/src/gdal/frmts/zarr/zarr_v3_codec_sequence.cpp
Line | Count | Source |
1 | | /****************************************************************************** |
2 | | * |
3 | | * Project: GDAL |
4 | | * Purpose: Zarr driver, ZarrV3CodecSequence class |
5 | | * Author: Even Rouault <even dot rouault at spatialys.com> |
6 | | * |
7 | | ****************************************************************************** |
8 | | * Copyright (c) 2023, Even Rouault <even dot rouault at spatialys.com> |
9 | | * |
10 | | * SPDX-License-Identifier: MIT |
11 | | ****************************************************************************/ |
12 | | |
13 | | #include "zarr_v3_codec.h" |
14 | | |
15 | | /************************************************************************/ |
16 | | /* ZarrV3CodecSequence::Clone() */ |
17 | | /************************************************************************/ |
18 | | |
19 | | std::unique_ptr<ZarrV3CodecSequence> ZarrV3CodecSequence::Clone() const |
20 | 0 | { |
21 | 0 | auto poClone = std::make_unique<ZarrV3CodecSequence>(m_oInputArrayMetadata); |
22 | 0 | for (const auto &poCodec : m_apoCodecs) |
23 | 0 | poClone->m_apoCodecs.emplace_back(poCodec->Clone()); |
24 | 0 | poClone->m_oCodecArray = m_oCodecArray.Clone(); |
25 | 0 | poClone->m_bPartialDecodingPossible = m_bPartialDecodingPossible; |
26 | 0 | return poClone; |
27 | 0 | } |
28 | | |
29 | | /************************************************************************/ |
30 | | /* ZarrV3CodecSequence::InitFromJson() */ |
31 | | /************************************************************************/ |
32 | | |
33 | | bool ZarrV3CodecSequence::InitFromJson(const CPLJSONObject &oCodecs, |
34 | | ZarrArrayMetadata &oOutputArrayMetadata) |
35 | 0 | { |
36 | 0 | if (oCodecs.GetType() != CPLJSONObject::Type::Array) |
37 | 0 | { |
38 | 0 | CPLError(CE_Failure, CPLE_AppDefined, "codecs is not an array"); |
39 | 0 | return false; |
40 | 0 | } |
41 | 0 | auto oCodecsArray = oCodecs.ToArray(); |
42 | |
|
43 | 0 | ZarrArrayMetadata oInputArrayMetadata = m_oInputArrayMetadata; |
44 | 0 | ZarrV3Codec::IOType eLastType = ZarrV3Codec::IOType::ARRAY; |
45 | 0 | std::string osLastCodec; |
46 | |
|
47 | 0 | const auto InsertImplicitEndianCodecIfNeeded = |
48 | 0 | [this, &oInputArrayMetadata, &eLastType, &osLastCodec]() |
49 | 0 | { |
50 | 0 | CPL_IGNORE_RET_VAL(this); |
51 | 0 | if (eLastType == ZarrV3Codec::IOType::ARRAY && |
52 | 0 | oInputArrayMetadata.oElt.nativeSize > 1) |
53 | 0 | { |
54 | 0 | CPLError(CE_Warning, CPLE_AppDefined, |
55 | 0 | "'bytes' codec missing. Assuming little-endian storage, " |
56 | 0 | "but such tolerance may be removed in future versions"); |
57 | 0 | auto poEndianCodec = std::make_unique<ZarrV3CodecBytes>(); |
58 | 0 | ZarrArrayMetadata oTmpOutputArrayMetadata; |
59 | 0 | poEndianCodec->InitFromConfiguration( |
60 | 0 | ZarrV3CodecBytes::GetConfiguration(true), oInputArrayMetadata, |
61 | 0 | oTmpOutputArrayMetadata, /* bEmitWarnings = */ true); |
62 | 0 | oInputArrayMetadata = std::move(oTmpOutputArrayMetadata); |
63 | 0 | eLastType = poEndianCodec->GetOutputType(); |
64 | 0 | osLastCodec = poEndianCodec->GetName(); |
65 | | if constexpr (!CPL_IS_LSB) |
66 | | { |
67 | | // Insert a little endian codec if we are on a big endian target |
68 | | m_apoCodecs.emplace_back(std::move(poEndianCodec)); |
69 | | } |
70 | 0 | } |
71 | 0 | }; |
72 | |
|
73 | 0 | bool bShardingFound = false; |
74 | 0 | std::vector<size_t> anBlockSizesBeforeSharding; |
75 | 0 | for (const auto &oCodec : oCodecsArray) |
