1 /*
2 * Copyright (c) 2018, Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included
12 * in all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
15 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 */
22 //!
23 //! \file cm_ssh.cpp
24 //! \brief Contains Class CmSSH definitions
25 //!
26
27 #include "cm_media_state.h"
28 #include "cm_kernel_ex.h"
29 #include <string>
30 #include <iostream>
31 #include <sstream>
32
33 using namespace CMRT_UMD;
34 using namespace std;
35
CmMediaState(CM_HAL_STATE * cmhal)36 CmMediaState::CmMediaState(CM_HAL_STATE *cmhal):
37 m_cmhal(cmhal),
38 m_heapMgr(nullptr),
39 m_curbeOffsetInternal(0),
40 m_mediaIDOffsetInternal(0),
41 m_samplerHeapOffsetInternal(0),
42 m_scratchSpaceOffsetExternal(0),
43 m_totalCurbeSize(0),
44 m_totalMediaIDSize(0),
45 m_totalSamplerHeapSize(0),
46 m_totalScratchSpaceSize(0),
47 m_mediaIDSize(0),
48 m_scratchSizePerThread(0),
49 m_state(_Empty)
50 {
51 }
52
~CmMediaState()53 CmMediaState::~CmMediaState()
54 {
55 if (m_state == _Allocated)
56 {
57 Submit();
58 }
59 }
60
Initialize(HeapManager * heapMgr)61 MOS_STATUS CmMediaState::Initialize(HeapManager *heapMgr)
62 {
63 if (heapMgr == nullptr || m_cmhal == nullptr)
64 {
65 return MOS_STATUS_NULL_POINTER;
66 }
67 m_heapMgr = heapMgr;
68
69 m_mediaIDSize = m_cmhal->renderHal->pHwSizes->dwSizeInterfaceDescriptor;
70
71 MOS_ZeroMemory(m_curbeOffsets, sizeof(m_curbeOffsets));
72 MOS_ZeroMemory(m_samplerOffsets, sizeof(m_samplerOffsets));
73 MOS_ZeroMemory(m_next3dSamplerOffsets, sizeof(m_next3dSamplerOffsets));
74 MOS_ZeroMemory(m_nextAvsSamplerOffsets, sizeof(m_nextAvsSamplerOffsets));
75 MOS_ZeroMemory(m_nextIndStateOffsets, sizeof(m_nextIndStateOffsets));
76 MOS_ZeroMemory(m_samplerCount, sizeof(m_samplerCount));
77
78 return MOS_STATUS_SUCCESS;
79 }
80
Submit()81 MOS_STATUS CmMediaState::Submit()
82 {
83 std::vector<MemoryBlock> blocks;
84 blocks.push_back(m_memoryBlock);
85 CM_CHK_MOSSTATUS_RETURN(m_heapMgr->SubmitBlocks(blocks));
86 m_state = _Submitted;
87
88 return MOS_STATUS_SUCCESS;
89 }
90
91 #if defined(ANDROID) || defined(LINUX)
92 #define PLATFORM_DIR_SEPERATOR "/"
93 #else
94 #define PLATFORM_DIR_SEPERATOR "\\"
95 #endif
96
Dump()97 void CmMediaState::Dump()
98 {
99 #if MDF_CURBE_DATA_DUMP
100 if (m_cmhal->dumpCurbeData)
101 {
102 char curbeFileNamePrefix[MAX_PATH];
103 char idFileNamePrefix[MAX_PATH];
104 static int fileCount = 0;
105 stringstream curbeFilename;
106 curbeFilename << "HALCM_Curbe_Data_Dumps" << PLATFORM_DIR_SEPERATOR << "curbe_" << fileCount << ".fast.log";
107 stringstream idFilename;
108 idFilename << "HALCM_Curbe_Data_Dumps" << PLATFORM_DIR_SEPERATOR << "id_" << fileCount << ".fast.log";
109
110 ++fileCount;
111
112 GetLogFileLocation(curbeFilename.str().c_str(), curbeFileNamePrefix,
113 m_cmhal->osInterface->pOsContext);
114 GetLogFileLocation(idFilename.str().c_str(), idFileNamePrefix,
115 m_cmhal->osInterface->pOsContext);
116
117 m_memoryBlock.Dump(curbeFileNamePrefix, m_curbeOffsetInternal, m_totalCurbeSize);
118 m_memoryBlock.Dump(idFilename.str(), m_mediaIDOffsetInternal, m_totalMediaIDSize);
119 }
120 #endif
121 }
122
Allocate(CmKernelEx ** kernels,int count,uint32_t trackerIndex,uint32_t trackerID)123 MOS_STATUS CmMediaState::Allocate(CmKernelEx **kernels, int count, uint32_t trackerIndex, uint32_t trackerID)
