Buddy Gemmini: IR dumps for matmul + batch_matmul + conv

experiments/gemmini/SUPPORT.md

@@ -0,0 +1,11 @@
+# Buddy Gemmini lowering coverage (Sparsh)
+
+This table is a quick view of what we’ve stress-tested and what Buddy lowers into.
+
+| Test | Input dialect/op | Layout | Proof of Gemmini match | Proof of Gemmini command expansion | Notes |
+|---|---|---|---|---|---|
+| matmul | linalg.matmul | (varies) | `gemmini.tile_matmul` | `gemmini.intr.loop_ws_config*` + `gemmini.intr.loop_ws` | matmul lowered end-to-end |
+| batch_matmul | linalg.batch_matmul | (varies) | `gemmini.tile_*` | `gemmini.intr.*` | batched path works |
+| conv (NHWC/HWCF) | linalg.conv_2d_nhwc_hwcf | NHWC x HWCF | `gemmini.tile_conv` | `gemmini.intr.loop_conv_ws_config*` + `gemmini.intr.loop_conv_ws` | conv lowered to WS loop |
+| conv (NCHW/FCHW) | linalg.conv_2d_nchw_fchw | NCHW x FCHW | `gemmini.tile_conv` | `gemmini.intr.loop_conv_ws_config*` + `gemmini.intr.loop_conv_ws` | alternate layout works |
+| mini CNN block | 2x conv + copy | NCHW/FCHW | 2x `gemmini.tile_conv` | `gemmini.intr.loop_conv_ws*` appears | multi-layer block lowers |

experiments/gemmini/inputs/batch_matmul.mlir

@@ -0,0 +1,30 @@
+// RUN: buddy-opt %s \
+// RUN:     --convert-linalg-to-gemmini | \
+// RUN: FileCheck %s
+
+func.func @main() -> i8 {
+  %0 = arith.constant 0 : i8 
+  %1 = arith.constant 1 : i8
+  %2 = arith.constant 2 : i8 
+  %input0 = memref.alloc() : memref<3x3x3xi8> 
+  %input1 = memref.alloc() : memref<3x3x3xi8> 
+  %output = memref.alloc() : memref<3x3x3xi8>  
+  linalg.fill
+    ins(%1 : i8)
+  outs(%input0 : memref<3x3x3xi8>)
+  linalg.fill
+    ins(%2 : i8)
+  outs(%input1 : memref<3x3x3xi8>)
+  // CHECK: gemmini.tile_matmul %subview %subview_2 %subview_3 %alloc_4 : 
+  // CHECK-SAME: memref<3x3xi8, strided<[3, 1]>> memref<3x3xi8, strided<[3, 1]>> memref<3x3xi8, strided<[3, 1]>> memref<3x3xi32>
+  // CHECK: gemmini.tile_matmul %subview_5 %subview_6 %subview_7 %alloc_8 : 
+  // CHECK-SAME: memref<3x3xi8, strided<[3, 1], offset: 9>> memref<3x3xi8, strided<[3, 1], offset: 9>> memref<3x3xi8, strided<[3, 1], offset: 9>> memref<3x3xi32>
+  // CHECK: gemmini.tile_matmul %subview_10 %subview_11 %subview_12 %alloc_13 : 
+  // CHECK-SAME: memref<3x3xi8, strided<[3, 1], offset: 18>> memref<3x3xi8, strided<[3, 1], offset: 18>> memref<3x3xi8, strided<[3, 1], offset: 18>> memref<3x3xi32>
+  linalg.batch_matmul
+    ins(%input0, %input1: memref<3x3x3xi8>, memref<3x3x3xi8>)
+  outs(%output : memref<3x3x3xi8>)
+  gemmini.print %output : memref<3x3x3xi8>
+  memref.dealloc %output : memref<3x3x3xi8> 
+  return %0 : i8
+}

