fix(rfi): Optimize RFI processing slice logic to ensure coverage of all samples required for dedispersion

- Extend RFI processing slice range to avoid boundary issues
- Optimize comments in the downsampling function for better readability
- Add dtrend configuration option to allow users to choose whether to detrend the subband matrix
- Optimize resource cleanup logic to ensure proper release of origin_data

Author: lintian233

include/cpucal.hpp

@@ -284,25 +284,27 @@ Spectrum<T> dedispered_fil_with_dm(
         throw std::invalid_argument("Time window too short for this DM and band (no valid samples after dedispersion).");
     }
 
-    // ---- RFI：只在“局部输入切片”上运行（长度 = t_len_eff + delay_max_idx）----
-    // 这样解色散访问到的所有原始样本都被一致地标注/掩膜。
+    // ---- RFI：只在“局部输入切片”上运行 ----
+    // 扩展窗口以覆盖解色散所需的所有样本
     T* origin_data = static_cast<T*>(fil->data);
-    const size_t slice_len_for_rfi = t_len_eff + (size_t)delay_max_idx;
-    T* slice_ptr = origin_data + t_start_idx * fil->nchans;
+    size_t rfi_t_start_idx = (t_start_idx > (size_t)delay_max_idx) ? (t_start_idx - delay_max_idx) : 0;
+    size_t rfi_offset_from_t_start = t_start_idx - rfi_t_start_idx;
+    T* slice_ptr_for_rfi = origin_data + rfi_t_start_idx * fil->nchans;
+    size_t slice_len_for_rfi = std::min(t_len_eff + rfi_offset_from_t_start + 2 * delay_max_idx, (size_t)fil->ndata - rfi_t_start_idx);
 
     RfiMarkerCPU<T> rfi_marker(maskfile);
     if (rficfg.use_iqrm) {
         auto win_masks = iqrm_cuda::rfi_iqrm_gpu_host<T>(
-            slice_ptr,               // 指向局部起点
+            slice_ptr_for_rfi,       // 指向扩展切片的起点
             chan_start, chan_end_excl,
             slice_len_for_rfi,       // 覆盖局部 + 最大延时
             fil->nchans,
             fil->tsamp, rficfg);
-        rfi_marker.mask(slice_ptr, fil->nchans, slice_len_for_rfi, win_masks);
+        rfi_marker.mask(slice_ptr_for_rfi, fil->nchans, slice_len_for_rfi, win_masks);
     }
     if (rficfg.use_mask) {
-        // 静态掩膜同样只作用在切片上（不必动全局）
-        rfi_marker.mark_rfi(slice_ptr, fil->nchans, slice_len_for_rfi);
+        // 静态掩膜同样只作用在扩展切片上
+        rfi_marker.mark_rfi(slice_ptr_for_rfi, fil->nchans, slice_len_for_rfi);
     }
 
     // ---- 输出光谱 ----
@@ -326,10 +328,9 @@ Spectrum<T> dedispered_fil_with_dm(
 #pragma omp simd
         for (ptrdiff_t ch = (ptrdiff_t)chan_start; ch < (ptrdiff_t)chan_end_excl; ++ch) {
             const int d = dm_delays[ch - chan_start];
-            const size_t src_idx = (size_t)ti + (size_t)d;  // 相对于 slice_ptr 的偏移
-            // 由 t_len_eff 的定义，src_idx < slice_len_for_rfi 恒成立，无需额外边界判断
+            const size_t src_idx = t_start_idx + (size_t)ti + (size_t)d;  // 相对于 origin_data 的绝对偏移
             result.data[(size_t)ti * (size_t)result.nchans + (size_t)(ch - chan_start)]
-                = slice_ptr[src_idx * fil->nchans + (size_t)ch];
+                = origin_data[src_idx * fil->nchans + (size_t)ch];
         }
     }
 
