[examples] LeNet E2E pipeline uses batchmatmul-optimize pass.

buddy-compiler · Aug 16, 2024 · 0424893 · 0424893
1 parent 1585e2a
commit 0424893
Show file tree

Hide file tree

Showing 3 changed files with 57 additions and 22 deletions.
diff --git a/examples/BuddyLeNet/CMakeLists.txt b/examples/BuddyLeNet/CMakeLists.txt
@@ -23,7 +23,8 @@ add_custom_command(
             -pass-pipeline "builtin.module(func.func(tosa-to-linalg-named, tosa-to-linalg, tosa-to-tensor, tosa-to-arith))" |
           ${BUDDY_BINARY_DIR}/buddy-opt
             -eliminate-empty-tensors
-            -convert-tensor-to-linalg 
+            -convert-tensor-to-linalg
+            -batchmatmul-optimize
             -linalg-bufferize
             -convert-linalg-to-affine-loops
             -lower-affine

diff --git a/examples/BuddyLeNet/fake-lenet.mlir b/examples/BuddyLeNet/fake-lenet.mlir
@@ -1,5 +1,6 @@
 module {
   func.func private @printMemrefF32(%ptr : tensor<*xf32>)
+  func.func private @rtclock() -> f64
 
   func.func @forward(%arg0: tensor<44426xf32>, %arg1: tensor<1x1x28x28xf32>) -> tensor<1x10xf32> {
     %extracted_slice = tensor.extract_slice %arg0[0] [150] [1] : tensor<44426xf32> to tensor<150xf32>
@@ -81,10 +82,16 @@ module {
     %fake_params = arith.constant dense<1.0> : tensor<44426xf32>
     %fake_input = arith.constant dense<2.0> : tensor<1x1x28x28xf32>
 
+    %t_start = call @rtclock() : () -> f64
     %fake_output = call @forward(%fake_params, %fake_input) : (tensor<44426xf32>, tensor<1x1x28x28xf32>) -> tensor<1x10xf32>
+    %t_end = call @rtclock() : () -> f64
 
     %tensor_unranked = tensor.cast %fake_output : tensor<1x10xf32> to tensor<*xf32>
     call @printMemrefF32(%tensor_unranked) : (tensor<*xf32>) -> ()
+
+    %time = arith.subf %t_end, %t_start : f64
+    vector.print %time : f64
+
     return
   }
 }
diff --git a/examples/BuddyLeNet/makefile b/examples/BuddyLeNet/makefile
@@ -1,30 +1,33 @@
 #!/bin/bash
-BUDDY_OPT := ../../build/bin/buddy-opt
-MLIR_OPT := ../../llvm/build/bin/mlir-opt
-MLIR_TRANSLATE := ../../llvm/build/bin/mlir-translate
-MLIR_CPU_RUNNER := ../../llvm/build/bin/mlir-cpu-runner
-LLC := ../../llvm/build/bin/llc
-OPT_FLAG := -O0
+BUDDY_BUILD_DIR := ../../build/
+LLVM_BUILD_DIR := ../../llvm/build/
+BUDDY_OPT := ${BUDDY_BUILD_DIR}/bin/buddy-opt
+MLIR_OPT := ${LLVM_BUILD_DIR}/bin/mlir-opt
+MLIR_TRANSLATE := ${LLVM_BUILD_DIR}/bin/mlir-translate
+MLIR_CPU_RUNNER := ${LLVM_BUILD_DIR}/bin/mlir-cpu-runner
+LLC := ${LLVM_BUILD_DIR}/bin/llc
+OPT_FLAG := -O3
 
 ifeq ($(shell uname),Linux)
-MLIR_RUNNER_UTILS := ../../llvm/build/lib/libmlir_runner_utils.so
-MLIR_C_RUNNER_UTILS := ../../llvm/build/lib/libmlir_c_runner_utils.so
-MLIR_ASYNC_RUNTIME := ../../llvm/build/lib/libmlir_async_runtime.so
+MLIR_RUNNER_UTILS := ${LLVM_BUILD_DIR}/lib/libmlir_runner_utils.so
+MLIR_C_RUNNER_UTILS := ${LLVM_BUILD_DIR}/lib/libmlir_c_runner_utils.so
+MLIR_ASYNC_RUNTIME := ${LLVM_BUILD_DIR}/lib/libmlir_async_runtime.so
 MTRIPLE := x86_64-unknown-linux-gnu
 else ifeq ($(shell uname),Darwin)
-MLIR_RUNNER_UTILS := ../../llvm/build/lib/libmlir_runner_utils.dylib
-MLIR_C_RUNNER_UTILS := ../../llvm/build/lib/libmlir_c_runner_utils.dylib
-MLIR_ASYNC_RUNTIME := ./../llvm/build/lib/libmlir_async_runtime.dylib
+MLIR_RUNNER_UTILS := ${LLVM_BUILD_DIR}/lib/libmlir_runner_utils.dylib
+MLIR_C_RUNNER_UTILS := ${LLVM_BUILD_DIR}/lib/libmlir_c_runner_utils.dylib
+MLIR_ASYNC_RUNTIME := ${LLVM_BUILD_DIR}/lib/libmlir_async_runtime.dylib
 MTRIPLE := x86_64-apple-darwin
 endif
 