124 {
125 // calculate the curbe size
126 m_curbeOffsetInternal = 0;
127 uint32_t offset = 0;
128 uint32_t totalCurbeSize = 0;
129 for (int i = 0; i < count; i++)
130 {
131 CmKernelEx *kernel = kernels[i];
132 uint32_t curbeSize = kernel->GetCurbeSize();
133 m_curbeOffsets[i] = totalCurbeSize;
134 totalCurbeSize += curbeSize;
135 }
136 m_totalCurbeSize = totalCurbeSize;
137
138 // calculate the sampler
139 m_samplerHeapOffsetInternal = MOS_ALIGN_CEIL(m_totalCurbeSize, MHW_SAMPLER_STATE_ALIGN);
140 uint32_t totalHeapSize = 0;
141 uint32_t maxSpillSize = 0;
142 for (int i = 0; i < count; i++)
143 {
144 CmKernelEx *kernel = kernels[i];
145 uint32_t heapSize = UpdateHeapSizeAndOffsets(kernel, i);
146 m_samplerOffsets[i] = totalHeapSize;
147 totalHeapSize += heapSize;
148
149 // get the spill size
150 maxSpillSize = MOS_MAX(maxSpillSize, kernel->GetSpillMemUsed());
151 }
152 m_totalSamplerHeapSize = totalHeapSize;
153
154 // calculate the media id
155 m_mediaIDOffsetInternal = m_samplerHeapOffsetInternal + m_totalSamplerHeapSize;
156 m_totalMediaIDSize = count * m_mediaIDSize;
157
158 // caculate the scratch space
159 uint32_t tempScratchOffset = m_mediaIDOffsetInternal + m_totalMediaIDSize;
160 if (maxSpillSize > 0 && (!m_cmhal->cmHalInterface->IsSeparateScratch()))
161 {
162 uint32_t perThreadScratchSpace = 1024;
163 for (perThreadScratchSpace; perThreadScratchSpace < maxSpillSize; perThreadScratchSpace <<= 1);
164
165 // get max thread number
166 MEDIA_SYSTEM_INFO *gtSystemInfo = m_cmhal->osInterface->pfnGetGtSystemInfo(m_cmhal->osInterface);
167 uint32_t numHWThreadsPerEU = gtSystemInfo->ThreadCount / gtSystemInfo->EUCount;
168 uint32_t maxHWThreads = gtSystemInfo->MaxEuPerSubSlice * numHWThreadsPerEU * gtSystemInfo->MaxSubSlicesSupported;
169 // add additional 1k, because the offset of scratch space needs to be 1k aligned
170 m_totalScratchSpaceSize = maxHWThreads * perThreadScratchSpace + MHW_SCRATCH_SPACE_ALIGN;
171 m_scratchSizePerThread = perThreadScratchSpace;
172 // change the extend step in gdsh
173 uint32_t currentExtendSize = m_heapMgr->GetExtendSize();
174 if (currentExtendSize < m_totalScratchSpaceSize)
175 {
176 // update extend size for scratch space
177 m_heapMgr->SetExtendHeapSize(m_totalScratchSpaceSize);
178 }
179 }
180
181 // allocate the memory block
182 uint32_t totalSize = tempScratchOffset + m_totalScratchSpaceSize;
183 PrepareMemoryBlock(totalSize, trackerIndex, trackerID);
184
185 // adjust the offset of scratch space to be 1k aligned
186 if (maxSpillSize > 0 && (!m_cmhal->cmHalInterface->IsSeparateScratch()))
187 {
188 uint32_t scratchOffsetExternal = m_memoryBlock.GetOffset() + tempScratchOffset;
189 m_scratchSpaceOffsetExternal = MOS_ALIGN_CEIL(scratchOffsetExternal, MHW_SCRATCH_SPACE_ALIGN);
190 }
191
192 return MOS_STATUS_SUCCESS;
193 }
194
PrepareMemoryBlock(uint32_t size,uint32_t trackerIndex,uint32_t trackerID)195 MOS_STATUS CmMediaState::PrepareMemoryBlock(uint32_t size, uint32_t trackerIndex, uint32_t trackerID)
196 {
197 uint32_t spaceNeeded = 0;
198 std::vector<MemoryBlock> blocks;
199 std::vector<uint32_t> blockSizes;
200 MemoryBlockManager::AcquireParams acquireParams =
201 MemoryBlockManager::AcquireParams(trackerID, blockSizes);
202 acquireParams.m_trackerIndex = trackerIndex;
