/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7p7x-minmax-wasmsimd-c16-acc2.c | 48 const v128_t vxi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 67 v128_t vacc0x89ABCDEF = wasm_i16x8_add(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 110 const v128_t vxi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 129 v128_t vacc0x89ABCDEF = wasm_i16x8_add(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 192 const v128_t vxi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 211 v128_t vacc0x89ABCDEF = wasm_i16x8_add(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
|
D | 7p7x-minmax-sse41-c16-acc2.c | 48 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 68 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() 111 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 131 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() 193 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 213 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
|
D | 7p7x-minmax-ssse3-c16-acc2.c | 69 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 82 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() 148 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 161 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() 246 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 259 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
D | 7p7x-minmax-wasmsimd-c24-acc2.c | 49 const v128_t vxi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 74 v128_t vacc0x89ABCDEF = wasm_i16x8_add(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 165 const v128_t vxi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 190 v128_t vacc0x89ABCDEF = wasm_i16x8_add(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 301 const v128_t vxi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 326 v128_t vacc0x89ABCDEF = wasm_i16x8_add(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
|
D | 7p7x-minmax-sse2-c16-acc2.c | 69 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 82 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 148 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 161 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 246 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 259 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
|
D | 7p7x-minmax-sse41-c24-acc2.c | 49 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 75 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() 167 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 193 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() 304 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 330 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
|
D | 7p7x-minmax-sse2-c24-acc2.c | 77 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 96 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 227 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 246 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 396 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 415 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
|
D | 7p7x-minmax-ssse3-c24-acc2.c | 77 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 96 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() 227 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 246 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() 396 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 415 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
|
D | 7x-minmax-sse41-c16-acc2.c | 65 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local 85 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
|
D | 7x-minmax-wasmsimd-c16-acc2.c | 66 const v128_t vxi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local 85 v128_t vacc0x89ABCDEF = wasm_i16x8_add(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
|
D | 7x-minmax-sse41-c24-acc2.c | 66 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local 92 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
D | 7x-minmax-ssse3-c16-acc2.c | 86 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local 99 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
|
D | 7x-minmax-wasmsimd-c24-acc2.c | 67 const v128_t vxi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local 92 v128_t vacc0x89ABCDEF = wasm_i16x8_add(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
|
D | 7x-minmax-sse2-c16-acc2.c | 86 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local 99 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
D | 7x-minmax-ssse3-c24-acc2.c | 94 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local 113 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
|
D | 7x-minmax-sse2-c24-acc2.c | 94 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 113 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up16x9-minmax-sse41-mul16.c | 115 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local 123 const __m128i vp1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() 124 const __m128i vp1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
|
D | up16x9-minmax-ssse3-mul16.c | 118 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local 123 const __m128i vp1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() 124 const __m128i vp1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
|
D | up16x9-minmax-sse2-mul16.c | 118 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local 123 const __m128i vp1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() 124 const __m128i vp1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
|
D | up24x9-minmax-sse41-mul16.c | 125 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local 137 const __m128i vp1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() 138 const __m128i vp1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
|
D | up24x9-minmax-ssse3-mul16.c | 130 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local 137 const __m128i vp1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() 138 const __m128i vp1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
|
D | up24x9-minmax-sse2-mul16.c | 130 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local 137 const __m128i vp1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() 138 const __m128i vp1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
|