experiments/gemmini/inputs/conv_2d_nchw_fchw_f32.mlir

@@ -0,0 +1,51 @@
+// RUN: buddy-opt %s \
+// RUN:     --convert-linalg-to-gemmini="acc_t=f32" | \
+// RUN: FileCheck %s
+
+memref.global "private" @input : memref<2x2x5x5xf32> = dense<[[[[1., 0., -1., 0., 1.],
+                                                               [1., 0., -1., 0., 1.],
+                                                               [1., 0., -1., 0., 1.],
+                                                               [1., 0., -1., 0., 1.],
+                                                               [-1., 0., 1., 0., -1.]],
+                                                              [[-1., 0., 1., 0., -1.],
+                                                               [-1., 0., 1., 0., -1.],
+                                                               [-1., 0., 1., 0., -1.],
+                                                               [-1., 0., 1., 0., -1.],
+                                                               [-1., 0., 1., 0., -1.]]],
+                                                             [[[1., 0., 2., 0., 1.],
+                                                               [1., 0., 2., 0., 1.],
+                                                               [1., 0., 2., 0., 1.],
+                                                               [1., 0., 2., 0., 1.],
+                                                               [-1., 0., 2., 0., -1.]],
+                                                              [[-1., 0., 2., 0., -1.],
+                                                               [-1., 0., 2., 0., -1.],
+                                                               [-1., 0., 2., 0., -1.],
+                                                               [-1., 0., 2., 0., -1.],
+                                                               [-1., 0., 2., 0., -1.]]]]>
+
+memref.global "private" @weight : memref<2x2x3x3xf32> = dense<[[[[1., 2., 3.],
+                                                                [3., 2., 1.],
+                                                                [1., 2., 3.]],
+                                                               [[3., 2., 1.],
+                                                                [1., 2., 3.],
+                                                                [3., 2., 1.]]],
+                                                                [[[1., 2., 3.],
+                                                                [3., 2., 1.],
+                                                                [1., 2., 3.]],
+                                                               [[3., 2., 1.],
+                                                                [1., 2., 3.],
+                                                                [3., 2., 1.]]]]>
+
+func.func @main() -> i8 {
+  %0 = arith.constant 0 : i8
+  %mem0 = memref.get_global @input  : memref<2x2x5x5xf32> 
+  %mem1 = memref.get_global @weight : memref<2x2x3x3xf32>
+  %mem2 = memref.alloc() : memref<2x2x3x3xf32> 
+  // CHECK: gemmini.tile_conv %alloc_{{[0-9]+}} %alloc_{{[0-9]+}} %alloc_{{[0-9]+}} %alloc_{{[0-9]+}} %{{.+}} %{{.+}} : 
+  // CHECK-SAME: memref<2x5x5x2xf32> memref<18x2xf32> memref<2xf32> memref<18x2xf32> i64 i64
+  linalg.conv_2d_nchw_fchw 
+    ins (%mem0, %mem1 : memref<2x2x5x5xf32>, memref<2x2x3x3xf32>)
+  outs(%mem2 : memref<2x2x3x3xf32>)
+  gemmini.print %mem2 : memref<2x2x3x3xf32>
+  return %0 : i8
+}

experiments/gemmini/inputs/matmul.mlir

@@ -0,0 +1,7 @@
+module {
+  func.func @matmul(%A: memref<64x64xf16>, %B: memref<64x64xf16>, %C: memref<64x64xf32>) {
+    linalg.matmul ins(%A, %B : memref<64x64xf16>, memref<64x64xf16>)
+                 outs(%C : memref<64x64xf32>)
+    return
+  }
+}