@@ -420,20 +421,23 @@ Spectrum<T> dedisperse_spec_with_dm(
     }
 
     // ---- RFI（局部+最大延时）----
-    T* slice_ptr = spec + t_start_idx * header.nchans;
-    size_t slice_len_for_rfi = t_len_eff + (size_t)delay_max_idx;
+    size_t rfi_t_start_idx = (t_start_idx > (size_t)delay_max_idx) ? (t_start_idx - delay_max_idx) : 0;
+    size_t rfi_offset_from_t_start = t_start_idx - rfi_t_start_idx;
+    T* slice_ptr_for_rfi = spec + rfi_t_start_idx * header.nchans;
+    size_t slice_len_for_rfi = std::min(t_len_eff + rfi_offset_from_t_start + 2 * delay_max_idx, header.ndata - rfi_t_start_idx);
+
     RfiMarkerCPU<T> rfi_marker(maskfile);
     if (rficfg.use_iqrm) {
         auto win_masks = iqrm_cuda::rfi_iqrm_gpu_host<T>(
-            slice_ptr,
+            slice_ptr_for_rfi,
             chan_start, chan_end_excl,
             slice_len_for_rfi,
             header.nchans,
             header.tsamp, rficfg);
-        rfi_marker.mask(slice_ptr, header.nchans, slice_len_for_rfi, win_masks);
+        rfi_marker.mask(slice_ptr_for_rfi, header.nchans, slice_len_for_rfi, win_masks);
     }
     if (rficfg.use_mask) {
-        rfi_marker.mark_rfi(slice_ptr, header.nchans, slice_len_for_rfi);
+        rfi_marker.mark_rfi(slice_ptr_for_rfi, header.nchans, slice_len_for_rfi);
     }
 
     // ---- 构造输出 ----
@@ -457,9 +461,9 @@ Spectrum<T> dedisperse_spec_with_dm(
 #pragma omp simd
         for (ptrdiff_t ch = (ptrdiff_t)chan_start; ch < (ptrdiff_t)chan_end_excl; ++ch) {
             int d = dm_delays[ch - chan_start];
-            size_t src_idx = (size_t)ti + (size_t)d; // 相对于 slice_ptr
+            size_t src_idx = t_start_idx + (size_t)ti + (size_t)d; // 相对于 spec
             result.data[(size_t)ti * result.nchans + (size_t)(ch - chan_start)]
-                = slice_ptr[src_idx * header.nchans + (size_t)ch];
+                = spec[src_idx * header.nchans + (size_t)ch];
         }
     }
 

python/astroflow/plotter.py

@@ -788,9 +788,15 @@ def downsample_freq_weighted_vec(spec_data, freq_axis, n_out):
     """
     完全向量化的频率方向降采样。
     保证能量守恒 & 无跑频。
-    spec_data: [ntime, nfreq_in]
-    freq_axis: 频率中心(升序)
-    n_out:     目标子带数
+
+    参数
+    ------
+    spec_data : ndarray
+        [ntime, nfreq_in] 的动态谱
+    freq_axis : ndarray
+        频率中心(升序)。
+    n_out : int
+        目标子带数。
     """
     ntime, nfreq_in = spec_data.shape
 
@@ -848,8 +854,8 @@ def _setup_subband_spectrum_plots(fig, gs, spec_data, spec_time_axis, spec_freq_
     )
     # c_end_time = time.time()s
     # print(f"Subband processing time: {c_end_time - curr_time:.3f} seconds")
-
-    subband_matrix = _detrend(subband_matrix, axis=0, type='linear')
+    if specconfig.get("dtrend", False):
+        subband_matrix = _detrend(subband_matrix, axis=0, type='linear')
     # subband_matrix = _detrend_frequency(subband_matrix.T, poly_order=6).T
 
     if specconfig.get("norm", True):
@@ -973,6 +979,7 @@ def plot_candidate(
         ValueError: If file_path has unsupported extension
         Exception: If data loading or processing fails
     """
+    origin_data = None
     try:
         # Parse candidate information
         dm, toa, freq_start, freq_end, dmt_idx, ref_toa, bbox = _parse_candidate_info(candinfo)
@@ -1000,7 +1007,6 @@ def plot_candidate(
         ax_time, ax_main, ax_dm = _setup_dm_plots(fig, gs, dm_data, time_axis, dm_axis, dm_vmin, dm_vmax, dm, toa)
 
         # Load and process spectrum data
-        origin_data = None
         try:
             origin_data = _load_data_file(file_path)
             header = origin_data.header()
@@ -1051,12 +1057,13 @@ def plot_candidate(
             if snr < snrhold:
                 plt.close('all')
                 if origin_data is not None:
-                    if hasattr(origin_data, "close"):
+                    close_method = getattr(origin_data, "close", None)
+                    if callable(close_method):
                         try:
-                            origin_data.close()
+                            close_method()
                         except Exception:
                             pass
-                    del origin_data
+                    origin_data = None
 
                 del spectrum, spec_data, initial_spectrum, initial_spec_data
                 gc.collect()
@@ -1137,11 +1144,12 @@ def plot_candidate(
     finally:
         # Cleanup
         plt.close('all')
-        if 'origin_data' in locals() and origin_data is not None:
-            if hasattr(origin_data, "close"):
+        if origin_data is not None:
+            close_method = getattr(origin_data, "close", None)
+            if callable(close_method):
                 try:
-                    origin_data.close()
+                    close_method()
                 except Exception:
                     pass
-            del origin_data
+            origin_data = None
         gc.collect()