Clang-format; make the test smaller to avoid timeout

test/xrt/16_matmul_8x16_core_transform_bf16/gen.py

@@ -22,14 +22,13 @@
 #map3 = affine_map<()[s0] -> (s0 * 8)>
 #map4 = affine_map<(d0) -> (d0 * 4)>
 module {
-  func.func @forward(%arg0: memref<512x1024xbf16>, %arg1: memref<128x8x8x4x16xbf16>, %arg2: memref<512x512xf32>) -> memref<512x512xf32> {
+  func.func @forward(%arg0: memref<512x256xbf16>, %arg1: memref<32x8x8x4x16xbf16>, %arg2: memref<512x512xf32>) -> memref<512x512xf32> {
     %c8 = arith.constant 8 : index
     %c16 = arith.constant 16 : index
     %c64 = arith.constant 64 : index
     %c32 = arith.constant 32 : index
     %c128 = arith.constant 128 : index
     %c256 = arith.constant 256 : index
-    %c1024 = arith.constant 1024 : index
     %c0 = arith.constant 0 : index
     %c1 = arith.constant 1 : index
     %cst = arith.constant 0.000000e+00 : f32
@@ -40,16 +39,16 @@
       %0 = affine.apply #map(%arg3)
       %1 = affine.apply #map(%arg4)
       %subview = memref.subview %arg2[%0, %1] [128, 128] [1, 1] : memref<512x512xf32> to memref<128x128xf32, strided<[512, 1], offset: ?>>
-      %alloc_0 = memref.alloc() : memref<128x1024xbf16, 1>
-      scf.for %arg5 = %c0 to %c1024 step %c256 {
-        %subview_3 = memref.subview %arg0[%0, %arg5] [128, 256] [1, 1] : memref<512x1024xbf16> to memref<128x256xbf16, strided<[1024, 1], offset: ?>>
-        %subview_4 = memref.subview %alloc_0[0, %arg5] [128, 256] [1, 1] : memref<128x1024xbf16, 1> to memref<128x256xbf16, strided<[1024, 1], offset: ?>, 1>
-        linalg.copy ins(%subview_3 : memref<128x256xbf16, strided<[1024, 1], offset: ?>>) outs(%subview_4 : memref<128x256xbf16, strided<[1024, 1], offset: ?>, 1>)
+      %alloc_0 = memref.alloc() : memref<128x256xbf16, 1>
+      scf.for %arg5 = %c0 to %c256 step %c256 {
+        %subview_3 = memref.subview %arg0[%0, %arg5] [128, 256] [1, 1] : memref<512x256xbf16> to memref<128x256xbf16, strided<[256, 1], offset: ?>>
+        %subview_4 = memref.subview %alloc_0[0, %arg5] [128, 256] [1, 1] : memref<128x256xbf16, 1> to memref<128x256xbf16, strided<[256, 1], offset: ?>, 1>
+        linalg.copy ins(%subview_3 : memref<128x256xbf16, strided<[256, 1], offset: ?>>) outs(%subview_4 : memref<128x256xbf16, strided<[256, 1], offset: ?>, 1>)
       }
-      %alloc_1 = memref.alloc() : memref<128x8x8x16xbf16, 1>
-      scf.for %arg5 = %c0 to %c128 step %c32 {
-        %subview_3 = memref.subview %arg1[%arg5, 0, 0, %arg4, 0] [32, 8, 8, 1, 16] [1, 1, 1, 1, 1] : memref<128x8x8x4x16xbf16> to memref<32x8x8x16xbf16, strided<[4096, 512, 64, 1], offset: ?>>
-        %subview_4 = memref.subview %alloc_1[%arg5, 0, 0, 0] [32, 8, 8, 16] [1, 1, 1, 1] : memref<128x8x8x16xbf16, 1> to memref<32x8x8x16xbf16, strided<[1024, 128, 16, 1], offset: ?>, 1>
+      %alloc_1 = memref.alloc() : memref<32x8x8x16xbf16, 1>
+      scf.for %arg5 = %c0 to %c32 step %c32 {
+        %subview_3 = memref.subview %arg1[%arg5, 0, 0, %arg4, 0] [32, 8, 8, 1, 16] [1, 1, 1, 1, 1] : memref<32x8x8x4x16xbf16> to memref<32x8x8x16xbf16, strided<[4096, 512, 64, 1], offset: ?>>
+        %subview_4 = memref.subview %alloc_1[%arg5, 0, 0, 0] [32, 8, 8, 16] [1, 1, 1, 1] : memref<32x8x8x16xbf16, 1> to memref<32x8x8x16xbf16, strided<[1024, 128, 16, 1], offset: ?>, 1>
