/aosp_15_r20/external/pytorch/aten/src/ATen/native/quantized/cpu/qnnpack/src/x8zip/ |
H A D | x4-sse2.c | 33 const __m128i vzw_hi = _mm_unpackhi_epi8(vz, vw); in pytorch_qnnp_x8zip_x4__sse2() local 36 const __m128i vxyzw2 = _mm_unpacklo_epi16(vxy_hi, vzw_hi); in pytorch_qnnp_x8zip_x4__sse2() 37 const __m128i vxyzw3 = _mm_unpackhi_epi16(vxy_hi, vzw_hi); in pytorch_qnnp_x8zip_x4__sse2() 58 const __m128i vzw_hi = _mm_unpackhi_epi8(vz, vw); in pytorch_qnnp_x8zip_x4__sse2() local 61 const __m128i vxyzw2 = _mm_unpacklo_epi16(vxy_hi, vzw_hi); in pytorch_qnnp_x8zip_x4__sse2() 62 const __m128i vxyzw3 = _mm_unpackhi_epi16(vxy_hi, vzw_hi); in pytorch_qnnp_x8zip_x4__sse2()
|
H A D | xm-sse2.c | 46 const __m128i vzw_hi = _mm_unpackhi_epi8(vz, vw); in pytorch_qnnp_x8zip_xm__sse2() local 49 __m128i vxyzw2 = _mm_unpacklo_epi16(vxy_hi, vzw_hi); in pytorch_qnnp_x8zip_xm__sse2() 50 __m128i vxyzw3 = _mm_unpackhi_epi16(vxy_hi, vzw_hi); in pytorch_qnnp_x8zip_xm__sse2()
|
/aosp_15_r20/external/XNNPACK/src/x8-zip/ |
H A D | x4-sse2.c | 38 const __m128i vzw_hi = _mm_unpackhi_epi8(vz, vw); in xnn_x8_zip_x4_ukernel__sse2() local 41 const __m128i vxyzw2 = _mm_unpacklo_epi16(vxy_hi, vzw_hi); in xnn_x8_zip_x4_ukernel__sse2() 42 const __m128i vxyzw3 = _mm_unpackhi_epi16(vxy_hi, vzw_hi); in xnn_x8_zip_x4_ukernel__sse2() 59 const __m128i vzw_hi = _mm_unpackhi_epi8(vz, vw); in xnn_x8_zip_x4_ukernel__sse2() local 62 const __m128i vxyzw2 = _mm_unpacklo_epi16(vxy_hi, vzw_hi); in xnn_x8_zip_x4_ukernel__sse2() 63 const __m128i vxyzw3 = _mm_unpackhi_epi16(vxy_hi, vzw_hi); in xnn_x8_zip_x4_ukernel__sse2()
|
H A D | xm-sse2.c | 49 const __m128i vzw_hi = _mm_unpackhi_epi8(vz, vw); in xnn_x8_zip_xm_ukernel__sse2() local 52 __m128i vxyzw2 = _mm_unpacklo_epi16(vxy_hi, vzw_hi); in xnn_x8_zip_xm_ukernel__sse2() 53 __m128i vxyzw3 = _mm_unpackhi_epi16(vxy_hi, vzw_hi); in xnn_x8_zip_xm_ukernel__sse2()
|
/aosp_15_r20/external/XNNPACK/src/x32-zip/ |
H A D | x4-wasmsimd.c | 40 const v128_t vzw_hi = wasm_v32x4_shuffle(vz, vw, 2, 6, 3, 7); in xnn_x32_zip_x4_ukernel__wasmsimd() local 44 const v128_t vxyzw2 = wasm_v32x4_shuffle(vxy_hi, vzw_hi, 0, 1, 4, 5); in xnn_x32_zip_x4_ukernel__wasmsimd() 45 const v128_t vxyzw3 = wasm_v32x4_shuffle(vxy_hi, vzw_hi, 2, 3, 6, 7); in xnn_x32_zip_x4_ukernel__wasmsimd()
|
H A D | x4-sse2.c | 40 const __m128i vzw_hi = _mm_unpackhi_epi32(vz, vw); in xnn_x32_zip_x4_ukernel__sse2() local 44 const __m128i vxyzw2 = _mm_unpacklo_epi64(vxy_hi, vzw_hi); in xnn_x32_zip_x4_ukernel__sse2() 45 const __m128i vxyzw3 = _mm_unpackhi_epi64(vxy_hi, vzw_hi); in xnn_x32_zip_x4_ukernel__sse2()
|
H A D | xm-wasmsimd.c | 54 const v128_t vzw_hi = wasm_v32x4_shuffle(vz, vw, 2, 6, 3, 7); in xnn_x32_zip_xm_ukernel__wasmsimd() local 58 const v128_t vxyzw2 = wasm_v32x4_shuffle(vxy_hi, vzw_hi, 0, 1, 4, 5); in xnn_x32_zip_xm_ukernel__wasmsimd() 59 const v128_t vxyzw3 = wasm_v32x4_shuffle(vxy_hi, vzw_hi, 2, 3, 6, 7); in xnn_x32_zip_xm_ukernel__wasmsimd()
|
H A D | xm-sse2.c | 53 const __m128i vzw_hi = _mm_unpackhi_epi32(vz, vw); in xnn_x32_zip_xm_ukernel__sse2() local 57 const __m128i vxyzw2 = _mm_unpacklo_epi64(vxy_hi, vzw_hi); in xnn_x32_zip_xm_ukernel__sse2() 58 const __m128i vxyzw3 = _mm_unpackhi_epi64(vxy_hi, vzw_hi); in xnn_x32_zip_xm_ukernel__sse2()
|
/aosp_15_r20/external/XNNPACK/src/amalgam/ |
H A D | sse2.c | 13059 const __m128i vzw_hi = _mm_unpackhi_epi32(vz, vw); in xnn_x32_zip_x4_ukernel__sse2() local 13063 const __m128i vxyzw2 = _mm_unpacklo_epi64(vxy_hi, vzw_hi); in xnn_x32_zip_x4_ukernel__sse2() 13064 const __m128i vxyzw3 = _mm_unpackhi_epi64(vxy_hi, vzw_hi); in xnn_x32_zip_x4_ukernel__sse2() 13147 const __m128i vzw_hi = _mm_unpackhi_epi32(vz, vw); in xnn_x32_zip_xm_ukernel__sse2() local 13151 const __m128i vxyzw2 = _mm_unpacklo_epi64(vxy_hi, vzw_hi); in xnn_x32_zip_xm_ukernel__sse2() 13152 const __m128i vxyzw3 = _mm_unpackhi_epi64(vxy_hi, vzw_hi); in xnn_x32_zip_xm_ukernel__sse2() 14129 const __m128i vzw_hi = _mm_unpackhi_epi8(vz, vw); in xnn_x8_zip_x4_ukernel__sse2() local 14132 const __m128i vxyzw2 = _mm_unpacklo_epi16(vxy_hi, vzw_hi); in xnn_x8_zip_x4_ukernel__sse2() 14133 const __m128i vxyzw3 = _mm_unpackhi_epi16(vxy_hi, vzw_hi); in xnn_x8_zip_x4_ukernel__sse2() 14150 const __m128i vzw_hi = _mm_unpackhi_epi8(vz, vw); in xnn_x8_zip_x4_ukernel__sse2() local [all …]
|