[matmul] fix single core design (#1721)

Co-authored-by: Joseph Melber <[email protected]>
Xilinx · Sep 10, 2024 · 89d1544 · 89d1544
1 parent 3903d97
commit 89d1544
Show file tree

Hide file tree

Showing 2 changed files with 14 additions and 0 deletions.
diff --git a/programming_examples/basic/matrix_multiplication/single_core/aie2.py b/programming_examples/basic/matrix_multiplication/single_core/aie2.py
@@ -297,6 +297,9 @@ def sequence(A, B, C):
                         )
                         bd_id_base = 8 * pingpong
                         num_tile_rows = min([rows_per_block // 2, M_div_m - row_base])
+                        if num_tile_rows <= 0:
+                            # At the very last iteration, we may not need a 'pong' iteration
+                            break
                         npu_dma_memcpy_nd(
                             metadata="outC",
                             bd_id=bd_id_base,

diff --git a/programming_examples/basic/matrix_multiplication/single_core/run_makefile_1.lit b/programming_examples/basic/matrix_multiplication/single_core/run_makefile_1.lit
@@ -0,0 +1,11 @@
+// (c) Copyright 2024 Advanced Micro Devices, Inc.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+// REQUIRES: ryzen_ai, chess
+//
+// RUN: mkdir -p %S/test_2
+// RUN: cd %S/test_2
+// RUN: make -f %S/Makefile clean
+// RUN: env M=768 K=512 N=512 m=64 k=64 n=64 dtype_in=i16 dtype_out=i16 make -f %S/Makefile 
+// RUN: %run_on_npu env M=768 K=512 N=512 m=64 k=64 n=64 dtype_in=i16 dtype_out=i16 make -f %S/Makefile run | FileCheck %s
+// CHECK: PASS!