/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up32x9-minmax-avx512skx-mul32.c | 96 …__m512i vaccGHIJKLMNOPQRSTUV = _mm512_loadu_si512((const void*) ((uintptr_t) w + 16 * sizeof(int32… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() local 106 …vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi0xGHIJKLMNOPQRS… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() 115 …vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi1xGHIJKLMNOPQRS… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() 124 …vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi2xGHIJKLMNOPQRS… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() 133 …vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi3xGHIJKLMNOPQRS… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() 142 …vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi4xGHIJKLMNOPQRS… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() 151 …vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi5xGHIJKLMNOPQRS… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() 160 …vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi6xGHIJKLMNOPQRS… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() 169 …vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi7xGHIJKLMNOPQRS… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() 178 …vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi8xGHIJKLMNOPQRS… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() [all …]
|
/external/XNNPACK/src/f32-hswish/gen/ |
D | hswish-avx512f-x32.c | 39 __m512 vaccGHIJKLMNOPQRSTUV = _mm512_fmadd_ps(vxGHIJKLMNOPQRSTUV, vsixth, vhalf); in xnn_f32_hswish_ukernel__avx512f_x32() local 42 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUV, vzero); in xnn_f32_hswish_ukernel__avx512f_x32() 45 vaccGHIJKLMNOPQRSTUV = _mm512_min_ps(vaccGHIJKLMNOPQRSTUV, vone); in xnn_f32_hswish_ukernel__avx512f_x32() 48 vaccGHIJKLMNOPQRSTUV = _mm512_mul_ps(vaccGHIJKLMNOPQRSTUV, vxGHIJKLMNOPQRSTUV); in xnn_f32_hswish_ukernel__avx512f_x32() 51 _mm512_storeu_ps(y + 16, vaccGHIJKLMNOPQRSTUV); in xnn_f32_hswish_ukernel__avx512f_x32()
|
/external/XNNPACK/src/f32-clamp/gen/ |
D | avx512f-x32.c | 35 __m512 vaccGHIJKLMNOPQRSTUV = _mm512_loadu_ps(x + 16); in xnn_f32_clamp_ukernel__avx512f_x32() local 39 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_clamp_ukernel__avx512f_x32() 42 vaccGHIJKLMNOPQRSTUV = _mm512_min_ps(vaccGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_clamp_ukernel__avx512f_x32() 45 _mm512_storeu_ps(y + 16, vaccGHIJKLMNOPQRSTUV); in xnn_f32_clamp_ukernel__avx512f_x32()
|
/external/XNNPACK/src/f32-vlrelu/gen/ |
D | vlrelu-avx512f-x32.c | 33 __m512 vaccGHIJKLMNOPQRSTUV = _mm512_loadu_ps(x + 16); in xnn_f32_vlrelu_ukernel__avx512f_x32() local 37 …const __mmask16 vsignGHIJKLMNOPQRSTUV = _mm512_cmp_ps_mask(vaccGHIJKLMNOPQRSTUV, vzero, _CMP_LT_OQ… in xnn_f32_vlrelu_ukernel__avx512f_x32() 40 …vaccGHIJKLMNOPQRSTUV = _mm512_mask_mul_ps(vaccGHIJKLMNOPQRSTUV, vsignGHIJKLMNOPQRSTUV, vaccGHIJKLM… in xnn_f32_vlrelu_ukernel__avx512f_x32() 43 _mm512_storeu_ps(y + 16, vaccGHIJKLMNOPQRSTUV); in xnn_f32_vlrelu_ukernel__avx512f_x32()
|
/external/XNNPACK/src/f32-relu/gen/ |
D | avx512f-x32.c | 34 __m512 vaccGHIJKLMNOPQRSTUV = _mm512_loadu_ps(x + 16); in xnn_f32_relu_ukernel__avx512f_x32() local 38 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUV, vzero); in xnn_f32_relu_ukernel__avx512f_x32() 41 _mm512_storeu_ps(y + 16, vaccGHIJKLMNOPQRSTUV); in xnn_f32_relu_ukernel__avx512f_x32()
|
/external/XNNPACK/src/f32-dwconv/gen/ |
D | up32x4-minmax-avx512f.c | 105 __m512 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUVp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f() local 107 vaccGHIJKLMNOPQRSTUV = _mm512_min_ps(vaccGHIJKLMNOPQRSTUV, vmax); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f() 110 _mm512_storeu_ps(output + 16, vaccGHIJKLMNOPQRSTUV); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
|
D | up32x4-minmax-avx512f-acc2.c | 108 __m512 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUVp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2() local 110 vaccGHIJKLMNOPQRSTUV = _mm512_min_ps(vaccGHIJKLMNOPQRSTUV, vmax); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2() 113 _mm512_storeu_ps(output + 16, vaccGHIJKLMNOPQRSTUV); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
|
D | up32x9-minmax-avx512f-acc2.c | 178 __m512 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUVp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() local 180 vaccGHIJKLMNOPQRSTUV = _mm512_min_ps(vaccGHIJKLMNOPQRSTUV, vmax); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() 183 _mm512_storeu_ps(output + 16, vaccGHIJKLMNOPQRSTUV); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
|
D | up32x9-minmax-avx512f.c | 175 __m512 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUVp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() local 177 vaccGHIJKLMNOPQRSTUV = _mm512_min_ps(vaccGHIJKLMNOPQRSTUV, vmax); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() 180 _mm512_storeu_ps(output + 16, vaccGHIJKLMNOPQRSTUV); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
|
D | up32x25-minmax-avx512f-acc2.c | 402 __m512 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUVp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 404 vaccGHIJKLMNOPQRSTUV = _mm512_min_ps(vaccGHIJKLMNOPQRSTUV, vmax); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() 407 _mm512_storeu_ps(output + 16, vaccGHIJKLMNOPQRSTUV); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
|
D | up32x25-minmax-avx512f.c | 399 __m512 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUVp0, vmin); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 401 vaccGHIJKLMNOPQRSTUV = _mm512_min_ps(vaccGHIJKLMNOPQRSTUV, vmax); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() 404 _mm512_storeu_ps(output + 16, vaccGHIJKLMNOPQRSTUV); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
|