/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7p7x-minmax-wasmsimd-c16-acc2.c | 54 const v128_t vxi3x89ABCDEF = wasm_i16x8_load_8x8(i3 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 69 v128_t vacc1x89ABCDEF = wasm_i16x8_add(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 116 const v128_t vxi3x89ABCDEF = wasm_i16x8_load_8x8(i3 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 131 v128_t vacc1x89ABCDEF = wasm_i16x8_add(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 198 const v128_t vxi3x89ABCDEF = wasm_i16x8_load_8x8(i3 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 213 v128_t vacc1x89ABCDEF = wasm_i16x8_add(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
|
D | 7p7x-minmax-sse41-c16-acc2.c | 54 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 70 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() 117 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 133 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() 199 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 215 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
|
D | 7p7x-minmax-ssse3-c16-acc2.c | 73 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 84 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() 152 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 163 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() 250 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 261 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
D | 7p7x-minmax-sse2-c16-acc2.c | 73 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 84 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 152 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 163 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 250 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 261 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
|
D | 7p7x-minmax-wasmsimd-c24-acc2.c | 57 const v128_t vxi3x89ABCDEF = wasm_i16x8_load_8x8(i3 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 77 v128_t vacc1x89ABCDEF = wasm_i16x8_add(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 173 const v128_t vxi3x89ABCDEF = wasm_i16x8_load_8x8(i3 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 193 v128_t vacc1x89ABCDEF = wasm_i16x8_add(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 309 const v128_t vxi3x89ABCDEF = wasm_i16x8_load_8x8(i3 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 329 v128_t vacc1x89ABCDEF = wasm_i16x8_add(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
|
D | 7p7x-minmax-sse41-c24-acc2.c | 57 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 78 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() 175 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 196 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() 312 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 333 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
|
D | 7p7x-minmax-sse2-c24-acc2.c | 83 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 99 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 233 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 249 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 402 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 418 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
|
D | 7p7x-minmax-ssse3-c24-acc2.c | 83 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 99 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() 233 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 249 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() 402 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 418 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
|
D | 7x-minmax-sse41-c16-acc2.c | 71 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local 87 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
|
D | 7x-minmax-wasmsimd-c16-acc2.c | 72 const v128_t vxi3x89ABCDEF = wasm_i16x8_load_8x8(i3 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local 87 v128_t vacc1x89ABCDEF = wasm_i16x8_add(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
|
D | 7x-minmax-ssse3-c16-acc2.c | 90 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local 101 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
|
D | 7x-minmax-sse41-c24-acc2.c | 74 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local 95 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
D | 7x-minmax-wasmsimd-c24-acc2.c | 75 const v128_t vxi3x89ABCDEF = wasm_i16x8_load_8x8(i3 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local 95 v128_t vacc1x89ABCDEF = wasm_i16x8_add(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
|
D | 7x-minmax-sse2-c16-acc2.c | 90 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local 101 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
D | 7x-minmax-sse2-c24-acc2.c | 100 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 116 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
|
D | 7x-minmax-ssse3-c24-acc2.c | 100 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local 116 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up16x9-minmax-sse41-mul16.c | 157 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local 165 const __m128i vp3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() 166 const __m128i vp3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
|
D | up16x9-minmax-ssse3-mul16.c | 160 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local 165 const __m128i vp3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() 166 const __m128i vp3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
|
D | up16x9-minmax-sse2-mul16.c | 160 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local 165 const __m128i vp3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() 166 const __m128i vp3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
|
D | up24x9-minmax-sse41-mul16.c | 183 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local 195 const __m128i vp3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() 196 const __m128i vp3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
|
D | up24x9-minmax-ssse3-mul16.c | 188 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local 195 const __m128i vp3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() 196 const __m128i vp3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
|
D | up24x9-minmax-sse2-mul16.c | 188 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local 195 const __m128i vp3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() 196 const __m128i vp3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
|