| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x25-minmax-fma3.c | 294 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
418 const __m256 vi21x01234567 = _mm256_maskload_ps(i21, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
|
| D | up8x25-minmax-avx.c | 294 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
418 const __m256 vi21x01234567 = _mm256_maskload_ps(i21, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
|
| D | up8x25-minmax-fma3-acc2.c | 294 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
420 const __m256 vi21x01234567 = _mm256_maskload_ps(i21, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
|
| D | up8x25-minmax-avx-acc2.c | 294 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
420 const __m256 vi21x01234567 = _mm256_maskload_ps(i21, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
|
| /external/XNNPACK/src/f16-dwconv/gen/ |
| D | up8x25-minmax-neonfp16arith-acc2.c | 254 const float16x8_t vi21x01234567 = vld1q_f16(i21); i21 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
366 const float16x8_t vi21x01234567 = vld1q_f16(i21); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
|
| D | up8x25-minmax-fma3-acc2.c | 297 const __m256 vi21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i21)); in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
443 const __m256 vi21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i21)); in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
|
| D | up8x25-minmax-fma3.c | 297 const __m256 vi21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i21)); in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3() local
441 const __m256 vi21x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i21)); in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3() local
|
| D | up8x25-minmax-neonfp16arith.c | 254 const float16x8_t vi21x01234567 = vld1q_f16(i21); i21 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
364 const float16x8_t vi21x01234567 = vld1q_f16(i21); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
|
| /external/XNNPACK/src/qu8-dwconv/gen/ |
| D | up8x25-minmax-rndnu-neon-mul16.c | 300 const int16x8_t vi21x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i21))); i21 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16() local
463 const int16x8_t vi21x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i21))); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16() local
|
| D | up8x25-minmax-fp32-neonv8-mul16.c | 299 const int16x8_t vi21x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i21))); i21 += 8; in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
462 const int16x8_t vi21x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i21))); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
|
| D | up8x25-minmax-rndnu-neon-mul8.c | 343 const uint8x8_t vi21x01234567 = vld1_u8(i21); i21 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul8() local
514 const int16x8_t vi21x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i21))); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul8() local
|
| D | up8x25-minmax-fp32-wasmsimd-mul16.c | 377 const v128_t vi21x01234567 = wasm_u16x8_load8x8(i21); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16() local
645 const v128_t vi21x01234567 = wasm_u16x8_load8x8(i21); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16() local
|
| /external/XNNPACK/src/qs8-dwconv/gen/ |
| D | up8x25-minmax-rndnu-neon-mul16.c | 299 const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21)); i21 += 8; in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16() local
462 const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16() local
|
| D | up8x25-minmax-fp32-neonv8-mul16.c | 298 const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21)); i21 += 8; in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
461 const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
|
| D | up8x25-minmax-fp32-wasmsimd-mul16-add16.c | 334 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16_add16() local
552 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16_add16() local
|
| D | up8x25-minmax-rndnu-neon-mul8-ld64.c | 319 const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8; in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul8_ld64() local
529 const int8x8_t vi21x01234567 = vld1_s8(i21); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul8_ld64() local
|
| D | up8x25-minmax-fp32-avx2-mul32.c | 293 const __m256i vi21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i21)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx2_mul32() local
450 const __m256i vi21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i21)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx2_mul32() local
|
| D | up8x25-minmax-fp32-wasmsimd-mul16.c | 356 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16() local
598 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16() local
|
| /external/XNNPACK/src/qc8-dwconv/gen/ |
| D | up8x25-minmax-fp32-neon-mla8-ld64.c | 295 const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mla8_ld64() local
485 const int8x8_t vi21x01234567 = vld1_s8(i21); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mla8_ld64() local
|
| D | up8x25-minmax-fp32-wasmsimd-mul16-add16.c | 334 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16_add16() local
555 const v128_t vi21x01234567 = wasm_i16x8_load8x8(i21); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16_add16() local
|
| D | up8x25-minmax-fp32-neonv8-mul16.c | 297 const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21)); i21 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
463 const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
|
| D | up8x25-minmax-fp32-neon-mul16.c | 297 const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21)); i21 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16() local
464 const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16() local
|
| D | up8x25-minmax-fp32-neonv8-mul8-ld64.c | 317 const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul8_ld64() local
530 const int8x8_t vi21x01234567 = vld1_s8(i21); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul8_ld64() local
|
| D | up8x25-minmax-fp32-avx2-mul32.c | 293 const __m256i vi21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i21)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx2_mul32() local
451 const __m256i vi21x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i21)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx2_mul32() local
|
| D | up8x25-minmax-fp32-neonv8-mla8-ld64.c | 295 const int8x8_t vi21x01234567 = vld1_s8(i21); i21 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mla8_ld64() local
484 const int8x8_t vi21x01234567 = vld1_s8(i21); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mla8_ld64() local
|