1 //===- RocmRuntimeWrappers.cpp - MLIR ROCM runtime wrapper library --------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // Implements C wrappers around the ROCM library for easy linking in ORC jit.
10 // Also adds some debugging helpers that are helpful when writing MLIR code to
11 // run on GPUs.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include <cassert>
16 #include <numeric>
17 
18 #include "mlir/ExecutionEngine/CRunnerUtils.h"
19 #include "llvm/ADT/ArrayRef.h"
20 
21 #include "hip/hip_runtime.h"
22 
23 #define HIP_REPORT_IF_ERROR(expr)                                              \
24   [](hipError_t result) {                                                      \
25     if (!result)                                                               \
26       return;                                                                  \
27     const char *name = hipGetErrorName(result);                                \
28     if (!name)                                                                 \
29       name = "<unknown>";                                                      \
30     fprintf(stderr, "'%s' failed with '%s'\n", #expr, name);                   \
31   }(expr)
32 
33 thread_local static int32_t defaultDevice = 0;
34 
35 // Sets the `Context` for the duration of the instance and restores the previous
36 // context on destruction.
37 class ScopedContext {
38 public:
ScopedContext()39   ScopedContext() {
40     // Static reference to HIP primary context for device ordinal defaultDevice.
41     static hipCtx_t context = [] {
42       HIP_REPORT_IF_ERROR(hipInit(/*flags=*/0));
43       hipDevice_t device;
44       HIP_REPORT_IF_ERROR(hipDeviceGet(&device, /*ordinal=*/defaultDevice));
45       hipCtx_t ctx;
46       HIP_REPORT_IF_ERROR(hipDevicePrimaryCtxRetain(&ctx, device));
47       return ctx;
48     }();
49 
50     HIP_REPORT_IF_ERROR(hipCtxPushCurrent(context));
51   }
52 
~ScopedContext()53   ~ScopedContext() { HIP_REPORT_IF_ERROR(hipCtxPopCurrent(nullptr)); }
54 };
55 
mgpuModuleLoad(void * data)56 extern "C" hipModule_t mgpuModuleLoad(void *data) {
57   ScopedContext scopedContext;
58   hipModule_t module = nullptr;
59   HIP_REPORT_IF_ERROR(hipModuleLoadData(&module, data));
60   return module;
61 }
62 
mgpuModuleUnload(hipModule_t module)63 extern "C" void mgpuModuleUnload(hipModule_t module) {
64   HIP_REPORT_IF_ERROR(hipModuleUnload(module));
65 }
66 
mgpuModuleGetFunction(hipModule_t module,const char * name)67 extern "C" hipFunction_t mgpuModuleGetFunction(hipModule_t module,
68                                                const char *name) {
69   hipFunction_t function = nullptr;
70   HIP_REPORT_IF_ERROR(hipModuleGetFunction(&function, module, name));
71   return function;
72 }
73 
74 // The wrapper uses intptr_t instead of ROCM's unsigned int to match
75 // the type of MLIR's index type. This avoids the need for casts in the
76 // generated MLIR code.
mgpuLaunchKernel(hipFunction_t function,intptr_t gridX,intptr_t gridY,intptr_t gridZ,intptr_t blockX,intptr_t blockY,intptr_t blockZ,int32_t smem,hipStream_t stream,void ** params,void ** extra)77 extern "C" void mgpuLaunchKernel(hipFunction_t function, intptr_t gridX,
78                                  intptr_t gridY, intptr_t gridZ,
79                                  intptr_t blockX, intptr_t blockY,
80                                  intptr_t blockZ, int32_t smem,
81                                  hipStream_t stream, void **params,
82                                  void **extra) {
83   ScopedContext scopedContext;
84   HIP_REPORT_IF_ERROR(hipModuleLaunchKernel(function, gridX, gridY, gridZ,
85                                             blockX, blockY, blockZ, smem,
86                                             stream, params, extra));
87 }
88 
mgpuStreamCreate()89 extern "C" hipStream_t mgpuStreamCreate() {
90   ScopedContext scopedContext;
91   hipStream_t stream = nullptr;
92   HIP_REPORT_IF_ERROR(hipStreamCreate(&stream));
93   return stream;
94 }
95 
mgpuStreamDestroy(hipStream_t stream)96 extern "C" void mgpuStreamDestroy(hipStream_t stream) {
97   HIP_REPORT_IF_ERROR(hipStreamDestroy(stream));
98 }
99 
mgpuStreamSynchronize(hipStream_t stream)100 extern "C" void mgpuStreamSynchronize(hipStream_t stream) {
101   return HIP_REPORT_IF_ERROR(hipStreamSynchronize(stream));
102 }
103 
mgpuStreamWaitEvent(hipStream_t stream,hipEvent_t event)104 extern "C" void mgpuStreamWaitEvent(hipStream_t stream, hipEvent_t event) {
105   HIP_REPORT_IF_ERROR(hipStreamWaitEvent(stream, event, /*flags=*/0));
106 }
107 
mgpuEventCreate()108 extern "C" hipEvent_t mgpuEventCreate() {
109   ScopedContext scopedContext;
110   hipEvent_t event = nullptr;
