Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[GPU] Add an example of running LeNet on GPU. #415

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
+141 −105
Open

[GPU] Add an example of running LeNet on GPU. #415

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4a9fbd9
[GPU] Fix ConvertMemcpyToGPUPass: Skip handling function declarations…
NaOHCC Oct 21, 2024
3f7bf26
[GPU] Add an example of running LeNet on GPU.
NaOHCC Oct 21, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
39 changes: 39 additions & 0 deletions examples/BuddyLeNet/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,6 +57,45 @@ SET_TARGET_PROPERTIES(LENET PROPERTIES LINKER_LANGUAGE C)
add_executable(buddy-lenet-run buddy-lenet-main.cpp)
target_link_directories(buddy-lenet-run PRIVATE ${LLVM_LIBRARY_DIR})

if(NOT DEFINED BUDDY_ENABLE_PNG)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We can use BMP image files, so using libpng is not a strict requirement.

message(FATAL_ERROR "To run LeNet inference, the png library is required. Please define BUDDY_ENABLE_PNG for CMake.")
endif()
set(BUDDY_LENET_LIBS LENET mlir_c_runner_utils ${PNG_LIBRARIES})

target_link_libraries(buddy-lenet-run ${BUDDY_LENET_LIBS})

set(ONE_SHOT_BUFFERIZE_OPTION "bufferize-function-boundaries=1 function-boundary-type-conversion=identity-layout-map")
set(LOWER_TO_NVVM_OPTION "cubin-chip=sm_80 cubin-features=+ptx71 cubin-format=fatbin")
add_custom_command(
OUTPUT subgraph0_gpu.o
COMMAND ${LLVM_TOOLS_BINARY_DIR}/mlir-opt ${BUDDY_EXAMPLES_DIR}/BuddyLeNet/subgraph0.mlir
-pass-pipeline "builtin.module(func.func(tosa-to-linalg-named, tosa-to-linalg, tosa-to-tensor, tosa-to-arith))" |
${LLVM_TOOLS_BINARY_DIR}/mlir-opt
-one-shot-bufferize=${ONE_SHOT_BUFFERIZE_OPTION}
-buffer-deallocation
-convert-linalg-to-parallel-loops
-canonicalize
-gpu-map-parallel-loops
-convert-parallel-loops-to-gpu
-gpu-kernel-outlining
-canonicalize
-cse |
${BUDDY_BINARY_DIR}/buddy-opt -convert-memcpy-to-gpu -gpu-async-region -canonicalize |
${LLVM_TOOLS_BINARY_DIR}/mlir-opt -llvm-request-c-wrappers --test-lower-to-nvvm=${LOWER_TO_NVVM_OPTION} |
${LLVM_TOOLS_BINARY_DIR}/mlir-translate -mlir-to-llvmir |
${LLVM_TOOLS_BINARY_DIR}/llvm-as |
${LLVM_TOOLS_BINARY_DIR}/llc -filetype=obj -relocation-model=pic -O0 -o ${BUDDY_BINARY_DIR}/../examples/BuddyLeNet/subgraph0_gpu.o
DEPENDS ${BUDDY_EXAMPLES_DIR}/BuddyLeNet/subgraph0.mlir
COMMENT "Building subgraph0_gpu.o"
VERBATIM)

add_library(LENET_GPU STATIC subgraph0_gpu.o forward.o)

SET_TARGET_PROPERTIES(LENET_GPU PROPERTIES LINKER_LANGUAGE C)

add_executable(buddy-lenet-run-gpu buddy-lenet-main.cpp)
target_link_directories(buddy-lenet-run-gpu PRIVATE ${LLVM_LIBRARY_DIR})

set(BUDDY_LENET_LIBS_GPU LENET_GPU mlir_c_runner_utils mlir_async_runtime mlir_runner_utils mlir_cuda_runtime ${PNG_LIBRARIES})

target_link_libraries(buddy-lenet-run-gpu ${BUDDY_LENET_LIBS_GPU})
131 changes: 41 additions & 90 deletions midend/lib/Conversion/MLIRGPU/ConvertMemcpyToGPU.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,21 +18,17 @@
//
//===---------------------------------------------------------------------===//