203 if (blockSizes.empty())
204 {
205 blockSizes.emplace_back(size);
206 }
207 else
208 {
209 blockSizes[0] = size;
210 }
211
212 m_heapMgr->AcquireSpace(acquireParams, blocks, spaceNeeded);
213
214 if (blocks.empty())
215 {
216 MHW_RENDERHAL_ASSERTMESSAGE("No blocks were acquired");
217 return MOS_STATUS_UNKNOWN;
218 }
219 if (!(blocks[0].IsValid()))
220 {
221 MHW_RENDERHAL_ASSERTMESSAGE("No blocks were acquired");
222 return MOS_STATUS_UNKNOWN;
223 }
224
225 m_memoryBlock = blocks[0];
226
227 // zero memory block
228 m_memoryBlock.AddData(nullptr, 0, 0, true);
229
230 m_state = _Allocated;
231 return MOS_STATUS_SUCCESS;
232 }
233
LoadCurbe(CmKernelEx * kernel,int index)234 MOS_STATUS CmMediaState::LoadCurbe(CmKernelEx *kernel, int index)
235 {
236 return LoadCurbe(kernel->GetCurbe(), kernel->GetCurbeSize(), index);
237 }
238
LoadCurbe(uint8_t * curbe,uint32_t size,int index)239 MOS_STATUS CmMediaState::LoadCurbe(uint8_t *curbe, uint32_t size, int index)
240 {
241 if (m_state != _Allocated)
242 {
243 MHW_RENDERHAL_ASSERTMESSAGE("Media State not allocated yet");
244 return MOS_STATUS_UNKNOWN;
245 }
246
247 m_memoryBlock.AddData(curbe, m_curbeOffsetInternal+m_curbeOffsets[index], size);
248 return MOS_STATUS_SUCCESS;
249 }
250
LoadMediaID(CmKernelEx * kernel,int index,uint32_t btOffset,CmThreadGroupSpace * threadGroupSpace)251 MOS_STATUS CmMediaState::LoadMediaID(CmKernelEx *kernel, int index, uint32_t btOffset, CmThreadGroupSpace *threadGroupSpace)
252 {
253 if (m_state != _Allocated)
254 {
255 MHW_RENDERHAL_ASSERTMESSAGE("Media State not allocated yet");
256 return MOS_STATUS_UNKNOWN;
257 }
258 MHW_ID_ENTRY_PARAMS params;
259 uint32_t mediaStateOffset = m_memoryBlock.GetOffset();
260
261 // Get states, params
262 params.dwMediaIdOffset = mediaStateOffset + m_mediaIDOffsetInternal;
263 params.iMediaId = index;
264 params.dwKernelOffset = kernel->GetOffsetInIsh();
265 params.dwSamplerOffset = GetSamplerHeapOffset(index);
266 params.dwSamplerCount = MOS_MIN(4, (GetSamplerCount(index) + 3 ) / 4);
267 params.dwBindingTableOffset = btOffset;
268 params.iCurbeOffset = m_curbeOffsets[index];
269 params.iCurbeLength = kernel->GetCurbeSizePerThread();
270 if (threadGroupSpace == nullptr)
271 {
272 params.bBarrierEnable = false;
273 params.bGlobalBarrierEnable = false; //It's only applied for BDW+
274 params.dwNumberofThreadsInGPGPUGroup = 1;
275 params.dwSharedLocalMemorySize = 0;
276 params.iCrsThdConDataRdLn = 0;
277 }
278 else
279 {
280 uint32_t threadW = 0;
281 uint32_t threadH = 0;
282 uint32_t threadD = 0;
283 uint32_t groupW = 0;
284 uint32_t groupH = 0;
285 uint32_t groupD = 0;
286 threadGroupSpace->GetThreadGroupSpaceSize(threadW,
287 threadH,
288 threadD,
289 groupW,
290 groupH,
291 groupD);
292 params.bBarrierEnable = (kernel->GetBarrierMode() != CM_NO_BARRIER);
293 params.bGlobalBarrierEnable = (kernel->GetBarrierMode() == CM_GLOBAL_BARRIER);
294 params.dwNumberofThreadsInGPGPUGroup = threadW * threadH * threadD;
295 params.dwSharedLocalMemorySize =
296 m_cmhal->renderHal->pfnEncodeSLMSize(m_cmhal->renderHal, kernel->GetSLMSize());;
297 params.iCrsThdConDataRdLn = kernel->GetCurbeSizeCrossThread();
298 }
299
300 params.memoryBlock = &m_memoryBlock;
301
302 CM_CHK_MOSSTATUS_RETURN(m_cmhal->renderHal->pMhwStateHeap->AddInterfaceDescriptorData(¶ms));
