/external/XNNPACK/src/f32-velu/gen/ |
D | velu-avx-rr2-lut16-p3-x40.c | 19 extern XNN_INTERNAL const int xnn_table_exp2minus_k_over_16[16]; 74 …__m128i vl0_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 75 …__m128i vl0_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 76 …__m128i vl0_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 77 …__m128i vl0_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() 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() 83 …__m128i vl0_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() [all …]
|
D | velu-avx-rr2-lut16-p3-x32.c | 19 extern XNN_INTERNAL const int xnn_table_exp2minus_k_over_16[16]; 71 …__m128i vl0_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 72 …__m128i vl0_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 73 …__m128i vl0_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 74 …__m128i vl0_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() 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() 80 …__m128i vl0_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() [all …]
|
D | velu-avx-rr2-lut16-p3-x48.c | 19 extern XNN_INTERNAL const int xnn_table_exp2minus_k_over_16[16]; 77 …__m128i vl0_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 78 …__m128i vl0_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 79 …__m128i vl0_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 80 …__m128i vl0_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() 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() 86 …__m128i vl0_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() [all …]
|
D | velu-avx-rr2-lut16-p3-x24.c | 19 extern XNN_INTERNAL const int xnn_table_exp2minus_k_over_16[16]; 68 …__m128i vl0_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 69 …__m128i vl0_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 70 …__m128i vl0_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 71 …__m128i vl0_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() 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() 77 …__m128i vl0_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() [all …]
|
D | velu-avx-rr2-lut16-p3-x16.c | 19 extern XNN_INTERNAL const int xnn_table_exp2minus_k_over_16[16]; 65 …__m128i vl0_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 66 …__m128i vl0_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 67 …__m128i vl0_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 68 …__m128i vl0_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() 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() 74 …__m128i vl0_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() [all …]
|
D | velu-avx-rr2-lut16-p3-x8.c | 19 extern XNN_INTERNAL const int xnn_table_exp2minus_k_over_16[16]; 62 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 63 …__m128i vl_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 64 …__m128i vl_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 65 …__m128i vl_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() 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() 71 …__m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() [all …]
|
D | velu-sse41-rr2-lut16-p3-x20.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 79 …__m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 80 …t __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 81 …128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 82 …128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 86 …__m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 87 …t __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 88 …128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 89 …128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() 93 …__m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20() [all …]
|
D | velu-sse41-rr2-lut16-p3-x24.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 84 …__m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 85 …t __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 86 …128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 87 …128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 91 …__m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 92 …t __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 93 …128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 94 …128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() 98 …__m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24() [all …]
|
D | velu-sse41-rr2-lut16-p3-x16.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 74 …__m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 75 …t __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 76 …128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 77 …128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 81 …__m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 82 …t __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 83 …128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 84 …128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() 88 …__m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x16() [all …]
|
D | velu-sse41-rr2-lut16-p3-x12.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 69 …__m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 70 …t __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 71 …128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 72 …128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 76 …__m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 77 …t __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 78 …128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 79 …128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() 83 …__m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x12() [all …]
|
D | velu-sse41-rr2-lut16-p3-x8.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 64 …__m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 65 …t __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 66 …128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 67 …128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 71 …__m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 72 …t __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 73 …128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 74 …128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() 81 …__m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x8() [all …]
|
D | velu-sse2-rr2-lut16-p3-x12.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 69 …__m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() 70 …t __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() 71 …t __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() 73 …t __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() 78 …__m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() 79 …t __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() 80 …t __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() 82 …t __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() 87 …__m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12() [all …]
|
D | velu-sse2-rr2-lut16-p3-x16.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 74 …__m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() 75 …t __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() 76 …t __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() 78 …t __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() 83 …__m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() 84 …t __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() 85 …t __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() 87 …t __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() 92 …__m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16() [all …]
|
D | velu-sse2-rr2-lut16-p3-x20.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 79 …__m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() 80 …t __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() 81 …t __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() 83 …t __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() 88 …__m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() 89 …t __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() 90 …t __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() 92 …t __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() 97 …__m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20() [all …]
|
D | velu-sse2-rr2-lut16-p3-x24.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 84 …__m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() 85 …t __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() 86 …t __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() 88 …t __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() 93 …__m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() 94 …t __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() 95 …t __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() 97 …t __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() 102 …__m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x24() [all …]