#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/GPU/IR/GPUDialect.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/IR/Builders.h"
#include "mlir/IR/BuiltinOps.h"
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/IR/OperationSupport.h"
#include "mlir/IR/TypeRange.h"
#include "mlir/IR/ValueRange.h"
#include "mlir/IR/Visitors.h"
#include "mlir/Support/LLVM.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include <mlir/Dialect/Affine/IR/AffineOps.h>
#include <mlir/Dialect/Func/IR/FuncOps.h>
#include <mlir/Dialect/Linalg/Transforms/Transforms.h>
Expand All @@ -42,11 +38,8 @@
#include <mlir/IR/Value.h>
#include <mlir/Pass/Pass.h>

#include <map>
#include <set>
#include <sstream>
#include <string>
#include <utility>
#include <vector>

using namespace mlir;
using namespace vector;

Expand Down Expand Up @@ -82,6 +75,9 @@ class ConvertMemcpyToGPUPass
void ConvertMemcpyToGPUPass::runOnOperation() {
auto funcOp = getOperation();

if (funcOp.isDeclaration() || funcOp.isExternal())
return;

// Make sure the gpu function is already outlined.
funcOp->walk<WalkOrder::PreOrder>([&](Operation *nestedOp) {
if (auto gpuLaunchOp = dyn_cast<gpu::LaunchOp>(nestedOp)) {
Expand All @@ -90,8 +86,9 @@ void ConvertMemcpyToGPUPass::runOnOperation() {
return WalkResult::advance();
});

std::set<gpu::AllocOp *> unDeallocatedOperations;
std::vector<Value> unDeallocatedValue;
OpBuilder builder(funcOp->getContext());

// Copy all function arguments to gpu, needs deallocation
if (processArgs) {
builder.setInsertionPointToStart(&(funcOp.getBody().front()));
Expand All @@ -103,23 +100,11 @@ void ConvertMemcpyToGPUPass::runOnOperation() {
auto memrefType = dyn_cast<MemRefType>(arg.getType());
auto gpuAllocOp = builder.create<gpu::AllocOp>(
builder.getUnknownLoc(), TypeRange({memrefType}), ValueRange({}));
unDeallocatedOperations.insert(&gpuAllocOp);
unDeallocatedValue.push_back(gpuAllocOp->getResult(0));
auto gpuMemcpyOp = builder.create<gpu::MemcpyOp>(
gpuAllocOp.getLoc(), TypeRange(), ValueRange(),
gpuAllocOp.getResult(0), arg);
// Replace all users with GPU memory
auto users = arg.getUsers();
std::vector<Operation *> usersVec(users.begin(), users.end());
for (auto user : usersVec) {
// Don't replace memcpy's operand
if (isa<gpu::MemcpyOp>(user))
continue;
for (size_t j = 0; j < user->getNumOperands(); j++) {
if (user->getOperand(j) == arg) {
user->setOperand(j, gpuAllocOp.getResult(0));
}
}
}
arg.replaceAllUsesExcept(gpuAllocOp->getResult(0), gpuMemcpyOp);
}
}

Expand Down Expand Up @@ -149,19 +134,18 @@ void ConvertMemcpyToGPUPass::runOnOperation() {

auto gpuAllocOp = builder.create<gpu::AllocOp>(
allocOp->getLoc(), TypeRange({memrefType}), ValueRange({}));
auto users = result.getUsers();
std::vector<Operation *> usersVec(users.begin(), users.end());
for (auto user : usersVec) {
for (size_t j = 0; j < user->getNumOperands(); j++) {
// Only the return value will not have dealloc op
if (auto deallocOp = dyn_cast<memref::DeallocOp>(user)) {
builder.setInsertionPointAfter(deallocOp);
auto gpuDeallocOp = builder.create<gpu::DeallocOp>(
deallocOp->getLoc(), TypeRange(), ValueRange(),
gpuAllocOp.getResult(0));
deallocOp->erase();
} else if (user->getOperand(j) == result) {
user->setOperand(j, gpuAllocOp.getResult(0));

