rsovAllocation.cpp revision 30dfb1368d0a3440cfedc882c27cc236d799f77c 
1/*
2 * Copyright (C) 2016 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 *      http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17#include "rsovAllocation.h"
18
19#include <map>
20
21#include "rsAllocation.h"
22#include "rsContext.h"
23#include "rsCppUtils.h"
24#include "rsElement.h"
25#include "rsType.h"
26#include "rsovContext.h"
27#include "rsovCore.h"
28
29namespace android {
30namespace renderscript {
31namespace rsov {
32
33namespace {
34
35using std::make_pair;
36
37// TODO: handle 8-bit, 16-bit, and 64-bit integers and floating point numbers
38const std::map<std::pair<RsDataType, uint32_t>, VkFormat> mapElementToFormat{
39    make_pair(make_pair(RS_TYPE_FLOAT_32, 1), VK_FORMAT_R32_SFLOAT),
40    make_pair(make_pair(RS_TYPE_FLOAT_32, 2), VK_FORMAT_R32G32_SFLOAT),
41    make_pair(make_pair(RS_TYPE_FLOAT_32, 3), VK_FORMAT_R32G32B32_SFLOAT),
42    make_pair(make_pair(RS_TYPE_FLOAT_32, 4), VK_FORMAT_R32G32B32A32_SFLOAT),
43
44    make_pair(make_pair(RS_TYPE_SIGNED_32, 1), VK_FORMAT_R32_SINT),
45    make_pair(make_pair(RS_TYPE_SIGNED_32, 2), VK_FORMAT_R32G32_SINT),
46    make_pair(make_pair(RS_TYPE_SIGNED_32, 3), VK_FORMAT_R32G32B32_SINT),
47    make_pair(make_pair(RS_TYPE_SIGNED_32, 4), VK_FORMAT_R32G32B32A32_SINT),
48
49    make_pair(make_pair(RS_TYPE_UNSIGNED_32, 1), VK_FORMAT_R32_UINT),
50    make_pair(make_pair(RS_TYPE_UNSIGNED_32, 2), VK_FORMAT_R32G32_UINT),
51    make_pair(make_pair(RS_TYPE_UNSIGNED_32, 3), VK_FORMAT_R32G32B32_UINT),
52    make_pair(make_pair(RS_TYPE_UNSIGNED_32, 4), VK_FORMAT_R32G32B32A32_UINT),
53};
54
55VkFormat VkFormatFromRSElement(const Element &elem) {
56  // TODO: reject struct, allocation, and other non-numeric element
57  rsAssert(!elem.getFieldCount());
58
59  RsDataType dataType = elem.getType();
60  uint32_t vectorWidth = elem.getVectorSize();
61
62  auto it = mapElementToFormat.find(make_pair(dataType, vectorWidth));
63  if (it != mapElementToFormat.end()) {
64    return it->second;
65  }
66
67  rsAssert(0 && "Unexpected RS Element to map to VkFormat");
68
69  return VK_FORMAT_R32G32B32A32_SFLOAT;
70}
71
72size_t DeriveYUVLayout(int yuv, Allocation::Hal::DrvState *state) {
73  // For the flexible YCbCr format, layout is initialized during call to
74  // Allocation::ioReceive.  Return early and avoid clobberring any
75  // pre-existing layout.
76  if (yuv == HAL_PIXEL_FORMAT_YCbCr_420_888) {
77    return 0;
78  }
79
80  // YUV only supports basic 2d
81  // so we can stash the plane pointers in the mipmap levels.
