| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x25-minmax-fma3.c | 276 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
406 const __m256 vi18x01234567 = _mm256_maskload_ps(i18, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
|
| D | up8x25-minmax-avx.c | 276 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
406 const __m256 vi18x01234567 = _mm256_maskload_ps(i18, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
|
| D | up8x25-minmax-fma3-acc2.c | 276 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
408 const __m256 vi18x01234567 = _mm256_maskload_ps(i18, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
|
| D | up8x25-minmax-avx-acc2.c | 276 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
408 const __m256 vi18x01234567 = _mm256_maskload_ps(i18, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
|
| /external/XNNPACK/src/f16-dwconv/gen/ |
| D | up8x25-minmax-neonfp16arith-acc2.c | 242 const float16x8_t vi18x01234567 = vld1q_f16(i18); i18 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
354 const float16x8_t vi18x01234567 = vld1q_f16(i18); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
|
| D | up8x25-minmax-fma3-acc2.c | 279 const __m256 vi18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i18)); in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
428 const __m256 vi18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i18)); in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
|
| D | up8x25-minmax-fma3.c | 279 const __m256 vi18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i18)); in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3() local
426 const __m256 vi18x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i18)); in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3() local
|
| D | up8x25-minmax-neonfp16arith.c | 242 const float16x8_t vi18x01234567 = vld1q_f16(i18); i18 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
352 const float16x8_t vi18x01234567 = vld1q_f16(i18); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
|
| /external/XNNPACK/src/qu8-dwconv/gen/ |
| D | up8x25-minmax-rndnu-neon-mul16.c | 282 const int16x8_t vi18x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i18))); i18 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16() local
448 const int16x8_t vi18x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i18))); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16() local
|
| D | up8x25-minmax-fp32-neonv8-mul16.c | 281 const int16x8_t vi18x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i18))); i18 += 8; in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
447 const int16x8_t vi18x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i18))); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
|
| D | up8x25-minmax-rndnu-neon-mul8.c | 319 const uint8x8_t vi18x01234567 = vld1_u8(i18); i18 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul8() local
499 const int16x8_t vi18x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i18))); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul8() local
|
| D | up8x25-minmax-fp32-wasmsimd-mul16.c | 347 const v128_t vi18x01234567 = wasm_u16x8_load8x8(i18); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16() local
618 const v128_t vi18x01234567 = wasm_u16x8_load8x8(i18); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16() local
|
| /external/XNNPACK/src/qs8-dwconv/gen/ |
| D | up8x25-minmax-rndnu-neon-mul16.c | 281 const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18)); i18 += 8; in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16() local
447 const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16() local
|
| D | up8x25-minmax-fp32-neonv8-mul16.c | 280 const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18)); i18 += 8; in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
446 const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
|
| D | up8x25-minmax-fp32-wasmsimd-mul16-add16.c | 311 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16_add16() local
532 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16_add16() local
|
| D | up8x25-minmax-rndnu-neon-mul8-ld64.c | 298 const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8; in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul8_ld64() local
508 const int8x8_t vi18x01234567 = vld1_s8(i18); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul8_ld64() local
|
| D | up8x25-minmax-fp32-avx2-mul32.c | 275 const __m256i vi18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i18)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx2_mul32() local
435 const __m256i vi18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i18)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__avx2_mul32() local
|
| D | up8x25-minmax-fp32-wasmsimd-mul16.c | 329 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16() local
574 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18); 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 | 278 const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mla8_ld64() local
468 const int8x8_t vi18x01234567 = vld1_s8(i18); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mla8_ld64() local
|
| D | up8x25-minmax-fp32-wasmsimd-mul16-add16.c | 311 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16_add16() local
535 const v128_t vi18x01234567 = wasm_i16x8_load8x8(i18); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16_add16() local
|
| D | up8x25-minmax-fp32-neonv8-mul16.c | 279 const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18)); i18 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
448 const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
|
| D | up8x25-minmax-fp32-neon-mul16.c | 279 const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18)); i18 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16() local
449 const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16() local
|
| D | up8x25-minmax-fp32-neonv8-mul8-ld64.c | 296 const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul8_ld64() local
509 const int8x8_t vi18x01234567 = vld1_s8(i18); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul8_ld64() local
|
| D | up8x25-minmax-fp32-avx2-mul32.c | 275 const __m256i vi18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i18)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx2_mul32() local
436 const __m256i vi18x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i18)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__avx2_mul32() local
|
| D | up8x25-minmax-fp32-neonv8-mla8-ld64.c | 278 const int8x8_t vi18x01234567 = vld1_s8(i18); i18 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mla8_ld64() local
467 const int8x8_t vi18x01234567 = vld1_s8(i18); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mla8_ld64() local
|