for (auto user : llvm::make_early_inc_range(result.getUsers())) {
if (auto deallocOp = dyn_cast<memref::DeallocOp>(user)) {
builder.setInsertionPointAfter(deallocOp);
builder.create<gpu::DeallocOp>(deallocOp->getLoc(), TypeRange(),
ValueRange(), gpuAllocOp.getResult(0));
deallocOp->erase();
} else {
for (auto &opOperand : user->getOpOperands()) {
if (opOperand.is(result)) {
opOperand.set(gpuAllocOp.getResult(0));
}
}
}
}
Expand All @@ -175,28 +159,8 @@ void ConvertMemcpyToGPUPass::runOnOperation() {
builder.setInsertionPointAfter(copyOp);
auto gpuMemcpyOp = builder.create<gpu::MemcpyOp>(
copyOp->getLoc(), TypeRange(), ValueRange(), dst, src);
{
auto users = src.getUsers();
std::vector<Operation *> usersVec(users.begin(), users.end());
for (auto user : usersVec) {
for (size_t j = 0; j < user->getNumOperands(); j++) {
if (user->getOperand(j) == src) {
user->setOperand(j, gpuMemcpyOp.getOperand(1));
}
}
}
}
{
auto users = dst.getUsers();
std::vector<Operation *> usersVec(users.begin(), users.end());
for (auto user : usersVec) {
for (size_t j = 0; j < user->getNumOperands(); j++) {
if (user->getOperand(j) == src) {
user->setOperand(j, gpuMemcpyOp.getOperand(0));
}
}
}
}
src.replaceAllUsesWith(gpuMemcpyOp->getResult(1));
dst.replaceAllUsesWith(gpuMemcpyOp->getResult(0));
copyOp->erase();
}
// Allocate space on GPU and copy global memrefs to GPU, needs deallocation
Expand All @@ -206,47 +170,34 @@ void ConvertMemcpyToGPUPass::runOnOperation() {
auto memrefType = dyn_cast<MemRefType>(result.getType());
auto gpuAllocOp = builder.create<gpu::AllocOp>(
getGlobalOp->getLoc(), TypeRange({memrefType}), ValueRange({}));
unDeallocatedOperations.insert(&gpuAllocOp);
unDeallocatedValue.push_back(gpuAllocOp->getResult(0));

auto src = result;
auto dst = gpuAllocOp->getResult(0);
auto gpuMemcpyOp = builder.create<gpu::MemcpyOp>(
gpuAllocOp->getLoc(), TypeRange(), ValueRange(), dst, src);
{
auto users = src.getUsers();
std::vector<Operation *> usersVec(users.begin(), users.end());
for (auto user : usersVec) {
if (isa<gpu::MemcpyOp>(user))
continue;
// TODO: replace with src.replaceAllUsesExcept()
for (size_t j = 0; j < user->getNumOperands(); j++) {
if (user->getOperand(j) == src) {
user->setOperand(j, dst);
}
}
}
}
src.replaceAllUsesExcept(dst, gpuMemcpyOp);
}
// Copy data back to CPU, deallocate GPU, then return
else if (auto returnOp = dyn_cast<func::ReturnOp>(nestedOp)) {
builder.setInsertionPoint(returnOp);

for (auto *gpuAllocOp : unDeallocatedOperations) {
auto gpuDeallocOp = builder.create<gpu::DeallocOp>(
builder.getUnknownLoc(), TypeRange(), ValueRange(),
gpuAllocOp->getResult(0));
}
builder.setInsertionPoint(returnOp);
for (unsigned i = 0; i < returnOp.getNumOperands(); ++i) {
auto val = returnOp->getOperand(i);
auto memRefType = dyn_cast<MemRefType>(val.getType());
auto allocOp = builder.create<memref::AllocOp>(builder.getUnknownLoc(),
memRefType);
auto gpuMemcpyOp = builder.create<gpu::MemcpyOp>(
allocOp.getLoc(), TypeRange(), ValueRange(), allocOp->getResult(0),
val);
auto gpuDeallocOp = builder.create<gpu::DeallocOp>(
gpuMemcpyOp->getLoc(), TypeRange(), ValueRange(), val);
returnOp->setOperand(i, allocOp->getResult(0));
if (auto memrefType = dyn_cast<MemRefType>(val.getType())) {
auto allocOp =
builder.create<memref::AllocOp>(returnOp->getLoc(), memrefType);
builder.create<gpu::MemcpyOp>(allocOp.getLoc(), TypeRange(),
ValueRange(), allocOp->getResult(0),
val);
// FIXME: may be leak memory
// auto gpuDeallocOp = builder.create<gpu::DeallocOp>(
// gpuMemcpyOp->getLoc(), TypeRange(), ValueRange(), val);
returnOp->setOperand(i, allocOp->getResult(0));
}
}
for (auto value : unDeallocatedValue) {
builder.create<gpu::DeallocOp>(returnOp->getLoc(), TypeRange(),
ValueRange(), value);
}
}
return WalkResult::advance();
Expand Down
76 changes: 61 additions & 15 deletions tests/Conversion/convert-memcpy-to-gpu.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,22 +1,68 @@
// RUN: buddy-opt -convert-memcpy-to-gpu -canonicalize %s | FileCheck %s
// RUN: buddy-opt -convert-memcpy-to-gpu="process-args=1" %s | FileCheck %s

