Home
last modified time | relevance | path

Searched refs:vyGHIJKLMNOPQRSTUV (Results 1 – 25 of 49) sorted by relevance

12

/external/XNNPACK/src/f32-vbinary/gen/
Dvsubc-minmax-avx512f-x32.c42 __m512 vyGHIJKLMNOPQRSTUV = _mm512_sub_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vsubc_minmax_ukernel__avx512f_x32() local
46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vsubc_minmax_ukernel__avx512f_x32()
49 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vsubc_minmax_ukernel__avx512f_x32()
52 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vsubc_minmax_ukernel__avx512f_x32()
Dvrsubc-minmax-avx512f-x32.c42 __m512 vyGHIJKLMNOPQRSTUV = _mm512_sub_ps(vb, vaGHIJKLMNOPQRSTUV); in xnn_f32_vrsubc_minmax_ukernel__avx512f_x32() local
46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vrsubc_minmax_ukernel__avx512f_x32()
49 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vrsubc_minmax_ukernel__avx512f_x32()
52 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vrsubc_minmax_ukernel__avx512f_x32()
Dvaddc-minmax-avx512f-x32.c42 __m512 vyGHIJKLMNOPQRSTUV = _mm512_add_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32() local
46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32()
49 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32()
52 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32()
Dvdivc-minmax-avx512f-x32.c42 __m512 vyGHIJKLMNOPQRSTUV = _mm512_div_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vdivc_minmax_ukernel__avx512f_x32() local
46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vdivc_minmax_ukernel__avx512f_x32()
49 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vdivc_minmax_ukernel__avx512f_x32()
52 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vdivc_minmax_ukernel__avx512f_x32()
Dvrdivc-minmax-avx512f-x32.c42 __m512 vyGHIJKLMNOPQRSTUV = _mm512_div_ps(vb, vaGHIJKLMNOPQRSTUV); in xnn_f32_vrdivc_minmax_ukernel__avx512f_x32() local
46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vrdivc_minmax_ukernel__avx512f_x32()
49 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vrdivc_minmax_ukernel__avx512f_x32()
52 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vrdivc_minmax_ukernel__avx512f_x32()
Dvmulc-minmax-avx512f-x32.c42 __m512 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vmulc_minmax_ukernel__avx512f_x32() local
46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vmulc_minmax_ukernel__avx512f_x32()
49 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vmulc_minmax_ukernel__avx512f_x32()
52 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vmulc_minmax_ukernel__avx512f_x32()
Dvdiv-minmax-avx512f-x32.c45 __m512 vyGHIJKLMNOPQRSTUV = _mm512_div_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32() local
49 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32()
52 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32()
55 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32()
Dvmul-minmax-avx512f-x32.c45 __m512 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vmul_minmax_ukernel__avx512f_x32() local
49 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vmul_minmax_ukernel__avx512f_x32()
52 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vmul_minmax_ukernel__avx512f_x32()
55 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vmul_minmax_ukernel__avx512f_x32()
Dvadd-minmax-avx512f-x32.c45 __m512 vyGHIJKLMNOPQRSTUV = _mm512_add_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vadd_minmax_ukernel__avx512f_x32() local
49 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vadd_minmax_ukernel__avx512f_x32()
52 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vadd_minmax_ukernel__avx512f_x32()
