/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x4c8-xw-minmax-ssse3.c | 87 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_xw_minmax_ukernel_1x4c8__ssse3() local 91 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_xw_minmax_ukernel_1x4c8__ssse3()
|
D | 1x4c8-minmax-ssse3-ld128.c | 89 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_1x4c8__ssse3_ld128() local 93 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_minmax_ukernel_1x4c8__ssse3_ld128()
|
D | 1x4c8-xw-minmax-sse2.c | 87 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_xw_minmax_ukernel_1x4c8__sse2() local 91 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_xw_minmax_ukernel_1x4c8__sse2()
|
D | 1x4c8-minmax-ssse3-ld64.c | 91 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_1x4c8__ssse3_ld64() local 95 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_minmax_ukernel_1x4c8__ssse3_ld64()
|
D | 1x4c8-minmax-sse2-ld128.c | 89 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_1x4c8__sse2_ld128() local 93 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_minmax_ukernel_1x4c8__sse2_ld128()
|
D | 1x4c8-minmax-sse2-ld64.c | 91 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_1x4c8__sse2_ld64() local 95 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_minmax_ukernel_1x4c8__sse2_ld64()
|
D | 1x4c2-xw-minmax-sse2.c | 110 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_xw_minmax_ukernel_1x4c2__sse2() local 114 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_xw_minmax_ukernel_1x4c2__sse2()
|
D | 1x4c2-xw-minmax-ssse3.c | 110 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_xw_minmax_ukernel_1x4c2__ssse3() local 114 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_xw_minmax_ukernel_1x4c2__ssse3()
|
D | 1x4c2-minmax-sse2-ld64.c | 117 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_1x4c2__sse2_ld64() local 121 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_minmax_ukernel_1x4c2__sse2_ld64()
|
D | 1x4c2-minmax-ssse3-ld64.c | 117 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_1x4c2__ssse3_ld64() local 121 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_minmax_ukernel_1x4c2__ssse3_ld64()
|
D | 1x4c2-minmax-ssse3-ld128.c | 117 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_1x4c2__ssse3_ld128() local 121 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_minmax_ukernel_1x4c2__ssse3_ld128()
|
D | 1x4c2-minmax-sse2-ld128.c | 117 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_1x4c2__sse2_ld128() local 121 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_minmax_ukernel_1x4c2__sse2_ld128()
|
D | 2x4c8-minmax-ssse3-ld128.c | 112 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_2x4c8__ssse3_ld128() local 118 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_minmax_ukernel_2x4c8__ssse3_ld128()
|
D | 2x4c8-xw-minmax-sse2.c | 110 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse2() local 116 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse2()
|
D | 2x4c8-minmax-ssse3-ld64.c | 114 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_2x4c8__ssse3_ld64() local 120 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_minmax_ukernel_2x4c8__ssse3_ld64()
|
D | 2x4c8-xw-minmax-ssse3.c | 110 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__ssse3() local 116 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__ssse3()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x4c8-minmax-ssse3-ld128.c | 102 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_1x4c8__ssse3_ld128() local 106 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_igemm_minmax_ukernel_1x4c8__ssse3_ld128()
|
D | 1x4c8-minmax-sse2-ld64.c | 104 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_1x4c8__sse2_ld64() local 108 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_igemm_minmax_ukernel_1x4c8__sse2_ld64()
|
D | 1x4c8-minmax-sse2-ld128.c | 102 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_1x4c8__sse2_ld128() local 106 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_igemm_minmax_ukernel_1x4c8__sse2_ld128()
|
D | 1x4c8-minmax-ssse3-ld64.c | 104 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_1x4c8__ssse3_ld64() local 108 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_igemm_minmax_ukernel_1x4c8__ssse3_ld64()
|
D | 1x4c2-minmax-ssse3-ld128.c | 130 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_1x4c2__ssse3_ld128() local 134 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_igemm_minmax_ukernel_1x4c2__ssse3_ld128()
|
D | 1x4c2-minmax-sse2-ld128.c | 130 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_1x4c2__sse2_ld128() local 134 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_igemm_minmax_ukernel_1x4c2__sse2_ld128()
|
D | 1x4c2-minmax-ssse3-ld64.c | 130 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_1x4c2__ssse3_ld64() local 134 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_igemm_minmax_ukernel_1x4c2__ssse3_ld64()
|
D | 1x4c2-minmax-sse2-ld64.c | 130 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_1x4c2__sse2_ld64() local 134 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_igemm_minmax_ukernel_1x4c2__sse2_ld64()
|
D | 2x4c8-minmax-ssse3-ld128.c | 127 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_2x4c8__ssse3_ld128() local 133 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02); in xnn_qs8_igemm_minmax_ukernel_2x4c8__ssse3_ld128()
|