 buddy-lenet-lower:
-	@${MLIR_OPT} ./fake-lenet.mlir \
+	@${BUDDY_OPT} ./fake-lenet.mlir \
 		-pass-pipeline "builtin.module(func.func(tosa-to-linalg-named, tosa-to-linalg, tosa-to-tensor, tosa-to-arith))" | \
-		${MLIR_OPT} \
+	${BUDDY_OPT} \
 		-eliminate-empty-tensors \
 		-convert-tensor-to-linalg \
 		-linalg-bufferize \
+		-batchmatmul-optimize \
 		-convert-linalg-to-affine-loops \
 		-lower-affine \
 		-func-bufferize \
@@ -38,16 +41,15 @@ buddy-lenet-lower:
 		-convert-arith-to-llvm \
 		-finalize-memref-to-llvm \
 		-convert-scf-to-cf \
-		-llvm-request-c-wrappers \
 		-convert-arith-to-llvm \
 		-convert-func-to-llvm \
 		-reconcile-unrealized-casts \
 		-o ./log.mlir
 
 buddy-lenet-translate:
-	@${MLIR_OPT} ./fake-lenet.mlir \
+	@${BUDDY_OPT} ./fake-lenet.mlir \
 		-pass-pipeline "builtin.module(func.func(tosa-to-linalg-named, tosa-to-linalg, tosa-to-tensor, tosa-to-arith))" | \
-		${MLIR_OPT} \
+	${BUDDY_OPT} \
 		-eliminate-empty-tensors \
 		-convert-tensor-to-linalg \
 		-linalg-bufferize \
@@ -64,17 +66,16 @@ buddy-lenet-translate:
 		-convert-arith-to-llvm \
 		-finalize-memref-to-llvm \
 		-convert-scf-to-cf \
-		-llvm-request-c-wrappers \
 		-convert-arith-to-llvm \
 		-convert-func-to-llvm \
 		-reconcile-unrealized-casts | \
 	${MLIR_TRANSLATE} -mlir-to-llvmir -o log.ll
 
 
 buddy-lenet-run:
-	@${MLIR_OPT} ./fake-lenet.mlir \
+	@${BUDDY_OPT} ./fake-lenet.mlir \
 		-pass-pipeline "builtin.module(func.func(tosa-to-linalg-named, tosa-to-linalg, tosa-to-tensor, tosa-to-arith))" | \
-		${MLIR_OPT} \
+	${BUDDY_OPT} \
 		-eliminate-empty-tensors \
 		-convert-tensor-to-linalg \
 		-linalg-bufferize \
@@ -91,7 +92,33 @@ buddy-lenet-run:
 		-convert-arith-to-llvm \
 		-finalize-memref-to-llvm \
 		-convert-scf-to-cf \
-		-llvm-request-c-wrappers \
+		-convert-arith-to-llvm \
+		-convert-func-to-llvm \
+		-reconcile-unrealized-casts | \
+	${MLIR_CPU_RUNNER} ${OPT_FLAG} -e main -entry-point-result=void \
+		-shared-libs=${MLIR_RUNNER_UTILS} -shared-libs=${MLIR_C_RUNNER_UTILS}
+
+buddy-lenet-opt-run:
+	@${BUDDY_OPT} ./fake-lenet.mlir \
+		-pass-pipeline "builtin.module(func.func(tosa-to-linalg-named, tosa-to-linalg, tosa-to-tensor, tosa-to-arith))" | \
+	${BUDDY_OPT} \
+		-eliminate-empty-tensors \
+		-convert-tensor-to-linalg \
+		-linalg-bufferize \
+		-batchmatmul-optimize \
+		-convert-linalg-to-affine-loops \
+		-lower-affine \
+		-func-bufferize \
+		-arith-bufferize \
+		-tensor-bufferize \
+		-buffer-deallocation \
+		-finalizing-bufferize \
+		-convert-vector-to-scf \
+		-expand-strided-metadata \
+		-convert-vector-to-llvm \
+		-convert-arith-to-llvm \
+		-finalize-memref-to-llvm \
+		-convert-scf-to-cf \
 		-convert-arith-to-llvm \
 		-convert-func-to-llvm \
 		-reconcile-unrealized-casts | \