Home
last modified time | relevance | path

Searched refs:vi4xGHIJKLMN (Results 1 – 16 of 16) sorted by relevance

/external/XNNPACK/src/qs8-gavgpool/gen/
D7p7x-minmax-sse2-c24-acc2.c62 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
87 …st __m128i vxi4xGHIJKLMN = _mm_unpacklo_epi8(vi4xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
212 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
237 …st __m128i vxi4xGHIJKLMN = _mm_unpacklo_epi8(vi4xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
381 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
406 …st __m128i vxi4xGHIJKLMN = _mm_unpacklo_epi8(vi4xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
D7p7x-minmax-ssse3-c24-acc2.c62 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
87 …st __m128i vxi4xGHIJKLMN = _mm_unpacklo_epi8(vi4xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
212 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
237 …st __m128i vxi4xGHIJKLMN = _mm_unpacklo_epi8(vi4xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
381 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
406 …st __m128i vxi4xGHIJKLMN = _mm_unpacklo_epi8(vi4xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
D7p7x-minmax-neon-c24-acc2.c58 const int8x8_t vi4xGHIJKLMN = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local
75 vacc0xGHIJKLMN = vaddw_s8(vacc0xGHIJKLMN, vi4xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2()
158 const int8x8_t vi4xGHIJKLMN = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local
175 vacc0xGHIJKLMN = vaddw_s8(vacc0xGHIJKLMN, vi4xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2()
293 const int8x8_t vi4xGHIJKLMN = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local
310 vacc0xGHIJKLMN = vaddw_s8(vacc0xGHIJKLMN, vi4xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2()
D7p7x-minmax-neon-c32-acc2.c62 const int8x8_t vi4xGHIJKLMN = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() local
84 vacc0xGHIJKLMN = vaddw_s8(vacc0xGHIJKLMN, vi4xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2()
179 const int8x8_t vi4xGHIJKLMN = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() local
201 vacc0xGHIJKLMN = vaddw_s8(vacc0xGHIJKLMN, vi4xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2()
333 const int8x8_t vi4xGHIJKLMN = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() local
355 vacc0xGHIJKLMN = vaddw_s8(vacc0xGHIJKLMN, vi4xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2()
D7x-minmax-ssse3-c24-acc2.c79 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local
104 …st __m128i vxi4xGHIJKLMN = _mm_unpacklo_epi8(vi4xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
D7x-minmax-sse2-c24-acc2.c79 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local
104 …st __m128i vxi4xGHIJKLMN = _mm_unpacklo_epi8(vi4xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
D7x-minmax-neon-c24-acc2.c81 const int8x8_t vi4xGHIJKLMN = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() local
98 vacc0xGHIJKLMN = vaddw_s8(vacc0xGHIJKLMN, vi4xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2()
D7x-minmax-neon-c32-acc2.c85 const int8x8_t vi4xGHIJKLMN = vld1_s8(i4); i4 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() local
107 vacc0xGHIJKLMN = vaddw_s8(vacc0xGHIJKLMN, vi4xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2()
/external/XNNPACK/src/qs8-dwconv/gen/
Dup24x9-minmax-neon-mul16.c157 const int16x8_t vi4xGHIJKLMN = vmovl_s8(vld1_s8(i4)); i4 += 8; in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16() local
164 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi4xGHIJKLMN), vget_low_s16(vk4xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
165 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi4xGHIJKLMN), vget_high_s16(vk4xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
Dup32x9-minmax-neon-mul16.c175 const int16x8_t vi4xGHIJKLMN = vmovl_s8(vld1_s8(i4)); i4 += 8; in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16() local
184 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi4xGHIJKLMN), vget_low_s16(vk4xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
185 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi4xGHIJKLMN), vget_high_s16(vk4xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
Dup24x9-minmax-avx2-mul32.c140 … const __m256i vi4xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 16))); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32() local
146 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi4xGHIJKLMN, vk4xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
Dup32x9-minmax-avx2-mul32.c153 … const __m256i vi4xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 16))); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32() local
161 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi4xGHIJKLMN, vk4xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
Dup24x9-minmax-wasmsimd-mul16.c171 const v128_t vi4xGHIJKLMN = wasm_i16x8_load_8x8(i4 + 16); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() local
177 const v128_t vprod4xGHIJKLMN = wasm_i16x8_mul(vi4xGHIJKLMN, vk4xGHIJKLMN); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
Dup24x9-minmax-ssse3-mul16.c211 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
219 …st __m128i vxi4xGHIJKLMN = _mm_unpacklo_epi8(vi4xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
Dup24x9-minmax-sse2-mul16.c211 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
219 …st __m128i vxi4xGHIJKLMN = _mm_unpacklo_epi8(vi4xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
Dup24x9-minmax-sse41-mul16.c215 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
216 const __m128i vxi4xGHIJKLMN = _mm_cvtepi8_epi16(vi4xGHIJKLMN); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()