AOT-biomaps 2.9.281__tar.gz → 2.9.312__tar.gz

{aot_biomaps-2.9.281 → aot_biomaps-2.9.312}/AOT_biomaps/AOT_Recon/AOT_Optimizers/LS.py

@@ -23,8 +23,7 @@ def LS(
     denominator_threshold=1e-6,
     max_saves=5000,
     show_logs=True,
-    smatrixType=SMatrixType.SELL,
-    Z=350,
+    smatrixType=SMatrixType.SELL
 ):
     """
     Least Squares reconstruction using Projected Gradient Descent (PGD) with non-negativity constraint.
@@ -44,7 +43,7 @@ def LS(
     # Dispatch to the appropriate implementation
     if use_gpu:
             if smatrixType == SMatrixType.CSR:
-                return _LS_CG_sparseCSR_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tumor_str, device, max_saves, denominator_threshold, Z, show_logs)
+                return _LS_CG_sparseCSR_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tumor_str, device, max_saves, denominator_threshold, show_logs)
             elif smatrixType == SMatrixType.SELL:
                 return _LS_CG_sparseSELL_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tumor_str, device, max_saves, denominator_threshold, show_logs)
             elif smatrixType == SMatrixType.DENSE:
@@ -181,13 +180,12 @@ def _LS_CG_sparseCSR_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tu
             print(f"Dim X: {X}, Dim Z: {Z}, TN: {TN}, ZX: {ZX}")
 
         stream = drv.Stream()
-        mod = drv.module_from_file('AOT_biomaps_kernels.cubin')
         
         # Récupération des Kernels
-        projection_kernel = mod.get_function('projection_kernel__CSR')
-        backprojection_kernel = mod.get_function('backprojection_kernel__CSR')
-        axpby_kernel = mod.get_function("vector_axpby_kernel")
-        minus_axpy_kernel = mod.get_function("vector_minus_axpy_kernel")
+        projection_kernel = SMatrix.sparse_mod.get_function('projection_kernel__CSR')
+        backprojection_kernel = SMatrix.sparse_mod.get_function('backprojection_kernel__CSR')
+        axpby_kernel = SMatrix.sparse_mod.get_function("vector_axpby_kernel")
+        minus_axpy_kernel = SMatrix.sparse_mod.get_function("vector_minus_axpy_kernel")
         
         # --- Allocation des buffers (Pointeurs Bruts) ---
         y = y.T.flatten().astype(dtype)
@@ -231,7 +229,7 @@ def _LS_CG_sparseCSR_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tu
         drv.memcpy_dtod(p_flat_gpu, r_flat_gpu, ZX * np.dtype(dtype).itemsize)
         
         # 6. rho_prev = ||r_0||^2
-        rho_prev = _dot_product_gpu(mod, r_flat_gpu, r_flat_gpu, ZX, stream)
+        rho_prev = _dot_product_gpu(SMatrix.sparse_mod, r_flat_gpu, r_flat_gpu, ZX, stream)
 
         # --- Boucle itérative ---
         saved_theta, saved_indices = [], []
@@ -258,7 +256,7 @@ def _LS_CG_sparseCSR_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tu
                                   block=(block_size, 1, 1), grid=((TN + block_size - 1) // block_size, 1, 1), stream=stream)
 
             # c. alpha = rho_prev / <p, z>
-            pAp = _dot_product_gpu(mod, p_flat_gpu, z_flat_gpu, ZX, stream)
+            pAp = _dot_product_gpu(SMatrix.sparse_mod, p_flat_gpu, z_flat_gpu, ZX, stream)
             
             if abs(pAp) < 1e-15: break
             alpha = rho_prev / pAp
@@ -273,7 +271,7 @@ def _LS_CG_sparseCSR_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tu
                               block=(block_size, 1, 1), grid=((ZX + block_size - 1) // block_size, 1, 1), stream=stream)
             
             # f. rho_curr = ||r||^2
-            rho_curr = _dot_product_gpu(mod, r_flat_gpu, r_flat_gpu, ZX, stream)
+            rho_curr = _dot_product_gpu(SMatrix.sparse_mod, r_flat_gpu, r_flat_gpu, ZX, stream)
             
             if rho_curr < tolerance: break
 
@@ -364,11 +362,10 @@ def _LS_CG_sparseSELL_pycuda(SMatrix, y, numIterations, isSavingEachIteration, t
         tolerance = 1e-12 
 