55 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vadd_minmax_ukernel__avx512f_x32()
Dvsub-minmax-avx512f-x32.c45 __m512 vyGHIJKLMNOPQRSTUV = _mm512_sub_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vsub_minmax_ukernel__avx512f_x32() local
49 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vsub_minmax_ukernel__avx512f_x32()
52 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vsub_minmax_ukernel__avx512f_x32()
55 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vsub_minmax_ukernel__avx512f_x32()
Dvsqrdiffc-avx512f-x32.c40 __m512 vyGHIJKLMNOPQRSTUV = _mm512_sub_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vsqrdiffc_ukernel__avx512f_x32() local
43 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vyGHIJKLMNOPQRSTUV); in xnn_f32_vsqrdiffc_ukernel__avx512f_x32()
47 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vsqrdiffc_ukernel__avx512f_x32()
Dvsqrdiff-avx512f-x32.c43 __m512 vyGHIJKLMNOPQRSTUV = _mm512_sub_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vsqrdiff_ukernel__avx512f_x32() local
46 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vyGHIJKLMNOPQRSTUV); in xnn_f32_vsqrdiff_ukernel__avx512f_x32()
50 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vsqrdiff_ukernel__avx512f_x32()
/external/XNNPACK/src/amalgam/
Davx512f.c1462 __m512 vyGHIJKLMNOPQRSTUV = _mm512_add_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vadd_minmax_ukernel__avx512f_x32() local
1466 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vadd_minmax_ukernel__avx512f_x32()
1469 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vadd_minmax_ukernel__avx512f_x32()
1472 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vadd_minmax_ukernel__avx512f_x32()
1528 __m512 vyGHIJKLMNOPQRSTUV = _mm512_add_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32() local
1532 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32()
1535 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32()
1538 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32()
1593 __m512 vyGHIJKLMNOPQRSTUV = _mm512_div_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32() local
1597 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32()
[all …]
/external/XNNPACK/src/f32-qs8-vcvt/gen/
Dvcvt-neonv8-x32.c73 int8x16_t vyGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV)); in xnn_f32_qs8_vcvt_ukernel__neonv8_x32() local
76 vyGHIJKLMNOPQRSTUV = vmaxq_s8(vyGHIJKLMNOPQRSTUV, voutput_min); in xnn_f32_qs8_vcvt_ukernel__neonv8_x32()
79 vyGHIJKLMNOPQRSTUV = vminq_s8(vyGHIJKLMNOPQRSTUV, voutput_max); in xnn_f32_qs8_vcvt_ukernel__neonv8_x32()
82 vst1q_s8(y, vyGHIJKLMNOPQRSTUV); y += 16; in xnn_f32_qs8_vcvt_ukernel__neonv8_x32()
Dvcvt-wasmsimd-cvt-x32.c83 v128_t vyGHIJKLMNOPQRSTUV = wasm_i8x16_narrow_i16x8(vaccGHIJKLMN, vaccOPQRSTUV); in xnn_f32_qs8_vcvt_ukernel__wasmsimd_cvt_x32() local
86 vyGHIJKLMNOPQRSTUV = wasm_i8x16_max(vyGHIJKLMNOPQRSTUV, voutput_min); in xnn_f32_qs8_vcvt_ukernel__wasmsimd_cvt_x32()
89 vyGHIJKLMNOPQRSTUV = wasm_i8x16_min(vyGHIJKLMNOPQRSTUV, voutput_max); in xnn_f32_qs8_vcvt_ukernel__wasmsimd_cvt_x32()
92 wasm_v128_store(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_qs8_vcvt_ukernel__wasmsimd_cvt_x32()
Dvcvt-neon-x32.c78 int8x16_t vyGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV)); in xnn_f32_qs8_vcvt_ukernel__neon_x32() local
81 vyGHIJKLMNOPQRSTUV = vmaxq_s8(vyGHIJKLMNOPQRSTUV, voutput_min); in xnn_f32_qs8_vcvt_ukernel__neon_x32()
84 vyGHIJKLMNOPQRSTUV = vminq_s8(vyGHIJKLMNOPQRSTUV, voutput_max); in xnn_f32_qs8_vcvt_ukernel__neon_x32()
87 vst1q_s8(y, vyGHIJKLMNOPQRSTUV); y += 16; in xnn_f32_qs8_vcvt_ukernel__neon_x32()
/external/XNNPACK/src/f32-qu8-vcvt/gen/
