| /external/XNNPACK/src/x8-zip/ |
| D | xm-sse2.c | 54 *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0); in xnn_x8_zip_xm_ukernel__sse2()
57 *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0); in xnn_x8_zip_xm_ukernel__sse2()
60 *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0); in xnn_x8_zip_xm_ukernel__sse2()
63 *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0); in xnn_x8_zip_xm_ukernel__sse2()
66 *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw1); in xnn_x8_zip_xm_ukernel__sse2()
69 *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw1); in xnn_x8_zip_xm_ukernel__sse2()
72 *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw1); in xnn_x8_zip_xm_ukernel__sse2()
75 *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw1); in xnn_x8_zip_xm_ukernel__sse2()
78 *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw2); in xnn_x8_zip_xm_ukernel__sse2()
81 *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw2); in xnn_x8_zip_xm_ukernel__sse2()
[all …]
|
| /external/libmpeg2/common/x86/ |
| D | ideint_cac_ssse3.c | 136 sum_1 = _mm_cvtsi128_si32(sum_t[i + 0]); in ideint_cac_8x8_ssse3()
137 sum_2 = _mm_cvtsi128_si32(sum_b[i + 0]); in ideint_cac_8x8_ssse3()
142 sum_3 = _mm_cvtsi128_si32(sum_t[i + 1]); in ideint_cac_8x8_ssse3()
143 sum_4 = _mm_cvtsi128_si32(sum_b[i + 1]); in ideint_cac_8x8_ssse3()
151 sum_1 = _mm_cvtsi128_si32(_mm_srli_si128(sum_t[i + 0], 8)); in ideint_cac_8x8_ssse3()
152 sum_2 = _mm_cvtsi128_si32(_mm_srli_si128(sum_b[i + 0], 8)); in ideint_cac_8x8_ssse3()
157 sum_3 = _mm_cvtsi128_si32(_mm_srli_si128(sum_t[i + 1], 8)); in ideint_cac_8x8_ssse3()
158 sum_4 = _mm_cvtsi128_si32(_mm_srli_si128(sum_b[i + 1], 8)); in ideint_cac_8x8_ssse3()
221 diff_sum = _mm_cvtsi128_si32(diff); in ideint_cac_8x8_ssse3()
224 diff_sum = _mm_cvtsi128_si32(_mm_srli_si128(diff, 8)); in ideint_cac_8x8_ssse3()
|
| D | ideint_spatial_filter_ssse3.c | 161 adiff[0] = _mm_cvtsi128_si32(diffs[0]); in ideint_spatial_filter_ssse3()
162 adiff[1] = _mm_cvtsi128_si32(diffs[1]); in ideint_spatial_filter_ssse3()
163 adiff[2] = _mm_cvtsi128_si32(diffs[2]); in ideint_spatial_filter_ssse3()
164 adiff[3] = _mm_cvtsi128_si32(_mm_srli_si128(diffs[0], 8)); in ideint_spatial_filter_ssse3()
165 adiff[4] = _mm_cvtsi128_si32(_mm_srli_si128(diffs[1], 8)); in ideint_spatial_filter_ssse3()
166 adiff[5] = _mm_cvtsi128_si32(_mm_srli_si128(diffs[2], 8)); in ideint_spatial_filter_ssse3()
|
| /external/libaom/libaom/aom_dsp/x86/ |
| D | jnt_sad_ssse3.c | 57 _mm_cvtsi128_si32(sad) + _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)); in aom_sad4xh_sse2()
88 _mm_cvtsi128_si32(sad) + _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)); in aom_sad8xh_sse2()
112 _mm_cvtsi128_si32(sad) + _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)); in aom_sad16xh_sse2()
138 _mm_cvtsi128_si32(sad) + _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)); in aom_sad32xh_sse2()
164 _mm_cvtsi128_si32(sad) + _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)); in aom_sad64xh_sse2()
190 _mm_cvtsi128_si32(sad) + _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)); in aom_sad128xh_sse2()
|
| D | intrapred_sse2.c | 115 uint32_t sum = _mm_cvtsi128_si32(sum_above); in aom_dc_predictor_4x8_sse2()
120 const uint32_t pred = _mm_cvtsi128_si32(row); in aom_dc_predictor_4x8_sse2()
130 uint32_t sum = _mm_cvtsi128_si32(sum_above); in aom_dc_predictor_4x16_sse2()
135 const uint32_t pred = _mm_cvtsi128_si32(row); in aom_dc_predictor_4x16_sse2()
145 uint32_t sum = _mm_cvtsi128_si32(sum_above); in aom_dc_predictor_8x4_sse2()
159 uint32_t sum = _mm_cvtsi128_si32(sum_above); in aom_dc_predictor_8x16_sse2()
172 uint32_t sum = _mm_cvtsi128_si32(sum_above); in aom_dc_predictor_8x32_sse2()
185 uint32_t sum = _mm_cvtsi128_si32(sum_above); in aom_dc_predictor_16x4_sse2()