82  size_t uvSize = 0;
83  state->lod[1].dimX = state->lod[0].dimX / 2;
84  state->lod[1].dimY = state->lod[0].dimY / 2;
85  state->lod[2].dimX = state->lod[0].dimX / 2;
86  state->lod[2].dimY = state->lod[0].dimY / 2;
87  state->yuv.shift = 1;
88  state->yuv.step = 1;
89  state->lodCount = 3;
90
91  switch (yuv) {
92    case HAL_PIXEL_FORMAT_YV12:
93      state->lod[2].stride = rsRound(state->lod[0].stride >> 1, 16);
94      state->lod[2].mallocPtr = ((uint8_t *)state->lod[0].mallocPtr) +
95                                (state->lod[0].stride * state->lod[0].dimY);
96      uvSize += state->lod[2].stride * state->lod[2].dimY;
97
98      state->lod[1].stride = state->lod[2].stride;
99      state->lod[1].mallocPtr = ((uint8_t *)state->lod[2].mallocPtr) +
100                                (state->lod[2].stride * state->lod[2].dimY);
101      uvSize += state->lod[1].stride * state->lod[2].dimY;
102      break;
103    case HAL_PIXEL_FORMAT_YCrCb_420_SP:  // NV21
104      // state->lod[1].dimX = state->lod[0].dimX;
105      state->lod[1].stride = state->lod[0].stride;
106      state->lod[2].stride = state->lod[0].stride;
107      state->lod[2].mallocPtr = ((uint8_t *)state->lod[0].mallocPtr) +
108                                (state->lod[0].stride * state->lod[0].dimY);
109      state->lod[1].mallocPtr = ((uint8_t *)state->lod[2].mallocPtr) + 1;
110      uvSize += state->lod[1].stride * state->lod[1].dimY;
111      state->yuv.step = 2;
112      break;
113    default:
114      rsAssert(0);
115  }
116
117  return uvSize;
118}
119
120// TODO: Dedup this with the same code under frameworks/rs/driver
121size_t AllocationBuildPointerTable(const Context *rsc, const Allocation *alloc,
122                                   const Type *type, uint8_t *ptr,
123                                   size_t requiredAlignment) {
124  alloc->mHal.drvState.lod[0].dimX = type->getDimX();
125  alloc->mHal.drvState.lod[0].dimY = type->getDimY();
126  alloc->mHal.drvState.lod[0].dimZ = type->getDimZ();
127  alloc->mHal.drvState.lod[0].mallocPtr = 0;
128  // Stride needs to be aligned to a boundary defined by requiredAlignment!
129  size_t stride =
130      alloc->mHal.drvState.lod[0].dimX * type->getElementSizeBytes();
131  alloc->mHal.drvState.lod[0].stride = rsRound(stride, requiredAlignment);
132  alloc->mHal.drvState.lodCount = type->getLODCount();
133  alloc->mHal.drvState.faceCount = type->getDimFaces();
134
135  size_t offsets[Allocation::MAX_LOD];
136  memset(offsets, 0, sizeof(offsets));
137
138  size_t o = alloc->mHal.drvState.lod[0].stride *
139             rsMax(alloc->mHal.drvState.lod[0].dimY, 1u) *
140             rsMax(alloc->mHal.drvState.lod[0].dimZ, 1u);
141  if (alloc->mHal.state.yuv) {
142    o += DeriveYUVLayout(alloc->mHal.state.yuv, &alloc->mHal.drvState);
143
144    for (uint32_t ct = 1; ct < alloc->mHal.drvState.lodCount; ct++) {
145      offsets[ct] = (size_t)alloc->mHal.drvState.lod[ct].mallocPtr;
146    }
147  } else if (alloc->mHal.drvState.lodCount > 1) {
148    uint32_t tx = alloc->mHal.drvState.lod[0].dimX;
149    uint32_t ty = alloc->mHal.drvState.lod[0].dimY;
150    uint32_t tz = alloc->mHal.drvState.lod[0].dimZ;
151    for (uint32_t lod = 1; lod < alloc->mHal.drvState.lodCount; lod++) {
152      alloc->mHal.drvState.lod[lod].dimX = tx;
153      alloc->mHal.drvState.lod[lod].dimY = ty;
154      alloc->mHal.drvState.lod[lod].dimZ = tz;
155      alloc->mHal.drvState.lod[lod].stride =
156          rsRound(tx * type->getElementSizeBytes(), requiredAlignment);
157      offsets[lod] = o;
158      o += alloc->mHal.drvState.lod[lod].stride * rsMax(ty, 1u) * rsMax(tz, 1u);
159      if (tx > 1) tx >>= 1;
160      if (ty > 1) ty >>= 1;
161      if (tz > 1) tz >>= 1;
162    }
163  }
164
165  alloc->mHal.drvState.faceOffset = o;
166
167  alloc->mHal.drvState.lod[0].mallocPtr = ptr;
168  for (uint32_t lod = 1; lod < alloc->mHal.drvState.lodCount; lod++) {
169    alloc->mHal.drvState.lod[lod].mallocPtr = ptr + offsets[lod];
170  }
171
172  size_t allocSize = alloc->mHal.drvState.faceOffset;
173  if (alloc->mHal.drvState.faceCount) {
174    allocSize *= 6;
175  }
176
177  return allocSize;
178}
179
180size_t AllocationBuildPointerTable(const Context *rsc, const Allocation *alloc,
181                                   const Type *type, uint8_t *ptr) {
182  return AllocationBuildPointerTable(rsc, alloc, type, ptr,
183                                     Allocation::kMinimumRSAlignment);
184}
185