experiments/gemmini/inputs/tile-matmul-ws-softmax.mlir

@@ -0,0 +1,49 @@
+// RUN: buddy-opt %s \
+// RUN:     --lower-gemmini | \
+// RUN: FileCheck %s
+
+memref.global "private" @g1 : memref<5x5xi8> = dense<[[1, 0, 0, 1, 0], [1, -1, 1, 0, 0], [-1, 0, 1, -1, 1], [1, 0, 0, 1, 0], [-1, 0, 0, -1, 0]]>
+memref.global "private" @g2 : memref<5x5xi8> = dense<[[1, -1, 0, 0, 1], [1, 0, -1, 0, -1], [-1, -1, 0, -1, 1], [-1, 0, 0, 1, 0], [1, 0, 0, -1, 0]]>
+
+
+func.func @main() -> i8 {
+  %i0 = arith.constant 0 : i8
+  %i1I8 = arith.constant 1 : i8
+  %minus1 = arith.constant -2 : i8
+  %i2I8 = arith.constant 2 : i8
+  %i2I32 = arith.constant 2 : i32
+  %dI32 = arith.constant 0 : i32
+  %c0 = arith.constant 0 : index 
+  %c1 = arith.constant 1 : index
+  %aArray = memref.get_global @g1 : memref<5x5xi8>
+  %bArray = memref.get_global @g2 : memref<5x5xi8>
+  %cArray = memref.alloc()  : memref<5x5xi8>
+  %dArray = memref.alloc()  : memref<5x5xi32>
+  %dim_I = memref.dim %aArray, %c0 : memref<5x5xi8>
+  %dim_J = memref.dim %bArray, %c1 : memref<5x5xi8>
+  %dim_K = memref.dim %aArray, %c1 : memref<5x5xi8>
+
+  scf.for %i3 = %c0 to %dim_I step %c1 {
+    scf.for %j3 = %c0 to %dim_J step %c1 {
+      memref.store %dI32, %dArray[%i3, %j3] : memref<5x5xi32>
+    }
+  }
+
+  gemmini.tile_matmul %aArray %bArray %cArray %dArray {dataflow=1}: memref<5x5xi8> memref<5x5xi8> memref<5x5xi8> memref<5x5xi32>
+  gemmini.print %cArray : memref<5x5xi8>
+
+  // CHECK: "gemmini.intr.config_ex"
+  // CHECK: "gemmini.intr.config_st"
+  // CHECK: "gemmini.intr.config_ld"
+  // CHECK: "gemmini.intr.config_norm"
+  // CHECK: "gemmini.intr.loop_ws_config_bounds"
+  // CHECK: "gemmini.intr.loop_ws_config_addrs_ab"
+  // CHECK: "gemmini.intr.loop_ws_config_addrs_dc"
+  // CHECK: "gemmini.intr.loop_ws_config_strides_ab"
+  // CHECK: "gemmini.intr.loop_ws_config_strides_dc"
+  // CHECK: "gemmini.intr.loop_ws"
+  // CHECk: "gemmini.intr.flush"
+  gemmini.tile_matmul %aArray %bArray %cArray %dArray {dataflow=1, act=4, bertScale=0.05:f32}: memref<5x5xi8> memref<5x5xi8> memref<5x5xi8> memref<5x5xi32>
+  gemmini.print %cArray : memref<5x5xi8>
+  return %i0 : i8
+}

