/external/XNNPACK/src/f32-dwconv/gen/ |
D | up16x9-minmax-fma3-acc2.c | 147 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local 227 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local 285 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local
|
D | up16x9-minmax-fma3.c | 147 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local 224 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local 280 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local
|
D | up16x9-minmax-avx.c | 147 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local 224 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local 280 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local
|
D | up16x9-minmax-avx-acc2.c | 147 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx_acc2() local 227 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx_acc2() local 285 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx_acc2() local
|
D | up8x9-minmax-fma3.c | 127 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__fma3() local 183 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__fma3() local
|
D | up8x9-minmax-avx.c | 127 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__avx() local 183 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__avx() local
|
D | up8x9-minmax-fma3-acc2.c | 127 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__fma3_acc2() local 185 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__fma3_acc2() local
|
D | up8x9-minmax-avx-acc2.c | 127 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__avx_acc2() local 185 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up8x9__avx_acc2() local
|
/external/XNNPACK/src/f16-dwconv/gen/ |
D | up16x9-minmax-neonfp16arith-acc2.c | 135 const float16x8_t vk6x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local 195 const float16x8_t vk6x01234567 = vld1q_f16(w + 104); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local 243 const float16x8_t vk6x01234567 = vld1q_f16(w + 112); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local
|
D | up16x9-minmax-fma3.c | 150 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 112))); in xnn_f16_dwconv_minmax_ukernel_up16x9__fma3() local 227 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 112))); in xnn_f16_dwconv_minmax_ukernel_up16x9__fma3() local 289 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 112))); in xnn_f16_dwconv_minmax_ukernel_up16x9__fma3() local
|
D | up16x9-minmax-fma3-acc2.c | 150 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 112))); in xnn_f16_dwconv_minmax_ukernel_up16x9__fma3_acc2() local 230 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 112))); in xnn_f16_dwconv_minmax_ukernel_up16x9__fma3_acc2() local 294 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 112))); in xnn_f16_dwconv_minmax_ukernel_up16x9__fma3_acc2() local
|
D | up16x9-minmax-neonfp16arith.c | 135 const float16x8_t vk6x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local 192 const float16x8_t vk6x01234567 = vld1q_f16(w + 104); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local 238 const float16x8_t vk6x01234567 = vld1q_f16(w + 112); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local
|
D | up8x9-minmax-neonfp16arith.c | 115 const float16x8_t vk6x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith() local 161 const float16x8_t vk6x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith() local
|
D | up8x9-minmax-fma3.c | 130 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 56))); in xnn_f16_dwconv_minmax_ukernel_up8x9__fma3() local 192 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 56))); in xnn_f16_dwconv_minmax_ukernel_up8x9__fma3() local
|
D | up8x9-minmax-fma3-acc2.c | 130 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 56))); in xnn_f16_dwconv_minmax_ukernel_up8x9__fma3_acc2() local 194 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 56))); in xnn_f16_dwconv_minmax_ukernel_up8x9__fma3_acc2() local
|
D | up8x9-minmax-neonfp16arith-acc2.c | 115 const float16x8_t vk6x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2() local 163 const float16x8_t vk6x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2() local
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up8x9-minmax-rndnu-neon-mul16.c | 130 … const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mul16() local 212 const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 48))); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mul16() local
|
D | up8x9-minmax-fp32-avx2-mul32.c | 124 …const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32() local 200 …const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32() local
|
D | up8x9-minmax-rndnu-neon-mla8-ld64.c | 129 const int8x8_t vk6x01234567 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mla8_ld64() local 219 const int8x8_t vk6x01234567 = vld1_s8((const void*) ((const int8_t*) w + 48)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mla8_ld64() local
|
D | up8x9-minmax-fp32-wasmsimd-mul16.c | 142 …const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) … in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__wasmsimd_mul16() local 255 …const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) … in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__wasmsimd_mul16() local
|
D | up8x9-minmax-fp32-wasmsimd-mul16-add16.c | 136 …const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) … in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__wasmsimd_mul16_add16() local 241 …const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) … in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__wasmsimd_mul16_add16() local
|
/external/XNNPACK/src/qc8-dwconv/gen/ |
D | up8x9-minmax-fp32-avx2-mul32.c | 124 …const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32() local 201 …const __m256i vk6x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32() local
|
/external/XNNPACK/src/qu8-dwconv/gen/ |
D | up8x9-minmax-fp32-avx2-mul32.c | 125 …const __m256i vk6x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*… in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32() local 201 …const __m256i vk6x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*… in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32() local
|
D | up8x9-minmax-fp32-neonv8-mul16.c | 130 …const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__neonv8_mul16() local 212 …const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_… in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__neonv8_mul16() local
|
D | up8x9-minmax-rndnu-neon-mul16.c | 131 …const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mul16() local 213 …const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8_… in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mul16() local
|