Home
last modified time | relevance | path

Searched refs:vk5x89ABCDEF (Results 1 – 25 of 25) sorted by relevance

/external/XNNPACK/src/qs8-dwconv/gen/
Dup16x9-minmax-neon-mul16.c148 …const int16x8_t vk5x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof… in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16() local
152 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi5x89ABCDEF), vget_low_s16(vk5x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
153 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi5x89ABCDEF), vget_high_s16(vk5x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
Dup24x9-minmax-neon-mul16.c170 …const int16x8_t vk5x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof… in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16() local
176 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi5x89ABCDEF), vget_low_s16(vk5x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
177 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi5x89ABCDEF), vget_high_s16(vk5x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
Dup16x9-minmax-avx2-mul32.c135 …const __m256i vk5x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32() local
139 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi5x89ABCDEF, vk5x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
Dup32x9-minmax-neon-mul16.c192 …const int16x8_t vk5x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof… in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16() local
200 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi5x89ABCDEF), vget_low_s16(vk5x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
201 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi5x89ABCDEF), vget_high_s16(vk5x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
Dup24x9-minmax-avx2-mul32.c151 …const __m256i vk5x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32() local
157 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi5x89ABCDEF, vk5x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
Dup16x9-minmax-wasmsimd-mul16.c162 …const v128_t vk5x89ABCDEF = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() local
166 const v128_t vprod5x89ABCDEF = wasm_i16x8_mul(vi5x89ABCDEF, vk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
Dup16x9-minmax-ssse3-mul16.c197 …const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
203 …st __m128i vxk5x89ABCDEF = _mm_unpacklo_epi8(vk5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk5… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
Dup16x9-minmax-sse2-mul16.c197 …const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
203 …st __m128i vxk5x89ABCDEF = _mm_unpacklo_epi8(vk5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk5… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
Dup32x9-minmax-avx2-mul32.c167 …const __m256i vk5x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32() local
175 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi5x89ABCDEF, vk5x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
Dup24x9-minmax-wasmsimd-mul16.c189 …const v128_t vk5x89ABCDEF = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() local
195 const v128_t vprod5x89ABCDEF = wasm_i16x8_mul(vi5x89ABCDEF, vk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
Dup16x9-minmax-sse41-mul16.c200 …const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
201 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
Dup24x9-minmax-ssse3-mul16.c239 …const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
247 …st __m128i vxk5x89ABCDEF = _mm_unpacklo_epi8(vk5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk5… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
Dup24x9-minmax-sse2-mul16.c239 …const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
247 …st __m128i vxk5x89ABCDEF = _mm_unpacklo_epi8(vk5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk5… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
Dup24x9-minmax-sse41-mul16.c242 …const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
243 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
/external/XNNPACK/src/f32-dwconv/gen/
Dup16x9-minmax-fma3.c141 const __m256 vk5x89ABCDEF = _mm256_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local
143 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi5x89ABCDEF, vk5x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
Dup16x9-minmax-fma3-acc2.c141 const __m256 vk5x89ABCDEF = _mm256_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local
143 vacc89ABCDEFp1 = _mm256_fmadd_ps(vi5x89ABCDEF, vk5x89ABCDEF, vacc89ABCDEFp1); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
Dup16x9-minmax-avx.c141 const __m256 vk5x89ABCDEF = _mm256_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local
143 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi5x89ABCDEF, vk5x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx()
Dup16x25-minmax-fma3.c221 const __m256 vk5x89ABCDEF = _mm256_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
223 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi5x89ABCDEF, vk5x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c221 const __m256 vk5x89ABCDEF = _mm256_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
223 vacc89ABCDEFp1 = _mm256_fmadd_ps(vi5x89ABCDEF, vk5x89ABCDEF, vacc89ABCDEFp1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c221 const __m256 vk5x89ABCDEF = _mm256_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
223 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi5x89ABCDEF, vk5x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup16x25-minmax-avx-acc2.c221 const __m256 vk5x89ABCDEF = _mm256_load_ps(w + 104); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
223 vacc89ABCDEFp1 = _mm256_add_ps(vacc89ABCDEFp1, _mm256_mul_ps(vi5x89ABCDEF, vk5x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup16x9-minmax-neonfp16arith.c129 const float16x8_t vk5x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local
131 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi5x89ABCDEF, vk5x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
Dup16x9-minmax-neonfp16arith-acc2.c129 const float16x8_t vk5x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local
131 vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi5x89ABCDEF, vk5x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith-acc2.c209 const float16x8_t vk5x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
211 vacc89ABCDEFp1 = vfmaq_f16(vacc89ABCDEFp1, vi5x89ABCDEF, vk5x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c209 const float16x8_t vk5x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
211 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi5x89ABCDEF, vk5x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()