         # Accès aux paramètres SELL
-        mod = SMatrix.sparse_mod
-        projection_kernel = mod.get_function("projection_kernel__SELL")
-        backprojection_kernel = mod.get_function("backprojection_kernel__SELL")
-        axpby_kernel = mod.get_function("vector_axpby_kernel")
-        minus_axpy_kernel = mod.get_function("vector_minus_axpy_kernel")
+        projection_kernel = SMatrix.sparse_mod.get_function("projection_kernel__SELL")
+        backprojection_kernel = SMatrix.sparse_mod.get_function("backprojection_kernel__SELL")
+        axpby_kernel = SMatrix.sparse_mod.get_function("vector_axpby_kernel")
+        minus_axpy_kernel = SMatrix.sparse_mod.get_function("vector_minus_axpy_kernel")
         slice_height = np.int32(SMatrix.slice_height)
         grid_rows = ((TN + block_size - 1) // block_size, 1, 1)
 
@@ -416,7 +413,7 @@ def _LS_CG_sparseSELL_pycuda(SMatrix, y, numIterations, isSavingEachIteration, t
         drv.memcpy_dtod(p_flat_gpu, r_flat_gpu, ZX * np.dtype(dtype).itemsize)
         
         # 6. rho_prev = ||r_0||^2
-        rho_prev = _dot_product_gpu(mod, r_flat_gpu, r_flat_gpu, ZX, stream)
+        rho_prev = _dot_product_gpu(SMatrix.sparse_mod, r_flat_gpu, r_flat_gpu, ZX, stream)
 
         # --- Boucle itérative ---
         saved_theta, saved_indices = [], []
@@ -443,7 +440,7 @@ def _LS_CG_sparseSELL_pycuda(SMatrix, y, numIterations, isSavingEachIteration, t
                                   block=(block_size, 1, 1), grid=grid_rows, stream=stream)
 
             # c. alpha = rho_prev / <p, z>
-            pAp = _dot_product_gpu(mod, p_flat_gpu, z_flat_gpu, ZX, stream)
+            pAp = _dot_product_gpu(SMatrix.sparse_mod, p_flat_gpu, z_flat_gpu, ZX, stream)
             
             if abs(pAp) < 1e-15: break
             alpha = rho_prev / pAp
@@ -458,7 +455,7 @@ def _LS_CG_sparseSELL_pycuda(SMatrix, y, numIterations, isSavingEachIteration, t
                               block=(block_size, 1, 1), grid=((ZX + block_size - 1) // block_size, 1, 1), stream=stream)
             
             # f. rho_curr = ||r||^2
-            rho_curr = _dot_product_gpu(mod, r_flat_gpu, r_flat_gpu, ZX, stream)
+            rho_curr = _dot_product_gpu(SMatrix.sparse_mod, r_flat_gpu, r_flat_gpu, ZX, stream)
             
             if rho_curr < tolerance: break
 

{aot_biomaps-2.9.281 → aot_biomaps-2.9.312}/AOT_biomaps/AOT_Recon/AOT_Optimizers/MLEM.py

@@ -26,7 +26,6 @@ def MLEM(
     max_saves=5000,
     show_logs=True,
     smatrixType=SMatrixType.SELL,
-    Z=350,
 ):
     """
     Unified MLEM algorithm for Acousto-Optic Tomography.
@@ -59,11 +58,11 @@ def MLEM(
     # Dispatch to the appropriate implementation
     if use_gpu:
             if smatrixType == SMatrixType.CSR:
-                return MLEM_sparseCSR_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tumor_str, device, max_saves, denominator_threshold, Z, show_logs)
+                return MLEM_sparseCSR_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tumor_str, max_saves, denominator_threshold, show_logs)
             elif smatrixType == SMatrixType.SELL:
-                return MLEM_sparseSELL_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tumor_str, device, max_saves, denominator_threshold, show_logs)
+                return MLEM_sparseSELL_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tumor_str, max_saves, denominator_threshold, show_logs)
             elif smatrixType == SMatrixType.DENSE:
-                return _MLEM_single_GPU(SMatrix, y, numIterations, isSavingEachIteration, tumor_str, device, max_saves, denominator_threshold,show_logs)
+                return _MLEM_single_GPU(SMatrix, y, numIterations, isSavingEachIteration, tumor_str, max_saves, denominator_threshold,show_logs)
             else:
                 raise ValueError("Unsupported SMatrixType for GPU MLEM.")
     else:
@@ -229,49 +228,60 @@ def _MLEM_CPU_opti(SMatrix, y, numIterations, isSavingEachIteration, tumor_str,
         print(f"Error in optimized CPU MLEM: {type(e).__name__}: {e}")
         return None, None
     