76 | 0 | { |
77 | 0 | if (oCodec.GetType() != CPLJSONObject::Type::Object) |
78 | 0 | { |
79 | 0 | CPLError(CE_Failure, CPLE_AppDefined, "codecs[] is not an object"); |
80 | 0 | return false; |
81 | 0 | } |
82 | 0 | const auto osName = oCodec["name"].ToString(); |
83 | 0 | std::unique_ptr<ZarrV3Codec> poCodec; |
84 | 0 | if (osName == ZarrV3CodecGZip::NAME) |
85 | 0 | poCodec = std::make_unique<ZarrV3CodecGZip>(); |
86 | 0 | else if (osName == ZarrV3CodecBlosc::NAME) |
87 | 0 | poCodec = std::make_unique<ZarrV3CodecBlosc>(); |
88 | 0 | else if (osName == ZarrV3CodecZstd::NAME) |
89 | 0 | poCodec = std::make_unique<ZarrV3CodecZstd>(); |
90 | 0 | else if (osName == ZarrV3CodecBytes::NAME || |
91 | 0 | osName == "endian" /* endian is the old name */) |
92 | 0 | poCodec = std::make_unique<ZarrV3CodecBytes>(); |
93 | 0 | else if (osName == ZarrV3CodecTranspose::NAME) |
94 | 0 | poCodec = std::make_unique<ZarrV3CodecTranspose>(); |
95 | 0 | else if (osName == ZarrV3CodecCRC32C::NAME) |
96 | 0 | poCodec = std::make_unique<ZarrV3CodecCRC32C>(); |
97 | 0 | else if (osName == ZarrV3CodecShardingIndexed::NAME) |
98 | 0 | { |
99 | 0 | bShardingFound = true; |
100 | 0 | poCodec = std::make_unique<ZarrV3CodecShardingIndexed>(); |
101 | 0 | } |
102 | 0 | else |
103 | 0 | { |
104 | 0 | CPLError(CE_Failure, CPLE_NotSupported, "Unsupported codec: %s", |
105 | 0 | osName.c_str()); |
106 | 0 | return false; |
107 | 0 | } |
108 | | |
109 | 0 | if (poCodec->GetInputType() == ZarrV3Codec::IOType::ARRAY) |
110 | 0 | { |
111 | 0 | if (eLastType == ZarrV3Codec::IOType::BYTES) |
112 | 0 | { |
113 | 0 | CPLError(CE_Failure, CPLE_AppDefined, |
114 | 0 | "Cannot chain codec %s with %s", |
115 | 0 | poCodec->GetName().c_str(), osLastCodec.c_str()); |
116 | 0 | return false; |
117 | 0 | } |
118 | 0 | } |
119 | 0 | else |
120 | 0 | { |
121 | 0 | InsertImplicitEndianCodecIfNeeded(); |
122 | 0 | } |
123 | | |
124 | 0 | ZarrArrayMetadata oStepOutputArrayMetadata; |
125 | 0 | if (osName == ZarrV3CodecShardingIndexed::NAME) |
126 | 0 | { |
127 | 0 | anBlockSizesBeforeSharding = oInputArrayMetadata.anBlockSizes; |
128 | 0 | } |
129 | 0 | if (!poCodec->InitFromConfiguration(oCodec["configuration"], |
130 | 0 | oInputArrayMetadata, |
131 | 0 | oStepOutputArrayMetadata, |
132 | 0 | /* bEmitWarnings = */ true)) |
133 | 0 | { |
134 | 0 | return false; |
135 | 0 | } |
136 | 0 | oInputArrayMetadata = std::move(oStepOutputArrayMetadata); |
137 | 0 | eLastType = poCodec->GetOutputType(); |
138 | 0 | osLastCodec = poCodec->GetName(); |
139 | |
|
140 | 0 | if (!poCodec->IsNoOp()) |
141 | 0 | m_apoCodecs.emplace_back(std::move(poCodec)); |
142 | 0 | } |
143 | | |
144 | 0 | if (bShardingFound) |
145 | 0 | { |
146 | 0 | m_bPartialDecodingPossible = |
147 | 0 | (m_apoCodecs.back()->GetName() == ZarrV3CodecShardingIndexed::NAME); |
148 | 0 | if (!m_bPartialDecodingPossible) |
149 | 0 | { |
150 | 0 | m_bPartialDecodingPossible = false; |
151 | | // This is not an implementation limitation, but the result of a |
152 | | // badly thought dataset. Zarr-Python also emits a similar warning. |
153 | 0 | CPLError( |
154 | 0 | CE_Warning, CPLE_AppDefined, |
155 | 0 | "Sharding codec found, but not in last position. Consequently " |
156 | 0 | "partial shard decoding will not be possible"); |
157 | 0 | oInputArrayMetadata.anBlockSizes = |
158 | 0 | std::move(anBlockSizesBeforeSharding); |
159 | 0 | } |
160 | 0 | } |
161 | |
|
162 | 0 | InsertImplicitEndianCodecIfNeeded(); |
163 | |
|
164 | 0 | m_oCodecArray = oCodecs.Clone(); |