         linalg.copy ins(%subview_3 : memref<32x8x8x16xbf16, strided<[4096, 512, 64, 1], offset: ?>>) outs(%subview_4 : memref<32x8x8x16xbf16, strided<[1024, 128, 16, 1], offset: ?>, 1>)
       }
       %alloc_2 = memref.alloc() : memref<128x128xf32, 1>
@@ -60,14 +59,14 @@
         %subview_3 = memref.subview %alloc_2[%2, %3] [64, 64] [1, 1] : memref<128x128xf32, 1> to memref<64x64xf32, strided<[128, 1], offset: ?>, 1>
         %alloc_4 = memref.alloc() : memref<64x64xf32, 2>
         linalg.fill ins(%cst : f32) outs(%alloc_4 : memref<64x64xf32, 2>)
-        scf.for %arg7 = %c0 to %c128 step %c8 {
+        scf.for %arg7 = %c0 to %c32 step %c8 {
           %4 = affine.apply #map3()[%arg7]
-          %subview_5 = memref.subview %alloc_0[%2, %4] [64, 64] [1, 1] : memref<128x1024xbf16, 1> to memref<64x64xbf16, strided<[1024, 1], offset: ?>, 1>
-          %subview_6 = memref.subview %alloc_1[%arg7, %map, 0, 0] [8, 4, 8, 16] [1, 1, 1, 1] : memref<128x8x8x16xbf16, 1> to memref<8x4x8x16xbf16, strided<[1024, 128, 16, 1], offset: ?>, 1>
+          %subview_5 = memref.subview %alloc_0[%2, %4] [64, 64] [1, 1] : memref<128x256xbf16, 1> to memref<64x64xbf16, strided<[256, 1], offset: ?>, 1>
+          %subview_6 = memref.subview %alloc_1[%arg7, %map, 0, 0] [8, 4, 8, 16] [1, 1, 1, 1] : memref<32x8x8x16xbf16, 1> to memref<8x4x8x16xbf16, strided<[1024, 128, 16, 1], offset: ?>, 1>
           %transpose = memref.transpose %subview_6 (d0, d1, d2, d3) -> (d0, d2, d1, d3) : memref<8x4x8x16xbf16, strided<[1024, 128, 16, 1], offset: ?>, 1> to memref<8x8x4x16xbf16, strided<[1024, 16, 128, 1], offset: ?>, 1>
           %alloc_7 = memref.alloc() : memref<64x64xbf16, 2>
           %alloc_8 = memref.alloc() : memref<8x8x4x16xbf16, 2>
-          memref.copy %subview_5, %alloc_7 : memref<64x64xbf16, strided<[1024, 1], offset: ?>, 1> to memref<64x64xbf16, 2>
+          memref.copy %subview_5, %alloc_7 : memref<64x64xbf16, strided<[256, 1], offset: ?>, 1> to memref<64x64xbf16, 2>
           memref.copy %transpose, %alloc_8 : memref<8x8x4x16xbf16, strided<[1024, 16, 128, 1], offset: ?>, 1> to memref<8x8x4x16xbf16, 2>
           %reshape = memref.reshape %alloc_8(%alloc) : (memref<8x8x4x16xbf16, 2>, memref<2xindex>) -> memref<64x64xbf16, 2>
           linalg.matmul {cast = #linalg.type_fn<cast_signed>} ins(%alloc_7, %reshape : memref<64x64xbf16, 2>, memref<64x64xbf16, 2>) outs(%alloc_4 : memref<64x64xf32, 2>)
@@ -78,8 +77,8 @@
         memref.dealloc %alloc_4 : memref<64x64xf32, 2>
       } {mapping = [#gpu.thread<y>, #gpu.thread<x>]}
       linalg.copy ins(%alloc_2 : memref<128x128xf32, 1>) outs(%subview : memref<128x128xf32, strided<[512, 1], offset: ?>>)
-      memref.dealloc %alloc_0 : memref<128x1024xbf16, 1>
-      memref.dealloc %alloc_1 : memref<128x8x8x16xbf16, 1>
+      memref.dealloc %alloc_0 : memref<128x256xbf16, 1>
+      memref.dealloc %alloc_1 : memref<32x8x8x16xbf16, 1>
       memref.dealloc %alloc_2 : memref<128x128xf32, 1>
     }
     return %arg2 : memref<512x512xf32>

test/xrt/16_matmul_8x16_core_transform_bf16/kernel.cpp

@@ -16,7 +16,7 @@
 
 #include <aie_api/aie.hpp>
 
-template <typename T, int M, int N> 
+template <typename T, int M, int N>
 void zero_scalar(T *__restrict c) {
   for (int i = 0; i < M * N; i++) {
     c[i] = 0.0f;

test/xrt/16_matmul_8x16_core_transform_bf16/run.py

@@ -11,7 +11,7 @@
 
 M = 512
 N = 512
-K = 1024
+K = 256
 
 Tx = 16
 Ty = 8