experiments/gemmini/iree_inputs/mini_cnn_block.gemmini_tile.mlir

@@ -0,0 +1,191 @@
+module {
+  func.func @mini_cnn_block(%arg0: memref<1x3x32x32xf32>, %arg1: memref<16x3x3x3xf32>, %arg2: memref<32x16x3x3xf32>, %arg3: memref<1x32x26x26xf32>) {
+    %alloc = memref.alloc() : memref<1x16x30x30xf32>
+    %alloc_0 = memref.alloc() : memref<1x32x26x26xf32>
+    %alloc_1 = memref.alloc() : memref<1x32x32x3xf32>
+    %alloc_2 = memref.alloc() : memref<27x16xf32>
+    %alloc_3 = memref.alloc() : memref<16xi32>
+    %alloc_4 = memref.alloc() : memref<900x16xf32>
+    %c30_i64 = arith.constant 30 : i64
+    %c3 = arith.constant 3 : index
+    %c3_5 = arith.constant 3 : index
+    %c0 = arith.constant 0 : index
+    %c1 = arith.constant 1 : index
+    %c1_6 = arith.constant 1 : index
+    scf.for %arg4 = %c0 to %c1 step %c1_6 {
+      %c0_27 = arith.constant 0 : index
+      %c3_28 = arith.constant 3 : index
+      %c1_29 = arith.constant 1 : index
+      scf.for %arg5 = %c0_27 to %c3_28 step %c1_29 {
+        %c0_30 = arith.constant 0 : index
+        %c32_31 = arith.constant 32 : index
+        %c1_32 = arith.constant 1 : index
+        scf.for %arg6 = %c0_30 to %c32_31 step %c1_32 {
+          %c0_33 = arith.constant 0 : index
+          %c32_34 = arith.constant 32 : index
+          %c1_35 = arith.constant 1 : index
+          scf.for %arg7 = %c0_33 to %c32_34 step %c1_35 {
+            %0 = memref.load %arg0[%arg4, %arg5, %arg6, %arg7] : memref<1x3x32x32xf32>
+            memref.store %0, %alloc_1[%arg4, %arg6, %arg7, %arg5] : memref<1x32x32x3xf32>
+          }
+        }
+      }
+    }
+    %c0_7 = arith.constant 0 : index
+    %c16 = arith.constant 16 : index
+    %c1_8 = arith.constant 1 : index
+    scf.for %arg4 = %c0_7 to %c16 step %c1_8 {
+      %c0_27 = arith.constant 0 : index
+      %c3_28 = arith.constant 3 : index
+      %c1_29 = arith.constant 1 : index
+      scf.for %arg5 = %c0_27 to %c3_28 step %c1_29 {
+        %c0_30 = arith.constant 0 : index
+        %c3_31 = arith.constant 3 : index
+        %c1_32 = arith.constant 1 : index
+        scf.for %arg6 = %c0_30 to %c3_31 step %c1_32 {
+          %c0_33 = arith.constant 0 : index
+          %c3_34 = arith.constant 3 : index
+          %c1_35 = arith.constant 1 : index
+          scf.for %arg7 = %c0_33 to %c3_34 step %c1_35 {
+            %0 = arith.muli %arg6, %c3 : index
+            %1 = arith.muli %0, %c3_5 : index
+            %2 = arith.muli %arg7, %c3_5 : index
+            %3 = arith.addi %1, %2 : index
+            %4 = arith.addi %3, %arg5 : index
+            %5 = memref.load %arg1[%arg4, %arg5, %arg6, %arg7] : memref<16x3x3x3xf32>
+            memref.store %5, %alloc_2[%4, %arg4] : memref<27x16xf32>
+          }
+        }
+      }
+    }
+    %c3_i64 = arith.constant 3 : i64
+    gemmini.tile_conv %alloc_1 %alloc_2 %alloc_3 %alloc_4 %c30_i64 %c30_i64 %c3_i64 : memref<1x32x32x3xf32> memref<27x16xf32> memref<16xi32> memref<900x16xf32> i64 i64 i64
+    %c0_9 = arith.constant 0 : index
+    %c1_10 = arith.constant 1 : index
+    %c1_11 = arith.constant 1 : index
+    scf.for %arg4 = %c0_9 to %c1_10 step %c1_11 {
+      %c0_27 = arith.constant 0 : index
