Searched refs:dst_4 (Results 1 – 6 of 6) sorted by relevance
| /third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/fp16_grad/ |
| D | arithmetic_self_grad.c | 29 float16x8_t dst_4 = vmulq_f16(src0_4, vrecpeq_f16(vmulq_f16(src1_4, log_10))); in Fp16LogGrad() local
30 vst1q_f16(dst + i, dst_4); in Fp16LogGrad()
|
| D | activation_grad.c | 135 float16x4_t dst_4 = vbsl_f16(mask_4, src0_4, vmul_f16(alpha_4, vmul_f16(expm1_4, src0_4))); in EluFp16Grad() local
136 vst1_f16(dst + i, dst_4); in EluFp16Grad()
|
| /third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/fp32_grad/ |
| D | activation_grad.c | 32 float32x4_t dst_4 = vbslq_f32(mask_4, src0_4, zero_4); in ReluGrad() local
33 vst1q_f32(dst + i, dst_4); in ReluGrad()
53 float32x4_t dst_4 = vbslq_f32(mask_4, src0_4, zero_4); in Relu6Grad() local
54 vst1q_f32(dst + i, dst_4); in Relu6Grad()
|
| /third_party/mindspore/mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/fp32/ |
| D | conv_common_fp32.c | 440 float *dst_4 = dst + out_step * C3NUM; in SWConv3x32Kernel() local
587 : "a"(act_flag), "r"(out_step), "r"(dst), "r"(write_mode), "r"(dst_4) in SWConv3x32Kernel()
599 float *dst_4 = dst + out_step * C3NUM; in SWConv1x32Kernel() local
683 : "a"(act_flag), "r"(out_step), "r"(dst), "r"(write_mode), "r"(dst_4) in SWConv1x32Kernel()
|
| D | conv_depthwise_fp32.c | 877 float32x4_t dst_4 = vld1q_f32(dst_kw); in DeconvDwBorderPixel() local
878 dst_4 = vfmaq_f32(dst_4, src_4, weight_4); in DeconvDwBorderPixel()
879 vst1q_f32(dst_kw, dst_4); in DeconvDwBorderPixel()
|
| /third_party/mesa3d/src/intel/compiler/ |
| D | brw_vec4_generator.cpp | 1134 struct brw_reg dst_4 = suboffset(vec1(dst), 4); in generate_unpack_flags() local
1137 brw_AND(p, dst_4, flags, brw_imm_ud(0xf0)); in generate_unpack_flags()
1138 brw_SHR(p, dst_4, dst_4, brw_imm_ud(4)); in generate_unpack_flags()
|