|
D | velu-sse2-rr2-lut16-p3-x8.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 64 …__m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() 65 …t __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() 66 …t __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() 68 …t __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() 73 …__m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() 74 …t __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() 75 …t __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() 77 …t __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() 85 …__m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x8() [all …]
|
D | velu-sse41-rr2-lut16-p3-x4.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 58 …m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 59 …__m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 60 …i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 61 …i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 63 …__m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 64 …__m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 65 …i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 66 …i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() 101 …m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x4() [all …]
|
D | velu-wasmsimd-arm-rr2-lut16-p3-x24.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 83 …const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx0… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x24() 84 …const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x24() 85 …const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23)… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x24() 86 …const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x24() 90 …const float vl4 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx4… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x24() 91 …const float vl5 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x24() 92 …const float vl6 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67)… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x24() 93 …const float vl7 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x24() 97 …const float vl8 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx8… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x24() [all …]
|
D | velu-sse2-rr2-lut16-p3-x4.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 58 …m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() 59 …__m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() 60 …__m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() 62 …__m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() 65 …__m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() 66 …__m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() 67 …__m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() 69 …__m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() 106 …m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x4() [all …]
|
D | velu-wasmsimd-arm-rr2-lut16-p3-x20.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 78 …const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx0… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x20() 79 …const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x20() 80 …const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23)… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x20() 81 …const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x20() 85 …const float vl4 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx4… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x20() 86 …const float vl5 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x20() 87 …const float vl6 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67)… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x20() 88 …const float vl7 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x20() 92 …const float vl8 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx8… in xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x20() [all …]
|
D | velu-neonfma-rr1-lut16-p3-x24.c | 18 extern XNN_INTERNAL const int32_t xnn_table_exp2minus_k_over_16[16]; 81 …int32x2_t vl01 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24() 82 …int32x2_t vl23 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24() 83 …vl01 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vid… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24() 84 …vl23 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vid… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24() 88 …int32x2_t vl45 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24() 89 …int32x2_t vl67 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24() 90 …vl45 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vid… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24() 91 …vl67 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vid… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24() 95 …int32x2_t vl89 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x24() [all …]
|
D | velu-wasmsimd-x86-rr2-lut16-p3-x24.c | 18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 83 …const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx0… in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() 84 …const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01… in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() 85 …const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23)… in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() 86 …const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23… in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() 90 …const float vl4 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx4… in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() 91 …const float vl5 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45… in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() 92 …const float vl6 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67)… in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() 93 …const float vl7 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67… in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() 97 …const float vl8 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx8… in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() [all …]
|
D | velu-neon-rr2-lut16-p3-x24.c | 18 extern XNN_INTERNAL const int32_t xnn_table_exp2minus_k_over_16[16]; 82 …int32x2_t vl01 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24() 83 …int32x2_t vl23 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24() 84 …vl01 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vid… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24() 85 …vl23 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vid… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24() 89 …int32x2_t vl45 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24() 90 …int32x2_t vl67 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24() 91 …vl45 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vid… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24() 92 …vl67 = vld1_lane_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vid… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24() 96 …int32x2_t vl89 = vld1_dup_s32((const int32_t*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint3… in xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x24() [all …]
|
/external/XNNPACK/src/math/ |
D | expm1minus-avx-rr2-lut16-p3.c | 16 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; 80 …__m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() 81 …__m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() 82 …__m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() 83 …__m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() 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() 89 …__m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() [all …]
|
/external/XNNPACK/src/f32-velu/ |
D | avx-rr2-lut16-p3.c.in | 18 extern XNN_INTERNAL const int xnn_table_exp2minus_k_over_16[16]; 67 …__m128i vl${N}_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… 68 …__m128i vl${N}_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… 69 …__m128i vl${N}_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… 70 …__m128i vl${N}_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… 71 …_ll = _mm_insert_epi32(vl${N}_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… 72 …_lh = _mm_insert_epi32(vl${N}_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… 73 …_hl = _mm_insert_epi32(vl${N}_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… 74 …_hh = _mm_insert_epi32(vl${N}_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… 76 …__m128i vl${N}_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uin… [all …]
|