip_neon() local
|
H A D | 4x4-multi-multi-zip-neon.c | 59 const uint16x4_t v2_1 = vld1_u16(i1); i1 = (uint16_t*) ((uintptr_t) i1 + input_offset); in xnn_x16_transposec_ukernel__4x4_multi_multi_zip_neon() local 80 const uint16x4_t v2_1 = vld1_u16(i1); in xnn_x16_transposec_ukernel__4x4_multi_multi_zip_neon() local
|
H A D | 4x4-reuse-dec-zip-neon.c | 47 const uint16x4_t v2_1 = vld1_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); in xnn_x16_transposec_ukernel__4x4_reuse_dec_zip_neon() local 80 const uint16x4_t v2_1 = vld1_u16(i1); in xnn_x16_transposec_ukernel__4x4_reuse_dec_zip_neon() local
|
H A D | 4x4-reuse-switch-zip-neon.c | 46 const uint16x4_t v2_1 = vld1_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); in xnn_x16_transposec_ukernel__4x4_reuse_switch_zip_neon() local 78 const uint16x4_t v2_1 = vld1_u16(i1); in xnn_x16_transposec_ukernel__4x4_reuse_switch_zip_neon() local
|
H A D | 4x4-multi-dec-zip-neon.c | 51 const uint16x4_t v2_1 = vld1_u16(i1); i1 = (uint16_t*) ((uintptr_t) i1 + input_offset); in xnn_x16_transposec_ukernel__4x4_multi_dec_zip_neon() local 83 const uint16x4_t v2_1 = vld1_u16(i1); in xnn_x16_transposec_ukernel__4x4_multi_dec_zip_neon() local
|
H A D | 4x4-reuse-mov-zip-neon.c | 47 const uint16x4_t v2_1 = vld1_u16(i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); in xnn_x16_transposec_ukernel__4x4_reuse_mov_zip_neon() local 83 const uint16x4_t v2_1 = vld1_u16(i1); in xnn_x16_transposec_ukernel__4x4_reuse_mov_zip_neon() local
|
H A D | 4x4-multi-mov-zip-neon.c | 51 const uint16x4_t v2_1 = vld1_u16(i1); i1 = (uint16_t*) ((uintptr_t) i1 + input_offset); in xnn_x16_transposec_ukernel__4x4_multi_mov_zip_neon() local 86 const uint16x4_t v2_1 = vld1_u16(i1); in xnn_x16_transposec_ukernel__4x4_multi_mov_zip_neon() local
|
H A D | 4x4-multi-switch-zip-neon.c | 50 const uint16x4_t v2_1 = vld1_u16(i1); i1 = (uint16_t*) ((uintptr_t) i1 + input_offset); in xnn_x16_transposec_ukernel__4x4_multi_switch_zip_neon() local 81 const uint16x4_t v2_1 = vld1_u16(i1); in xnn_x16_transposec_ukernel__4x4_multi_switch_zip_neon() local
|