198 uint32_t sum = _mm_cvtsi128_si32(sum_above); in aom_dc_predictor_16x8_sse2()
211 uint32_t sum = _mm_cvtsi128_si32(sum_above); in aom_dc_predictor_16x32_sse2()
[all …]
|
| D | sad_avx2.c | 41 res = _mm_cvtsi128_si32(sum_sad128); \
74 res = _mm_cvtsi128_si32(sum_sad128); \
128 res = _mm_cvtsi128_si32(sum_sad128); \
167 res = _mm_cvtsi128_si32(sum_sad128); \
|
| D | obmc_variance_avx2.c | 79 *sum = _mm_cvtsi128_si32(v_d); in obmc_variance_w8n()
80 *sse = _mm_cvtsi128_si32(_mm_srli_si128(v_d, 4)); in obmc_variance_w8n()
149 *sum = _mm_cvtsi128_si32(res0); in obmc_variance_w16n()
150 *sse = _mm_cvtsi128_si32(_mm_srli_si128(res0, 4)); in obmc_variance_w16n()
|
| D | masked_sad_intrin_ssse3.c | 136 _mm_cvtsi128_si32(res) + _mm_cvtsi128_si32(_mm_srli_si128(res, 8)); in masked_sad_ssse3()
181 _mm_cvtsi128_si32(res) + _mm_cvtsi128_si32(_mm_srli_si128(res, 8)); in aom_masked_sad8xh_ssse3()
225 int32_t sad = _mm_cvtsi128_si32(res); in aom_masked_sad4xh_ssse3()
340 int sad = _mm_cvtsi128_si32(res); in highbd_masked_sad_ssse3()
400 int sad = _mm_cvtsi128_si32(res); in aom_highbd_masked_sad4xh_ssse3()
|
| /external/neven/Embedded/common/src/b_BasicEm/ |
| D | MathSSE2.c | 99 resultL = _mm_cvtsi128_si32( m_XMM7 ); in bbs_dotProduct_64SSE2()
190 resultL = _mm_cvtsi128_si32( m_XMM0 ); /* 1st 32bits */ in bbs_dotProduct_128SSE2()
194 resultL += _mm_cvtsi128_si32( m_XMM0 ); /* 2nd 32bits */ in bbs_dotProduct_128SSE2()
198 resultL += _mm_cvtsi128_si32( m_XMM0 ); /* 3rd 32bits */ in bbs_dotProduct_128SSE2()
202 resultL += _mm_cvtsi128_si32( m_XMM0 ); /* 4th 32bits */ in bbs_dotProduct_128SSE2()
294 resultL = _mm_cvtsi128_si32( m_XMM0 ); /* 1st 32bits */ in bbs_dotProduct_u128SSE2()
298 resultL += _mm_cvtsi128_si32( m_XMM0 ); /* 2nd 32bits */ in bbs_dotProduct_u128SSE2()
302 resultL += _mm_cvtsi128_si32( m_XMM0 ); /* 3rd 32bits */ in bbs_dotProduct_u128SSE2()
306 resultL += _mm_cvtsi128_si32( m_XMM0 ); /* 4th 32bits */ in bbs_dotProduct_u128SSE2()
|
| /external/webp/src/dsp/ |
| D | dec_sse2.c | 190 WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(dst0)); in Transform_SSE2()
191 WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(dst1)); in Transform_SSE2()
192 WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(dst2)); in Transform_SSE2()
193 WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(dst3)); in Transform_SSE2()
236 WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(dst0)); in TransformAC3()
237 WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(dst1)); in TransformAC3()
238 WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(dst2)); in TransformAC3()
239 WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(dst3)); in TransformAC3()
543 WebPUint32ToMem(dst, _mm_cvtsi128_si32(*x)); in Store4x4_SSE2()
911 const uint32_t vals = _mm_cvtsi128_si32(avg); in VE4_SSE2()
[all …]
|
| D | enc_sse2.c | 188 WebPUint32ToMem(&dst[0 * BPS], _mm_cvtsi128_si32(ref0)); in ITransform_SSE2()
189 WebPUint32ToMem(&dst[1 * BPS], _mm_cvtsi128_si32(ref1)); in ITransform_SSE2()
190 WebPUint32ToMem(&dst[2 * BPS], _mm_cvtsi128_si32(ref2)); in ITransform_SSE2()
191 WebPUint32ToMem(&dst[3 * BPS], _mm_cvtsi128_si32(ref3)); in ITransform_SSE2()
633 const int DC = _mm_cvtsi128_si32(sum) + 4; in DC8uvNoLeft_SSE2()
725 const uint32_t vals = _mm_cvtsi128_si32(avg); in VE4_SSE2()
763 WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32( abcdefg )); in LD4_SSE2()
764 WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1))); in LD4_SSE2()
765 WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2))); in LD4_SSE2()
766 WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3))); in LD4_SSE2()
[all …]
|
| /external/libvpx/libvpx/vpx_dsp/x86/ |
| D | mem_sse2.h | 36 const uint32_t val = _mm_cvtsi128_si32(v); in store_unaligned_u32()
