/external/XNNPACK/src/f32-velu/gen/ |
D | velu-avx-rr2-lut16-p3-x40.c | 78 …vl0_ll = _mm_insert_epi32(vl0_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 79 …vl0_lh = _mm_insert_epi32(vl0_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 80 …vl0_hl = _mm_insert_epi32(vl0_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 81 …vl0_hh = _mm_insert_epi32(vl0_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 87 …vl0_ll = _mm_insert_epi32(vl0_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 88 …vl0_lh = _mm_insert_epi32(vl0_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 89 …vl0_hl = _mm_insert_epi32(vl0_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 90 …vl0_hh = _mm_insert_epi32(vl0_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 107 …vl1_ll = _mm_insert_epi32(vl1_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 108 …vl1_lh = _mm_insert_epi32(vl1_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() [all …]
|
D | velu-avx-rr2-lut16-p3-x32.c | 75 …vl0_ll = _mm_insert_epi32(vl0_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 76 …vl0_lh = _mm_insert_epi32(vl0_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 77 …vl0_hl = _mm_insert_epi32(vl0_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 78 …vl0_hh = _mm_insert_epi32(vl0_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 84 …vl0_ll = _mm_insert_epi32(vl0_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 85 …vl0_lh = _mm_insert_epi32(vl0_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 86 …vl0_hl = _mm_insert_epi32(vl0_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 87 …vl0_hh = _mm_insert_epi32(vl0_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 104 …vl1_ll = _mm_insert_epi32(vl1_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 105 …vl1_lh = _mm_insert_epi32(vl1_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() [all …]
|
D | velu-avx-rr2-lut16-p3-x48.c | 81 …vl0_ll = _mm_insert_epi32(vl0_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 82 …vl0_lh = _mm_insert_epi32(vl0_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 83 …vl0_hl = _mm_insert_epi32(vl0_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 84 …vl0_hh = _mm_insert_epi32(vl0_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 90 …vl0_ll = _mm_insert_epi32(vl0_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 91 …vl0_lh = _mm_insert_epi32(vl0_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 92 …vl0_hl = _mm_insert_epi32(vl0_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 93 …vl0_hh = _mm_insert_epi32(vl0_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 110 …vl1_ll = _mm_insert_epi32(vl1_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 111 …vl1_lh = _mm_insert_epi32(vl1_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() [all …]
|
D | velu-avx-rr2-lut16-p3-x24.c | 72 …vl0_ll = _mm_insert_epi32(vl0_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 73 …vl0_lh = _mm_insert_epi32(vl0_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 74 …vl0_hl = _mm_insert_epi32(vl0_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 75 …vl0_hh = _mm_insert_epi32(vl0_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 81 …vl0_ll = _mm_insert_epi32(vl0_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 82 …vl0_lh = _mm_insert_epi32(vl0_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 83 …vl0_hl = _mm_insert_epi32(vl0_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 84 …vl0_hh = _mm_insert_epi32(vl0_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 101 …vl1_ll = _mm_insert_epi32(vl1_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 102 …vl1_lh = _mm_insert_epi32(vl1_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() [all …]
|
D | velu-avx-rr2-lut16-p3-x16.c | 69 …vl0_ll = _mm_insert_epi32(vl0_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 70 …vl0_lh = _mm_insert_epi32(vl0_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 71 …vl0_hl = _mm_insert_epi32(vl0_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 72 …vl0_hh = _mm_insert_epi32(vl0_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 78 …vl0_ll = _mm_insert_epi32(vl0_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 79 …vl0_lh = _mm_insert_epi32(vl0_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 80 …vl0_hl = _mm_insert_epi32(vl0_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 81 …vl0_hh = _mm_insert_epi32(vl0_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 98 …vl1_ll = _mm_insert_epi32(vl1_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 99 …vl1_lh = _mm_insert_epi32(vl1_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() [all …]
|
D | velu-avx-rr2-lut16-p3-x8.c | 66 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 67 …vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 68 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 69 …vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 75 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 76 …vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 77 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 78 …vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 132 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 133 …vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() [all …]
|
D | velu-sse41-rr2-lut16-p3-x20.c | 81 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 82 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 88 …const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 89 …const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 95 …const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 96 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 102 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 103 …const __m128i vlEF = _mm_insert_epi32(vlE, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 109 …const __m128i vlGH = _mm_insert_epi32(vlG, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 110 …const __m128i vlIJ = _mm_insert_epi32(vlI, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() [all …]
|