186uint8_t *GetOffsetPtr(const Allocation *alloc, uint32_t xoff, uint32_t yoff,
187                      uint32_t zoff, uint32_t lod,
188                      RsAllocationCubemapFace face) {
189  uint8_t *ptr = (uint8_t *)alloc->mHal.drvState.lod[lod].mallocPtr;
190  ptr += face * alloc->mHal.drvState.faceOffset;
191  ptr += zoff * alloc->mHal.drvState.lod[lod].dimY *
192         alloc->mHal.drvState.lod[lod].stride;
193  ptr += yoff * alloc->mHal.drvState.lod[lod].stride;
194  ptr += xoff * alloc->mHal.state.elementSizeBytes;
195  return ptr;
196}
197
198void mip565(const Allocation *alloc, int lod, RsAllocationCubemapFace face) {
199  uint32_t w = alloc->mHal.drvState.lod[lod + 1].dimX;
200  uint32_t h = alloc->mHal.drvState.lod[lod + 1].dimY;
201
202  for (uint32_t y = 0; y < h; y++) {
203    uint16_t *oPtr = (uint16_t *)GetOffsetPtr(alloc, 0, y, 0, lod + 1, face);
204    const uint16_t *i1 =
205        (uint16_t *)GetOffsetPtr(alloc, 0, 0, y * 2, lod, face);
206    const uint16_t *i2 =
207        (uint16_t *)GetOffsetPtr(alloc, 0, 0, y * 2 + 1, lod, face);
208
209    for (uint32_t x = 0; x < w; x++) {
210      *oPtr = rsBoxFilter565(i1[0], i1[1], i2[0], i2[1]);
211      oPtr++;
212      i1 += 2;
213      i2 += 2;
214    }
215  }
216}
217
218void mip8888(const Allocation *alloc, int lod, RsAllocationCubemapFace face) {
219  uint32_t w = alloc->mHal.drvState.lod[lod + 1].dimX;
220  uint32_t h = alloc->mHal.drvState.lod[lod + 1].dimY;
221
222  for (uint32_t y = 0; y < h; y++) {
223    uint32_t *oPtr = (uint32_t *)GetOffsetPtr(alloc, 0, y, 0, lod + 1, face);
224    const uint32_t *i1 =
225        (uint32_t *)GetOffsetPtr(alloc, 0, y * 2, 0, lod, face);
226    const uint32_t *i2 =
227        (uint32_t *)GetOffsetPtr(alloc, 0, y * 2 + 1, 0, lod, face);
228
229    for (uint32_t x = 0; x < w; x++) {
230      *oPtr = rsBoxFilter8888(i1[0], i1[1], i2[0], i2[1]);
231      oPtr++;
232      i1 += 2;
233      i2 += 2;
234    }
235  }
236}
237
238void mip8(const Allocation *alloc, int lod, RsAllocationCubemapFace face) {
239  uint32_t w = alloc->mHal.drvState.lod[lod + 1].dimX;
240  uint32_t h = alloc->mHal.drvState.lod[lod + 1].dimY;
241
242  for (uint32_t y = 0; y < h; y++) {
243    uint8_t *oPtr = GetOffsetPtr(alloc, 0, y, 0, lod + 1, face);
244    const uint8_t *i1 = GetOffsetPtr(alloc, 0, y * 2, 0, lod, face);
245    const uint8_t *i2 = GetOffsetPtr(alloc, 0, y * 2 + 1, 0, lod, face);
246
247    for (uint32_t x = 0; x < w; x++) {
248      *oPtr = (uint8_t)(((uint32_t)i1[0] + i1[1] + i2[0] + i2[1]) * 0.25f);
249      oPtr++;
250      i1 += 2;
251      i2 += 2;
252    }
253  }
254}
255
256}  // anonymous namespace
257
258RSoVAllocation::RSoVAllocation(RSoVContext *context, const Type *type,
259                               size_t size)
260    : mRSoV(context),
261      mDevice(context->getDevice()),
262      mType(type),
263      mWidth(type->getDimX()),
264      mHeight(type->getDimY()),
265      mDepth(type->getDimZ()) {
266  InitBuffer(size);
267}
268
269RSoVAllocation::~RSoVAllocation() {
270  vkDestroyBuffer(mDevice, mBuf, nullptr);
271  vkFreeMemory(mDevice, mMem, nullptr);
272}
273
274void RSoVAllocation::InitBuffer(size_t bufferSize) {
275  VkResult res;
276
277  VkBufferCreateInfo buf_info = {
278      .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
279      .pNext = nullptr,
280      .usage = VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
281      .size = bufferSize,
282      .queueFamilyIndexCount = 0,
283      .pQueueFamilyIndices = nullptr,
284      .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
285      .flags = 0,
286  };
287  res = vkCreateBuffer(mDevice, &buf_info, nullptr, &mBuf);
288  rsAssert(res == VK_SUCCESS);
289
290  VkMemoryRequirements mem_reqs;
291  vkGetBufferMemoryRequirements(mDevice, mBuf, &mem_reqs);
292
293  VkMemoryAllocateInfo allocateInfo = {
294      .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
295      .pNext = nullptr,
296      .memoryTypeIndex = 0,
297      .allocationSize = mem_reqs.size,
298  };
299
300  bool pass;
301  pass = mRSoV->MemoryTypeFromProperties(
302      mem_reqs.memoryTypeBits, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
303                                   VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
304      &allocateInfo.memoryTypeIndex);
305  ALOGV("TypeBits = 0x%08X", mem_reqs.memoryTypeBits);
306  rsAssert(pass);
307
308  // TODO: Make this aligned
309  res = vkAllocateMemory(mDevice, &allocateInfo, nullptr, &mMem);