+      %c16_28 = arith.constant 16 : index
+      %c1_29 = arith.constant 1 : index
+      scf.for %arg5 = %c0_27 to %c16_28 step %c1_29 {
+        %c0_30 = arith.constant 0 : index
+        %c30 = arith.constant 30 : index
+        %c1_31 = arith.constant 1 : index
+        scf.for %arg6 = %c0_30 to %c30 step %c1_31 {
+          %c0_32 = arith.constant 0 : index
+          %c30_33 = arith.constant 30 : index
+          %c1_34 = arith.constant 1 : index
+          scf.for %arg7 = %c0_32 to %c30_33 step %c1_34 {
+            %c30_35 = arith.constant 30 : index
+            %0 = arith.muli %arg4, %c30_35 : index
+            %1 = arith.muli %0, %c30_35 : index
+            %2 = arith.muli %arg6, %c30_35 : index
+            %3 = arith.addi %1, %2 : index
+            %4 = arith.addi %3, %arg7 : index
+            %5 = memref.load %alloc_4[%4, %arg5] : memref<900x16xf32>
+            memref.store %5, %alloc[%arg4, %arg5, %arg6, %arg7] : memref<1x16x30x30xf32>
+          }
+        }
+      }
+    }
+    memref.dealloc %alloc_1 : memref<1x32x32x3xf32>
+    memref.dealloc %alloc_2 : memref<27x16xf32>
+    memref.dealloc %alloc_4 : memref<900x16xf32>
+    memref.dealloc %alloc_3 : memref<16xi32>
+    %alloc_12 = memref.alloc() : memref<1x30x30x16xf32>
+    %alloc_13 = memref.alloc() : memref<144x32xf32>
+    %alloc_14 = memref.alloc() : memref<32xi32>
+    %alloc_15 = memref.alloc() : memref<676x32xf32>
+    %c26_i64 = arith.constant 26 : i64
+    %c3_16 = arith.constant 3 : index
+    %c16_17 = arith.constant 16 : index
+    %c0_18 = arith.constant 0 : index
+    %c1_19 = arith.constant 1 : index
+    %c1_20 = arith.constant 1 : index
+    scf.for %arg4 = %c0_18 to %c1_19 step %c1_20 {
+      %c0_27 = arith.constant 0 : index
+      %c16_28 = arith.constant 16 : index
+      %c1_29 = arith.constant 1 : index
+      scf.for %arg5 = %c0_27 to %c16_28 step %c1_29 {
+        %c0_30 = arith.constant 0 : index
+        %c30 = arith.constant 30 : index
+        %c1_31 = arith.constant 1 : index
+        scf.for %arg6 = %c0_30 to %c30 step %c1_31 {
+          %c0_32 = arith.constant 0 : index
+          %c30_33 = arith.constant 30 : index
+          %c1_34 = arith.constant 1 : index
+          scf.for %arg7 = %c0_32 to %c30_33 step %c1_34 {
+            %0 = memref.load %alloc[%arg4, %arg5, %arg6, %arg7] : memref<1x16x30x30xf32>
+            memref.store %0, %alloc_12[%arg4, %arg6, %arg7, %arg5] : memref<1x30x30x16xf32>
+          }
+        }
+      }
+    }
+    %c0_21 = arith.constant 0 : index
+    %c32 = arith.constant 32 : index
+    %c1_22 = arith.constant 1 : index
+    scf.for %arg4 = %c0_21 to %c32 step %c1_22 {
+      %c0_27 = arith.constant 0 : index
+      %c16_28 = arith.constant 16 : index
+      %c1_29 = arith.constant 1 : index
+      scf.for %arg5 = %c0_27 to %c16_28 step %c1_29 {
+        %c0_30 = arith.constant 0 : index
+        %c3_31 = arith.constant 3 : index
+        %c1_32 = arith.constant 1 : index
+        scf.for %arg6 = %c0_30 to %c3_31 step %c1_32 {
+          %c0_33 = arith.constant 0 : index
+          %c3_34 = arith.constant 3 : index
+          %c1_35 = arith.constant 1 : index
