Home
last modified time | relevance | path

Searched refs:_mm256_cvtepi16_epi32 (Results 1 – 21 of 21) sorted by relevance

/external/XNNPACK/src/qs8-dwconv/gen/
Dup32x9-minmax-avx2-mul16.c97 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
98 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
101 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
102 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod0xOPQRSTUV)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
112 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
113 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
116 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
117 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_cvtepi16_epi32(vprod1xOPQRSTUV)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
127 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
128 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
[all …]
Dup16x9-minmax-avx2-mul16.c93 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16()
94 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16()
102 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16()
103 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16()
111 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16()
112 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16()
120 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16()
121 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16()
129 …vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16()
130 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16()
[all …]
/external/libaom/libaom/aom_dsp/x86/
Dhighbd_adaptive_quantize_avx2.c26 *zbin = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)zbin_ptr)); in highbd_load_b_values_avx2()
28 *round = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)round_ptr)); in highbd_load_b_values_avx2()
29 *quant = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)quant_ptr)); in highbd_load_b_values_avx2()
31 _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)dequant_ptr)); in highbd_load_b_values_avx2()
32 *shift = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)shift_ptr)); in highbd_load_b_values_avx2()
312 zbin = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)zbin_ptr)); in aom_highbd_quantize_b_32x32_adaptive_avx2()
313 round = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)round_ptr)); in aom_highbd_quantize_b_32x32_adaptive_avx2()
314 quant = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)quant_ptr)); in aom_highbd_quantize_b_32x32_adaptive_avx2()
315 dequant = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)dequant_ptr)); in aom_highbd_quantize_b_32x32_adaptive_avx2()
317 _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)quant_shift_ptr)); in aom_highbd_quantize_b_32x32_adaptive_avx2()
Davg_intrin_avx2.c350 src32[0] = _mm256_cvtepi16_epi32(src16[0]); in aom_highbd_hadamard_8x8_avx2()
351 src32[1] = _mm256_cvtepi16_epi32(src16[1]); in aom_highbd_hadamard_8x8_avx2()
352 src32[2] = _mm256_cvtepi16_epi32(src16[2]); in aom_highbd_hadamard_8x8_avx2()
353 src32[3] = _mm256_cvtepi16_epi32(src16[3]); in aom_highbd_hadamard_8x8_avx2()
354 src32[4] = _mm256_cvtepi16_epi32(src16[4]); in aom_highbd_hadamard_8x8_avx2()
355 src32[5] = _mm256_cvtepi16_epi32(src16[5]); in aom_highbd_hadamard_8x8_avx2()
356 src32[6] = _mm256_cvtepi16_epi32(src16[6]); in aom_highbd_hadamard_8x8_avx2()
357 src32[7] = _mm256_cvtepi16_epi32(src16[7]); in aom_highbd_hadamard_8x8_avx2()
Dhighbd_variance_avx2.c42 __m256i v_sum00 = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(v_sum_d)); in aom_highbd_calc8x8var_avx2()
43 __m256i v_sum01 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(v_sum_d, 1)); in aom_highbd_calc8x8var_avx2()
Dvariance_avx2.c86 const __m256i sum_lo = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(sum)); in sum_to_32bit_avx2()
88 _mm256_cvtepi16_epi32(_mm256_extractf128_si256(sum, 1)); in sum_to_32bit_avx2()
/external/libvpx/libvpx/vpx_dsp/x86/
Davg_intrin_avx2.c109 src32[0] = _mm256_cvtepi16_epi32(src16[0]); in vpx_highbd_hadamard_8x8_avx2()
110 src32[1] = _mm256_cvtepi16_epi32(src16[1]); in vpx_highbd_hadamard_8x8_avx2()
111 src32[2] = _mm256_cvtepi16_epi32(src16[2]); in vpx_highbd_hadamard_8x8_avx2()