310  rsAssert(res == VK_SUCCESS);
311
312  res = vkBindBufferMemory(mDevice, mBuf, mMem, 0);
313  rsAssert(res == VK_SUCCESS);
314
315  mBufferInfo.buffer = mBuf;
316  mBufferInfo.offset = 0;
317  mBufferInfo.range = bufferSize;
318
319  res = vkMapMemory(mDevice, mMem, 0, mem_reqs.size, 0, (void **)&mPtr);
320  rsAssert(res == VK_SUCCESS);
321}
322
323void RSoVAllocation::InitImage() {
324  VkResult res;
325
326  mFormat = VkFormatFromRSElement(*mType->getElement());
327
328  const uint32_t width = mWidth;
329  const uint32_t height = mHeight;
330  const uint32_t depth = mDepth;
331
332  VkImageType imageType =
333      depth > 0 ? VK_IMAGE_TYPE_3D
334                : (height > 0 ? VK_IMAGE_TYPE_2D : VK_IMAGE_TYPE_1D);
335
336  VkImageCreateInfo createInfo = {
337      .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
338      .pNext = nullptr,
339      .flags = 0,
340      .imageType = imageType,
341      .format = mFormat,
342      .extent = {width, rsMax(height, 1U), rsMax(depth, 1U)},
343      .mipLevels = 1,
344      .arrayLayers = 1,
345      .samples = VK_SAMPLE_COUNT_1_BIT,
346      .tiling = VK_IMAGE_TILING_LINEAR,
347      .usage = VK_IMAGE_USAGE_STORAGE_BIT,
348      .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
349      .queueFamilyIndexCount = 0,
350      .pQueueFamilyIndices = nullptr,
351      .initialLayout = VK_IMAGE_LAYOUT_PREINITIALIZED,
352  };
353
354  res = vkCreateImage(mDevice, &createInfo, nullptr, &mImage);
355  rsAssert(res == VK_SUCCESS);
356
357  VkMemoryRequirements mem_reqs;
358  vkGetImageMemoryRequirements(mDevice, mImage, &mem_reqs);
359
360  ALOGI("size of memory needed = %u", (uint)mem_reqs.size);
361
362  VkMemoryAllocateInfo allocateInfo = {
363      .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
364      .pNext = nullptr,
365      .allocationSize = mem_reqs.size,
366  };
367
368  /* Use the memory properties to determine the type of memory required */
369  bool pass;
370  pass = mRSoV->MemoryTypeFromProperties(
371      mem_reqs.memoryTypeBits, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
372                                   VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
373      &allocateInfo.memoryTypeIndex);
374  ALOGI("TypeBits = 0x%08X", mem_reqs.memoryTypeBits);
375  rsAssert(pass);
376
377  // TODO: Make this aligned
378  res = vkAllocateMemory(mDevice, &allocateInfo, nullptr, &mMem);
379  rsAssert(res == VK_SUCCESS);
380
381  res = vkBindImageMemory(mDevice, mImage, mMem, 0);
382  rsAssert(res == VK_SUCCESS);
383
384  VkImageViewType viewType =
385      depth > 0 ? VK_IMAGE_VIEW_TYPE_3D
386                : (height > 0 ? VK_IMAGE_VIEW_TYPE_2D : VK_IMAGE_VIEW_TYPE_1D);
387
388  VkImageViewCreateInfo view_info = {
389      .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
390      .pNext = nullptr,
391      .image = mImage,
392      .viewType = viewType,
393      .format = mFormat,
394      .components =
395          {
396              .r = VK_COMPONENT_SWIZZLE_IDENTITY,
397              .g = VK_COMPONENT_SWIZZLE_IDENTITY,
398              .b = VK_COMPONENT_SWIZZLE_IDENTITY,
399              .a = VK_COMPONENT_SWIZZLE_IDENTITY,
400          },
401      .subresourceRange =
402          {
403              .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
404              .baseMipLevel = 0,
405              .levelCount = 1,
406              .baseArrayLayer = 0,
407              .layerCount = 1,
408          },
409  };
410
411  res = vkCreateImageView(mDevice, &view_info, nullptr, &mImageView);
412  rsAssert(res == VK_SUCCESS);
413
414  mImageLayout = VK_IMAGE_LAYOUT_GENERAL;
415
416  mImageInfo = {
417      .imageView = mImageView, .imageLayout = mImageLayout,
418  };
419
420  res = vkMapMemory(mDevice, mMem, 0, mem_reqs.size, 0, (void **)&mPtr);
421  rsAssert(res == VK_SUCCESS);
422}
423
424}  // namespace rsov
425}  // namespace renderscript
426}  // namespace android
427
428using android::renderscript::Allocation;
429using android::renderscript::Context;
430using android::renderscript::Element;
431using android::renderscript::Type;
432using android::renderscript::rs_allocation;
433using android::renderscript::rsMax;
434using namespace android::renderscript::rsov;
435
436bool rsovAllocationInit(const Context *rsc, Allocation *alloc, bool forceZero) {
437  RSoVHal *hal = static_cast<RSoVHal *>(rsc->mHal.drv);
438  RSoVContext *rsov = hal->mRSoV;
439  const Type *type = alloc->getType();
440
441  // Calculate the object size.
