/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7p7x-minmax-wasmsimd-c16-acc2.c | 60 const v128_t vxi5x89ABCDEF = wasm_i16x8_load_8x8(i5 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 74 vacc1x89ABCDEF = wasm_i16x8_add(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 122 const v128_t vxi5x89ABCDEF = wasm_i16x8_load_8x8(i5 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 136 vacc1x89ABCDEF = wasm_i16x8_add(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 204 const v128_t vxi5x89ABCDEF = wasm_i16x8_load_8x8(i5 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 218 vacc1x89ABCDEF = wasm_i16x8_add(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
|
D | 7p7x-minmax-sse41-c16-acc2.c | 60 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 75 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() 123 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 138 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() 205 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 220 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
|
D | 7p7x-minmax-ssse3-c16-acc2.c | 77 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 89 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() 156 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 168 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() 254 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 266 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
D | 7p7x-minmax-wasmsimd-c24-acc2.c | 65 const v128_t vxi5x89ABCDEF = wasm_i16x8_load_8x8(i5 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 84 vacc1x89ABCDEF = wasm_i16x8_add(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 181 const v128_t vxi5x89ABCDEF = wasm_i16x8_load_8x8(i5 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 200 vacc1x89ABCDEF = wasm_i16x8_add(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 317 const v128_t vxi5x89ABCDEF = wasm_i16x8_load_8x8(i5 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 336 vacc1x89ABCDEF = wasm_i16x8_add(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
|
D | 7p7x-minmax-sse2-c16-acc2.c | 77 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 89 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 156 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 168 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 254 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 266 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
|
D | 7p7x-minmax-sse41-c24-acc2.c | 65 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 85 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() 183 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 203 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() 320 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 340 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
|
D | 7p7x-minmax-sse2-c24-acc2.c | 89 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 106 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 239 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 256 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 408 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 425 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
|
D | 7p7x-minmax-ssse3-c24-acc2.c | 89 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 106 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() 239 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 256 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() 408 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 425 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
|
D | 7x-minmax-sse41-c16-acc2.c | 77 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local 92 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
|
D | 7x-minmax-wasmsimd-c16-acc2.c | 78 const v128_t vxi5x89ABCDEF = wasm_i16x8_load_8x8(i5 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local 92 vacc1x89ABCDEF = wasm_i16x8_add(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
|
D | 7x-minmax-sse41-c24-acc2.c | 82 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local 102 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
D | 7x-minmax-ssse3-c16-acc2.c | 94 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local 106 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
|
D | 7x-minmax-wasmsimd-c24-acc2.c | 83 const v128_t vxi5x89ABCDEF = wasm_i16x8_load_8x8(i5 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local 102 vacc1x89ABCDEF = wasm_i16x8_add(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
|
D | 7x-minmax-sse2-c16-acc2.c | 94 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local 106 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
D | 7x-minmax-ssse3-c24-acc2.c | 106 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local 123 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
|
D | 7x-minmax-sse2-c24-acc2.c | 106 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 123 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up16x9-minmax-sse41-mul16.c | 199 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local 207 const __m128i vp5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() 208 const __m128i vp5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
|
D | up16x9-minmax-ssse3-mul16.c | 202 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local 207 const __m128i vp5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() 208 const __m128i vp5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
|
D | up16x9-minmax-sse2-mul16.c | 202 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local 207 const __m128i vp5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() 208 const __m128i vp5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
|
D | up24x9-minmax-sse41-mul16.c | 241 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local 253 const __m128i vp5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() 254 const __m128i vp5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
|
D | up24x9-minmax-ssse3-mul16.c | 246 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local 253 const __m128i vp5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() 254 const __m128i vp5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
|
D | up24x9-minmax-sse2-mul16.c | 246 …const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local 253 const __m128i vp5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() 254 const __m128i vp5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
|