112 src32[3] = _mm256_cvtepi16_epi32(src16[3]); in vpx_highbd_hadamard_8x8_avx2()
113 src32[4] = _mm256_cvtepi16_epi32(src16[4]); in vpx_highbd_hadamard_8x8_avx2()
114 src32[5] = _mm256_cvtepi16_epi32(src16[5]); in vpx_highbd_hadamard_8x8_avx2()
115 src32[6] = _mm256_cvtepi16_epi32(src16[6]); in vpx_highbd_hadamard_8x8_avx2()
116 src32[7] = _mm256_cvtepi16_epi32(src16[7]); in vpx_highbd_hadamard_8x8_avx2()
Dvariance_avx2.c95 const __m256i sum_lo = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(sum)); in sum_to_32bit_avx2()
97 _mm256_cvtepi16_epi32(_mm256_extractf128_si256(sum, 1)); in sum_to_32bit_avx2()
/external/llvm-project/clang/lib/Headers/
Davx512vlbf16intrin.h432 (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), 16)); in _mm256_cvtpbh_ps()
467 (__m256i)__S, (__mmask8)__U, (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), in _mm256_mask_cvtpbh_ps()
Davx2intrin.h356 _mm256_cvtepi16_epi32(__m128i __V) in _mm256_cvtepi16_epi32() function
Davx512vlintrin.h4094 (__v8si)_mm256_cvtepi16_epi32(__A), in _mm256_mask_cvtepi16_epi32()
4102 (__v8si)_mm256_cvtepi16_epi32(__A), in _mm256_maskz_cvtepi16_epi32()
/external/XNNPACK/src/qs8-dwconv/
Dunipass-avx2-mul16.c.in61 …vacc${ABC[C:C+8]} = _mm256_add_epi32(vacc${ABC[C:C+8]}, _mm256_cvtepi16_epi32(_mm256_castsi256_si1…
62 …vacc${ABC[C+8:C+16]} = _mm256_add_epi32(vacc${ABC[C+8:C+16]}, _mm256_cvtepi16_epi32(vprod${K}x${AB…
133 …vacc${ABC[0:8]} = _mm256_add_epi32(vacc${ABC[0:8]}, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(v…
134 …vacc${ABC[8:16]} = _mm256_add_epi32(vacc${ABC[8:16]}, _mm256_cvtepi16_epi32(vprod${K}x${ABC[8:16]}…
/external/libaom/libaom/av1/encoder/x86/
Dhighbd_fwd_txfm_avx2.c48 out[0] = _mm256_cvtepi16_epi32(out1[0]); in load_buffer_8x8_avx2()
49 out[1] = _mm256_cvtepi16_epi32(out1[1]); in load_buffer_8x8_avx2()
50 out[2] = _mm256_cvtepi16_epi32(out1[2]); in load_buffer_8x8_avx2()
51 out[3] = _mm256_cvtepi16_epi32(out1[3]); in load_buffer_8x8_avx2()
52 out[4] = _mm256_cvtepi16_epi32(out1[4]); in load_buffer_8x8_avx2()
53 out[5] = _mm256_cvtepi16_epi32(out1[5]); in load_buffer_8x8_avx2()
54 out[6] = _mm256_cvtepi16_epi32(out1[6]); in load_buffer_8x8_avx2()
55 out[7] = _mm256_cvtepi16_epi32(out1[7]); in load_buffer_8x8_avx2()
58 out[0] = _mm256_cvtepi16_epi32(mm_reverse_epi16(out1[0])); in load_buffer_8x8_avx2()
59 out[1] = _mm256_cvtepi16_epi32(mm_reverse_epi16(out1[1])); in load_buffer_8x8_avx2()
[all …]
Dav1_fwd_txfm2d_avx2.c1470 _mm256_cvtepi16_epi32(_mm256_castsi256_si128(in[i]))); in store_buffer_16bit_to_32bit_w16_avx2()
1473 _mm256_cvtepi16_epi32(_mm256_extracti128_si256(in[i], 1))); in store_buffer_16bit_to_32bit_w16_avx2()
1702 bufA[j] = _mm256_cvtepi16_epi32(buf[j * 2]); in lowbd_fwd_txfm2d_64x64_avx2()
1703 bufB[j] = _mm256_cvtepi16_epi32(buf[j * 2 + 1]); in lowbd_fwd_txfm2d_64x64_avx2()
1841 bufA[j] = _mm256_cvtepi16_epi32(buf[j * 2]); in lowbd_fwd_txfm2d_64x32_avx2()
1842 bufB[j] = _mm256_cvtepi16_epi32(buf[j * 2 + 1]); in lowbd_fwd_txfm2d_64x32_avx2()
1891 bufA[j] = _mm256_cvtepi16_epi32(buf[j * 2]); in lowbd_fwd_txfm2d_32x64_avx2()
1892 bufB[j] = _mm256_cvtepi16_epi32(buf[j * 2 + 1]); in lowbd_fwd_txfm2d_32x64_avx2()
/external/ruy/ruy/
Dpack_avx2_fma.cc652 sum = _mm256_add_epi32(sum, _mm256_cvtepi16_epi32(new_sum16));
/external/clang/test/CodeGen/
Davx2-builtins.c316 return _mm256_cvtepi16_epi32(a); in test_mm256_cvtepi16_epi32()
/external/llvm-project/clang/test/CodeGen/X86/
Davx2-builtins.c322 return _mm256_cvtepi16_epi32(a); in test_mm256_cvtepi16_epi32()
/external/libaom/libaom/aom_dsp/simd/
Dv256_intrinsics_x86.h316 return _mm256_cvtepi16_epi32(a); in v256_unpack_s16_s32()
/external/mesa3d/src/gallium/drivers/swr/rasterizer/core/
Dformat_types.h458 _mm256_cvtepi16_epi32(_mm_castps_si128(_mm256_castps256_ps128(in))));
/external/clang/lib/Headers/
Davx2intrin.h385 _mm256_cvtepi16_epi32(__m128i __V) in _mm256_cvtepi16_epi32() function
/external/libaom/libaom/av1/common/x86/
Dhighbd_inv_txfm_avx2.c94 __m256i x0 = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(pred)); in highbd_get_recon_16x8_avx2()
95 __m256i x1 = _mm256_cvtepi16_epi32(_mm256_extractf128_si256(pred, 1)); in highbd_get_recon_16x8_avx2()
135 __m256i v = _mm256_cvtepi16_epi32(temp); in highbd_write_buffer_8xn_avx2()