442  size_t allocSize = AllocationBuildPointerTable(rsc, alloc, type, nullptr);
443  RSoVAllocation *rsovAlloc = new RSoVAllocation(rsov, type, allocSize);
444  alloc->mHal.drv = rsovAlloc;
445  AllocationBuildPointerTable(rsc, alloc, type,
446                              (uint8_t *)rsovAlloc->getHostPtr());
447  return true;
448}
449
450void rsovAllocationDestroy(const Context *rsc, Allocation *alloc) {
451  RSoVAllocation *rsovAlloc = static_cast<RSoVAllocation *>(alloc->mHal.drv);
452  delete rsovAlloc;
453  alloc->mHal.drv = nullptr;
454}
455
456void rsovAllocationData1D(const Context *rsc, const Allocation *alloc,
457                          uint32_t xoff, uint32_t lod, size_t count,
458                          const void *data, size_t sizeBytes) {
459  const size_t eSize = alloc->mHal.state.type->getElementSizeBytes();
460  uint8_t *ptr =
461      GetOffsetPtr(alloc, xoff, 0, 0, 0, RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X);
462  size_t size = count * eSize;
463  if (ptr != data) {
464    // Skip the copy if we are the same allocation. This can arise from
465    // our Bitmap optimization, where we share the same storage.
466    if (alloc->mHal.state.hasReferences) {
467      alloc->incRefs(data, count);
468      alloc->decRefs(ptr, count);
469    }
470    memcpy(ptr, data, size);
471  }
472}
473
474void rsovAllocationData2D(const Context *rsc, const Allocation *alloc,
475                          uint32_t xoff, uint32_t yoff, uint32_t lod,
476                          RsAllocationCubemapFace face, uint32_t w, uint32_t h,
477                          const void *data, size_t sizeBytes, size_t stride) {
478  size_t eSize = alloc->mHal.state.elementSizeBytes;
479  size_t lineSize = eSize * w;
480  if (!stride) {
481    stride = lineSize;
482  }
483
484  if (alloc->mHal.drvState.lod[0].mallocPtr) {
485    const uint8_t *src = static_cast<const uint8_t *>(data);
486    uint8_t *dst = GetOffsetPtr(alloc, xoff, yoff, 0, lod, face);
487
488    for (uint32_t line = yoff; line < (yoff + h); line++) {
489      if (alloc->mHal.state.hasReferences) {
490        alloc->incRefs(src, w);
491        alloc->decRefs(dst, w);
492      }
493      memcpy(dst, src, lineSize);
494      src += stride;
495      dst += alloc->mHal.drvState.lod[lod].stride;
496    }
497    // TODO: handle YUV Allocations
498    if (alloc->mHal.state.yuv) {
499      size_t clineSize = lineSize;
500      int lod = 1;
501      int maxLod = 2;
502      if (alloc->mHal.state.yuv == HAL_PIXEL_FORMAT_YV12) {
503        maxLod = 3;
504        clineSize >>= 1;
505      } else if (alloc->mHal.state.yuv == HAL_PIXEL_FORMAT_YCrCb_420_SP) {
506        lod = 2;
507        maxLod = 3;
508      }
509
510      while (lod < maxLod) {
511        uint8_t *dst = GetOffsetPtr(alloc, xoff, yoff, 0, lod, face);
512
513        for (uint32_t line = (yoff >> 1); line < ((yoff + h) >> 1); line++) {
514          memcpy(dst, src, clineSize);
515          // When copying from an array to an Allocation, the src pointer
516          // to the array should just move by the number of bytes copied.
