/external/XNNPACK/src/f32-vbinary/gen/ |
D | vmax-wasmsimd-x86-x16.c | 42 const v128_t vbCDEF = wasm_v128_load(b + 12); in xnn_f32_vmax_ukernel__wasmsimd_x86_x16() local 48 const v128_t vmCDEF = wasm_f32x4_le(vaCDEF, vbCDEF); in xnn_f32_vmax_ukernel__wasmsimd_x86_x16() 53 v128_t vyCDEF = wasm_v128_bitselect(vbCDEF, vaCDEF, vmCDEF); in xnn_f32_vmax_ukernel__wasmsimd_x86_x16()
|
D | vmin-wasmsimd-x86-x16.c | 42 const v128_t vbCDEF = wasm_v128_load(b + 12); in xnn_f32_vmin_ukernel__wasmsimd_x86_x16() local 48 const v128_t vmCDEF = wasm_f32x4_lt(vaCDEF, vbCDEF); in xnn_f32_vmin_ukernel__wasmsimd_x86_x16() 53 v128_t vyCDEF = wasm_v128_bitselect(vaCDEF, vbCDEF, vmCDEF); in xnn_f32_vmin_ukernel__wasmsimd_x86_x16()
|
D | vsub-wasmsimd-x16.c | 42 const v128_t vbCDEF = wasm_v128_load(b + 12); in xnn_f32_vsub_ukernel__wasmsimd_x16() local 48 v128_t vyCDEF = wasm_f32x4_sub(vaCDEF, vbCDEF); in xnn_f32_vsub_ukernel__wasmsimd_x16()
|
D | vmul-wasmsimd-x16.c | 42 const v128_t vbCDEF = wasm_v128_load(b + 12); in xnn_f32_vmul_ukernel__wasmsimd_x16() local 48 v128_t vyCDEF = wasm_f32x4_mul(vaCDEF, vbCDEF); in xnn_f32_vmul_ukernel__wasmsimd_x16()
|
D | vmax-wasmsimd-arm-x16.c | 42 const v128_t vbCDEF = wasm_v128_load(b + 12); in xnn_f32_vmax_ukernel__wasmsimd_arm_x16() local 48 v128_t vyCDEF = wasm_f32x4_max(vaCDEF, vbCDEF); in xnn_f32_vmax_ukernel__wasmsimd_arm_x16()
|
D | vadd-wasmsimd-x16.c | 42 const v128_t vbCDEF = wasm_v128_load(b + 12); in xnn_f32_vadd_ukernel__wasmsimd_x16() local 48 v128_t vyCDEF = wasm_f32x4_add(vaCDEF, vbCDEF); in xnn_f32_vadd_ukernel__wasmsimd_x16()
|
D | vdiv-wasmsimd-x16.c | 42 const v128_t vbCDEF = wasm_v128_load(b + 12); in xnn_f32_vdiv_ukernel__wasmsimd_x16() local 48 v128_t vyCDEF = wasm_f32x4_div(vaCDEF, vbCDEF); in xnn_f32_vdiv_ukernel__wasmsimd_x16()
|
D | vmin-wasmsimd-arm-x16.c | 42 const v128_t vbCDEF = wasm_v128_load(b + 12); in xnn_f32_vmin_ukernel__wasmsimd_arm_x16() local 48 v128_t vyCDEF = wasm_f32x4_min(vaCDEF, vbCDEF); in xnn_f32_vmin_ukernel__wasmsimd_arm_x16()
|
D | vsqrdiff-wasmsimd-x16.c | 42 const v128_t vbCDEF = wasm_v128_load(b + 12); in xnn_f32_vsqrdiff_ukernel__wasmsimd_x16() local 48 v128_t vyCDEF = wasm_f32x4_sub(vaCDEF, vbCDEF); in xnn_f32_vsqrdiff_ukernel__wasmsimd_x16()
|
D | vadd-relu-wasmsimd-x16.c | 43 const v128_t vbCDEF = wasm_v128_load(b + 12); in xnn_f32_vadd_relu_ukernel__wasmsimd_x16() local 49 v128_t vyCDEF = wasm_f32x4_add(vaCDEF, vbCDEF); in xnn_f32_vadd_relu_ukernel__wasmsimd_x16()
|
D | vsub-relu-wasmsimd-x16.c | 43 const v128_t vbCDEF = wasm_v128_load(b + 12); in xnn_f32_vsub_relu_ukernel__wasmsimd_x16() local 49 v128_t vyCDEF = wasm_f32x4_sub(vaCDEF, vbCDEF); in xnn_f32_vsub_relu_ukernel__wasmsimd_x16()
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 4x16c8-minmax-fp32-avx512skx.c | 115 …const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w … in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 117 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF)); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 118 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF)); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 119 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF)); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 120 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF)); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
|
D | 3x16c8-minmax-fp32-avx512skx.c | 100 …const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w … in xnn_qs8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() local 102 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF)); in xnn_qs8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() 103 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF)); in xnn_qs8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() 104 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF)); in xnn_qs8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx()
|
D | 2x16c8-minmax-fp32-avx512skx.c | 85 …const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w … in xnn_qs8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx() local 87 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF)); in xnn_qs8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx() 88 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF)); in xnn_qs8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx()
|
/external/XNNPACK/src/qc8-gemm/gen/ |
D | 4x16c8-minmax-fp32-avx512skx.c | 114 …const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w … in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 116 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 117 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 118 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 119 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
|
