/external/XNNPACK/src/f32-vbinary/gen/ |
D | vaddc-minmax-avx-x16.c | 41 __m256 vy89ABCDEF = _mm256_add_ps(va89ABCDEF, vb); in xnn_f32_vaddc_minmax_ukernel__avx_x16() local 45 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min); in xnn_f32_vaddc_minmax_ukernel__avx_x16() 48 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max); in xnn_f32_vaddc_minmax_ukernel__avx_x16() 51 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_f32_vaddc_minmax_ukernel__avx_x16()
|
D | vdivc-minmax-avx-x16.c | 41 __m256 vy89ABCDEF = _mm256_div_ps(va89ABCDEF, vb); in xnn_f32_vdivc_minmax_ukernel__avx_x16() local 45 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min); in xnn_f32_vdivc_minmax_ukernel__avx_x16() 48 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max); in xnn_f32_vdivc_minmax_ukernel__avx_x16() 51 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_f32_vdivc_minmax_ukernel__avx_x16()
|
D | vsubc-minmax-avx-x16.c | 41 __m256 vy89ABCDEF = _mm256_sub_ps(va89ABCDEF, vb); in xnn_f32_vsubc_minmax_ukernel__avx_x16() local 45 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min); in xnn_f32_vsubc_minmax_ukernel__avx_x16() 48 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max); in xnn_f32_vsubc_minmax_ukernel__avx_x16() 51 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_f32_vsubc_minmax_ukernel__avx_x16()
|
D | vrdivc-minmax-avx-x16.c | 41 __m256 vy89ABCDEF = _mm256_div_ps(vb, va89ABCDEF); in xnn_f32_vrdivc_minmax_ukernel__avx_x16() local 45 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min); in xnn_f32_vrdivc_minmax_ukernel__avx_x16() 48 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max); in xnn_f32_vrdivc_minmax_ukernel__avx_x16() 51 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_f32_vrdivc_minmax_ukernel__avx_x16()
|
D | vrsubc-minmax-avx-x16.c | 41 __m256 vy89ABCDEF = _mm256_sub_ps(vb, va89ABCDEF); in xnn_f32_vrsubc_minmax_ukernel__avx_x16() local 45 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min); in xnn_f32_vrsubc_minmax_ukernel__avx_x16() 48 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max); in xnn_f32_vrsubc_minmax_ukernel__avx_x16() 51 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_f32_vrsubc_minmax_ukernel__avx_x16()
|
D | vmulc-minmax-avx-x16.c | 41 __m256 vy89ABCDEF = _mm256_mul_ps(va89ABCDEF, vb); in xnn_f32_vmulc_minmax_ukernel__avx_x16() local 45 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min); in xnn_f32_vmulc_minmax_ukernel__avx_x16() 48 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max); in xnn_f32_vmulc_minmax_ukernel__avx_x16() 51 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_f32_vmulc_minmax_ukernel__avx_x16()
|
D | vdiv-minmax-avx-x16.c | 44 __m256 vy89ABCDEF = _mm256_div_ps(va89ABCDEF, vb89ABCDEF); in xnn_f32_vdiv_minmax_ukernel__avx_x16() local 48 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min); in xnn_f32_vdiv_minmax_ukernel__avx_x16() 51 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max); in xnn_f32_vdiv_minmax_ukernel__avx_x16() 54 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_f32_vdiv_minmax_ukernel__avx_x16()
|
D | vmul-minmax-avx-x16.c | 44 __m256 vy89ABCDEF = _mm256_mul_ps(va89ABCDEF, vb89ABCDEF); in xnn_f32_vmul_minmax_ukernel__avx_x16() local 48 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min); in xnn_f32_vmul_minmax_ukernel__avx_x16() 51 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max); in xnn_f32_vmul_minmax_ukernel__avx_x16() 54 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_f32_vmul_minmax_ukernel__avx_x16()
|
D | vsub-minmax-avx-x16.c | 44 __m256 vy89ABCDEF = _mm256_sub_ps(va89ABCDEF, vb89ABCDEF); in xnn_f32_vsub_minmax_ukernel__avx_x16() local 48 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min); in xnn_f32_vsub_minmax_ukernel__avx_x16() 51 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max); in xnn_f32_vsub_minmax_ukernel__avx_x16() 54 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_f32_vsub_minmax_ukernel__avx_x16()
|
D | vadd-minmax-avx-x16.c | 44 __m256 vy89ABCDEF = _mm256_add_ps(va89ABCDEF, vb89ABCDEF); in xnn_f32_vadd_minmax_ukernel__avx_x16() local 48 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min); in xnn_f32_vadd_minmax_ukernel__avx_x16() 51 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max); in xnn_f32_vadd_minmax_ukernel__avx_x16() 54 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_f32_vadd_minmax_ukernel__avx_x16()
|
D | vsqrdiffc-avx-x16.c | 39 __m256 vy89ABCDEF = _mm256_sub_ps(va89ABCDEF, vb); in xnn_f32_vsqrdiffc_ukernel__avx_x16() local 42 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vy89ABCDEF); in xnn_f32_vsqrdiffc_ukernel__avx_x16() 46 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_f32_vsqrdiffc_ukernel__avx_x16()
