Home
last modified time | relevance | path

Searched refs:vk2x89ABCDEF (Results 1 – 25 of 31) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv/gen/
Dup16x4-minmax-fma3.c89 const __m256 vk2x89ABCDEF = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3() local
91 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
Dup16x4-minmax-fma3-acc2.c89 const __m256 vk2x89ABCDEF = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2() local
91 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2()
Dup16x4-minmax-avx-acc2.c89 const __m256 vk2x89ABCDEF = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2() local
91 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi2x89ABCDEF, vk2x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2()
Dup16x4-minmax-avx.c89 const __m256 vk2x89ABCDEF = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx() local
91 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi2x89ABCDEF, vk2x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx()
Dup16x9-minmax-fma3.c114 const __m256 vk2x89ABCDEF = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3() local
116 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
Dup16x9-minmax-fma3-acc2.c114 const __m256 vk2x89ABCDEF = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2() local
116 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
Dup16x9-minmax-avx.c114 const __m256 vk2x89ABCDEF = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx() local
116 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi2x89ABCDEF, vk2x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx()
Dup16x25-minmax-fma3.c194 const __m256 vk2x89ABCDEF = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
196 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c194 const __m256 vk2x89ABCDEF = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
196 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c194 const __m256 vk2x89ABCDEF = _mm256_load_ps(w + 56); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
196 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi2x89ABCDEF, vk2x89ABCDEF)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
/external/XNNPACK/src/f16-dwconv/gen/
Dup16x4-minmax-neonfp16arith-acc2.c83 const float16x8_t vk2x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2() local
85 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2()
Dup16x4-minmax-neonfp16arith.c83 const float16x8_t vk2x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith() local
85 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith()
Dup16x9-minmax-neonfp16arith.c108 const float16x8_t vk2x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith() local
110 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith()
Dup16x9-minmax-neonfp16arith-acc2.c108 const float16x8_t vk2x89ABCDEF = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2() local
110 vacc89ABCDEFp0 = vfmaq_f16(vacc89ABCDEFp0, vi2x89ABCDEF, vk2x89ABCDEF); in xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2()
/external/XNNPACK/src/qs8-dwconv/gen/
Dup16x9-minmax-neon-mul16.c118 …const int16x8_t vk2x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof… in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16() local
122 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
123 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
Dup24x9-minmax-neon-mul16.c128 …const int16x8_t vk2x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof… in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16() local
134 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
135 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
Dup16x9-minmax-avx2-mul32.c108 …const __m256i vk2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32() local
112 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi2x89ABCDEF, vk2x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
Dup32x9-minmax-neon-mul16.c138 …const int16x8_t vk2x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((uintptr_t) w + 8 * sizeof… in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16() local
146 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
147 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
Dup24x9-minmax-avx2-mul32.c115 …const __m256i vk2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32() local
121 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi2x89ABCDEF, vk2x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
Dup16x9-minmax-wasmsimd-mul16.c120 …const v128_t vk2x89ABCDEF = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() local
124 const v128_t vprod2x89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
Dup16x9-minmax-ssse3-mul16.c134 …const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
140 …st __m128i vxk2x89ABCDEF = _mm_unpacklo_epi8(vk2x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk2… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
Dup16x9-minmax-sse2-mul16.c134 …const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
140 …st __m128i vxk2x89ABCDEF = _mm_unpacklo_epi8(vk2x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk2… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
Dup32x9-minmax-avx2-mul32.c122 …const __m256i vk2x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32() local
130 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi2x89ABCDEF, vk2x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
Dup24x9-minmax-wasmsimd-mul16.c132 …const v128_t vk2x89ABCDEF = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() local
138 const v128_t vprod2x89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
Dup16x9-minmax-sse41-mul16.c137 …const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
138 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()

12