517          src += clineSize;
518          dst += alloc->mHal.drvState.lod[lod].stride;
519        }
520        lod++;
521      }
522    }
523  }
524}
525
526void rsovAllocationData3D(const Context *rsc, const Allocation *alloc,
527                          uint32_t xoff, uint32_t yoff, uint32_t zoff,
528                          uint32_t lod, uint32_t w, uint32_t h, uint32_t d,
529                          const void *data, size_t sizeBytes, size_t stride) {
530  uint32_t eSize = alloc->mHal.state.elementSizeBytes;
531  uint32_t lineSize = eSize * w;
532  if (!stride) {
533    stride = lineSize;
534  }
535
536  if (alloc->mHal.drvState.lod[0].mallocPtr) {
537    const uint8_t *src = static_cast<const uint8_t *>(data);
538    for (uint32_t z = zoff; z < (d + zoff); z++) {
539      uint8_t *dst = GetOffsetPtr(alloc, xoff, yoff, z, lod,
540                                  RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X);
541      for (uint32_t line = yoff; line < (yoff + h); line++) {
542        if (alloc->mHal.state.hasReferences) {
543          alloc->incRefs(src, w);
544          alloc->decRefs(dst, w);
545        }
546        memcpy(dst, src, lineSize);
547        src += stride;
548        dst += alloc->mHal.drvState.lod[lod].stride;
549      }
550    }
551  }
552}
553
554void rsovAllocationRead1D(const Context *rsc, const Allocation *alloc,
555                          uint32_t xoff, uint32_t lod, size_t count, void *data,
556                          size_t sizeBytes) {
557  const size_t eSize = alloc->mHal.state.type->getElementSizeBytes();
558  const uint8_t *ptr =
559      GetOffsetPtr(alloc, xoff, 0, 0, 0, RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X);
560  if (data != ptr) {
561    // Skip the copy if we are the same allocation. This can arise from
562    // our Bitmap optimization, where we share the same storage.
563    memcpy(data, ptr, count * eSize);
564  }
565}
566
567void rsovAllocationRead2D(const Context *rsc, const Allocation *alloc,
568                          uint32_t xoff, uint32_t yoff, uint32_t lod,
569                          RsAllocationCubemapFace face, uint32_t w, uint32_t h,
570                          void *data, size_t sizeBytes, size_t stride) {
571  size_t eSize = alloc->mHal.state.elementSizeBytes;
572  size_t lineSize = eSize * w;
573  if (!stride) {
574    stride = lineSize;
575  }
576
577  if (alloc->mHal.drvState.lod[0].mallocPtr) {
578    uint8_t *dst = static_cast<uint8_t *>(data);
579    const uint8_t *src = GetOffsetPtr(alloc, xoff, yoff, 0, lod, face);
580    if (dst == src) {
581      // Skip the copy if we are the same allocation. This can arise from
582      // our Bitmap optimization, where we share the same storage.
583      return;
584    }
585
586    for (uint32_t line = yoff; line < (yoff + h); line++) {
587      memcpy(dst, src, lineSize);
588      dst += stride;
589      src += alloc->mHal.drvState.lod[lod].stride;
590    }
591  } else {
592    ALOGE("Add code to readback from non-script memory");
593  }
594}
595
596void rsovAllocationRead3D(const Context *rsc, const Allocation *alloc,
597                          uint32_t xoff, uint32_t yoff, uint32_t zoff,
598                          uint32_t lod, uint32_t w, uint32_t h, uint32_t d,
599                          void *data, size_t sizeBytes, size_t stride) {
600  uint32_t eSize = alloc->mHal.state.elementSizeBytes;
601  uint32_t lineSize = eSize * w;
602  if (!stride) {
603    stride = lineSize;
604  }
605
606  if (alloc->mHal.drvState.lod[0].mallocPtr) {
607    uint8_t *dst = static_cast<uint8_t *>(data);
608    for (uint32_t z = zoff; z < (d + zoff); z++) {
609      const uint8_t *src = GetOffsetPtr(alloc, xoff, yoff, z, lod,
610                                        RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X);
611      if (dst == src) {
612        // Skip the copy if we are the same allocation. This can arise from
613        // our Bitmap optimization, where we share the same storage.