D | velu-sse41-rr2-lut16-p3-x24.c | 86 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 87 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 93 …const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 94 …const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 100 …const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 101 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 107 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 108 …const __m128i vlEF = _mm_insert_epi32(vlE, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 114 …const __m128i vlGH = _mm_insert_epi32(vlG, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 115 …const __m128i vlIJ = _mm_insert_epi32(vlI, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() [all …]
|
D | velu-sse41-rr2-lut16-p3-x16.c | 76 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 77 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 83 …const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 84 …const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 90 …const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 91 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 97 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 98 …const __m128i vlEF = _mm_insert_epi32(vlE, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 107 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 108 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() [all …]
|
D | velu-sse41-rr2-lut16-p3-x12.c | 71 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 72 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 78 …const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 79 …const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 85 …const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 86 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 95 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 96 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 104 …const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 105 …const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() [all …]
|
D | velu-sse41-rr2-lut16-p3-x8.c | 66 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 67 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 73 …const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 74 …const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 83 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 84 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 92 …const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 93 …const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 150 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 151 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() [all …]
|
D | velu-sse41-rr2-lut16-p3-x4.c | 60 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 61 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 65 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 66 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 103 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 104 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 108 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 109 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4()
|
/external/XNNPACK/src/f32-sigmoid/gen/ |
D | sse41-lut64-p2-div-x24.c | 80 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() 81 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() 87 …const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() 88 …const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() 94 …const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() 95 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() 101 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() 102 …const __m128i vlEF = _mm_insert_epi32(vlE, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() 108 …const __m128i vlGH = _mm_insert_epi32(vlG, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() 109 …const __m128i vlIJ = _mm_insert_epi32(vlI, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24() [all …]
|
D | sse41-lut64-p2-div-x16.c | 70 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() 71 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() 77 …const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() 78 …const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() 84 …const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() 85 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() 91 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() 92 …const __m128i vlEF = _mm_insert_epi32(vlE, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() 101 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() 102 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16() [all …]
|
D | sse41-lut64-p2-div-x20.c | 75 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() 76 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() 82 …const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() 83 …const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() 89 …const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() 90 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() 96 …const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() 97 …const __m128i vlEF = _mm_insert_epi32(vlE, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() 103 …const __m128i vlGH = _mm_insert_epi32(vlG, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() 104 …const __m128i vlIJ = _mm_insert_epi32(vlI, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20() [all …]
|
D | sse41-lut64-p2-div-x12.c | 65 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() 66 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() 72 …const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() 73 …const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() 79 …const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() 80 …const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() 89 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() 90 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() 98 …const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() 99 …const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12() [all …]
|
