| /external/rust/crates/quiche/deps/boringssl/src/crypto/poly1305/ |
| D | poly1305_vec.c | 232 st->H[1] = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_first_block()
233 T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12)); in poly1305_first_block()
235 st->H[3] = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_first_block()
236 st->H[4] = _mm_or_si128(_mm_srli_epi64(T6, 40), HIBIT); in poly1305_first_block()
312 M1 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_blocks()
313 T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12)); in poly1305_blocks()
315 M3 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_blocks()
316 M4 = _mm_or_si128(_mm_srli_epi64(T6, 40), HIBIT); in poly1305_blocks()
376 M1 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_blocks()
377 T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12)); in poly1305_blocks()
[all …]
|
| /external/boringssl/src/crypto/poly1305/ |
| D | poly1305_vec.c | 236 st->H[1] = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_first_block()
237 T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12)); in poly1305_first_block()
239 st->H[3] = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_first_block()
240 st->H[4] = _mm_or_si128(_mm_srli_epi64(T6, 40), HIBIT); in poly1305_first_block()
316 M1 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_blocks()
317 T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12)); in poly1305_blocks()
319 M3 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_blocks()
320 M4 = _mm_or_si128(_mm_srli_epi64(T6, 40), HIBIT); in poly1305_blocks()
380 M1 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_blocks()
381 T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12)); in poly1305_blocks()
[all …]
|
| /external/rust/crates/ring/crypto/poly1305/ |
| D | poly1305_vec.c | 240 st->H[1] = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_first_block()
241 T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12)); in poly1305_first_block()
243 st->H[3] = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_first_block()
244 st->H[4] = _mm_or_si128(_mm_srli_epi64(T6, 40), HIBIT); in poly1305_first_block()
320 M1 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_blocks()
321 T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12)); in poly1305_blocks()
323 M3 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_blocks()
324 M4 = _mm_or_si128(_mm_srli_epi64(T6, 40), HIBIT); in poly1305_blocks()
384 M1 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26)); in poly1305_blocks()
385 T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12)); in poly1305_blocks()
[all …]
|
| /external/XNNPACK/src/qu8-requantization/ |
| D | q31-ssse3.c | 117 const __m128i x_q31product_even = _mm_srli_epi64(x_rounded_product_even, 31); in xnn_qu8_requantize_q31__ssse3()
118 const __m128i x_q31product_odd = _mm_srli_epi64(x_rounded_product_odd, 31); in xnn_qu8_requantize_q31__ssse3()
119 const __m128i y_q31product_even = _mm_srli_epi64(y_rounded_product_even, 31); in xnn_qu8_requantize_q31__ssse3()
120 const __m128i y_q31product_odd = _mm_srli_epi64(y_rounded_product_odd, 31); in xnn_qu8_requantize_q31__ssse3()
121 const __m128i z_q31product_even = _mm_srli_epi64(z_rounded_product_even, 31); in xnn_qu8_requantize_q31__ssse3()
122 const __m128i z_q31product_odd = _mm_srli_epi64(z_rounded_product_odd, 31); in xnn_qu8_requantize_q31__ssse3()
123 const __m128i w_q31product_even = _mm_srli_epi64(w_rounded_product_even, 31); in xnn_qu8_requantize_q31__ssse3()
124 const __m128i w_q31product_odd = _mm_srli_epi64(w_rounded_product_odd, 31); in xnn_qu8_requantize_q31__ssse3()
|
| D | q31-sse2.c | 117 const __m128i x_q31product_even = _mm_srli_epi64(x_rounded_product_even, 31); in xnn_qu8_requantize_q31__sse2()
118 const __m128i x_q31product_odd = _mm_srli_epi64(x_rounded_product_odd, 31); in xnn_qu8_requantize_q31__sse2()
119 const __m128i y_q31product_even = _mm_srli_epi64(y_rounded_product_even, 31); in xnn_qu8_requantize_q31__sse2()
120 const __m128i y_q31product_odd = _mm_srli_epi64(y_rounded_product_odd, 31); in xnn_qu8_requantize_q31__sse2()
121 const __m128i z_q31product_even = _mm_srli_epi64(z_rounded_product_even, 31); in xnn_qu8_requantize_q31__sse2()
122 const __m128i z_q31product_odd = _mm_srli_epi64(z_rounded_product_odd, 31); in xnn_qu8_requantize_q31__sse2()
123 const __m128i w_q31product_even = _mm_srli_epi64(w_rounded_product_even, 31); in xnn_qu8_requantize_q31__sse2()
124 const __m128i w_q31product_odd = _mm_srli_epi64(w_rounded_product_odd, 31); in xnn_qu8_requantize_q31__sse2()