-def MLEM_sparseCSR_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tumor_str, device, max_saves, denominator_threshold, show_logs=True):
+def MLEM_sparseCSR_pycuda(
+    SMatrix,
+    y,
+    numIterations,
+    isSavingEachIteration,
+    tumor_str,
+    max_saves,
+    denominator_threshold,
+    show_logs=True,
+):
     """
-    SMatrix: instance of SparseMatrixGPU (already allocated)
-    y: measured data (1D np.float32 of length TN)
-
-    Assumptions:
-    - SMatrix.values_gpu and SMatrix.col_ind_gpu and SMatrix.row_ptr_gpu are device pointers
-    - SMatrix.norm_factor_inv_gpu exists
-    - SMatrix.ctx is the PyCUDA context for the target GPU.
+    Robust MLEM implementation for CSR SMatrix using PyCUDA kernels.
+    Expects SMatrix to be SparseSMatrix_CSR with attributes:
+      - values_gpu, col_ind_gpu, row_ptr_gpu (device pointers)
+      - norm_factor_inv_gpu (device pointer)
+      - sparse_mod (loaded module with kernels)
+      - ctx (PyCUDA context)
+    Returns (saved_theta_list, saved_indices) if isSavingEachIteration else (final_theta, None)
     """
-    
-    # We use a final_result placeholder to ensure it's defined outside the try block
     final_result = None
-    
+
+    # Local holders to free in finally
+    y_gpu = q_flat_gpu = e_flat_gpu = c_flat_gpu = theta_flat_gpu = None
+
     try:
         if not isinstance(SMatrix, SparseSMatrix_CSR):
             raise TypeError("SMatrix must be a SparseSMatrix_CSR object")
 
-        # --- CONTEXT FIX: Push the context associated with SMatrix ---
-        # This ensures all subsequent PyCUDA operations use the correct GPU/context.
-        if SMatrix.ctx:
+        # push context (if provided)
+        popped_ctx = False
+        if getattr(SMatrix, "ctx", None):
             SMatrix.ctx.push()
-        # -----------------------------------------------------------
+            popped_ctx = True
 
         dtype = np.float32
-        TN = SMatrix.N * SMatrix.T
-        ZX = SMatrix.Z * SMatrix.X
-        # Ensure Z and X are correctly defined for reshaping
-        Z = SMatrix.Z 
-        X = SMatrix.X
-
-        if show_logs:
-            # We assume SMatrix was initialized using the correct device index.
-            print(f"Executing on GPU device index: {SMatrix.device.primary_context.device.name()}")
-            print(f"Dim X: {X}, Dim Z: {Z}, TN: {TN}, ZX: {ZX}")
-
-        # streams
+        TN = int(SMatrix.N * SMatrix.T)
+        ZX = int(SMatrix.Z * SMatrix.X)
+        Z = int(SMatrix.Z)
+        X = int(SMatrix.X)
+
+        # Make sure required GPU pointers exist
+        if getattr(SMatrix, "values_gpu", None) is None or getattr(SMatrix, "col_ind_gpu", None) is None or getattr(SMatrix, "row_ptr_gpu", None) is None:
+            raise RuntimeError("SMatrix is missing GPU buffers (values_gpu / col_ind_gpu / row_ptr_gpu)")
+
+        if getattr(SMatrix, "norm_factor_inv_gpu", None) is None:
+            raise RuntimeError("SMatrix.norm_factor_inv_gpu not available on GPU")
+
+        # stream for async operations
         stream = drv.Stream()
 
-        # allocate device buffers
-        y = y.T.flatten().astype(np.float32)
-        y_gpu = drv.mem_alloc(y.nbytes)
-        drv.memcpy_htod_async(y_gpu, y.astype(dtype), stream)
+        # prepare device buffers
+        y_arr = np.ascontiguousarray(y.T.flatten().astype(np.float32))
+        y_gpu = drv.mem_alloc(y_arr.nbytes)
+        drv.memcpy_htod_async(y_gpu, y_arr, stream)
 
         theta_flat_gpu = drv.mem_alloc(ZX * np.dtype(dtype).itemsize)
         initial_theta = np.full(ZX, 0.1, dtype=dtype)
@@ -283,61 +293,111 @@ def MLEM_sparseCSR_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tumo
         e_flat_gpu = drv.mem_alloc(TN * np.dtype(dtype).itemsize)
         c_flat_gpu = drv.mem_alloc(ZX * np.dtype(dtype).itemsize)
 
