Home
last modified time | relevance | path

Searched refs:vi23x89ABCDEF (Results 1 – 25 of 88) sorted by relevance

1234

/external/XNNPACK/src/qs8-dwconv/gen/
Dup16x25-minmax-rndnu-neon-mul16.c407 const int16x8_t vi23x89ABCDEF = vmovl_s8(vld1_s8(i23)); i23 += 8; in xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16() local
412 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi23x89ABCDEF), vget_low_s16(vk23x89ABCDEF)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16()
413 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi23x89ABCDEF), vget_high_s16(vk23x89ABCDEF)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16()
Dup16x25-minmax-fp32-neonv8-mul16.c406 const int16x8_t vi23x89ABCDEF = vmovl_s8(vld1_s8(i23)); i23 += 8; in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16() local
411 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi23x89ABCDEF), vget_low_s16(vk23x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
412 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi23x89ABCDEF), vget_high_s16(vk23x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
Dup16x25-minmax-fp32-neon-mul16.c406 const int16x8_t vi23x89ABCDEF = vmovl_s8(vld1_s8(i23)); i23 += 8; in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16() local
411 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi23x89ABCDEF), vget_low_s16(vk23x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
412 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi23x89ABCDEF), vget_high_s16(vk23x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
Dup16x25-minmax-fp32-wasmsimd-mul16-add16.c445 const v128_t vi23x89ABCDEF = wasm_i16x8_load8x8(i23 + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__wasmsimd_mul16_add16() local
450 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi23x89ABCDEF, vk23x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__wasmsimd_mul16_add16()
Dup16x25-minmax-rndnu-neon-mla8-ld64.c404 const int8x8_t vi23x89ABCDEF = vld1_s8(i23); i23 += 8; in xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mla8_ld64() local
408 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi23x89ABCDEF, vk23x89ABCDEF); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mla8_ld64()
Dup16x25-minmax-fp32-avx2-mul32.c377 … const __m256i vi23x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i23 + 8))); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul32() local
382 … vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi23x89ABCDEF, vk23x89ABCDEF)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul32()
/external/XNNPACK/src/qc8-dwconv/gen/
Dup16x25-minmax-fp32-neon-mul16.c405 const int16x8_t vi23x89ABCDEF = vmovl_s8(vld1_s8(i23)); i23 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16() local
410 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi23x89ABCDEF), vget_low_s16(vk23x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
411 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi23x89ABCDEF), vget_high_s16(vk23x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
Dup16x25-minmax-fp32-neonv8-mul16.c405 const int16x8_t vi23x89ABCDEF = vmovl_s8(vld1_s8(i23)); i23 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16() local
410 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi23x89ABCDEF), vget_low_s16(vk23x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
411 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi23x89ABCDEF), vget_high_s16(vk23x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
Dup16x25-minmax-fp32-avx2-mul32.c377 … const __m256i vi23x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i23 + 8))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul32() local
382 … vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi23x89ABCDEF, vk23x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul32()
Dup16x25-minmax-fp32-neon-mla8-ld64.c402 const int8x8_t vi23x89ABCDEF = vld1_s8(i23); i23 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neon_mla8_ld64() local
406 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi23x89ABCDEF, vk23x89ABCDEF); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neon_mla8_ld64()
Dup16x25-minmax-fp32-neonv8-mla8-ld64.c402 const int8x8_t vi23x89ABCDEF = vld1_s8(i23); i23 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mla8_ld64() local
406 vprod89ABCDEF = vmlal_s8(vprod89ABCDEF, vi23x89ABCDEF, vk23x89ABCDEF); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mla8_ld64()
Dup16x25-minmax-fp32-wasmsimd-mul16-add16.c445 const v128_t vi23x89ABCDEF = wasm_i16x8_load8x8(i23 + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__wasmsimd_mul16_add16() local
450 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi23x89ABCDEF, vk23x89ABCDEF)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__wasmsimd_mul16_add16()
Dup16x25-minmax-fp32-sse2-mul16-add16.c583 const __m128i vi23x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i23 + 8)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__sse2_mul16_add16() local
589 … const __m128i vxi23x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x89ABCDEF, vi23x89ABCDEF), 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__sse2_mul16_add16()
/external/XNNPACK/src/qu8-dwconv/gen/
Dup16x25-minmax-fp32-wasmsimd-mul16.c538 const v128_t vi23x89ABCDEF = wasm_u16x8_load8x8(i23 + 8); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__wasmsimd_mul16() local
541 vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi23x89ABCDEF); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__wasmsimd_mul16()
545 vprod89ABCDEF = wasm_i16x8_mul(vi23x89ABCDEF, vk23x89ABCDEF); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__wasmsimd_mul16()
Dup16x25-minmax-fp32-neon-mul16.c407 const int16x8_t vi23x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i23))); i23 += 8; in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16() local
412 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi23x89ABCDEF), vget_low_s16(vk23x89ABCDEF)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
413 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi23x89ABCDEF), vget_high_s16(vk23x89ABCDEF)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16()
Dup16x25-minmax-fp32-neonv8-mul16.c407 const int16x8_t vi23x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i23))); i23 += 8; in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16() local
412 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi23x89ABCDEF), vget_low_s16(vk23x89ABCDEF)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
413 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi23x89ABCDEF), vget_high_s16(vk23x89ABCDEF)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul16()
Dup16x25-minmax-rndnu-neon-mul16.c408 const int16x8_t vi23x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i23))); i23 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16() local
413 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi23x89ABCDEF), vget_low_s16(vk23x89ABCDEF)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16()
414 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi23x89ABCDEF), vget_high_s16(vk23x89ABCDEF)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul16()
Dup16x25-minmax-rndnu-neon-mul8.c500 const uint8x8_t vi23x89ABCDEF = vld1_u8(i23); i23 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul8() local
505 vprod89ABCDEF = vmull_u8(vi23x89ABCDEF, vk23x89ABCDEF); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul8()
506 vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi23x89ABCDEF); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul8()
Dup16x25-minmax-fp32-avx2-mul32.c378 … const __m256i vi23x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i23 + 8))); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul32() local
383 … vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi23x89ABCDEF, vk23x89ABCDEF)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul32()
/external/XNNPACK/src/f32-dwconv/gen/
Dup16x25-minmax-avx.c377 const __m256 vi23x89ABCDEF = _mm256_loadu_ps(i23 + 8); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
383 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi23x89ABCDEF, vk23x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup16x25-minmax-fma3-acc2.c377 const __m256 vi23x89ABCDEF = _mm256_loadu_ps(i23 + 8); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
383 vacc89ABCDEFp1 = _mm256_fmadd_ps(vi23x89ABCDEF, vk23x89ABCDEF, vacc89ABCDEFp1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-fma3.c377 const __m256 vi23x89ABCDEF = _mm256_loadu_ps(i23 + 8); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
383 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi23x89ABCDEF, vk23x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-avx-acc2.c377 const __m256 vi23x89ABCDEF = _mm256_loadu_ps(i23 + 8); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
383 vacc89ABCDEFp1 = _mm256_add_ps(vacc89ABCDEFp1, _mm256_mul_ps(vi23x89ABCDEF, vk23x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup16x25-minmax-neonfp16arith-acc2.c333 const float16x8_t vi23x89ABCDEF = vld1q_f16(i23); i23 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
337 vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi23x89ABCDEF, vk23x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c333 const float16x8_t vi23x89ABCDEF = vld1q_f16(i23); i23 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
337 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi23x89ABCDEF, vk23x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()

1234