|
| D | q31-sse4.c | 77 const __m128i x_q31product_even = _mm_srli_epi64(x_product_even, 31); in xnn_qu8_requantize_q31__sse4()
79 const __m128i y_q31product_even = _mm_srli_epi64(y_product_even, 31); in xnn_qu8_requantize_q31__sse4()
81 const __m128i z_q31product_even = _mm_srli_epi64(z_product_even, 31); in xnn_qu8_requantize_q31__sse4()
83 const __m128i w_q31product_even = _mm_srli_epi64(w_product_even, 31); in xnn_qu8_requantize_q31__sse4()
|
| /external/XNNPACK/src/qs8-requantization/ |
| D | q31-ssse3.c | 117 const __m128i x_q31product_even = _mm_srli_epi64(x_rounded_product_even, 31); in xnn_qs8_requantize_q31__ssse3()
118 const __m128i x_q31product_odd = _mm_srli_epi64(x_rounded_product_odd, 31); in xnn_qs8_requantize_q31__ssse3()
119 const __m128i y_q31product_even = _mm_srli_epi64(y_rounded_product_even, 31); in xnn_qs8_requantize_q31__ssse3()
120 const __m128i y_q31product_odd = _mm_srli_epi64(y_rounded_product_odd, 31); in xnn_qs8_requantize_q31__ssse3()
121 const __m128i z_q31product_even = _mm_srli_epi64(z_rounded_product_even, 31); in xnn_qs8_requantize_q31__ssse3()
122 const __m128i z_q31product_odd = _mm_srli_epi64(z_rounded_product_odd, 31); in xnn_qs8_requantize_q31__ssse3()
123 const __m128i w_q31product_even = _mm_srli_epi64(w_rounded_product_even, 31); in xnn_qs8_requantize_q31__ssse3()
124 const __m128i w_q31product_odd = _mm_srli_epi64(w_rounded_product_odd, 31); in xnn_qs8_requantize_q31__ssse3()
|
| D | q31-sse2.c | 117 const __m128i x_q31product_even = _mm_srli_epi64(x_rounded_product_even, 31); in xnn_qs8_requantize_q31__sse2()
118 const __m128i x_q31product_odd = _mm_srli_epi64(x_rounded_product_odd, 31); in xnn_qs8_requantize_q31__sse2()
119 const __m128i y_q31product_even = _mm_srli_epi64(y_rounded_product_even, 31); in xnn_qs8_requantize_q31__sse2()
120 const __m128i y_q31product_odd = _mm_srli_epi64(y_rounded_product_odd, 31); in xnn_qs8_requantize_q31__sse2()
121 const __m128i z_q31product_even = _mm_srli_epi64(z_rounded_product_even, 31); in xnn_qs8_requantize_q31__sse2()
122 const __m128i z_q31product_odd = _mm_srli_epi64(z_rounded_product_odd, 31); in xnn_qs8_requantize_q31__sse2()
123 const __m128i w_q31product_even = _mm_srli_epi64(w_rounded_product_even, 31); in xnn_qs8_requantize_q31__sse2()
124 const __m128i w_q31product_odd = _mm_srli_epi64(w_rounded_product_odd, 31); in xnn_qs8_requantize_q31__sse2()
|
| D | q31-sse4.c | 77 const __m128i x_q31product_even = _mm_srli_epi64(x_product_even, 31); in xnn_qs8_requantize_q31__sse4()
79 const __m128i y_q31product_even = _mm_srli_epi64(y_product_even, 31); in xnn_qs8_requantize_q31__sse4()
81 const __m128i z_q31product_even = _mm_srli_epi64(z_product_even, 31); in xnn_qs8_requantize_q31__sse4()
83 const __m128i w_q31product_even = _mm_srli_epi64(w_product_even, 31); in xnn_qs8_requantize_q31__sse4()
|
| /external/webp/src/dsp/ |
| D | rescaler_sse2.c | 152 const __m128i D0 = _mm_srli_epi64(frac, 32); in RescalerImportRowShrink_SSE2()
179 const __m128i A2 = _mm_srli_epi64(A0, 32); in LoadDispatchAndMult_SSE2()
180 const __m128i A3 = _mm_srli_epi64(A1, 32); in LoadDispatchAndMult_SSE2()
210 const __m128i D0 = _mm_srli_epi64(C0, WEBP_RESCALER_RFIX); in ProcessRow_SSE2()
211 const __m128i D1 = _mm_srli_epi64(C1, WEBP_RESCALER_RFIX); in ProcessRow_SSE2()
269 const __m128i E0 = _mm_srli_epi64(D0, WEBP_RESCALER_RFIX); in RescalerExportRowExpand_SSE2()
270 const __m128i E1 = _mm_srli_epi64(D1, WEBP_RESCALER_RFIX); in RescalerExportRowExpand_SSE2()
271 const __m128i E2 = _mm_srli_epi64(D2, WEBP_RESCALER_RFIX); in RescalerExportRowExpand_SSE2()
272 const __m128i E3 = _mm_srli_epi64(D3, WEBP_RESCALER_RFIX); in RescalerExportRowExpand_SSE2()
305 const __m128i D0 = _mm_srli_epi64(B0, WEBP_RESCALER_RFIX); // = frac in RescalerExportRowShrink_SSE2()
[all …]
|
| /external/libopus/silk/fixed/x86/ |
| D | prefilter_FIX_sse.c | 97 xmm_product1 = _mm_srli_epi64( xmm_product1, 16 ); /* >> 16, zero extending works */ in silk_warped_LPC_analysis_filter_FIX_sse4_1()
98 xmm_product2 = _mm_srli_epi64( xmm_product2, 16 ); in silk_warped_LPC_analysis_filter_FIX_sse4_1()
103 xmm_tempa = _mm_srli_epi64( xmm_tempa, 16 ); in silk_warped_LPC_analysis_filter_FIX_sse4_1()
104 xmm_tempb = _mm_srli_epi64( xmm_tempb, 16 ); in silk_warped_LPC_analysis_filter_FIX_sse4_1()
|
| /external/XNNPACK/src/qu8-gemm/ |
| D | 4x4c2-minmax-sse2.c | 227 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
228 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
229 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