108 *(int *)(d + 0 * stride) = _mm_cvtsi128_si32(s[0]); in store_8bit_4x4()
109 *(int *)(d + 1 * stride) = _mm_cvtsi128_si32(s[1]); in store_8bit_4x4()
110 *(int *)(d + 2 * stride) = _mm_cvtsi128_si32(s[2]); in store_8bit_4x4()
111 *(int *)(d + 3 * stride) = _mm_cvtsi128_si32(s[3]); in store_8bit_4x4()
|
| D | sad_avx2.c | 38 res = _mm_cvtsi128_si32(sum_sad128); \
70 res = _mm_cvtsi128_si32(sum_sad128); \
121 res = _mm_cvtsi128_si32(sum_sad128); \
159 res = _mm_cvtsi128_si32(sum_sad128); \
|
| /external/libaom/libaom/aom_dsp/simd/ |
| D | v256_intrinsics_x86.h | 35 return (uint32_t)_mm_cvtsi128_si32(_mm256_extracti128_si256(a, 0)); in v256_low_u32()
399 return (int64_t)_mm_cvtsi128_si32(l) + in v256_dotp_s16()
400 (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 4)) + in v256_dotp_s16()
401 (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 8)) + in v256_dotp_s16()
402 (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 12)) + in v256_dotp_s16()
403 (int64_t)_mm_cvtsi128_si32(h) + in v256_dotp_s16()
404 (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 4)) + in v256_dotp_s16()
405 (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 8)) + in v256_dotp_s16()
406 (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 12)); in v256_dotp_s16()
422 return (int64_t)_mm_cvtsi128_si32(l) + in v256_dotp_s32()
[all …]
|
| /external/flac/src/libFLAC/ |
| D | fixed_intrin_ssse3.c | 108 total_error_0 = _mm_cvtsi128_si32(total_err0); in FLAC__fixed_compute_best_predictor_intrin_ssse3()
111 total_error_4 = _mm_cvtsi128_si32(total_err2); in FLAC__fixed_compute_best_predictor_intrin_ssse3()
112 total_error_2 = _mm_cvtsi128_si32(total_err1); in FLAC__fixed_compute_best_predictor_intrin_ssse3()
115 total_error_3 = _mm_cvtsi128_si32(total_err2); in FLAC__fixed_compute_best_predictor_intrin_ssse3()
116 total_error_1 = _mm_cvtsi128_si32(total_err1); in FLAC__fixed_compute_best_predictor_intrin_ssse3()
|
| D | fixed_intrin_sse2.c | 114 total_error_0 = _mm_cvtsi128_si32(total_err0); in FLAC__fixed_compute_best_predictor_intrin_sse2()
117 total_error_4 = _mm_cvtsi128_si32(total_err2); in FLAC__fixed_compute_best_predictor_intrin_sse2()
118 total_error_2 = _mm_cvtsi128_si32(total_err1); in FLAC__fixed_compute_best_predictor_intrin_sse2()
121 total_error_3 = _mm_cvtsi128_si32(total_err2); in FLAC__fixed_compute_best_predictor_intrin_sse2()
122 total_error_1 = _mm_cvtsi128_si32(total_err1); in FLAC__fixed_compute_best_predictor_intrin_sse2()
|
| /external/libhevc/common/x86/ |
| D | ihevc_chroma_intra_pred_filters_sse42_intr.c | 345 acc_dc_u = _mm_cvtsi128_si32(src_temp4); in ihevc_intra_pred_chroma_dc_sse42()
346 acc_dc_v = _mm_cvtsi128_si32(temp_sad); in ihevc_intra_pred_chroma_dc_sse42()
374 acc_dc_u = _mm_cvtsi128_si32(src_temp4); in ihevc_intra_pred_chroma_dc_sse42()
375 acc_dc_v = _mm_cvtsi128_si32(temp_sad); in ihevc_intra_pred_chroma_dc_sse42()
395 acc_dc_u = _mm_cvtsi128_si32(src_temp4); in ihevc_intra_pred_chroma_dc_sse42()
396 acc_dc_v = _mm_cvtsi128_si32(temp_sad); in ihevc_intra_pred_chroma_dc_sse42()
|
| D | ihevc_deblk_ssse3_intr.c | 187 d0 = _mm_cvtsi128_si32(mask_d_result_4x32b); in ihevc_deblk_luma_vert_ssse3()
188 d3 = _mm_cvtsi128_si32(temp_coef0_8x16b); in ihevc_deblk_luma_vert_ssse3()
189 dp = _mm_cvtsi128_si32(temp_coef1_8x16b); in ihevc_deblk_luma_vert_ssse3()
190 dq = _mm_cvtsi128_si32(mask_16x8b); in ihevc_deblk_luma_vert_ssse3()
199 de_0 = _mm_cvtsi128_si32(mask_de_result_8x16b); in ihevc_deblk_luma_vert_ssse3()
200 de_1 = _mm_cvtsi128_si32(temp_coef0_8x16b); in ihevc_deblk_luma_vert_ssse3()
201 de_2 = _mm_cvtsi128_si32(temp_coef1_8x16b); in ihevc_deblk_luma_vert_ssse3()
202 de_3 = _mm_cvtsi128_si32(mask_16x8b); in ihevc_deblk_luma_vert_ssse3()
647 d0 = _mm_cvtsi128_si32(mask_d_result_4x32b); in ihevc_deblk_luma_horz_ssse3()
