| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 3x4c8-minmax-xop-ld128.c | 187 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld128() local
192 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld128()
194 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld128()
|
| D | 3x4c8-minmax-xop-ld64.c | 189 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld64() local
194 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld64()
196 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld64()
|
| D | 3x4c8-minmax-sse41-ld64.c | 184 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld64() local
189 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld64()
191 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld64()
|
| D | 3x4c8-minmax-sse41-ld128.c | 182 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld128() local
187 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld128()
189 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld128()
|
| D | 3x4c8-minmax-ssse3-ld64.c | 216 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld64() local
221 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld64()
223 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld64()
|
| D | 3x4c8-minmax-ssse3-ld128.c | 214 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld128() local
219 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld128()
221 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld128()
|
| D | 3x4c8-minmax-sse2-ld128.c | 214 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128() local
219 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
221 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
|
| D | 3x4c8-minmax-sse2-ld64.c | 216 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64() local
221 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64()
223 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64()
|
| D | 3x4c8-minmax-wasmsimd-ld128.c | 199 …v128_t vacc22x0123 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123), vout… in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld128() local
201 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc22x0123); in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld128()
|
| D | 3x4c8-minmax-wasmsimd-ld64.c | 195 …v128_t vacc22x0123 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123), vout… in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld64() local
197 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc22x0123); in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld64()
|
| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 3x4c8-minmax-xop-ld64.c | 172 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld64() local
177 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld64()
179 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld64()
|
| D | 3x4c8-xw-minmax-xop.c | 168 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__xop() local
173 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__xop()
175 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__xop()
|
| D | 3x4c8-minmax-sse41-ld64.c | 167 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld64() local
172 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld64()
174 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld64()
|
| D | 3x4c8-xw-minmax-sse41.c | 163 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse41() local
168 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse41()
170 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse41()
|
| D | 3x4c8-minmax-sse41-ld128.c | 165 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld128() local
170 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld128()
172 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld128()
|
| D | 3x4c8-minmax-xop-ld128.c | 170 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld128() local
175 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld128()
177 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_minmax_ukernel_3x4c8__xop_ld128()
|
| D | 3x4c8-xw-minmax-ssse3.c | 195 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3() local
200 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3()
202 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__ssse3()
|
| D | 3x4c8-minmax-sse2-ld128.c | 197 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128() local
202 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
204 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
|
| D | 3x4c8-xw-minmax-sse2.c | 195 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2() local
200 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
202 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
|
| D | 3x4c8-minmax-sse2-ld64.c | 199 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64() local
204 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
206 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
|
| D | 3x4c8-minmax-ssse3-ld128.c | 197 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld128() local
202 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld128()
204 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld128()
|
| D | 3x4c8-minmax-ssse3-ld64.c | 199 … __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld64() local
204 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld64()
206 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld64()
|
| D | 3x4c8-minmax-wasmsimd-ld64.c | 178 …v128_t vacc22x0123 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123), vout… in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld64() local
180 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld64()
|
| D | 3x4c8-xw-minmax-wasmsimd.c | 178 …v128_t vacc22x0123 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123), vout… in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__wasmsimd() local
180 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__wasmsimd()
|
| D | 3x4c8-minmax-wasmsimd-ld128.c | 182 …v128_t vacc22x0123 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123), vout… in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld128() local
184 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc22x0123); in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld128()
|