230 const __m128i vq31prod3x02 = _mm_srli_epi64(_mm_add_epi64(vprod3x02, vrounding), 31); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
247 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
248 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
249 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
250 const __m128i vq31prod3x13 = _mm_srli_epi64(_mm_add_epi64(vprod3x13, vrounding), 31); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
298 *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(_mm_srli_epi64(vout, 32)); in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
|
| D | 2x4c8-minmax-sse2.c | 132 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31); in xnn_qu8_gemm_minmax_ukernel_2x4c8__sse2()
133 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31); in xnn_qu8_gemm_minmax_ukernel_2x4c8__sse2()
144 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31); in xnn_qu8_gemm_minmax_ukernel_2x4c8__sse2()
145 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qu8_gemm_minmax_ukernel_2x4c8__sse2()
178 *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(_mm_srli_epi64(vout, 32)); in xnn_qu8_gemm_minmax_ukernel_2x4c8__sse2()
|
| /external/XNNPACK/src/qu8-igemm/ |
| D | 4x4c2-minmax-sse2.c | 211 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2()
212 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2()
213 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2()
214 const __m128i vq31prod3x02 = _mm_srli_epi64(_mm_add_epi64(vprod3x02, vrounding), 31); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2()
231 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2()
232 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2()
233 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2()
234 const __m128i vq31prod3x13 = _mm_srli_epi64(_mm_add_epi64(vprod3x13, vrounding), 31); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2()
281 *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(_mm_srli_epi64(vout, 32)); in xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2()
|
| /external/pdfium/third_party/libopenjpeg20/ |
| D | mct.c | 250 lo = _mm_srli_epi64(lo, 13); in opj_mct_encode_real()
260 lo = _mm_srli_epi64(lo, 13); in opj_mct_encode_real()
270 lo = _mm_srli_epi64(lo, 13); in opj_mct_encode_real()
285 lo = _mm_srli_epi64(lo, 13); in opj_mct_encode_real()
295 lo = _mm_srli_epi64(lo, 13); in opj_mct_encode_real()
305 lo = _mm_srli_epi64(lo, 13); in opj_mct_encode_real()
320 lo = _mm_srli_epi64(lo, 13); in opj_mct_encode_real()
330 lo = _mm_srli_epi64(lo, 13); in opj_mct_encode_real()
340 lo = _mm_srli_epi64(lo, 13); in opj_mct_encode_real()
|
| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 3x4c8-xw-minmax-ssse3.c | 145 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3()
146 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3()
147 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3()
161 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3()
162 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3()
163 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3()
|
| D | 4x4c2-xw-minmax-sse2.c | 215 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2()
216 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2()
217 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2()
218 const __m128i vq31prod3x02 = _mm_srli_epi64(_mm_add_epi64(vprod3x02, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2()
235 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2()
236 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2()
237 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2()
238 const __m128i vq31prod3x13 = _mm_srli_epi64(_mm_add_epi64(vprod3x13, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__sse2()
|
| D | 4x4c2-minmax-sse2-ld128.c | 222 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128()
223 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128()
224 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128()
225 const __m128i vq31prod3x02 = _mm_srli_epi64(_mm_add_epi64(vprod3x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128()
242 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128()
243 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128()
244 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128()
245 const __m128i vq31prod3x13 = _mm_srli_epi64(_mm_add_epi64(vprod3x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld128()
|
| D | 4x4c2-minmax-sse2-ld64.c | 222 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
223 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
