Home
last modified time | relevance | path

Searched refs:vi4x89ABCDEF (Results 1 – 25 of 40) sorted by relevance

12

/external/XNNPACK/src/qs8-gavgpool/gen/
D7p7x-minmax-ssse3-c16-acc2.c57 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
75 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
136 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
154 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
234 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
252 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
D7p7x-minmax-sse2-c16-acc2.c57 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
75 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
136 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
154 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
234 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
252 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
D7p7x-minmax-neon-c16-acc2.c53 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() local
65 vacc0x89ABCDEF = vaddw_s8(vacc0x89ABCDEF, vi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2()
107 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() local
119 vacc0x89ABCDEF = vaddw_s8(vacc0x89ABCDEF, vi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2()
191 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() local
203 vacc0x89ABCDEF = vaddw_s8(vacc0x89ABCDEF, vi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2()
D7p7x-minmax-sse2-c24-acc2.c61 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
86 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
211 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
236 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
380 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
405 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
D7p7x-minmax-ssse3-c24-acc2.c61 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
86 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
211 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
236 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
380 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
405 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
D7p7x-minmax-neon-c24-acc2.c57 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local
74 vacc0x89ABCDEF = vaddw_s8(vacc0x89ABCDEF, vi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2()
157 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local
174 vacc0x89ABCDEF = vaddw_s8(vacc0x89ABCDEF, vi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2()
292 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local
309 vacc0x89ABCDEF = vaddw_s8(vacc0x89ABCDEF, vi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2()
D7x-minmax-ssse3-c16-acc2.c74 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local
92 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
D7x-minmax-sse2-c16-acc2.c74 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local
92 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
D7p7x-minmax-neon-c32-acc2.c61 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() local
83 vacc0x89ABCDEF = vaddw_s8(vacc0x89ABCDEF, vi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2()
178 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() local
200 vacc0x89ABCDEF = vaddw_s8(vacc0x89ABCDEF, vi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2()
332 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() local
354 vacc0x89ABCDEF = vaddw_s8(vacc0x89ABCDEF, vi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2()
D7x-minmax-ssse3-c24-acc2.c78 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local
103 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
D7x-minmax-sse2-c24-acc2.c78 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local
103 …st __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
D7x-minmax-neon-c16-acc2.c76 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c16_acc2() local
88 vacc0x89ABCDEF = vaddw_s8(vacc0x89ABCDEF, vi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c16_acc2()
D7x-minmax-neon-c24-acc2.c80 const int8x8_t vi4x89ABCDEF = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() local
97 vacc0x89ABCDEF = vaddw_s8(vacc0x89ABCDEF, vi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2()
/external/XNNPACK/src/qs8-dwconv/gen/
Dup16x9-minmax-neon-mul16.c137 const int16x8_t vi4x89ABCDEF = vmovl_s8(vld1_s8(i4)); i4 += 8; in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16() local
142 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
143 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
Dup24x9-minmax-neon-mul16.c155 const int16x8_t vi4x89ABCDEF = vmovl_s8(vld1_s8(i4)); i4 += 8; in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16() local
162 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
163 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
Dup16x9-minmax-avx2-mul32.c125 const __m256i vi4x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 8))); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32() local
130 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi4x89ABCDEF, vk4x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
Dup32x9-minmax-neon-mul16.c173 const int16x8_t vi4x89ABCDEF = vmovl_s8(vld1_s8(i4)); i4 += 8; in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16() local
182 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
183 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
Dup24x9-minmax-avx2-mul32.c138 const __m256i vi4x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 8))); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32() local
145 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi4x89ABCDEF, vk4x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
Dup16x9-minmax-wasmsimd-mul16.c147 const v128_t vi4x89ABCDEF = wasm_i16x8_load_8x8(i4 + 8); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() local
152 const v128_t vprod4x89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
/external/XNNPACK/src/f32-dwconv/gen/
Dup16x9-minmax-fma3.c128 const __m256 vi4x89ABCDEF = _mm256_loadu_ps(i4 + 8); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local
134 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi4x89ABCDEF, vk4x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
Dup16x9-minmax-fma3-acc2.c128 const __m256 vi4x89ABCDEF = _mm256_loadu_ps(i4 + 8); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local
134 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi4x89ABCDEF, vk4x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
Dup16x9-minmax-avx-acc2.c128 const __m256 vi4x89ABCDEF = _mm256_loadu_ps(i4 + 8); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx_acc2() local
134 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi4x89ABCDEF, vk4x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx_acc2()
Dup16x9-minmax-avx.c128 const __m256 vi4x89ABCDEF = _mm256_loadu_ps(i4 + 8); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local
134 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi4x89ABCDEF, vk4x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx()
/external/XNNPACK/src/f16-dwconv/gen/
Dup16x9-minmax-neonfp16arith-acc2.c120 const float16x8_t vi4x89ABCDEF = vld1q_f16(i4); i4 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local
124 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi4x89ABCDEF, vk4x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
Dup16x9-minmax-neonfp16arith.c120 const float16x8_t vi4x89ABCDEF = vld1q_f16(i4); i4 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local
124 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi4x89ABCDEF, vk4x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()

12