303
304 return MOS_STATUS_SUCCESS;
305 }
306
UpdateHeapSizeAndOffsets(CmKernelEx * kernel,uint32_t kernelIdx)307 uint32_t CmMediaState::UpdateHeapSizeAndOffsets(CmKernelEx *kernel, uint32_t kernelIdx)
308 {
309 uint32_t count3D; // not include the reserved
310 uint32_t countAVS; // not include the reserved
311 uint32_t reservedCount3D = 0;
312 kernel->GetSamplerCount(&count3D, &countAVS);
313 std::map<int, void *>reservedSamplers = kernel->GetReservedSamplerBteIndex();
314
315 m_samplerCount[kernelIdx] = count3D + countAVS + reservedSamplers.size();
316
317 if (m_samplerCount[kernelIdx] ==0)
318 {
319 // no sampler in the kernel
320 return 0;
321 }
322
323 // simplified the sampler allocations
324 // reserved samplers
325 // avs samplers
326 // 3d samplers
327 // 3d indirect states
328
329 // get the area of reserved samplers
330 uint32_t reservedEnd = 0;
331 uint32_t heapSize = 0;
332 for (auto it = reservedSamplers.begin(); it != reservedSamplers.end(); it ++)
333 {
334 int bteIndex = it->first;
335 MHW_SAMPLER_STATE_PARAM *param = (MHW_SAMPLER_STATE_PARAM *)it->second;
336 uint32_t elementSize;
337 if (param->SamplerType == MHW_SAMPLER_TYPE_3D)
338 {
339 reservedCount3D ++;
340 elementSize = m_3dSamplerElementSize;
341 }
342 else
343 {
344 elementSize = m_avsSamplerElementSize;
345 }
346 uint32_t end = (bteIndex + 1) * elementSize;
347 reservedEnd = MOS_MAX(reservedEnd, end);
348 }
349 heapSize = MOS_ALIGN_CEIL(reservedEnd, MHW_SAMPLER_STATE_ALIGN);
350 m_nextAvsSamplerOffsets[kernelIdx] = heapSize;
351 heapSize += countAVS*m_avsSamplerElementSize;
352 m_next3dSamplerOffsets[kernelIdx] = heapSize;
353 heapSize += count3D*m_3dSamplerElementSize;
354 heapSize = MOS_ALIGN_CEIL(heapSize, 1 << MHW_SAMPLER_INDIRECT_SHIFT);
355 m_nextIndStateOffsets[kernelIdx] = heapSize;
356 heapSize += (reservedCount3D + count3D) * m_cmhal->renderHal->pMhwStateHeap->m_HwSizes.dwSizeSamplerIndirectState;
357 heapSize = MOS_ALIGN_CEIL(heapSize, MHW_SAMPLER_STATE_ALIGN);
358
359 return heapSize;
360 }
361
AddSampler(void * samplerParam,int index,int bteIndex)362 int CmMediaState::AddSampler(void *samplerParam, int index, int bteIndex)
363 {
364 uint32_t offset;
365 MHW_SAMPLER_STATE_PARAM *param = (MHW_SAMPLER_STATE_PARAM *)samplerParam;
366 uint32_t elementSize = (param->SamplerType == MHW_SAMPLER_TYPE_3D)?m_3dSamplerElementSize:m_avsSamplerElementSize;
367 if (bteIndex == -1)
368 {
369 uint32_t *nextOffset = (param->SamplerType == MHW_SAMPLER_TYPE_3D)?m_next3dSamplerOffsets:m_nextAvsSamplerOffsets;
370 offset = nextOffset[index];
371 nextOffset[index] += elementSize;
372 }
373 else
374 {
375 offset = bteIndex * elementSize;
376 }
377
378 if (param->SamplerType == MHW_SAMPLER_TYPE_3D)
379 {
380 param->Unorm.IndirectStateOffset = m_samplerHeapOffsetInternal + m_samplerOffsets[index] + m_nextIndStateOffsets[index];
381 m_nextIndStateOffsets[index] += m_cmhal->renderHal->pMhwStateHeap->m_HwSizes.dwSizeSamplerIndirectState;
382 }
383
384 uint32_t heapOffset = m_samplerHeapOffsetInternal + m_samplerOffsets[index] + offset;
385 MOS_STATUS ret = m_cmhal->renderHal->pMhwStateHeap->AddSamplerStateData(heapOffset, &m_memoryBlock, param);
386 if (ret != MOS_STATUS_SUCCESS)
387 {
388 return -1;
389 }
390 return offset/elementSize;
391 }
392