/external/XNNPACK/src/f16-dwconv/gen/ |
D | up16x9-minmax-neonfp16arith.c | 140 const float16x8_t vi7x01234567 = vld1q_f16(i7); i7 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local 195 const float16x8_t vi7x01234567 = vld1q_f16(i7); i7 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local 241 const float16x8_t vi7x01234567 = vld1q_f16(i7); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local
|
D | up16x9-minmax-neonfp16arith-acc2.c | 140 const float16x8_t vi7x01234567 = vld1q_f16(i7); i7 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local 198 const float16x8_t vi7x01234567 = vld1q_f16(i7); i7 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local 246 const float16x8_t vi7x01234567 = vld1q_f16(i7); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local
|
D | up16x9-minmax-fma3-acc2.c | 154 const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7)); in xnn_f16_dwconv_minmax_ukernel_up16x9__fma3_acc2() local 232 const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7)); in xnn_f16_dwconv_minmax_ukernel_up16x9__fma3_acc2() local 296 const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7)); in xnn_f16_dwconv_minmax_ukernel_up16x9__fma3_acc2() local
|
D | up16x9-minmax-fma3.c | 154 const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7)); in xnn_f16_dwconv_minmax_ukernel_up16x9__fma3() local 229 const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7)); in xnn_f16_dwconv_minmax_ukernel_up16x9__fma3() local 291 const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7)); in xnn_f16_dwconv_minmax_ukernel_up16x9__fma3() local
|
D | up8x9-minmax-neonfp16arith.c | 118 const float16x8_t vi7x01234567 = vld1q_f16(i7); i7 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith() local 164 const float16x8_t vi7x01234567 = vld1q_f16(i7); in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith() local
|
D | up8x9-minmax-fma3.c | 132 const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7)); in xnn_f16_dwconv_minmax_ukernel_up8x9__fma3() local 194 const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7)); in xnn_f16_dwconv_minmax_ukernel_up8x9__fma3() local
|
D | up8x9-minmax-fma3-acc2.c | 132 const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7)); in xnn_f16_dwconv_minmax_ukernel_up8x9__fma3_acc2() local 196 const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7)); in xnn_f16_dwconv_minmax_ukernel_up8x9__fma3_acc2() local
|
D | up8x9-minmax-neonfp16arith-acc2.c | 118 const float16x8_t vi7x01234567 = vld1q_f16(i7); i7 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2() local 166 const float16x8_t vi7x01234567 = vld1q_f16(i7); in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2() local
|
/external/XNNPACK/src/f32-dwconv/gen/ |
D | up16x9-minmax-avx.c | 152 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local 227 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local 283 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local
|
D | up16x9-minmax-avx-acc2.c | 152 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx_acc2() local 230 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx_acc2() local 288 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx_acc2() local
|
D | up16x9-minmax-fma3.c | 152 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local 227 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local 283 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local
|
D | up16x9-minmax-fma3-acc2.c | 152 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local 230 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local 288 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local
|
D | up8x9-minmax-avx.c | 130 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__avx() local 186 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x9__avx() local
|
D | up8x9-minmax-fma3.c | 130 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__fma3() local 186 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x9__fma3() local
|
D | up8x9-minmax-fma3-acc2.c | 130 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__fma3_acc2() local 188 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x9__fma3_acc2() local
|
D | up8x9-minmax-avx-acc2.c | 130 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__avx_acc2() local 188 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x9__avx_acc2() local
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up8x9-minmax-rndnu-neon-mul16.c | 135 const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8; in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mul16() local 216 const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mul16() local
|
D | up8x9-minmax-fp32-avx2-mul32.c | 128 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32() local 203 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32() local
|
D | up8x9-minmax-rndnu-neon-mul8-ld64.c | 141 const int8x8_t vi7x01234567 = vld1_s8(i7); i7 += 8; in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mul8_ld64() local 239 const int8x8_t vi7x01234567 = vld1_s8(i7); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mul8_ld64() local
|
D | up8x9-minmax-fp32-wasmsimd-mul16-add16.c | 142 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__wasmsimd_mul16_add16() local 246 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__wasmsimd_mul16_add16() local
|
/external/XNNPACK/src/qu8-dwconv/gen/ |
D | up8x9-minmax-fp32-avx2-mul32.c | 129 const __m256i vi7x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i7)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32() local 204 const __m256i vi7x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i7)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32() local
|
D | up8x9-minmax-rndnu-neon-mul16.c | 136 const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mul16() local 217 const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mul16() local
|
D | up8x9-minmax-fp32-neon-mul16.c | 135 const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8; in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__neon_mul16() local 217 const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__neon_mul16() local
|
D | up8x9-minmax-fp32-neonv8-mul16.c | 135 const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8; in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__neonv8_mul16() local 216 const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__neonv8_mul16() local
|
/external/XNNPACK/src/qc8-dwconv/gen/ |
D | up8x9-minmax-fp32-avx2-mul32.c | 128 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32() local 204 const __m256i vi7x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i7)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32() local
|