Home
last modified time | relevance | path

Searched refs:vprod01234567lo (Results 1 – 25 of 27) sorted by relevance

12

/aosp_15_r20/external/XNNPACK/src/qs8-vmulc/gen/
H A Dminmax-fp32-sse41-mul16-ld64-x8.c42 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
45 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
46 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
76 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
79 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
80 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
H A Dminmax-fp32-sse2-mul16-ld64-x8.c43 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
46 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
47 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
78 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
81 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
82 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
H A Dminmax-fp32-avx-mul16-ld64-x8.c42 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
45 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
46 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
76 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
79 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
80 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
H A Dminmax-fp32-avx-mul16-ld64-x16.c44 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16() local
49 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16()
50 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16()
90 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16() local
93 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16()
94 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16()
H A Dminmax-fp32-sse41-mul16-ld64-x16.c44 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16() local
49 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
50 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
90 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16() local
93 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
94 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
H A Dminmax-fp32-sse2-mul16-ld64-x16.c46 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x16() local
51 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
52 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
95 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x16() local
98 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
99 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
/aosp_15_r20/external/XNNPACK/src/qu8-vmulc/gen/
H A Dminmax-fp32-sse41-mul16-ld64-x8.c42 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
45 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
46 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
76 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
79 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
80 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
H A Dminmax-fp32-avx-mul16-ld64-x8.c42 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
45 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
46 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
76 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
79 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
80 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
H A Dminmax-fp32-sse2-mul16-ld64-x8.c44 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
47 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
48 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
80 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
83 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
84 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
H A Dminmax-fp32-avx-mul16-ld64-x16.c44 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16() local
49 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16()
50 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16()
90 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16() local
93 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16()
94 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16()
H A Dminmax-fp32-sse2-mul16-ld64-x16.c47 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x16() local
52 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
53 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
95 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x16() local
98 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
99 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
H A Dminmax-fp32-sse41-mul16-ld64-x16.c44 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16() local
49 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
50 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
90 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16() local
93 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
94 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
/aosp_15_r20/external/XNNPACK/src/qs8-vmul/gen/
H A Dminmax-fp32-sse41-mul16-ld64-x8.c43 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
46 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
47 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
79 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
82 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
83 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
H A Dminmax-fp32-avx-mul16-ld64-x8.c43 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
46 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
47 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
79 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
82 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
83 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
H A Dminmax-fp32-sse2-mul16-ld64-x8.c45 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
48 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
49 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
83 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
86 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
87 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
H A Dminmax-fp32-avx-mul16-ld64-x16.c47 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16() local
52 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16()
53 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16()
96 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16() local
99 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16()
100 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16()
H A Dminmax-fp32-sse41-mul16-ld64-x16.c47 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16() local
52 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
53 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
96 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16() local
99 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
100 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
H A Dminmax-fp32-sse2-mul16-ld64-x16.c51 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16() local
56 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
57 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
104 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16() local
107 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
108 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
/aosp_15_r20/external/XNNPACK/src/qu8-vmul/gen/
H A Dminmax-fp32-sse41-mul16-ld64-x8.c43 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
46 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
47 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
79 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
82 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
83 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
H A Dminmax-fp32-avx-mul16-ld64-x8.c43 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
46 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
47 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
79 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
82 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
83 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
H A Dminmax-fp32-sse2-mul16-ld64-x8.c46 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
49 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
50 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
85 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
88 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
89 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
H A Dminmax-fp32-sse41-mul16-ld64-x16.c47 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16() local
52 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
53 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
96 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16() local
99 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
100 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
H A Dminmax-fp32-avx-mul16-ld64-x16.c47 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16() local
52 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16()
53 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16()
96 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16() local
99 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16()
100 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16()
H A Dminmax-fp32-sse2-mul16-ld64-x16.c52 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16() local
57 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
58 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
104 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16() local
107 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
108 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x16()
/aosp_15_r20/external/XNNPACK/src/amalgam/
H A Dsse41.c5476 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16() local
5481 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
5482 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
5525 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16() local
5528 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
5529 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
5596 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16() local
5601 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
5602 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
5642 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16() local
[all …]

12