614        return;
615      }
616
617      for (uint32_t line = yoff; line < (yoff + h); line++) {
618        memcpy(dst, src, lineSize);
619        dst += stride;
620        src += alloc->mHal.drvState.lod[lod].stride;
621      }
622    }
623  }
624}
625
626void *rsovAllocationLock1D(const Context *rsc, const Allocation *alloc) {
627  return alloc->mHal.drvState.lod[0].mallocPtr;
628}
629
630void rsovAllocationUnlock1D(const Context *rsc, const Allocation *alloc) {}
631
632void rsovAllocationData1D_alloc(const Context *rsc, const Allocation *dstAlloc,
633                                uint32_t dstXoff, uint32_t dstLod, size_t count,
634                                const Allocation *srcAlloc, uint32_t srcXoff,
635                                uint32_t srcLod) {}
636
637void rsovAllocationData2D_alloc_script(
638    const Context *rsc, const Allocation *dstAlloc, uint32_t dstXoff,
639    uint32_t dstYoff, uint32_t dstLod, RsAllocationCubemapFace dstFace,
640    uint32_t w, uint32_t h, const Allocation *srcAlloc, uint32_t srcXoff,
641    uint32_t srcYoff, uint32_t srcLod, RsAllocationCubemapFace srcFace) {
642  size_t elementSize = dstAlloc->getType()->getElementSizeBytes();
643  for (uint32_t i = 0; i < h; i++) {
644    uint8_t *dstPtr =
645        GetOffsetPtr(dstAlloc, dstXoff, dstYoff + i, 0, dstLod, dstFace);
646    uint8_t *srcPtr =
647        GetOffsetPtr(srcAlloc, srcXoff, srcYoff + i, 0, srcLod, srcFace);
648    memcpy(dstPtr, srcPtr, w * elementSize);
649  }
650}
651
652void rsovAllocationData3D_alloc_script(
653    const Context *rsc, const Allocation *dstAlloc, uint32_t dstXoff,
654    uint32_t dstYoff, uint32_t dstZoff, uint32_t dstLod, uint32_t w, uint32_t h,
655    uint32_t d, const Allocation *srcAlloc, uint32_t srcXoff, uint32_t srcYoff,
656    uint32_t srcZoff, uint32_t srcLod) {
657  uint32_t elementSize = dstAlloc->getType()->getElementSizeBytes();
658  for (uint32_t j = 0; j < d; j++) {
659    for (uint32_t i = 0; i < h; i++) {
660      uint8_t *dstPtr =
661          GetOffsetPtr(dstAlloc, dstXoff, dstYoff + i, dstZoff + j, dstLod,
662                       RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X);
663      uint8_t *srcPtr =
664          GetOffsetPtr(srcAlloc, srcXoff, srcYoff + i, srcZoff + j, srcLod,
665                       RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X);
666      memcpy(dstPtr, srcPtr, w * elementSize);
667    }
668  }
669}
670
671void rsovAllocationData2D_alloc(
672    const Context *rsc, const Allocation *dstAlloc, uint32_t dstXoff,
673    uint32_t dstYoff, uint32_t dstLod, RsAllocationCubemapFace dstFace,
674    uint32_t w, uint32_t h, const Allocation *srcAlloc, uint32_t srcXoff,
675    uint32_t srcYoff, uint32_t srcLod, RsAllocationCubemapFace srcFace) {
676  if (!dstAlloc->getIsScript() && !srcAlloc->getIsScript()) {
677    rsc->setError(RS_ERROR_FATAL_DRIVER,
678                  "Non-script allocation copies not "
679                  "yet implemented.");
680    return;
681  }
682  rsovAllocationData2D_alloc_script(rsc, dstAlloc, dstXoff, dstYoff, dstLod,
683                                    dstFace, w, h, srcAlloc, srcXoff, srcYoff,
684                                    srcLod, srcFace);
685}
686
687void rsovAllocationData3D_alloc(const Context *rsc, const Allocation *dstAlloc,
688                                uint32_t dstXoff, uint32_t dstYoff,
689                                uint32_t dstZoff, uint32_t dstLod, uint32_t w,
690                                uint32_t h, uint32_t d,
691                                const Allocation *srcAlloc, uint32_t srcXoff,
692                                uint32_t srcYoff, uint32_t srcZoff,
693                                uint32_t srcLod) {
694  if (!dstAlloc->getIsScript() && !srcAlloc->getIsScript()) {
695    rsc->setError(RS_ERROR_FATAL_DRIVER,
696                  "Non-script allocation copies not "
697                  "yet implemented.");
698    return;
699  }
700  rsovAllocationData3D_alloc_script(rsc, dstAlloc, dstXoff, dstYoff, dstZoff,
701                                    dstLod, w, h, d, srcAlloc, srcXoff, srcYoff,
702                                    srcZoff, srcLod);
703}
704
705void rsovAllocationAdapterOffset(const Context *rsc, const Allocation *alloc) {
706  // Get a base pointer to the new LOD
707  const Allocation *base = alloc->mHal.state.baseAlloc;
708  const Type *type = alloc->mHal.state.type;
709  if (base == nullptr) {
710    return;
711  }
712
713  const int lodBias = alloc->mHal.state.originLOD;
714  uint32_t lodCount = rsMax(alloc->mHal.drvState.lodCount, (uint32_t)1);
715  for (uint32_t lod = 0; lod < lodCount; lod++) {
716    alloc->mHal.drvState.lod[lod] = base->mHal.drvState.lod[lod + lodBias];
717    alloc->mHal.drvState.lod[lod].mallocPtr = GetOffsetPtr(
718        alloc, alloc->mHal.state.originX, alloc->mHal.state.originY,
719        alloc->mHal.state.originZ, lodBias,
720        (RsAllocationCubemapFace)alloc->mHal.state.originFace);
721  }
722}
723
724bool rsovAllocationAdapterInit(const Context *rsc, Allocation *alloc) {
725// TODO: may need a RSoV Allocation here
726#if 0
727    DrvAllocation *drv = (DrvAllocation *)calloc(1, sizeof(DrvAllocation));
728    if (!drv) {
729        return false;
730    }
731    alloc->mHal.drv = drv;
732#endif
733  // We need to build an allocation that looks like a subset of the parent
734  // allocation
735  rsovAllocationAdapterOffset(rsc, alloc);
736
737  return true;
738}
739
740void rsovAllocationSyncAll(const Context *rsc, const Allocation *alloc,
741                           RsAllocationUsageType src) {
742  // TODO: anything to do here?
