/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7p7x-minmax-wasmsimd-c24-acc2.c | 50 const v128_t vxi1xGHIJKLMN = wasm_i16x8_load_8x8(i1 + 16); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 75 v128_t vacc0xGHIJKLMN = wasm_i16x8_add(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 166 const v128_t vxi1xGHIJKLMN = wasm_i16x8_load_8x8(i1 + 16); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 191 v128_t vacc0xGHIJKLMN = wasm_i16x8_add(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 302 const v128_t vxi1xGHIJKLMN = wasm_i16x8_load_8x8(i1 + 16); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 327 v128_t vacc0xGHIJKLMN = wasm_i16x8_add(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
|
D | 7p7x-minmax-sse41-c24-acc2.c | 50 const __m128i vxi1xGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 16))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 76 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() 168 const __m128i vxi1xGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 16))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 194 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() 305 const __m128i vxi1xGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 16))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 331 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
|
D | 7p7x-minmax-sse2-c24-acc2.c | 78 …const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 97 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 228 …const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 247 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 397 …const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 416 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
|
D | 7p7x-minmax-ssse3-c24-acc2.c | 78 …const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 97 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() 228 …const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 247 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() 397 …const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 416 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
|
D | 7x-minmax-sse41-c24-acc2.c | 67 const __m128i vxi1xGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 16))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local 93 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
D | 7x-minmax-wasmsimd-c24-acc2.c | 68 const v128_t vxi1xGHIJKLMN = wasm_i16x8_load_8x8(i1 + 16); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local 93 v128_t vacc0xGHIJKLMN = wasm_i16x8_add(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
|
D | 7x-minmax-sse2-c24-acc2.c | 95 …const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 114 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
|
D | 7x-minmax-ssse3-c24-acc2.c | 95 …const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local 114 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up24x9-minmax-sse41-mul16.c | 129 const __m128i vxi1xGHIJKLMN = _mm_cvtepi8_epi16(vi1xGHIJKLMN); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local 139 const __m128i vp1xGHIJKLMNlo = _mm_mullo_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() 140 const __m128i vp1xGHIJKLMNhi = _mm_mulhi_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
|
D | up24x9-minmax-ssse3-mul16.c | 132 …const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local 139 const __m128i vp1xGHIJKLMNlo = _mm_mullo_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() 140 const __m128i vp1xGHIJKLMNhi = _mm_mulhi_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
|
D | up24x9-minmax-sse2-mul16.c | 132 …const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local 139 const __m128i vp1xGHIJKLMNlo = _mm_mullo_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() 140 const __m128i vp1xGHIJKLMNhi = _mm_mulhi_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
|