D | sse41-lut64-p2-div-x8.c | 60 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() 61 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() 67 …const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() 68 …const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() 77 …const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() 78 …const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() 86 …const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() 87 …const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() 140 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() 141 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8() [all …]
|
D | sse41-lut64-p2-div-x4.c | 53 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() 54 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() 58 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() 59 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() 95 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() 96 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() 100 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4() 101 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4()
|
/external/XNNPACK/src/math/ |
D | expm1minus-avx-rr2-lut16-p3.c | 84 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() 85 …vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() 86 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() 87 …vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() 93 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() 94 …vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() 95 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() 96 …vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3()
|
D | sigmoid-avx-rr2-lut64-p2-div.c | 88 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint3… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div() 89 …vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint3… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div() 90 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint3… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div() 91 …vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint3… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div() 97 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint3… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div() 98 …vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint3… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div() 99 …vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint3… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div() 100 …vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint3… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div()
|
/external/XNNPACK/src/f32-velu/ |
D | avx-rr2-lut16-p3.c.in | 71 …vl${N}_ll = _mm_insert_epi32(vl${N}_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … 72 …vl${N}_lh = _mm_insert_epi32(vl${N}_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … 73 …vl${N}_hl = _mm_insert_epi32(vl${N}_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … 74 …vl${N}_hh = _mm_insert_epi32(vl${N}_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … 80 …vl${N}_ll = _mm_insert_epi32(vl${N}_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … 81 …vl${N}_lh = _mm_insert_epi32(vl${N}_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … 82 …vl${N}_hl = _mm_insert_epi32(vl${N}_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … 83 …vl${N}_hh = _mm_insert_epi32(vl${N}_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 … 148 …vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… 149 …vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… [all …]
|
D | sse-rr2-lut16-p3.c.in | 73 …const __m128i vl${ABC[N:N+2]} = _mm_insert_epi32(vl${ABC[N]}, *((const int*) ((uintptr_t) xnn_tabl… 78 …const __m128i vl${ABC[N+2:N+4]} = _mm_insert_epi32(vl${ABC[N+2]}, *((const int*) ((uintptr_t) xnn_… 92 …const __m128i vl${ABC[N:N+2]} = _mm_insert_epi32(vl${ABC[N]}, *((const int*) ((uintptr_t) xnn_tabl… 97 …const __m128i vl${ABC[N+2:N+4]} = _mm_insert_epi32(vl${ABC[N+2]}, *((const int*) ((uintptr_t) xnn_… 165 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… 170 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… 178 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… 183 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… 232 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… 237 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… [all …]
|
/external/XNNPACK/src/f32-sigmoid/ |
D | sse-lut64-p2-div.c.in | 68 …const __m128i vl${ABC[N:N+2]} = _mm_insert_epi32(vl${ABC[N]}, *((const int*) ((uintptr_t) xnn_tabl… 73 …const __m128i vl${ABC[N+2:N+4]} = _mm_insert_epi32(vl${ABC[N+2]}, *((const int*) ((uintptr_t) xnn_… 87 …const __m128i vl${ABC[N:N+2]} = _mm_insert_epi32(vl${ABC[N]}, *((const int*) ((uintptr_t) xnn_tabl… 92 …const __m128i vl${ABC[N+2:N+4]} = _mm_insert_epi32(vl${ABC[N+2]}, *((const int*) ((uintptr_t) xnn_… 160 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… 165 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… 173 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… 178 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… 225 …const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… 230 …const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_ov… [all …]
|
/external/mesa3d/src/gallium/drivers/swr/rasterizer/common/ |
D | simdlib_256_avx.inl | 251 vAHi = _mm_insert_epi32(vAHi, aHi, 0); 256 vALow = _mm_insert_epi32(vALow, aLow, 0); 261 vAHi = _mm_insert_epi32(vAHi, aHi, 1); 266 vALow = _mm_insert_epi32(vALow, aLow, 1); 271 vAHi = _mm_insert_epi32(vAHi, aHi, 2); 276 vALow = _mm_insert_epi32(vALow, aLow, 2); 281 vAHi = _mm_insert_epi32(vAHi, aHi, 3); 286 vALow = _mm_insert_epi32(vALow, aLow, 3); 316 vAHi = _mm_insert_epi32(vAHi, aHi, 0); 321 vALow = _mm_insert_epi32(vALow, aLow, 0); [all …]
|
/external/libgav1/libgav1/src/dsp/x86/ |
D | motion_vector_search_sse4.cc | 93 lookup = _mm_insert_epi32( in MvProjectionSingleClip() 96 lookup = _mm_insert_epi32( in MvProjectionSingleClip() 99 lookup = _mm_insert_epi32( in MvProjectionSingleClip()
|