-        # Assuming the cubin file is found globally or managed by the caller
-        projection_kernel = SMatrix.sparse_mod.get_function('projection_kernel__CSR')
-        backprojection_kernel = SMatrix.sparse_mod.get_function('backprojection_kernel__CSR')
-        ratio_kernel = SMatrix.sparse_mod.get_function('ratio_kernel')
-        update_kernel = SMatrix.sparse_mod.get_function('update_theta_kernel')
+        # Ensure kernels exist
+        projection_kernel = SMatrix.sparse_mod.get_function("projection_kernel__CSR")
+        backprojection_kernel = SMatrix.sparse_mod.get_function("backprojection_kernel__CSR")
+        ratio_kernel = SMatrix.sparse_mod.get_function("ratio_kernel")
+        update_kernel = SMatrix.sparse_mod.get_function("update_theta_kernel")
         block_size = 256
 
-        saved_theta, saved_indices = [], []
+        # prepare save indices once
         if numIterations <= max_saves:
             save_indices = list(range(numIterations))
         else:
-            save_indices = list(range(0, numIterations, max(1, numIterations // max_saves)))
+            step = max(1, numIterations // max_saves)
+            save_indices = list(range(0, numIterations, step))
             if save_indices[-1] != numIterations - 1:
                 save_indices.append(numIterations - 1)
 
+        saved_theta = []
+        saved_indices = []
+
         description = f"AOT-BioMaps -- ML-EM (CSR-sparse SMatrix) ---- {tumor_str} TUMOR ---- GPU {torch.cuda.current_device()}"
         iterator = trange(numIterations, desc=description) if show_logs else range(numIterations)
+
+        # grid sizes
+        grid_rows = ((TN + block_size - 1) // block_size, 1, 1)
+        grid_cols = ((ZX + block_size - 1) // block_size, 1, 1)
+
         for it in iterator:
             # projection: q = A * theta
-            projection_kernel(q_flat_gpu, SMatrix.values_gpu, SMatrix.row_ptr_gpu, SMatrix.col_ind_gpu,
-                              theta_flat_gpu, np.int32(TN),
-                              block=(block_size, 1, 1), grid=((TN + block_size - 1) // block_size, 1, 1),
-                              stream=stream)
+            projection_kernel(
+                q_flat_gpu,
+                SMatrix.values_gpu,
+                SMatrix.row_ptr_gpu,
+                SMatrix.col_ind_gpu,
+                theta_flat_gpu,
+                np.int32(TN),
+                block=(block_size, 1, 1),
+                grid=grid_rows,
+                stream=stream,
+            )
 
             # ratio: e = y / max(q, threshold)
-            ratio_kernel(e_flat_gpu, y_gpu, q_flat_gpu, np.float32(denominator_threshold), np.int32(TN),
-                              block=(block_size, 1, 1), grid=((TN + block_size - 1) // block_size, 1, 1), stream=stream)
-
-            # backprojection: c = A^T * e
+            ratio_kernel(
+                e_flat_gpu,
+                y_gpu,
+                q_flat_gpu,
+                np.float32(denominator_threshold),
+                np.int32(TN),
+                block=(block_size, 1, 1),
+                grid=grid_rows,
+                stream=stream,
+            )
+
+            # backprojection: c = A^T * e  (zero c first)
             drv.memset_d32_async(c_flat_gpu, 0, ZX, stream)
-            backprojection_kernel(c_flat_gpu, SMatrix.values_gpu, SMatrix.row_ptr_gpu, SMatrix.col_ind_gpu,
-                                  e_flat_gpu, np.int32(TN),
-                                  block=(block_size, 1, 1), grid=((TN + block_size - 1) // block_size, 1, 1), stream=stream)
+            backprojection_kernel(
+                c_flat_gpu,
+                SMatrix.values_gpu,
+                SMatrix.row_ptr_gpu,
+                SMatrix.col_ind_gpu,
+                e_flat_gpu,
+                np.int32(TN),
+                block=(block_size, 1, 1),
+                grid=grid_rows,
+                stream=stream,
+            )
 
