| /external/XNNPACK/src/f16-prelu/gen/ |
| D | neonfp16arith-2x16.c | 78 const float16x8_t vi0x01234567 = vld1q_f16(i0); in xnn_f16_prelu_ukernel__neonfp16arith_2x16() local
83 float16x8_t vacc0x01234567 = vmulq_f16(vi0x01234567, vw01234567); in xnn_f16_prelu_ukernel__neonfp16arith_2x16()
84 … const uint16x8_t vm0x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi0x01234567), vmovq_n_s16(0)); in xnn_f16_prelu_ukernel__neonfp16arith_2x16()
88 vacc0x01234567 = vbslq_f16(vm0x01234567, vacc0x01234567, vi0x01234567); in xnn_f16_prelu_ukernel__neonfp16arith_2x16()
97 const float16x8_t vi0x01234567 = vld1q_f16(i0); in xnn_f16_prelu_ukernel__neonfp16arith_2x16() local
102 float16x8_t vacc0x01234567 = vmulq_f16(vi0x01234567, vw01234567); in xnn_f16_prelu_ukernel__neonfp16arith_2x16()
103 … const uint16x8_t vm0x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi0x01234567), vmovq_n_s16(0)); in xnn_f16_prelu_ukernel__neonfp16arith_2x16()
107 vacc0x01234567 = vbslq_f16(vm0x01234567, vacc0x01234567, vi0x01234567); in xnn_f16_prelu_ukernel__neonfp16arith_2x16()
|
| D | neonfp16arith-2x8.c | 67 const float16x8_t vi0x01234567 = vld1q_f16(i0); in xnn_f16_prelu_ukernel__neonfp16arith_2x8() local
72 float16x8_t vacc0x01234567 = vmulq_f16(vi0x01234567, vw01234567); in xnn_f16_prelu_ukernel__neonfp16arith_2x8()
73 … const uint16x8_t vm0x01234567 = vcltq_s16(vreinterpretq_s16_f16(vi0x01234567), vmovq_n_s16(0)); in xnn_f16_prelu_ukernel__neonfp16arith_2x8()
77 vacc0x01234567 = vbslq_f16(vm0x01234567, vacc0x01234567, vi0x01234567); in xnn_f16_prelu_ukernel__neonfp16arith_2x8()
|
| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up16x4-minmax-fma3.c | 66 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3() local
72 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
117 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3() local
121 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
157 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3() local
159 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
|
| D | up16x4-minmax-fma3-acc2.c | 66 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2() local
72 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2()
120 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2() local
124 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2()
162 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2() local
164 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2()
|
| D | up16x4-minmax-avx-acc2.c | 66 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2() local
72 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2()
120 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2() local
124 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2()
162 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2() local
164 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2()
|
| D | up16x4-minmax-avx.c | 66 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx() local
72 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx()
117 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx() local
121 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx()
157 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx() local
159 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx()
|
| D | up8x4-minmax-fma3.c | 65 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up8x4__fma3() local
69 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x4__fma3()
105 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x4__fma3() local
107 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x4__fma3()
|
| D | up8x4-minmax-avx.c | 65 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up8x4__avx() local
69 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x4__avx()
105 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x4__avx() local
107 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x4__avx()
|
| D | up8x4-minmax-avx-acc2.c | 65 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up8x4__avx_acc2() local
69 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x4__avx_acc2()
107 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x4__avx_acc2() local
109 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x4__avx_acc2()
|
| D | up8x4-minmax-fma3-acc2.c | 65 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up8x4__fma3_acc2() local
69 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x4__fma3_acc2()
107 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x4__fma3_acc2() local
109 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x4__fma3_acc2()
|
| D | up16x9-minmax-fma3.c | 91 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local
97 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
187 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local
191 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
257 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local
259 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
|
| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7p7x-minmax-sse2-c8-acc2.c | 44 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
59 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
98 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
113 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
171 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
186 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
255 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
270 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
|
| D | 7p7x-minmax-ssse3-c8-acc2.c | 44 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
59 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
98 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
113 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
171 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
186 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
252 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
267 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
|
| D | 7p7x-minmax-sse2-c24-acc2.c | 44 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
73 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
137 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
152 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
194 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
223 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
287 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
302 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
363 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
392 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
[all …]
|
| D | 7p7x-minmax-ssse3-c24-acc2.c | 44 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
73 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
137 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
152 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
194 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
223 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
287 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
302 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
363 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
392 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
[all …]
|
| D | 7p7x-minmax-neon-c8-acc2.c | 44 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() local
52 int16x8_t vacc0x01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2()
81 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() local
89 int16x8_t vacc0x01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2()
146 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() local
154 int16x8_t vacc0x01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2()
226 const int8x8_t vi0x01234567 = vld1_s8(i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() local
234 int16x8_t vacc0x01234567 = vaddl_s8(vi0x01234567, vi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2()
|
| D | 7x-minmax-ssse3-c8-acc2.c | 61 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local
76 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
141 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local
156 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
|
| D | 7p7x-minmax-ssse3-c16-acc2.c | 44 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
66 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
123 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
145 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
221 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
243 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
347 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
362 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
| D | 7x-minmax-sse2-c8-acc2.c | 61 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local
76 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
144 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local
159 …st __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
|
| /external/XNNPACK/src/f16-dwconv/gen/ |
| D | up16x4-minmax-neonfp16arith.c | 66 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith() local
70 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
107 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith() local
109 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
133 const float16x8_t vi0x01234567 = vld1q_f16(i0); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith() local
135 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
|
| D | up16x4-minmax-neonfp16arith-acc2.c | 66 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2() local
70 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
110 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2() local
112 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
138 const float16x8_t vi0x01234567 = vld1q_f16(i0); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2() local
140 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
|
| D | up8x4-minmax-neonfp16arith-acc2.c | 65 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2() local
67 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2()
93 const float16x8_t vi0x01234567 = vld1q_f16(i0); in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2() local
95 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2()
|
| D | up8x4-minmax-neonfp16arith.c | 65 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith() local
67 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith()
91 const float16x8_t vi0x01234567 = vld1q_f16(i0); in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith() local
93 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith()
|
| /external/XNNPACK/src/f32-prelu/gen/ |
| D | avx-2x8.c | 52 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_prelu_ukernel__avx_2x8() local
57 const __m256 vprod0x01234567 = _mm256_mul_ps(vi0x01234567, vw01234567); in xnn_f32_prelu_ukernel__avx_2x8()
60 const __m256 vacc0x01234567 = _mm256_blendv_ps(vi0x01234567, vprod0x01234567, vi0x01234567); in xnn_f32_prelu_ukernel__avx_2x8()
|
| D | avx-2x16.c | 53 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_prelu_ukernel__avx_2x16() local
60 const __m256 vprod0x01234567 = _mm256_mul_ps(vi0x01234567, vw01234567); in xnn_f32_prelu_ukernel__avx_2x16()
65 const __m256 vacc0x01234567 = _mm256_blendv_ps(vi0x01234567, vprod0x01234567, vi0x01234567); in xnn_f32_prelu_ukernel__avx_2x16()
|