Home
last modified time | relevance | path

Searched refs:vi7xGHIJKLMN (Results 1 – 25 of 77) sorted by relevance

1234

/external/XNNPACK/src/qc8-dwconv/gen/
Dup24x9-minmax-fp32-neonv8-mul16.c197 const int16x8_t vi7xGHIJKLMN = vmovl_s8(vld1_s8(i7)); i7 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16() local
204 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16()
205 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16()
Dup24x9-minmax-fp32-neon-mul16.c197 const int16x8_t vi7xGHIJKLMN = vmovl_s8(vld1_s8(i7)); i7 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() local
204 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16()
205 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16()
Dup24x9-minmax-fp32-avx2-mul32.c177 … const __m256i vi7xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 16))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx2_mul32() local
183 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi7xGHIJKLMN, vk7xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx2_mul32()
Dup32x9-minmax-fp32-neon-mul16.c227 const int16x8_t vi7xGHIJKLMN = vmovl_s8(vld1_s8(i7)); i7 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16() local
236 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
237 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
Dup32x9-minmax-fp32-neonv8-mul16.c227 const int16x8_t vi7xGHIJKLMN = vmovl_s8(vld1_s8(i7)); i7 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16() local
236 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
237 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
/external/XNNPACK/src/qs8-dwconv/gen/
Dup24x9-minmax-rndnu-neon-mul16.c199 const int16x8_t vi7xGHIJKLMN = vmovl_s8(vld1_s8(i7)); i7 += 8; in xnn_qs8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul16() local
206 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul16()
207 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul16()
Dup24x9-minmax-fp32-neon-mul16.c198 const int16x8_t vi7xGHIJKLMN = vmovl_s8(vld1_s8(i7)); i7 += 8; in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() local
205 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16()
206 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16()
Dup24x9-minmax-fp32-neonv8-mul16.c198 const int16x8_t vi7xGHIJKLMN = vmovl_s8(vld1_s8(i7)); i7 += 8; in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16() local
205 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16()
206 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16()
Dup24x9-minmax-fp32-avx2-mul32.c177 … const __m256i vi7xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 16))); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx2_mul32() local
183 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi7xGHIJKLMN, vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__avx2_mul32()
Dup32x9-minmax-fp32-neonv8-mul16.c228 const int16x8_t vi7xGHIJKLMN = vmovl_s8(vld1_s8(i7)); i7 += 8; in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16() local
237 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
238 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
Dup32x9-minmax-rndnu-neon-mul16.c229 const int16x8_t vi7xGHIJKLMN = vmovl_s8(vld1_s8(i7)); i7 += 8; in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16() local
238 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16()
239 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16()
Dup32x9-minmax-fp32-neon-mul16.c228 const int16x8_t vi7xGHIJKLMN = vmovl_s8(vld1_s8(i7)); i7 += 8; in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16() local
237 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
238 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
Dup24x9-minmax-fp32-wasmsimd-mul16-add16.c204 const v128_t vi7xGHIJKLMN = wasm_i16x8_load8x8(i7 + 16); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__wasmsimd_mul16_add16() local
210 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi7xGHIJKLMN, vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__wasmsimd_mul16_add16()
/external/XNNPACK/src/qu8-dwconv/gen/
Dup24x9-minmax-fp32-neon-mul16.c199 const int16x8_t vi7xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8; in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() local
206 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16()
207 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16()
Dup24x9-minmax-fp32-neonv8-mul16.c199 const int16x8_t vi7xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8; in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16() local
206 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16()
207 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16()
Dup24x9-minmax-rndnu-neon-mul16.c200 const int16x8_t vi7xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul16() local
207 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul16()
208 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul16()
Dup24x9-minmax-fp32-wasmsimd-mul16.c247 const v128_t vi7xGHIJKLMN = wasm_u16x8_load8x8(i7 + 16); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__wasmsimd_mul16() local
251 vsumxGHIJKLMN = wasm_i16x8_add(vsumxGHIJKLMN, vi7xGHIJKLMN); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__wasmsimd_mul16()
256 vprodGHIJKLMN = wasm_i16x8_mul(vi7xGHIJKLMN, vk7xGHIJKLMN); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__wasmsimd_mul16()
Dup24x9-minmax-rndnu-neon-mul8.c241 const uint8x8_t vi7xGHIJKLMN = vld1_u8(i7); i7 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul8() local
248 vprodGHIJKLMN = vmull_u8(vi7xGHIJKLMN, vk7xGHIJKLMN); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul8()
249 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi7xGHIJKLMN); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul8()
Dup32x9-minmax-fp32-neonv8-mul16.c229 const int16x8_t vi7xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8; in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16() local
238 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
239 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
Dup32x9-minmax-fp32-neon-mul16.c229 const int16x8_t vi7xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8; in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16() local
238 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
239 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
Dup32x9-minmax-rndnu-neon-mul16.c230 const int16x8_t vi7xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16() local
239 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16()
240 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16()
Dup32x9-minmax-rndnu-neon-mul8.c284 const uint8x8_t vi7xGHIJKLMN = vld1_u8(i7); i7 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8() local
293 vprodGHIJKLMN = vmull_u8(vi7xGHIJKLMN, vk7xGHIJKLMN); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8()
294 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi7xGHIJKLMN); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8()
/external/XNNPACK/src/f16-dwconv2d-chw/gen/
D3x3p1-minmax-neonfp16arith-6x8.c119 const float16x8_t vi7xGHIJKLMN = vld1q_f16(i7); i7 += 8; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8() local
191 const float16x8_t vi7x9ABCDEFG = vextq_f16(vi7x89ABCDEF, vi7xGHIJKLMN, 1); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8()
221 vi7x89ABCDEF = vi7xGHIJKLMN; in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8()
/external/XNNPACK/src/f16-dwconv/gen/
Dup32x9-minmax-neonfp16arith-acc2.c186 const float16x8_t vi7xGHIJKLMN = vld1q_f16(i7); i7 += 8; in xnn_f16_dwconv_minmax_ukernel_up32x9__neonfp16arith_acc2() local
194 vaccGHIJKLMNp1 = vfmaq_f16(vaccGHIJKLMNp1, vi7xGHIJKLMN, vk7xGHIJKLMN); in xnn_f16_dwconv_minmax_ukernel_up32x9__neonfp16arith_acc2()
Dup32x9-minmax-neonfp16arith.c186 const float16x8_t vi7xGHIJKLMN = vld1q_f16(i7); i7 += 8; in xnn_f16_dwconv_minmax_ukernel_up32x9__neonfp16arith() local
194 vaccGHIJKLMNp0 = vfmaq_f16(vaccGHIJKLMNp0, vi7xGHIJKLMN, vk7xGHIJKLMN); in xnn_f16_dwconv_minmax_ukernel_up32x9__neonfp16arith()

1234