Home
last modified time | relevance | path

Searched refs:MS_F32X8_GETI (Results 1 – 2 of 2) sorted by relevance

/third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/intrinsics/
Dms_simd_instructions.h39 #define MS_F32X8_GETI(src, i) src.m256_f32[i] macro
41 #define MS_F32X8_GETI(src, i) src[i] macro
217 MS_F32X8_GETI(dst, 0) = sqrtf(MS_F32X8_GETI(src, 0)); in MS_SQRTFX8_F32()
218 MS_F32X8_GETI(dst, 1) = sqrtf(MS_F32X8_GETI(src, 1)); in MS_SQRTFX8_F32()
219 MS_F32X8_GETI(dst, 2) = sqrtf(MS_F32X8_GETI(src, 2)); in MS_SQRTFX8_F32()
220 MS_F32X8_GETI(dst, 3) = sqrtf(MS_F32X8_GETI(src, 3)); in MS_SQRTFX8_F32()
221 MS_F32X8_GETI(dst, 4) = sqrtf(MS_F32X8_GETI(src, 4)); in MS_SQRTFX8_F32()
222 MS_F32X8_GETI(dst, 5) = sqrtf(MS_F32X8_GETI(src, 5)); in MS_SQRTFX8_F32()
223 MS_F32X8_GETI(dst, 6) = sqrtf(MS_F32X8_GETI(src, 6)); in MS_SQRTFX8_F32()
224 MS_F32X8_GETI(dst, 7) = sqrtf(MS_F32X8_GETI(src, 7)); in MS_SQRTFX8_F32()
/third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/fp32/
Dwinograd_avx.c45 dst_data[i + dst_k_offset + k * out_c] = MS_F32X8_GETI(m[k + m_k_offset], i); in OutputTransform4x2AvxUnit()
80 dst_data[i + dst_k_offset + k * out_c] = MS_F32X8_GETI(m[k + m_k_offset], i); in OutputTransform4x2ReluAvxUnit()
118 dst_data[i + dst_k_offset + k * out_c] = MS_F32X8_GETI(m[k + m_k_offset], i); in OutputTransform4x2Relu6AvxUnit()
154 dst_data[i + dst_k_offset + k * out_c] = MS_F32X8_GETI(m[k + m_k_offset], i); in OutputTransform4x3AvxUnit()
194 dst_data[i + dst_k_offset + k * out_c] = MS_F32X8_GETI(m[k + m_k_offset], i); in OutputTransform4x3ReluAvxUnit()
238 dst_data[i + dst_k_offset + k * out_c] = MS_F32X8_GETI(m[k + m_k_offset], i); in OutputTransform4x3Relu6AvxUnit()
281 dst_data[i + dst_k_offset + k * out_c] = MS_F32X8_GETI(m[k + m_k_offset], i); in OutputTransform6x2AvxUnit()
327 dst_data[i + dst_k_offset + k * out_c] = MS_F32X8_GETI(m[k + m_k_offset], i); in OutputTransform6x2ReluAvxUnit()
376 dst_data[i + dst_k_offset + k * out_c] = MS_F32X8_GETI(m[k + m_k_offset], i); in OutputTransform6x2Relu6AvxUnit()
416 dst_data[i + dst_k_offset + k * out_c] = MS_F32X8_GETI(m[k + m_k_offset], i); in OutputTransform6x3AvxUnit()
[all …]