// CHECK: %memref = gpu.alloc () : memref<32x32xf32>
// CHECK: %memref_0 = gpu.alloc () : memref<32x32xf32>
// CHECK: gpu.dealloc %memref : memref<32x32xf32>
// CHECK: %alloc = memref.alloc() : memref<32x32xf32>
// CHECK: gpu.memcpy %alloc, %memref_0 : memref<32x32xf32>, memref<32x32xf32>
// CHECK: gpu.dealloc %memref_0 : memref<32x32xf32>
#map = affine_map<(d0)[s0, s1] -> (d0 * s0 + s1)>
module attributes {gpu.container_module} {
func.func @matmul(%arg0: memref<32x32xf32>, %arg1: memref<32x32xf32>) -> memref<32x32xf32> {
%c2 = arith.constant 2 : index
%c64 = arith.constant 64 : index
memref.global "private" constant @__constant_1x10x10xf32 : memref<1x10x10xf32> = dense<1.000000e+00> {alignment = 64 : i64}
func.func @matmul(%arg0: memref<1x10x10xf32>, %arg1: memref<1x10x10xf32>) -> memref<1x10x10xf32> {
// CHECK: %[[d_arg0:.*]] = gpu.alloc () : memref<1x10x10xf32>
// CHECK-NEXT: gpu.memcpy %[[d_arg0]], %arg0 : memref<1x10x10xf32>, memref<1x10x10xf32>
// CHECK: %[[d_arg1:.*]] = gpu.alloc () : memref<1x10x10xf32>
// CHECK-NEXT: gpu.memcpy %[[d_arg1:.*]], %arg1 : memref<1x10x10xf32>, memref<1x10x10xf32>
%c10 = arith.constant 10 : index
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%alloc = memref.alloc() {alignment = 64 : i64} : memref<32x32xf32>
gpu.launch_func @matmul_kernel::@matmul_kernel blocks in (%c1, %c1, %c1) threads in (%c64, %c2, %c1)
return %alloc : memref<32x32xf32>
%cst = arith.constant 0.000000e+00 : f32
// CHECK: %[[h_global_data:.*]] = memref.get_global @__constant_1x10x10xf32 : memref<1x10x10xf32>
// CHECK: %[[d_global_data:.*]] = gpu.alloc () : memref<1x10x10xf32>
// CHECK: gpu.memcpy %[[d_global_data]], %[[h_global_data]] : memref<1x10x10xf32>, memref<1x10x10xf32>
%0 = memref.get_global @__constant_1x10x10xf32 : memref<1x10x10xf32>
// CHECK: %[[d_alloc0:.*]] = gpu.alloc () : memref<1x10x10xf32>
%alloc = memref.alloc() {alignment = 64 : i64} : memref<1x10x10xf32>
// CHECK: gpu.launch_func
gpu.launch_func @kernel::@fill blocks in (%c10, %c10, %c1) threads in (%c1, %c1, %c1) args(%c1 : index, %c0 : index, %cst : f32, %alloc : memref<1x10x10xf32>)
// CHECK: gpu.launch_func
// CHECK-SAME: %[[d_arg0]]
// CHECK-SAME: %[[d_arg1]]
// CHECK-SAME: %[[d_alloc0]]
gpu.launch_func @kernel::@matmul blocks in (%c10, %c10, %c1) threads in (%c1, %c1, %c1) args(%c1 : index, %c0 : index, %arg0 : memref<1x10x10xf32>, %arg1 : memref<1x10x10xf32>, %alloc : memref<1x10x10xf32>, %c10 : index)
// CHECK: %[[d_alloc1:.*]] = gpu.alloc () : memref<1x10x10xf32>
%alloc_0 = memref.alloc() {alignment = 64 : i64} : memref<1x10x10xf32>
// CHECK: gpu.launch_func
gpu.launch_func @kernel::@fill blocks in (%c10, %c10, %c1) threads in (%c1, %c1, %c1) args(%c1 : index, %c0 : index, %cst : f32, %alloc_0 : memref<1x10x10xf32>)
// CHECK: gpu.launch_func
// CHECK-SAME: %[[d_global_data]]
// CHECK-SAME: %[[d_alloc0]]
// CHECK-SAME: %[[d_alloc1]]
gpu.launch_func @kernel::@matmul blocks in (%c10, %c10, %c1) threads in (%c1, %c1, %c1) args(%c1 : index, %c0 : index, %0 : memref<1x10x10xf32>, %alloc : memref<1x10x10xf32>, %alloc_0 : memref<1x10x10xf32>, %c10 : index)
// CHECK: %[[d_result:.*]] = gpu.alloc () : memref<1x10x10xf32>
%alloc_1 = memref.alloc() {alignment = 64 : i64} : memref<1x10x10xf32>
// CHECK: gpu.launch_func
gpu.launch_func @kernel::@fill blocks in (%c10, %c10, %c1) threads in (%c1, %c1, %c1) args(%c1 : index, %c0 : index, %cst : f32, %alloc_1 : memref<1x10x10xf32>)
// CHECK: gpu.launch_func
// CHECK-SAME: %[[d_alloc0]]
// CHECK-SAME: %[[d_alloc1]]
// CHECK-SAME: %[[d_result]]
gpu.launch_func @kernel::@matmul blocks in (%c10, %c10, %c1) threads in (%c1, %c1, %c1) args(%c1 : index, %c0 : index, %alloc : memref<1x10x10xf32>, %alloc_0 : memref<1x10x10xf32>, %alloc_1 : memref<1x10x10xf32>, %c10 : index)
// CHECK: gpu.dealloc %[[d_alloc1]] : memref<1x10x10xf32>
memref.dealloc %alloc_0 : memref<1x10x10xf32>
// CHECK: gpu.dealloc %[[d_alloc0]] : memref<1x10x10xf32>
memref.dealloc %alloc : memref<1x10x10xf32>