+          scf.for %arg7 = %c0_33 to %c3_34 step %c1_35 {
+            %0 = arith.muli %arg6, %c3_16 : index
+            %1 = arith.muli %0, %c16_17 : index
+            %2 = arith.muli %arg7, %c16_17 : index
+            %3 = arith.addi %1, %2 : index
+            %4 = arith.addi %3, %arg5 : index
+            %5 = memref.load %arg2[%arg4, %arg5, %arg6, %arg7] : memref<32x16x3x3xf32>
+            memref.store %5, %alloc_13[%4, %arg4] : memref<144x32xf32>
+          }
+        }
+      }
+    }
+    %c3_i64_23 = arith.constant 3 : i64
+    gemmini.tile_conv %alloc_12 %alloc_13 %alloc_14 %alloc_15 %c26_i64 %c26_i64 %c3_i64_23 : memref<1x30x30x16xf32> memref<144x32xf32> memref<32xi32> memref<676x32xf32> i64 i64 i64
+    %c0_24 = arith.constant 0 : index
+    %c1_25 = arith.constant 1 : index
+    %c1_26 = arith.constant 1 : index
+    scf.for %arg4 = %c0_24 to %c1_25 step %c1_26 {
+      %c0_27 = arith.constant 0 : index
+      %c32_28 = arith.constant 32 : index
+      %c1_29 = arith.constant 1 : index
+      scf.for %arg5 = %c0_27 to %c32_28 step %c1_29 {
+        %c0_30 = arith.constant 0 : index
+        %c26 = arith.constant 26 : index
+        %c1_31 = arith.constant 1 : index
+        scf.for %arg6 = %c0_30 to %c26 step %c1_31 {
+          %c0_32 = arith.constant 0 : index
+          %c26_33 = arith.constant 26 : index
+          %c1_34 = arith.constant 1 : index
+          scf.for %arg7 = %c0_32 to %c26_33 step %c1_34 {
+            %c26_35 = arith.constant 26 : index
+            %0 = arith.muli %arg4, %c26_35 : index
+            %1 = arith.muli %0, %c26_35 : index
+            %2 = arith.muli %arg6, %c26_35 : index
+            %3 = arith.addi %1, %2 : index
+            %4 = arith.addi %3, %arg7 : index
+            %5 = memref.load %alloc_15[%4, %arg5] : memref<676x32xf32>
+            memref.store %5, %alloc_0[%arg4, %arg5, %arg6, %arg7] : memref<1x32x26x26xf32>
+          }
+        }
+      }
+    }
+    memref.dealloc %alloc_12 : memref<1x30x30x16xf32>
+    memref.dealloc %alloc_13 : memref<144x32xf32>
+    memref.dealloc %alloc_15 : memref<676x32xf32>
+    memref.dealloc %alloc_14 : memref<32xi32>
+    linalg.copy ins(%alloc_0 : memref<1x32x26x26xf32>) outs(%arg3 : memref<1x32x26x26xf32>)
+    memref.dealloc %alloc : memref<1x16x30x30xf32>
+    memref.dealloc %alloc_0 : memref<1x32x26x26xf32>
+    return
+  }
+}
+

experiments/gemmini/libgemmini_status.txt

@@ -0,0 +1,13 @@
+Mac setup notes:
+
+- buddy-mlir Gemmini lowering works (matmul, batch_matmul, conv, matmul+softmax).
+- Generated LLVM IR (log.ll) and RISC-V asm (log.s) via Makefile targets, asm has Gemmini ops (config_ex, config_st, loop_ws, etc.).
+- Spike + pk + riscv64-unknown-elf-gcc work for a simple "hello" test.
+
+Blocked on:
+- Installing libgemmini (Spike extension) from https://github.com/ucb-bar/libgemmini.
+- `make libgemmini.so` fails on macOS with `ld: symbol(s) not found for architecture arm64` and RISCV-dependent paths that assume a full Chipyard/Gemmini tree.
+
+Plan:
+- Use Mac primarily for IR/pipeline experiments.
+- Do Spike+Gemmini execution on a SLICE Linux machine with Chipyard/Gemmini installed.