             # update: theta *= norm_factor_inv * c
-            update_kernel(theta_flat_gpu, c_flat_gpu, norm_factor_inv_gpu, np.int32(ZX),
-                          block=(block_size, 1, 1), grid=((ZX + block_size - 1) // block_size, 1, 1), stream=stream)
-
+            update_kernel(
+                theta_flat_gpu,
+                c_flat_gpu,
+                norm_factor_inv_gpu,
+                np.int32(ZX),
+                block=(block_size, 1, 1),
+                grid=grid_cols,
+                stream=stream,
+            )
+
+            # periodic synchronization for stability / logging
             if show_logs and (it % 10 == 0 or it == numIterations - 1):
-                drv.Context.synchronize()
+                stream.synchronize()
 
+            # save snapshot if required
             if isSavingEachIteration and it in save_indices:
+                # ensure kernels finished
+                stream.synchronize()
                 theta_host = np.empty(ZX, dtype=dtype)
                 drv.memcpy_dtoh(theta_host, theta_flat_gpu)
                 saved_theta.append(theta_host.reshape(Z, X))
-                saved_indices.append(it)
-
-        drv.Context.synchronize()
-
-        final_result = np.empty(ZX, dtype=dtype)
-        drv.memcpy_dtoh(final_result, theta_flat_gpu)
-        final_result = final_result.reshape(Z, X)
-
-        # free local allocations
-        y_gpu.free(); q_flat_gpu.free(); e_flat_gpu.free(); c_flat_gpu.free(); theta_flat_gpu.free()
+                saved_indices.append(int(it))
+
+        # make sure everything finished
+        stream.synchronize()
+        final_theta_host = np.empty(ZX, dtype=dtype)
+        drv.memcpy_dtoh(final_theta_host, theta_flat_gpu)
+        final_result = final_theta_host.reshape(Z, X)
+
+        # free local allocations (will also be freed in finally if exception)
+        try:
+            y_gpu.free()
+            q_flat_gpu.free()
+            e_flat_gpu.free()
+            c_flat_gpu.free()
+            theta_flat_gpu.free()
+        except Exception:
+            pass
 
         return (saved_theta, saved_indices) if isSavingEachIteration else (final_result, None)
 
@@ -345,32 +405,50 @@ def MLEM_sparseCSR_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tumo
         print(f"Error in MLEM_sparseCSR_pycuda: {type(e).__name__}: {e}")
         gc.collect()
         return None, None
-    
-    finally:
-        # --- CONTEXT FIX: Pop the context ---
-        if SMatrix and hasattr(SMatrix, 'ctx') and SMatrix.ctx:
-            SMatrix.ctx.pop()
-        # ------------------------------------
 
-def MLEM_sparseSELL_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tumor_str, device, max_saves, denominator_threshold, show_logs=True):
+    finally:
+        # free buffers if still allocated
+        for buf in ("y_gpu", "q_flat_gpu", "e_flat_gpu", "c_flat_gpu", "theta_flat_gpu"):
+            try:
+                val = locals().get(buf, None)
+                if val is not None:
+                    val.free()
+            except Exception:
+                pass
+        # pop context safely
+        try:
+            if SMatrix and hasattr(SMatrix, "ctx") and SMatrix.ctx and popped_ctx:
+                SMatrix.ctx.pop()
+        except Exception:
+            pass
+
+def MLEM_sparseSELL_pycuda(
+    SMatrix,
+    y,
+    numIterations,
+    isSavingEachIteration,
+    tumor_str,
+    max_saves,
+    denominator_threshold,
+    show_logs=True,
+):
     """
     MLEM using SELL-C-σ kernels already present on device.
     y must be float32 length TN.
+
+    Version propre : diagnostics retirés.
     """
     final_result = None
 
     try:
-        # check if SMatrix is SparseSMatrix_SELL object
         if not isinstance(SMatrix, SparseSMatrix_SELL):
             raise TypeError("SMatrix must be a SparseSMatrix_SELL object")
         if SMatrix.sell_values_gpu is None:
             raise RuntimeError("SELL not built. Call allocate_sell_c_sigma_direct() first.")
-            
-        # --- CONTEXT FIX: Push the context associated with SMatrix ---
-        # This ensures all subsequent PyCUDA operations use the correct GPU/context.
+
+        # Context
         if SMatrix.ctx:
             SMatrix.ctx.push()
-        # -----------------------------------------------------------
 
         TN = int(SMatrix.N * SMatrix.T)
         ZX = int(SMatrix.Z * SMatrix.X)
@@ -384,7 +462,7 @@ def MLEM_sparseSELL_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tum
 
         stream = drv.Stream()
 