Dvcvt-neonv8-x32.c73 … uint8x16_t vyGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vaccGHIJKLMN), vqmovun_s16(vaccOPQRSTUV)); in xnn_f32_qu8_vcvt_ukernel__neonv8_x32() local
76 vyGHIJKLMNOPQRSTUV = vmaxq_u8(vyGHIJKLMNOPQRSTUV, voutput_min); in xnn_f32_qu8_vcvt_ukernel__neonv8_x32()
79 vyGHIJKLMNOPQRSTUV = vminq_u8(vyGHIJKLMNOPQRSTUV, voutput_max); in xnn_f32_qu8_vcvt_ukernel__neonv8_x32()
82 vst1q_u8(y, vyGHIJKLMNOPQRSTUV); y += 16; in xnn_f32_qu8_vcvt_ukernel__neonv8_x32()
Dvcvt-wasmsimd-cvt-x32.c83 v128_t vyGHIJKLMNOPQRSTUV = wasm_u8x16_narrow_i16x8(vaccGHIJKLMN, vaccOPQRSTUV); in xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32() local
86 vyGHIJKLMNOPQRSTUV = wasm_u8x16_max(vyGHIJKLMNOPQRSTUV, voutput_min); in xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32()
89 vyGHIJKLMNOPQRSTUV = wasm_u8x16_min(vyGHIJKLMNOPQRSTUV, voutput_max); in xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32()
92 wasm_v128_store(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32()
Dvcvt-neon-x32.c78 … uint8x16_t vyGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vaccGHIJKLMN), vqmovun_s16(vaccOPQRSTUV)); in xnn_f32_qu8_vcvt_ukernel__neon_x32() local
81 vyGHIJKLMNOPQRSTUV = vmaxq_u8(vyGHIJKLMNOPQRSTUV, voutput_min); in xnn_f32_qu8_vcvt_ukernel__neon_x32()
84 vyGHIJKLMNOPQRSTUV = vminq_u8(vyGHIJKLMNOPQRSTUV, voutput_max); in xnn_f32_qu8_vcvt_ukernel__neon_x32()
87 vst1q_u8(y, vyGHIJKLMNOPQRSTUV); y += 16; in xnn_f32_qu8_vcvt_ukernel__neon_x32()
/external/XNNPACK/src/qu8-f32-vcvt/gen/
Dvcvt-avx512skx-x32.c41 __m512 vyGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vxGHIJKLMNOPQRSTUV); in xnn_qu8_f32_vcvt_ukernel__avx512skx_x32() local
44 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vscale); in xnn_qu8_f32_vcvt_ukernel__avx512skx_x32()
47 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_qu8_f32_vcvt_ukernel__avx512skx_x32()
Dvcvt-avx512skx-x48.c43 __m512 vyGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vxGHIJKLMNOPQRSTUV); in xnn_qu8_f32_vcvt_ukernel__avx512skx_x48() local
47 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vscale); in xnn_qu8_f32_vcvt_ukernel__avx512skx_x48()
51 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_qu8_f32_vcvt_ukernel__avx512skx_x48()
Dvcvt-avx512skx-x64.c45 __m512 vyGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vxGHIJKLMNOPQRSTUV); in xnn_qu8_f32_vcvt_ukernel__avx512skx_x64() local
50 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vscale); in xnn_qu8_f32_vcvt_ukernel__avx512skx_x64()
55 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_qu8_f32_vcvt_ukernel__avx512skx_x64()
/external/XNNPACK/src/qs8-f32-vcvt/gen/
Dvcvt-avx512skx-x32.c41 __m512 vyGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vxGHIJKLMNOPQRSTUV); in xnn_qs8_f32_vcvt_ukernel__avx512skx_x32() local
44 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vscale); in xnn_qs8_f32_vcvt_ukernel__avx512skx_x32()
47 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_qs8_f32_vcvt_ukernel__avx512skx_x32()
Dvcvt-avx512skx-x48.c43 __m512 vyGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vxGHIJKLMNOPQRSTUV); in xnn_qs8_f32_vcvt_ukernel__avx512skx_x48() local
47 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vscale); in xnn_qs8_f32_vcvt_ukernel__avx512skx_x48()
51 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_qs8_f32_vcvt_ukernel__avx512skx_x48()
Dvcvt-avx512skx-x64.c45 __m512 vyGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vxGHIJKLMNOPQRSTUV); in xnn_qs8_f32_vcvt_ukernel__avx512skx_x64() local
50 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vscale); in xnn_qs8_f32_vcvt_ukernel__avx512skx_x64()
55 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_qs8_f32_vcvt_ukernel__avx512skx_x64()

12