| /external/XNNPACK/src/x8-zip/ |
| D | x4-sse2.c | 35 const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy); in xnn_x8_zip_x4_ukernel__sse2() local
39 const __m128i vxyzw0 = _mm_unpacklo_epi16(vxy_lo, vzw_lo); in xnn_x8_zip_x4_ukernel__sse2()
40 const __m128i vxyzw1 = _mm_unpackhi_epi16(vxy_lo, vzw_lo); in xnn_x8_zip_x4_ukernel__sse2()
56 const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy); in xnn_x8_zip_x4_ukernel__sse2() local
60 const __m128i vxyzw0 = _mm_unpacklo_epi16(vxy_lo, vzw_lo); in xnn_x8_zip_x4_ukernel__sse2()
61 const __m128i vxyzw1 = _mm_unpackhi_epi16(vxy_lo, vzw_lo); in xnn_x8_zip_x4_ukernel__sse2()
|
| D | x2-sse2.c | 29 const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy); in xnn_x8_zip_x2_ukernel__sse2() local
31 _mm_storeu_si128((__m128i*) o, vxy_lo); in xnn_x8_zip_x2_ukernel__sse2()
40 const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy); in xnn_x8_zip_x2_ukernel__sse2() local
43 _mm_storeu_si128((__m128i*) o, vxy_lo); in xnn_x8_zip_x2_ukernel__sse2()
|
| D | xm-sse2.c | 46 const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy); in xnn_x8_zip_xm_ukernel__sse2() local
50 __m128i vxyzw0 = _mm_unpacklo_epi16(vxy_lo, vzw_lo); in xnn_x8_zip_xm_ukernel__sse2()
51 __m128i vxyzw1 = _mm_unpackhi_epi16(vxy_lo, vzw_lo); in xnn_x8_zip_xm_ukernel__sse2()
|
| /external/XNNPACK/src/x32-zip/ |
| D | x4-wasmsimd.c | 37 const v128_t vxy_lo = wasm_v32x4_shuffle(vx, vy, 0, 4, 1, 5); in xnn_x32_zip_x4_ukernel__wasmsimd() local
42 const v128_t vxyzw0 = wasm_v32x4_shuffle(vxy_lo, vzw_lo, 0, 1, 4, 5); in xnn_x32_zip_x4_ukernel__wasmsimd()
43 const v128_t vxyzw1 = wasm_v32x4_shuffle(vxy_lo, vzw_lo, 2, 3, 6, 7); in xnn_x32_zip_x4_ukernel__wasmsimd()
|
| D | x4-sse2.c | 37 const __m128i vxy_lo = _mm_unpacklo_epi32(vx, vy); in xnn_x32_zip_x4_ukernel__sse2() local
42 const __m128i vxyzw0 = _mm_unpacklo_epi64(vxy_lo, vzw_lo); in xnn_x32_zip_x4_ukernel__sse2()
43 const __m128i vxyzw1 = _mm_unpackhi_epi64(vxy_lo, vzw_lo); in xnn_x32_zip_x4_ukernel__sse2()
|
| D | xm-wasmsimd.c | 51 const v128_t vxy_lo = wasm_v32x4_shuffle(vx, vy, 0, 4, 1, 5); in xnn_x32_zip_xm_ukernel__wasmsimd() local
56 const v128_t vxyzw0 = wasm_v32x4_shuffle(vxy_lo, vzw_lo, 0, 1, 4, 5); in xnn_x32_zip_xm_ukernel__wasmsimd()
57 const v128_t vxyzw1 = wasm_v32x4_shuffle(vxy_lo, vzw_lo, 2, 3, 6, 7); in xnn_x32_zip_xm_ukernel__wasmsimd()
|
| D | x2-wasmsimd.c | 30 const v128_t vxy_lo = wasm_v32x4_shuffle(vx, vy, 0, 4, 1, 5); in xnn_x32_zip_x2_ukernel__wasmsimd() local
32 wasm_v128_store(o, vxy_lo); in xnn_x32_zip_x2_ukernel__wasmsimd()
|
| D | x2-sse2.c | 30 const __m128i vxy_lo = _mm_unpacklo_epi32(vx, vy); in xnn_x32_zip_x2_ukernel__sse2() local
32 _mm_storeu_si128((__m128i*) o, vxy_lo); in xnn_x32_zip_x2_ukernel__sse2()
|
| D | xm-sse2.c | 50 const __m128i vxy_lo = _mm_unpacklo_epi32(vx, vy); in xnn_x32_zip_xm_ukernel__sse2() local
55 const __m128i vxyzw0 = _mm_unpacklo_epi64(vxy_lo, vzw_lo); in xnn_x32_zip_xm_ukernel__sse2()
56 const __m128i vxyzw1 = _mm_unpackhi_epi64(vxy_lo, vzw_lo); in xnn_x32_zip_xm_ukernel__sse2()
|
| /external/XNNPACK/src/amalgam/ |
| D | sse2.c | 12928 const __m128i vxy_lo = _mm_unpacklo_epi32(vx, vy); in xnn_x32_zip_x2_ukernel__sse2() local
12930 _mm_storeu_si128((__m128i*) o, vxy_lo); in xnn_x32_zip_x2_ukernel__sse2()
13056 const __m128i vxy_lo = _mm_unpacklo_epi32(vx, vy); in xnn_x32_zip_x4_ukernel__sse2() local
13061 const __m128i vxyzw0 = _mm_unpacklo_epi64(vxy_lo, vzw_lo); in xnn_x32_zip_x4_ukernel__sse2()
13062 const __m128i vxyzw1 = _mm_unpackhi_epi64(vxy_lo, vzw_lo); in xnn_x32_zip_x4_ukernel__sse2()
13144 const __m128i vxy_lo = _mm_unpacklo_epi32(vx, vy); in xnn_x32_zip_xm_ukernel__sse2() local
13149 const __m128i vxyzw0 = _mm_unpacklo_epi64(vxy_lo, vzw_lo); in xnn_x32_zip_xm_ukernel__sse2()
13150 const __m128i vxyzw1 = _mm_unpackhi_epi64(vxy_lo, vzw_lo); in xnn_x32_zip_xm_ukernel__sse2()
13952 const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy); in xnn_x8_zip_x2_ukernel__sse2() local
13954 _mm_storeu_si128((__m128i*) o, vxy_lo); in xnn_x8_zip_x2_ukernel__sse2()
[all …]
|