-        # device buffers
+        # Device buffers
         y = y.T.flatten().astype(np.float32)
         y_gpu = drv.mem_alloc(y.nbytes)
         drv.memcpy_htod_async(y_gpu, y.astype(dtype), stream)
@@ -403,6 +481,7 @@ def MLEM_sparseSELL_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tum
         grid_rows = ((TN + block_size - 1) // block_size, 1, 1)
         grid_cols = ((ZX + block_size - 1) // block_size, 1, 1)
 
+        # Prepare save indices
         saved_theta, saved_indices = [], []
         if numIterations <= max_saves:
             save_indices = list(range(numIterations))
@@ -413,52 +492,59 @@ def MLEM_sparseSELL_pycuda(SMatrix, y, numIterations, isSavingEachIteration, tum
 
         description = f"AOT-BioMaps -- ML-EM (SELL-c-σ-sparse SMatrix) ---- {tumor_str} TUMOR ---- GPU {torch.cuda.current_device()}"
         iterator = trange(numIterations, desc=description) if show_logs else range(numIterations)
+
+        # --- MLEM Loop ---
         for it in iterator:
-            # projection
-            proj(q_gpu, SMatrix.sell_values_gpu, SMatrix.sell_colinds_gpu, slice_ptr_gpu, slice_len_gpu,
-                    theta_gpu, np.int32(TN), slice_height,
-                    block=(block_size,1,1), grid=grid_rows, stream=stream)
 
-            # ratio
+            proj(q_gpu, SMatrix.sell_values_gpu, SMatrix.sell_colinds_gpu,
+                 slice_ptr_gpu, slice_len_gpu,
+                 theta_gpu, np.int32(TN), slice_height,
+                 block=(block_size,1,1), grid=grid_rows, stream=stream)
+
             ratio(e_gpu, y_gpu, q_gpu, np.float32(denominator_threshold), np.int32(TN),
-                    block=(block_size,1,1), grid=grid_rows, stream=stream)
+                  block=(block_size,1,1), grid=grid_rows, stream=stream)
 
-            # zero c
             drv.memset_d32_async(c_gpu, 0, ZX, stream)
 
-            # backprojection accumulate
-            backproj(SMatrix.sell_values_gpu, SMatrix.sell_colinds_gpu, slice_ptr_gpu, slice_len_gpu,
-                        e_gpu, c_gpu, np.int32(TN), slice_height,
-                        block=(block_size,1,1), grid=grid_rows, stream=stream)
+            backproj(SMatrix.sell_values_gpu, SMatrix.sell_colinds_gpu,
+                     slice_ptr_gpu, slice_len_gpu,
+                     e_gpu, c_gpu, np.int32(TN), slice_height,
+                     block=(block_size,1,1), grid=grid_rows, stream=stream)
 
-            # update
             update(theta_gpu, c_gpu, SMatrix.norm_factor_inv_gpu, np.int32(ZX),
-                    block=(block_size,1,1), grid=grid_cols, stream=stream)
+                   block=(block_size,1,1), grid=grid_cols, stream=stream)
 
-            stream.synchronize()
             if isSavingEachIteration and it in save_indices:
                 out = np.empty(ZX, dtype=np.float32)
                 drv.memcpy_dtoh(out, theta_gpu)
                 saved_theta.append(out.reshape((SMatrix.Z, SMatrix.X)))
                 saved_indices.append(it)
 
-        # final copy
+        stream.synchronize()
         res = np.empty(ZX, dtype=np.float32)
         drv.memcpy_dtoh(res, theta_gpu)
 
-        # free temporaries
-        y_gpu.free(); q_gpu.free(); e_gpu.free(); c_gpu.free(); theta_gpu.free()
-        
+        # free
+        try:
+            y_gpu.free()
+            q_gpu.free()
+            e_gpu.free()
+            c_gpu.free()
+            theta_gpu.free()
+        except Exception:
+            pass
+
         final_result = res.reshape((SMatrix.Z, SMatrix.X))
         return (saved_theta, saved_indices) if isSavingEachIteration else (final_result, None)
-        
+
     except Exception as e:
         print(f"Error in MLEM_sparseSELL_pycuda: {type(e).__name__}: {e}")
         gc.collect()
         return None, None
-    
+
     finally:
-        # --- CONTEXT FIX: Pop the context ---
         if SMatrix and hasattr(SMatrix, 'ctx') and SMatrix.ctx:
-            SMatrix.ctx.pop()
-        # ------------------------------------
+            try:
+                SMatrix.ctx.pop()
+            except Exception:
+                pass