648 d3 = _mm_cvtsi128_si32(temp_coef0_8x16b); in ihevc_deblk_luma_horz_ssse3()
[all …]
|
| /external/libavc/common/x86/ |
| D | ih264_luma_intra_pred_filters_ssse3.c | 357 row1 = _mm_cvtsi128_si32(res_16x8b); in ih264_intra_pred_luma_4x4_mode_diag_dl_ssse3()
359 row2 = _mm_cvtsi128_si32(res_16x8b); in ih264_intra_pred_luma_4x4_mode_diag_dl_ssse3()
361 row3 = _mm_cvtsi128_si32(res_16x8b); in ih264_intra_pred_luma_4x4_mode_diag_dl_ssse3()
363 row4 = _mm_cvtsi128_si32(res_16x8b); in ih264_intra_pred_luma_4x4_mode_diag_dl_ssse3()
445 row1 = _mm_cvtsi128_si32(res2_16x8b); in ih264_intra_pred_luma_4x4_mode_diag_dr_ssse3()
447 row2 = _mm_cvtsi128_si32(res2_16x8b); in ih264_intra_pred_luma_4x4_mode_diag_dr_ssse3()
449 row3 = _mm_cvtsi128_si32(res2_16x8b); in ih264_intra_pred_luma_4x4_mode_diag_dr_ssse3()
450 row4 = _mm_cvtsi128_si32(res1_16x8b); in ih264_intra_pred_luma_4x4_mode_diag_dr_ssse3()
539 row1 = _mm_cvtsi128_si32(row1_16x8b); in ih264_intra_pred_luma_4x4_mode_vert_r_ssse3()
540 row2 = _mm_cvtsi128_si32(row2_16x8b); in ih264_intra_pred_luma_4x4_mode_vert_r_ssse3()
[all …]
|
| D | ih264_weighted_pred_sse42.c | 118 *((WORD32 *)(pu1_dst)) = _mm_cvtsi128_si32(y0_0_16x8b); in ih264_default_weighted_pred_luma_sse42()
119 *((WORD32 *)(pu1_dst + dst_strd)) = _mm_cvtsi128_si32(y0_1_16x8b); in ih264_default_weighted_pred_luma_sse42()
120 *((WORD32 *)(pu1_dst + (dst_strd << 1))) = _mm_cvtsi128_si32(y0_2_16x8b); in ih264_default_weighted_pred_luma_sse42()
121 *((WORD32 *)(pu1_dst + dst_strd * 3)) = _mm_cvtsi128_si32(y0_3_16x8b); in ih264_default_weighted_pred_luma_sse42()
273 *((WORD32 *)(pu1_dst)) = _mm_cvtsi128_si32(uv0_0_16x8b); in ih264_default_weighted_pred_chroma_sse42()
274 *((WORD32 *)(pu1_dst + dst_strd)) = _mm_cvtsi128_si32(uv0_1_16x8b); in ih264_default_weighted_pred_chroma_sse42()
436 *((WORD32 *)(pu1_dst)) = _mm_cvtsi128_si32(y_0_16x8b); in ih264_weighted_pred_luma_sse42()
437 *((WORD32 *)(pu1_dst + dst_strd)) = _mm_cvtsi128_si32(y_1_16x8b); in ih264_weighted_pred_luma_sse42()
438 *((WORD32 *)(pu1_dst + (dst_strd << 1))) = _mm_cvtsi128_si32(y_2_16x8b); in ih264_weighted_pred_luma_sse42()
439 *((WORD32 *)(pu1_dst + dst_strd * 3)) = _mm_cvtsi128_si32(y_3_16x8b); in ih264_weighted_pred_luma_sse42()
[all …]
|
| /external/libaom/libaom/av1/common/x86/ |
| D | convolve_avx2.c | 134 *(uint16_t *)p_0 = (uint16_t)_mm_cvtsi128_si32(res_0); in av1_convolve_y_sr_avx2()
135 *(uint16_t *)p_1 = (uint16_t)_mm_cvtsi128_si32(res_1); in av1_convolve_y_sr_avx2()
247 *(uint16_t *)p_0 = (uint16_t)_mm_cvtsi128_si32(res_0); in av1_convolve_y_sr_avx2()
248 *(uint16_t *)p_1 = (uint16_t)_mm_cvtsi128_si32(res_1); in av1_convolve_y_sr_avx2()
332 *(uint16_t *)p_0 = (uint16_t)_mm_cvtsi128_si32(res_0); in av1_convolve_x_sr_avx2()
333 *(uint16_t *)p_1 = (uint16_t)_mm_cvtsi128_si32(res_1); in av1_convolve_x_sr_avx2()
404 *(uint16_t *)p_0 = _mm_cvtsi128_si32(res_0); in av1_convolve_x_sr_avx2()
405 *(uint16_t *)p_1 = _mm_cvtsi128_si32(res_1); in av1_convolve_x_sr_avx2()
|
| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-avx-rr2-lut16-p3-x40.c | 83 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx0_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
85 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx0_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
112 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx1_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
114 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx1_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
141 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx2_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
143 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx2_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
170 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx3_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
172 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx3_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
199 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx4_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
201 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx4_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
[all …]
|
| D | velu-avx-rr2-lut16-p3-x32.c | 80 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx0_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
82 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx0_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
109 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx1_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
111 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx1_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
138 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx2_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
140 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx2_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
167 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx3_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
169 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx3_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
284 …(const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
286 …(const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
[all …]
|
| D | velu-avx-rr2-lut16-p3-x48.c | 86 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx0_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
88 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx0_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
115 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx1_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
117 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx1_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
144 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx2_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
146 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx2_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
173 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx3_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
175 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx3_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
202 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx4_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
204 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx4_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
[all …]
|
| D | velu-avx-rr2-lut16-p3-x24.c | 77 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx0_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
79 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx0_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
106 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx1_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
108 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx1_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
135 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx2_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
137 …const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx2_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
235 …(const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
237 …(const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
301 …(const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_lo))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
303 …(const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_hi))); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
|