Home
last modified time | relevance | path

Searched refs:vk6x89ABCDEF (Results 1 – 25 of 26) sorted by relevance

12

/external/XNNPACK/src/qs8-dwconv/gen/
Dup16x9-minmax-neon-mul16.c158 …const int16x8_t vk6x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof… in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16() local
162 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi6x89ABCDEF), vget_low_s16(vk6x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
163 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi6x89ABCDEF), vget_high_s16(vk6x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
Dup24x9-minmax-neon-mul16.c184 …const int16x8_t vk6x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof… in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16() local
190 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi6x89ABCDEF), vget_low_s16(vk6x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
191 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi6x89ABCDEF), vget_high_s16(vk6x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
Dup16x9-minmax-avx2-mul32.c144 …const __m256i vk6x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32() local
148 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi6x89ABCDEF, vk6x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
Dup32x9-minmax-neon-mul16.c210 …const int16x8_t vk6x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof… in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16() local
218 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi6x89ABCDEF), vget_low_s16(vk6x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
219 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi6x89ABCDEF), vget_high_s16(vk6x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
Dup24x9-minmax-avx2-mul32.c163 …const __m256i vk6x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32() local
169 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi6x89ABCDEF, vk6x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
Dup16x9-minmax-wasmsimd-mul16.c176 …const v128_t vk6x89ABCDEF = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() local
180 const v128_t vprod6x89ABCDEF = wasm_i16x8_mul(vi6x89ABCDEF, vk6x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
Dup16x9-minmax-ssse3-mul16.c218 …const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
224 …st __m128i vxk6x89ABCDEF = _mm_unpacklo_epi8(vk6x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk6… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
Dup16x9-minmax-sse2-mul16.c218 …const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
224 …st __m128i vxk6x89ABCDEF = _mm_unpacklo_epi8(vk6x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk6… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
Dup32x9-minmax-avx2-mul32.c182 …const __m256i vk6x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32() local
190 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi6x89ABCDEF, vk6x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
Dup24x9-minmax-wasmsimd-mul16.c208 …const v128_t vk6x89ABCDEF = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() local
214 const v128_t vprod6x89ABCDEF = wasm_i16x8_mul(vi6x89ABCDEF, vk6x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
Dup16x9-minmax-sse41-mul16.c221 …const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
222 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
Dup24x9-minmax-ssse3-mul16.c268 …const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
276 …st __m128i vxk6x89ABCDEF = _mm_unpacklo_epi8(vk6x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk6… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
Dup24x9-minmax-sse2-mul16.c268 …const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
276 …st __m128i vxk6x89ABCDEF = _mm_unpacklo_epi8(vk6x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk6… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
/external/XNNPACK/src/f32-dwconv/gen/
Dup16x9-minmax-fma3.c150 const __m256 vk6x89ABCDEF = _mm256_load_ps(w + 120); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local
152 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi6x89ABCDEF, vk6x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
Dup16x9-minmax-fma3-acc2.c150 const __m256 vk6x89ABCDEF = _mm256_load_ps(w + 120); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local
152 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi6x89ABCDEF, vk6x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
Dup16x9-minmax-avx-acc2.c150 const __m256 vk6x89ABCDEF = _mm256_load_ps(w + 120); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx_acc2() local
152 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi6x89ABCDEF, vk6x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx_acc2()
Dup16x9-minmax-avx.c150 const __m256 vk6x89ABCDEF = _mm256_load_ps(w + 120); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local
152 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi6x89ABCDEF, vk6x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx()
Dup16x25-minmax-fma3.c230 const __m256 vk6x89ABCDEF = _mm256_load_ps(w + 120); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
232 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi6x89ABCDEF, vk6x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c230 const __m256 vk6x89ABCDEF = _mm256_load_ps(w + 120); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
232 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi6x89ABCDEF, vk6x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c230 const __m256 vk6x89ABCDEF = _mm256_load_ps(w + 120); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
232 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi6x89ABCDEF, vk6x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup16x25-minmax-avx-acc2.c230 const __m256 vk6x89ABCDEF = _mm256_load_ps(w + 120); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
232 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi6x89ABCDEF, vk6x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup16x9-minmax-neonfp16arith-acc2.c136 const float16x8_t vk6x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local
138 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi6x89ABCDEF, vk6x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
Dup16x9-minmax-neonfp16arith.c136 const float16x8_t vk6x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local
138 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi6x89ABCDEF, vk6x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
Dup16x25-minmax-neonfp16arith-acc2.c216 const float16x8_t vk6x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
218 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi6x89ABCDEF, vk6x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c216 const float16x8_t vk6x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
218 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi6x89ABCDEF, vk6x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()

12