743}
744
745void rsovAllocationMarkDirty(const Context *rsc, const Allocation *alloc) {
746  // TODO: anything to do here?
747}
748
749void rsovAllocationResize(const Context *rsc, const Allocation *alloc,
750                          const Type *newType, bool zeroNew) {
751  // TODO: implement this
752  // can this be done without copying, if the new size is greater than the
753  // original?
754}
755
756void rsovAllocationGenerateMipmaps(const Context *rsc,
757                                   const Allocation *alloc) {
758  if (!alloc->mHal.drvState.lod[0].mallocPtr) {
759    return;
760  }
761  uint32_t numFaces = alloc->getType()->getDimFaces() ? 6 : 1;
762  for (uint32_t face = 0; face < numFaces; face++) {
763    for (uint32_t lod = 0; lod < (alloc->getType()->getLODCount() - 1); lod++) {
764      switch (alloc->getType()->getElement()->getSizeBits()) {
765        case 32:
766          mip8888(alloc, lod, (RsAllocationCubemapFace)face);
767          break;
768        case 16:
769          mip565(alloc, lod, (RsAllocationCubemapFace)face);
770          break;
771        case 8:
772          mip8(alloc, lod, (RsAllocationCubemapFace)face);
773          break;
774      }
775    }
776  }
777}
778
779uint32_t rsovAllocationGrallocBits(const Context *rsc, Allocation *alloc) {
780  return 0;
781}
782
783void rsovAllocationUpdateCachedObject(const Context *rsc,
784                                      const Allocation *alloc,
785                                      rs_allocation *obj) {
786  obj->p = alloc;
787#ifdef __LP64__
788  if (alloc != nullptr) {
789    obj->r = alloc->mHal.drvState.lod[0].mallocPtr;
790    obj->v1 = alloc->mHal.drv;
791    obj->v2 = (void *)alloc->mHal.drvState.lod[0].stride;
792  } else {
793    obj->r = nullptr;
794    obj->v1 = nullptr;
795    obj->v2 = nullptr;
796  }
797#endif
798}
799
800void rsovAllocationSetSurface(const Context *rsc, Allocation *alloc,
801                              ANativeWindow *nw) {
802  // TODO: implement this
803}
804
805void rsovAllocationIoSend(const Context *rsc, Allocation *alloc) {
806  // TODO: implement this
807}
808
809void rsovAllocationIoReceive(const Context *rsc, Allocation *alloc) {
810  // TODO: implement this
811}
812
813void rsovAllocationElementData(const Context *rsc, const Allocation *alloc,
814                               uint32_t x, uint32_t y, uint32_t z,
815                               const void *data, uint32_t cIdx,
816                               size_t sizeBytes) {
817  uint8_t *ptr =
818      GetOffsetPtr(alloc, x, y, z, 0, RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X);
819
820  const Element *e = alloc->mHal.state.type->getElement()->getField(cIdx);
821  ptr += alloc->mHal.state.type->getElement()->getFieldOffsetBytes(cIdx);
822
823  if (alloc->mHal.state.hasReferences) {
824    e->incRefs(data);
825    e->decRefs(ptr);
826  }
827
828  memcpy(ptr, data, sizeBytes);
829}
830
831void rsovAllocationElementRead(const Context *rsc, const Allocation *alloc,
832                               uint32_t x, uint32_t y, uint32_t z, void *data,
833                               uint32_t cIdx, size_t sizeBytes) {
834  uint8_t *ptr =
835      GetOffsetPtr(alloc, x, y, z, 0, RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X);
836
837  const Element *e = alloc->mHal.state.type->getElement()->getField(cIdx);
838  ptr += alloc->mHal.state.type->getElement()->getFieldOffsetBytes(cIdx);
839
840  memcpy(data, ptr, sizeBytes);
841}
842