/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7p7x-minmax-wasmsimd-c16-acc2.c | 57 const v128_t vxi4x89ABCDEF = wasm_i16x8_load_8x8(i4 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 72 vacc0x89ABCDEF = wasm_i16x8_add(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 119 const v128_t vxi4x89ABCDEF = wasm_i16x8_load_8x8(i4 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 134 vacc0x89ABCDEF = wasm_i16x8_add(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 201 const v128_t vxi4x89ABCDEF = wasm_i16x8_load_8x8(i4 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 216 vacc0x89ABCDEF = wasm_i16x8_add(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
|
D | 7p7x-minmax-sse41-c16-acc2.c | 57 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 73 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() 120 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 136 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() 202 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 218 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
|
D | 7p7x-minmax-ssse3-c16-acc2.c | 75 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 87 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() 154 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 166 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() 252 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 264 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
D | 7p7x-minmax-sse2-c16-acc2.c | 75 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 87 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 154 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 166 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 252 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 264 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
|
D | 7p7x-minmax-wasmsimd-c24-acc2.c | 61 const v128_t vxi4x89ABCDEF = wasm_i16x8_load_8x8(i4 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 81 vacc0x89ABCDEF = wasm_i16x8_add(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 177 const v128_t vxi4x89ABCDEF = wasm_i16x8_load_8x8(i4 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 197 vacc0x89ABCDEF = wasm_i16x8_add(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 313 const v128_t vxi4x89ABCDEF = wasm_i16x8_load_8x8(i4 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 333 vacc0x89ABCDEF = wasm_i16x8_add(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
|
D | 7p7x-minmax-sse41-c24-acc2.c | 61 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 82 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() 179 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 200 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() 316 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 337 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
|
D | 7p7x-minmax-sse2-c24-acc2.c | 86 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 103 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 236 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 253 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 405 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 422 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
|
D | 7p7x-minmax-ssse3-c24-acc2.c | 86 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 103 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() 236 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 253 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() 405 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 422 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
|
D | 7x-minmax-sse41-c16-acc2.c | 74 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local 90 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
|
D | 7x-minmax-wasmsimd-c16-acc2.c | 75 const v128_t vxi4x89ABCDEF = wasm_i16x8_load_8x8(i4 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local 90 vacc0x89ABCDEF = wasm_i16x8_add(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
|
D | 7x-minmax-ssse3-c16-acc2.c | 92 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local 104 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
|
D | 7x-minmax-sse41-c24-acc2.c | 78 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local 99 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
D | 7x-minmax-wasmsimd-c24-acc2.c | 79 const v128_t vxi4x89ABCDEF = wasm_i16x8_load_8x8(i4 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local 99 vacc0x89ABCDEF = wasm_i16x8_add(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
|
D | 7x-minmax-sse2-c16-acc2.c | 92 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local 104 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
D | 7x-minmax-sse2-c24-acc2.c | 103 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 120 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
|
D | 7x-minmax-ssse3-c24-acc2.c | 103 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local 120 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up16x9-minmax-sse41-mul16.c | 178 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local 186 const __m128i vp4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() 187 const __m128i vp4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
|
D | up16x9-minmax-ssse3-mul16.c | 181 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local 186 const __m128i vp4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() 187 const __m128i vp4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
|
D | up16x9-minmax-sse2-mul16.c | 181 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local 186 const __m128i vp4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() 187 const __m128i vp4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
|
D | up24x9-minmax-sse41-mul16.c | 212 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local 224 const __m128i vp4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() 225 const __m128i vp4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
|
D | up24x9-minmax-ssse3-mul16.c | 217 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local 224 const __m128i vp4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() 225 const __m128i vp4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
|
D | up24x9-minmax-sse2-mul16.c | 217 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local 224 const __m128i vp4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() 225 const __m128i vp4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
|