Home
last modified time | relevance | path

Searched refs:vk1x89ABCDEF (Results 1 – 25 of 31) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv/gen/
Dup16x4-minmax-fma3.c80 const __m256 vk1x89ABCDEF = _mm256_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3() local
82 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi1x89ABCDEF, vk1x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
Dup16x4-minmax-fma3-acc2.c80 const __m256 vk1x89ABCDEF = _mm256_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2() local
82 __m256 vacc89ABCDEFp1 = _mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2()
Dup16x4-minmax-avx-acc2.c80 const __m256 vk1x89ABCDEF = _mm256_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2() local
82 __m256 vacc89ABCDEFp1 = _mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2()
Dup16x4-minmax-avx.c80 const __m256 vk1x89ABCDEF = _mm256_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx() local
82 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx()
Dup16x9-minmax-fma3.c105 const __m256 vk1x89ABCDEF = _mm256_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local
107 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi1x89ABCDEF, vk1x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
Dup16x9-minmax-fma3-acc2.c105 const __m256 vk1x89ABCDEF = _mm256_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local
107 __m256 vacc89ABCDEFp1 = _mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
Dup16x9-minmax-avx.c105 const __m256 vk1x89ABCDEF = _mm256_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local
107 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx()
Dup16x25-minmax-fma3.c185 const __m256 vk1x89ABCDEF = _mm256_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
187 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi1x89ABCDEF, vk1x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c185 const __m256 vk1x89ABCDEF = _mm256_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
187 __m256 vacc89ABCDEFp1 = _mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c185 const __m256 vk1x89ABCDEF = _mm256_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
187 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
/external/XNNPACK/src/f16-dwconv/gen/
Dup16x4-minmax-neonfp16arith-acc2.c76 const float16x8_t vk1x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2() local
78 float16x8_t vacc89ABCDEFp1 = vmulq_f16(vi1x89ABCDEF, vk1x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
Dup16x4-minmax-neonfp16arith.c76 const float16x8_t vk1x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith() local
78 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi1x89ABCDEF, vk1x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
Dup16x9-minmax-neonfp16arith.c101 const float16x8_t vk1x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local
103 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi1x89ABCDEF, vk1x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
Dup16x9-minmax-neonfp16arith-acc2.c101 const float16x8_t vk1x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local
103 float16x8_t vacc89ABCDEFp1 = vmulq_f16(vi1x89ABCDEF, vk1x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
/external/XNNPACK/src/qs8-dwconv/gen/
Dup16x9-minmax-neon-mul16.c108 …const int16x8_t vk1x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof… in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16() local
112 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
113 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
Dup24x9-minmax-neon-mul16.c114 …const int16x8_t vk1x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof… in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16() local
120 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
121 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
Dup16x9-minmax-avx2-mul32.c99 …const __m256i vk1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32() local
103 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi1x89ABCDEF, vk1x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
Dup32x9-minmax-neon-mul16.c120 …const int16x8_t vk1x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof… in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16() local
128 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
129 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
Dup24x9-minmax-avx2-mul32.c103 …const __m256i vk1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32() local
109 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi1x89ABCDEF, vk1x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
Dup16x9-minmax-wasmsimd-mul16.c106 …const v128_t vk1x89ABCDEF = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() local
110 const v128_t vprod1x89ABCDEF = wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
Dup16x9-minmax-ssse3-mul16.c113 …const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
119 …st __m128i vxk1x89ABCDEF = _mm_unpacklo_epi8(vk1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk1… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
Dup16x9-minmax-sse2-mul16.c113 …const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
119 …st __m128i vxk1x89ABCDEF = _mm_unpacklo_epi8(vk1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk1… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
Dup32x9-minmax-avx2-mul32.c107 …const __m256i vk1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32() local
115 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi1x89ABCDEF, vk1x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
Dup24x9-minmax-wasmsimd-mul16.c113 …const v128_t vk1x89ABCDEF = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() local
119 const v128_t vprod1x89ABCDEF = wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
Dup16x9-minmax-sse41-mul16.c116 …const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
117 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()

12