224 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
225 const __m128i vq31prod3x02 = _mm_srli_epi64(_mm_add_epi64(vprod3x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
242 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
243 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
244 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
245 const __m128i vq31prod3x13 = _mm_srli_epi64(_mm_add_epi64(vprod3x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
|
| D | 4x4c2-minmax-ssse3-ld64.c | 222 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64()
223 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64()
224 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64()
225 const __m128i vq31prod3x02 = _mm_srli_epi64(_mm_add_epi64(vprod3x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64()
242 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64()
243 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64()
244 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64()
245 const __m128i vq31prod3x13 = _mm_srli_epi64(_mm_add_epi64(vprod3x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld64()
|
| D | 4x4c2-xw-minmax-ssse3.c | 215 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3()
216 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3()
217 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3()
218 const __m128i vq31prod3x02 = _mm_srli_epi64(_mm_add_epi64(vprod3x02, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3()
235 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3()
236 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3()
237 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3()
238 const __m128i vq31prod3x13 = _mm_srli_epi64(_mm_add_epi64(vprod3x13, vrounding), 31); in xnn_qs8_gemm_xw_minmax_ukernel_4x4c2__ssse3()
|
| D | 4x4c2-minmax-ssse3-ld128.c | 222 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128()
223 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128()
224 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128()
225 const __m128i vq31prod3x02 = _mm_srli_epi64(_mm_add_epi64(vprod3x02, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128()
242 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128()
243 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128()
244 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128()
245 const __m128i vq31prod3x13 = _mm_srli_epi64(_mm_add_epi64(vprod3x13, vrounding), 31); in xnn_qs8_gemm_minmax_ukernel_4x4c2__ssse3_ld128()
|
| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 4x4c2-minmax-sse2-ld128.c | 241 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128()
242 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128()
243 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128()
244 const __m128i vq31prod3x02 = _mm_srli_epi64(_mm_add_epi64(vprod3x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128()
261 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128()
262 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128()
263 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128()
264 const __m128i vq31prod3x13 = _mm_srli_epi64(_mm_add_epi64(vprod3x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__sse2_ld128()
|
| D | 4x4c2-minmax-ssse3-ld64.c | 241 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
242 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
243 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
244 const __m128i vq31prod3x02 = _mm_srli_epi64(_mm_add_epi64(vprod3x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
261 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
262 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
263 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
264 const __m128i vq31prod3x13 = _mm_srli_epi64(_mm_add_epi64(vprod3x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
|
| D | 4x4c2-minmax-ssse3-ld128.c | 241 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128()
242 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128()
243 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128()
244 const __m128i vq31prod3x02 = _mm_srli_epi64(_mm_add_epi64(vprod3x02, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128()
261 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128()
262 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128()
263 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128()
264 const __m128i vq31prod3x13 = _mm_srli_epi64(_mm_add_epi64(vprod3x13, vrounding), 31); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld128()
|