111   HIP_REPORT_IF_ERROR(hipEventCreateWithFlags(&event, hipEventDisableTiming));
112   return event;
113 }
114 
mgpuEventDestroy(hipEvent_t event)115 extern "C" void mgpuEventDestroy(hipEvent_t event) {
116   HIP_REPORT_IF_ERROR(hipEventDestroy(event));
117 }
118 
mgpuEventSynchronize(hipEvent_t event)119 extern "C" void mgpuEventSynchronize(hipEvent_t event) {
120   HIP_REPORT_IF_ERROR(hipEventSynchronize(event));
121 }
122 
mgpuEventRecord(hipEvent_t event,hipStream_t stream)123 extern "C" void mgpuEventRecord(hipEvent_t event, hipStream_t stream) {
124   HIP_REPORT_IF_ERROR(hipEventRecord(event, stream));
125 }
126 
mgpuMemAlloc(uint64_t sizeBytes,hipStream_t)127 extern "C" void *mgpuMemAlloc(uint64_t sizeBytes, hipStream_t /*stream*/) {
128   ScopedContext scopedContext;
129   void *ptr;
130   HIP_REPORT_IF_ERROR(hipMalloc(&ptr, sizeBytes));
131   return ptr;
132 }
133 
mgpuMemFree(void * ptr,hipStream_t)134 extern "C" void mgpuMemFree(void *ptr, hipStream_t /*stream*/) {
135   HIP_REPORT_IF_ERROR(hipFree(ptr));
136 }
137 
mgpuMemcpy(void * dst,void * src,size_t sizeBytes,hipStream_t stream)138 extern "C" void mgpuMemcpy(void *dst, void *src, size_t sizeBytes,
139                            hipStream_t stream) {
140   HIP_REPORT_IF_ERROR(
141       hipMemcpyAsync(dst, src, sizeBytes, hipMemcpyDefault, stream));
142 }
143 
mgpuMemset32(void * dst,int value,size_t count,hipStream_t stream)144 extern "C" void mgpuMemset32(void *dst, int value, size_t count,
145                              hipStream_t stream) {
146   HIP_REPORT_IF_ERROR(hipMemsetD32Async(reinterpret_cast<hipDeviceptr_t>(dst),
147                                         value, count, stream));
148 }
149 /// Helper functions for writing mlir example code
150 
151 // Allows to register byte array with the ROCM runtime. Helpful until we have
152 // transfer functions implemented.
mgpuMemHostRegister(void * ptr,uint64_t sizeBytes)153 extern "C" void mgpuMemHostRegister(void *ptr, uint64_t sizeBytes) {
154   ScopedContext scopedContext;
155   HIP_REPORT_IF_ERROR(hipHostRegister(ptr, sizeBytes, /*flags=*/0));
156 }
157 
158 // Allows to register a MemRef with the ROCm runtime. Helpful until we have
159 // transfer functions implemented.
160 extern "C" void
mgpuMemHostRegisterMemRef(int64_t rank,StridedMemRefType<char,1> * descriptor,int64_t elementSizeBytes)161 mgpuMemHostRegisterMemRef(int64_t rank, StridedMemRefType<char, 1> *descriptor,
162                           int64_t elementSizeBytes) {
163 
164   llvm::SmallVector<int64_t, 4> denseStrides(rank);
165   llvm::ArrayRef<int64_t> sizes(descriptor->sizes, rank);
166   llvm::ArrayRef<int64_t> strides(sizes.end(), rank);
167 
168   std::partial_sum(sizes.rbegin(), sizes.rend(), denseStrides.rbegin(),
169                    std::multiplies<int64_t>());
170   auto sizeBytes = denseStrides.front() * elementSizeBytes;
171 
172   // Only densely packed tensors are currently supported.
173   std::rotate(denseStrides.begin(), denseStrides.begin() + 1,
174               denseStrides.end());
175   denseStrides.back() = 1;
176   assert(strides == llvm::makeArrayRef(denseStrides));
177 
178   auto ptr = descriptor->data + descriptor->offset * elementSizeBytes;
179   mgpuMemHostRegister(ptr, sizeBytes);
180 }
181 
182 template <typename T>
mgpuMemGetDevicePointer(T * hostPtr,T ** devicePtr)183 void mgpuMemGetDevicePointer(T *hostPtr, T **devicePtr) {
184   HIP_REPORT_IF_ERROR(hipSetDevice(0));
185   HIP_REPORT_IF_ERROR(
186       hipHostGetDevicePointer((void **)devicePtr, hostPtr, /*flags=*/0));
187 }
188 
189 extern "C" StridedMemRefType<float, 1>
mgpuMemGetDeviceMemRef1dFloat(float * allocated,float * aligned,int64_t offset,int64_t size,int64_t stride)190 mgpuMemGetDeviceMemRef1dFloat(float *allocated, float *aligned, int64_t offset,
191                               int64_t size, int64_t stride) {
192   float *devicePtr = nullptr;
193   mgpuMemGetDevicePointer(aligned, &devicePtr);
194   return {devicePtr, devicePtr, offset, {size}, {stride}};
195 }
196 
197 extern "C" StridedMemRefType<int32_t, 1>
mgpuMemGetDeviceMemRef1dInt32(int32_t * allocated,int32_t * aligned,int64_t offset,int64_t size,int64_t stride)198 mgpuMemGetDeviceMemRef1dInt32(int32_t *allocated, int32_t *aligned,
199                               int64_t offset, int64_t size, int64_t stride) {
200   int32_t *devicePtr = nullptr;
201   mgpuMemGetDevicePointer(aligned, &devicePtr);
202   return {devicePtr, devicePtr, offset, {size}, {stride}};
203 }
204 
mgpuSetDefaultDevice(int32_t device)205 extern "C" void mgpuSetDefaultDevice(int32_t device) {
206   defaultDevice = device;
207   HIP_REPORT_IF_ERROR(hipSetDevice(device));
208 }
209