D | 3x16c8-minmax-fp32-avx512skx.c | 99 …const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w … in xnn_qc8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() local 101 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF)); in xnn_qc8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() 102 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF)); in xnn_qc8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() 103 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF)); in xnn_qc8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx()
|
D | 2x16c8-minmax-fp32-avx512skx.c | 84 …const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w … in xnn_qc8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx() local 86 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF)); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx() 87 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF)); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx()
|
/external/XNNPACK/src/qu8-igemm/gen/ |
D | 4x16c8-minmax-fp32-avx512skx.c | 131 …const __m512i vbCDEF = _mm512_sub_epi16(_mm512_cvtepu8_epi16(_mm256_load_si256((const __m256i*) ((… in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 133 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 134 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 135 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 136 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
|
D | 3x16c8-minmax-fp32-avx512skx.c | 114 …const __m512i vbCDEF = _mm512_sub_epi16(_mm512_cvtepu8_epi16(_mm256_load_si256((const __m256i*) ((… in xnn_qu8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() local 116 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF)); in xnn_qu8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() 117 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF)); in xnn_qu8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() 118 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF)); in xnn_qu8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 4x16c8-minmax-fp32-avx512skx.c | 130 …const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w … in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 132 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF)); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 133 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF)); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 134 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF)); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 135 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF)); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
|
D | 3x16c8-minmax-fp32-avx512skx.c | 113 …const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w … in xnn_qs8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() local 115 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF)); in xnn_qs8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() 116 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF)); in xnn_qs8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() 117 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF)); in xnn_qs8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx()
|
/external/XNNPACK/src/qu8-gemm/gen/ |
D | 4x16c8-minmax-fp32-avx512skx.c | 116 …const __m512i vbCDEF = _mm512_sub_epi16(_mm512_cvtepu8_epi16(_mm256_load_si256((const __m256i*) ((… in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 118 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 119 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 120 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 121 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
|
D | 3x16c8-minmax-fp32-avx512skx.c | 101 …const __m512i vbCDEF = _mm512_sub_epi16(_mm512_cvtepu8_epi16(_mm256_load_si256((const __m256i*) ((… in xnn_qu8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() local 103 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF)); in xnn_qu8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() 104 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF)); in xnn_qu8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() 105 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF)); in xnn_qu8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx()
|
/external/XNNPACK/src/qc8-igemm/gen/ |
D | 4x16c8-minmax-fp32-avx512skx.c | 129 …const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w … in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 131 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 132 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 133 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 134 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
|
D | 3x16c8-minmax-fp32-avx512skx.c | 112 …const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w … in xnn_qc8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() local 114 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF)); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() 115 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF)); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() 116 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF)); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx()
|