// CHECK: %[[h_alloc:.*]] = memref.alloc() : memref<1x10x10xf32>
// CHECK-NEXT: gpu.memcpy %[[h_alloc]], %[[d_result]] : memref<1x10x10xf32>, memref<1x10x10xf32>

// CHECK: gpu.dealloc %[[d_arg0]] : memref<1x10x10xf32>
// CHECK: gpu.dealloc %[[d_arg1]] : memref<1x10x10xf32>
// CHECK: gpu.dealloc %[[d_global_data]] : memref<1x10x10xf32>

// CHECK: return %[[h_alloc]] : memref<1x10x10xf32>
return %alloc_1 : memref<1x10x10xf32>
}
gpu.module @matmul_kernel {
gpu.func @matmul_kernel() kernel attributes {gpu.known_block_size = array<i32: 64, 2, 1>, gpu.known_grid_size = array<i32: 1, 1, 1>} {
gpu.module @kernel {
gpu.func @fill(%arg0: index, %arg1: index, %arg2: f32, %arg3: memref<1x10x10xf32>) kernel attributes {gpu.known_block_size = array<i32: 1, 1, 1>} {
gpu.return
}
gpu.func @matmul(%arg0: index, %arg1: index, %arg2: memref<1x10x10xf32>, %arg3: memref<1x10x10xf32>, %arg4: memref<1x10x10xf32>, %arg5: index) kernel attributes {gpu.known_block_size = array<i32: 1, 1, 1>} {
gpu.return
}
}
Expand Down