Bugfixes

SeisSol · Oct 27, 2024 · 6193c09 · 6193c09
1 parent aa95d58
commit 6193c09
Show file tree

Hide file tree

Showing 2 changed files with 6 additions and 4 deletions.
diff --git a/pspamm/codegen/architectures/arm_sve/generator.py b/pspamm/codegen/architectures/arm_sve/generator.py
@@ -273,12 +273,15 @@ def move_register_block(self,
                     addr, comment = cursor.look(cursor_ptr, block_offset, cell_offset)
                     addr.disp += self.precision.size() * load_offset
 
+                    offset = addr.disp - prev_disp
+
                     # count how many elements we have processed between last step and this step
-                    cont_counter = ((addr.disp - prev_disp) // mul_vl)
+                    cont_counter = (offset // mul_vl)
                     larger_max_offset = cont_counter > max_mem_ins_mult
+                    non_dividing_offset = offset % mul_vl != 0
 
-                    if larger_max_offset or (prev_overhead and addr.disp > 0):
-                        offset_comment = "disp > {}".format(max_offset) if larger_max_offset else "previous mem. instr. used p0"
+                    if larger_max_offset or (prev_overhead and addr.disp > 0) or non_dividing_offset:
+                        offset_comment = f"disp > {max_offset}" if larger_max_offset else ("disp % VL != 0" if non_dividing_offset else "previous mem. instr. used p0")
                         asm.add(add(addr.disp, additional_regs[0], offset_comment, addr.base))
                         prev_disp = addr.disp
                         addr.base = additional_regs[0]

diff --git a/pspamm/matmul.py b/pspamm/matmul.py
@@ -325,7 +325,6 @@ def kernelK(asm, Bki, A_ptr, B_ptr):
                                 store_block.add(fma(regs[ir, x + ic], self.alpha_reg[1], A_regs_cut[ir, ic], "C = C + alpha * AB", None, pred=pred_m))
                     store_block.add(self.generator.move_register_block(self.C, C_ptr, Coords(), A_regs_cut, self.v_size, self.additional_regs, None, True, self.prefetching, self.ldc * x))
                 asm.add(store_block)
-
             else:
                 asm.add(self.generator.move_register_block(self.C, C_ptr, Coords(), regs, self.v_size, self.additional_regs, None, True, self.prefetching))