|
D | vsqrdiff-avx-x16.c | 42 __m256 vy89ABCDEF = _mm256_sub_ps(va89ABCDEF, vb89ABCDEF); in xnn_f32_vsqrdiff_ukernel__avx_x16() local 45 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vy89ABCDEF); in xnn_f32_vsqrdiff_ukernel__avx_x16() 49 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_f32_vsqrdiff_ukernel__avx_x16()
|
/external/XNNPACK/src/f32-qs8-vcvt/gen/ |
D | vcvt-sse2-x16.c | 58 __m128i vy89ABCDEF = _mm_packs_epi32(vy89AB, vyCDEF); in xnn_f32_qs8_vcvt_ukernel__sse2_x16() local 61 vy89ABCDEF = _mm_adds_epi16(vy89ABCDEF, voutput_zero_point); in xnn_f32_qs8_vcvt_ukernel__sse2_x16() 64 vy89ABCDEF = _mm_max_epi16(vy89ABCDEF, voutput_min); in xnn_f32_qs8_vcvt_ukernel__sse2_x16() 66 __m128i vy0123456789ABCDEF = _mm_packs_epi16(vy01234567, vy89ABCDEF); in xnn_f32_qs8_vcvt_ukernel__sse2_x16()
|
D | vcvt-sse2-x24.c | 66 __m128i vy89ABCDEF = _mm_packs_epi32(vy89AB, vyCDEF); in xnn_f32_qs8_vcvt_ukernel__sse2_x24() local 70 vy89ABCDEF = _mm_adds_epi16(vy89ABCDEF, voutput_zero_point); in xnn_f32_qs8_vcvt_ukernel__sse2_x24() 74 vy89ABCDEF = _mm_max_epi16(vy89ABCDEF, voutput_min); in xnn_f32_qs8_vcvt_ukernel__sse2_x24() 77 __m128i vy0123456789ABCDEF = _mm_packs_epi16(vy01234567, vy89ABCDEF); in xnn_f32_qs8_vcvt_ukernel__sse2_x24()
|
D | vcvt-sse2-x32.c | 74 __m128i vy89ABCDEF = _mm_packs_epi32(vy89AB, vyCDEF); in xnn_f32_qs8_vcvt_ukernel__sse2_x32() local 79 vy89ABCDEF = _mm_adds_epi16(vy89ABCDEF, voutput_zero_point); in xnn_f32_qs8_vcvt_ukernel__sse2_x32() 84 vy89ABCDEF = _mm_max_epi16(vy89ABCDEF, voutput_min); in xnn_f32_qs8_vcvt_ukernel__sse2_x32() 88 __m128i vy0123456789ABCDEF = _mm_packs_epi16(vy01234567, vy89ABCDEF); in xnn_f32_qs8_vcvt_ukernel__sse2_x32()
|
/external/XNNPACK/src/qs8-f32-vcvt/gen/ |
D | vcvt-avx2-x16.c | 41 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__avx2_x16() local 44 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale); in xnn_qs8_f32_vcvt_ukernel__avx2_x16() 47 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__avx2_x16()
|
D | vcvt-avx-x16.c | 48 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__avx_x16() local 51 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale); in xnn_qs8_f32_vcvt_ukernel__avx_x16() 54 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__avx_x16()
|
D | vcvt-avx2-x24.c | 43 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__avx2_x24() local 47 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale); in xnn_qs8_f32_vcvt_ukernel__avx2_x24() 51 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__avx2_x24()
|
D | vcvt-avx2-x32.c | 45 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__avx2_x32() local 50 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale); in xnn_qs8_f32_vcvt_ukernel__avx2_x32() 55 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__avx2_x32()
|
D | vcvt-avx-x24.c | 53 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__avx_x24() local 57 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale); in xnn_qs8_f32_vcvt_ukernel__avx_x24() 61 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__avx_x24()
|
/external/XNNPACK/src/qu8-f32-vcvt/gen/ |
D | vcvt-avx2-x16.c | 41 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__avx2_x16() local 44 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale); in xnn_qu8_f32_vcvt_ukernel__avx2_x16() 47 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__avx2_x16()
|
D | vcvt-avx2-x24.c | 43 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__avx2_x24() local 47 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale); in xnn_qu8_f32_vcvt_ukernel__avx2_x24() 51 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__avx2_x24()
|
D | vcvt-avx-x16.c | 48 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__avx_x16() local 51 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale); in xnn_qu8_f32_vcvt_ukernel__avx_x16() 54 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__avx_x16()
|
D | vcvt-avx2-x32.c | 45 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__avx2_x32() local 50 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale); in xnn_qu8_f32_vcvt_ukernel__avx2_x32() 55 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__avx2_x32()
|
D | vcvt-avx-x24.c | 53 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__avx_x24() local 57 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale); in xnn_qu8_f32_vcvt_ukernel__avx_x24() 61 _mm256_storeu_ps(y + 8, vy89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__avx_x24()
|