| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 4x4c2-minmax-ssse3-ld128.c | 200 const __m128i vabsacc3x0123 = _mm_abs_epi32(vacc3x0123); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128() local
205 const __m128i vabsacc3x1133 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128()
210 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128()
|
| D | 4x4c2-minmax-sse2-ld64.c | 200 …const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64() local
205 const __m128i vabsacc3x1133 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
210 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
|
| D | 4x4c2-minmax-ssse3-ld64.c | 200 const __m128i vabsacc3x0123 = _mm_abs_epi32(vacc3x0123); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64() local
205 const __m128i vabsacc3x1133 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64()
210 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64()
|
| D | 4x4c2-xw-minmax-ssse3.c | 193 const __m128i vabsacc3x0123 = _mm_abs_epi32(vacc3x0123); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3() local
198 const __m128i vabsacc3x1133 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3()
203 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3()
|
| D | 4x4c2-minmax-sse2-ld128.c | 200 …const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128() local
205 const __m128i vabsacc3x1133 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128()
210 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128()
|
| D | 4x4c2-xw-minmax-sse2.c | 193 …const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2() local
198 const __m128i vabsacc3x1133 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2()
203 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2()
|
| /external/XNNPACK/src/qu8-gemm/ |
| D | 4x4c2-minmax-sse2.c | 205 …const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2() local
210 const __m128i vabsacc3x1032 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(2, 3, 0, 1)); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
215 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
|
| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 4x4c2-minmax-sse2-ld128.c | 219 …const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128() local
224 const __m128i vabsacc3x1133 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128()
229 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128()
|
| D | 4x4c2-minmax-ssse3-ld128.c | 219 const __m128i vabsacc3x0123 = _mm_abs_epi32(vacc3x0123); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128() local
224 const __m128i vabsacc3x1133 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128()
229 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128()
|
| D | 4x4c2-minmax-ssse3-ld64.c | 219 const __m128i vabsacc3x0123 = _mm_abs_epi32(vacc3x0123); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64() local
224 const __m128i vabsacc3x1133 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
229 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
|
| D | 4x4c2-minmax-sse2-ld64.c | 219 …const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld64() local
224 const __m128i vabsacc3x1133 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld64()
229 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld64()
|
| /external/XNNPACK/src/qu8-igemm/ |
| D | 4x4c2-minmax-sse2.c | 189 …const __m128i vabsacc3x0123 = _mm_sub_epi32(_mm_xor_si128(vacc3x0123, vnmask3x0123), vnmask3x0123); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2() local
194 const __m128i vabsacc3x1032 = _mm_shuffle_epi32(vabsacc3x0123, _MM_SHUFFLE(2, 3, 0, 1)); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2()
199 const __m128i vabsprod3x02 = _mm_mul_epu32(vabsacc3x0123, vmultiplier); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2()
|