/external/XNNPACK/src/f16-dwconv/gen/ |
D | up32x4-minmax-neonfp16arith.c | 113 const float16x8_t vk3xGHIJKLMN = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up32x4__neonfp16arith() local 117 vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi3xGHIJKLMN, vk3xGHIJKLMN); in xnn_f16_dwconv_minmax_ukernel_up32x4__neonfp16arith()
|
D | up32x4-minmax-neonfp16arith-acc2.c | 113 const float16x8_t vk3xGHIJKLMN = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up32x4__neonfp16arith_acc2() local 117 vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi3xGHIJKLMN, vk3xGHIJKLMN); in xnn_f16_dwconv_minmax_ukernel_up32x4__neonfp16arith_acc2()
|
D | up32x4-minmax-fma3.c | 122 const __m256 vk3xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 144))); in xnn_f16_dwconv_minmax_ukernel_up32x4__fma3() local 126 …vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3xGHIJKLMN, vk3xGHIJKLMN, vaccG… in xnn_f16_dwconv_minmax_ukernel_up32x4__fma3()
|
D | up32x4-minmax-fma3-acc2.c | 122 const __m256 vk3xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 144))); in xnn_f16_dwconv_minmax_ukernel_up32x4__fma3_acc2() local 126 …vaccGHIJKLMNp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3xGHIJKLMN, vk3xGHIJKLMN, vaccG… in xnn_f16_dwconv_minmax_ukernel_up32x4__fma3_acc2()
|
D | up32x9-minmax-neonfp16arith-acc2.c | 138 const float16x8_t vk3xGHIJKLMN = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up32x9__neonfp16arith_acc2() local 142 vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi3xGHIJKLMN, vk3xGHIJKLMN); in xnn_f16_dwconv_minmax_ukernel_up32x9__neonfp16arith_acc2()
|
D | up32x9-minmax-neonfp16arith.c | 138 const float16x8_t vk3xGHIJKLMN = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up32x9__neonfp16arith() local 142 vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi3xGHIJKLMN, vk3xGHIJKLMN); in xnn_f16_dwconv_minmax_ukernel_up32x9__neonfp16arith()
|
/external/XNNPACK/src/qc8-dwconv/gen/ |
D | up24x9-minmax-fp32-neonv8-mul16.c | 142 … const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16() local 148 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16() 149 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16()
|
D | up24x9-minmax-fp32-neon-mul16.c | 142 … const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() local 148 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() 149 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16()
|
D | up24x9-minmax-fp32-avx2-mul32.c | 130 …const __m256i vk3xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx2_mul32() local 135 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi3xGHIJKLMN, vk3xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx2_mul32()
|
D | up32x9-minmax-fp32-neon-mul16.c | 156 … const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16() local 164 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16() 165 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
|
D | up32x9-minmax-fp32-neonv8-mul16.c | 156 … const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16() local 164 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16() 165 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up24x9-minmax-rndnu-neon-mul16.c | 144 … const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul16() local 150 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul16() 151 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul16()
|
D | up24x9-minmax-fp32-neon-mul16.c | 143 … const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() local 149 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() 150 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16()
|
D | up24x9-minmax-fp32-neonv8-mul16.c | 143 … const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16() local 149 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16() 150 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16()
|
D | up24x9-minmax-fp32-avx2-mul32.c | 130 …const __m256i vk3xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx2_mul32() local 135 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi3xGHIJKLMN, vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx2_mul32()
|
D | up32x9-minmax-fp32-neonv8-mul16.c | 157 … const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16() local 165 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16() 166 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
|
D | up32x9-minmax-rndnu-neon-mul16.c | 158 … const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16() local 166 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16() 167 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16()
|
D | up32x9-minmax-fp32-neon-mul16.c | 157 … const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16() local 165 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16() 166 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
|
D | up24x9-minmax-fp32-wasmsimd-mul16-add16.c | 141 …const v128_t vk3xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t)… in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__wasmsimd_mul16_add16() local 146 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi3xGHIJKLMN, vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__wasmsimd_mul16_add16()
|
/external/XNNPACK/src/qu8-dwconv/gen/ |
D | up24x9-minmax-fp32-neon-mul16.c | 144 …const int16x8_t vk3xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() local 150 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() 151 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16()
|
D | up24x9-minmax-fp32-neonv8-mul16.c | 144 …const int16x8_t vk3xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16() local 150 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16() 151 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16()
|
D | up24x9-minmax-rndnu-neon-mul16.c | 145 …const int16x8_t vk3xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul16() local 151 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul16() 152 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul16()
|
D | up32x9-minmax-fp32-neonv8-mul16.c | 158 …const int16x8_t vk3xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16() local 166 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16() 167 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
|
D | up32x9-minmax-fp32-neon-mul16.c | 158 …const int16x8_t vk3xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16() local 166 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16() 167 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
|
D | up32x9-minmax-rndnu-neon-mul16.c | 159 …const int16x8_t vk3xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16() local 167 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16() 168 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16()
|