165 | 0 | oOutputArrayMetadata = std::move(oInputArrayMetadata); |
166 | 0 | return true; |
167 | 0 | } |
168 | | |
169 | | /************************************************************************/ |
170 | | /* ZarrV3CodecBytes::AllocateBuffer() */ |
171 | | /************************************************************************/ |
172 | | |
173 | | bool ZarrV3CodecSequence::AllocateBuffer(ZarrByteVectorQuickResize &abyBuffer, |
174 | | size_t nEltCount) |
175 | 0 | { |
176 | 0 | if (!m_apoCodecs.empty()) |
177 | 0 | { |
178 | 0 | const size_t nRawSize = |
179 | 0 | nEltCount * m_oInputArrayMetadata.oElt.nativeSize; |
180 | | // Grow the temporary buffer a bit beyond the uncompressed size |
181 | 0 | const size_t nMaxSize = nRawSize + nRawSize / 3 + 64; |
182 | 0 | try |
183 | 0 | { |
184 | 0 | m_abyTmp.resize(nMaxSize); |
185 | 0 | } |
186 | 0 | catch (const std::exception &e) |
187 | 0 | { |
188 | 0 | CPLError(CE_Failure, CPLE_OutOfMemory, "%s", e.what()); |
189 | 0 | return false; |
190 | 0 | } |
191 | 0 | m_abyTmp.resize(nRawSize); |
192 | | |
193 | | // Grow the input/output buffer too if we have several steps |
194 | 0 | if (m_apoCodecs.size() >= 2 && abyBuffer.capacity() < nMaxSize) |
195 | 0 | { |
196 | 0 | const size_t nSize = abyBuffer.size(); |
197 | 0 | try |
198 | 0 | { |
199 | 0 | abyBuffer.resize(nMaxSize); |
200 | 0 | } |
201 | 0 | catch (const std::exception &e) |
202 | 0 | { |
203 | 0 | CPLError(CE_Failure, CPLE_OutOfMemory, "%s", e.what()); |
204 | 0 | return false; |
205 | 0 | } |
206 | 0 | abyBuffer.resize(nSize); |
207 | 0 | } |
208 | 0 | } |
209 | 0 | return true; |
210 | 0 | } |
211 | | |
212 | | /************************************************************************/ |
213 | | /* ZarrV3CodecSequence::Encode() */ |
214 | | /************************************************************************/ |
215 | | |
216 | | bool ZarrV3CodecSequence::Encode(ZarrByteVectorQuickResize &abyBuffer) |
217 | 0 | { |
218 | 0 | if (!AllocateBuffer(abyBuffer, |
219 | 0 | MultiplyElements(m_oInputArrayMetadata.anBlockSizes))) |
220 | 0 | return false; |
221 | 0 | for (const auto &poCodec : m_apoCodecs) |
222 | 0 | { |
223 | 0 | if (!poCodec->Encode(abyBuffer, m_abyTmp)) |
224 | 0 | return false; |
225 | 0 | std::swap(abyBuffer, m_abyTmp); |
226 | 0 | } |
227 | 0 | return true; |
228 | 0 | } |
229 | | |
230 | | /************************************************************************/ |
231 | | /* ZarrV3CodecSequence::Decode() */ |
232 | | /************************************************************************/ |
233 | | |
234 | | bool ZarrV3CodecSequence::Decode(ZarrByteVectorQuickResize &abyBuffer) |
235 | 0 | { |
236 | 0 | if (!AllocateBuffer(abyBuffer, |
237 | 0 | MultiplyElements(m_oInputArrayMetadata.anBlockSizes))) |
238 | 0 | return false; |
239 | 0 | for (auto iter = m_apoCodecs.rbegin(); iter != m_apoCodecs.rend(); ++iter) |
240 | 0 | { |
241 | 0 | const auto &poCodec = *iter; |
242 | 0 | if (!poCodec->Decode(abyBuffer, m_abyTmp)) |
243 | 0 | return false; |
244 | 0 | std::swap(abyBuffer, m_abyTmp); |
245 | 0 | } |
246 | 0 | return true; |
247 | 0 | } |
248 | | |
249 | | /************************************************************************/ |
250 | | /* ZarrV3CodecSequence::DecodePartial() */ |
251 | | /************************************************************************/ |
252 | | |
253 | | bool ZarrV3CodecSequence::DecodePartial(VSIVirtualHandle *poFile, |
254 | | ZarrByteVectorQuickResize &abyBuffer, |
255 | | const std::vector<size_t> &anStartIdxIn, |
256 | | const std::vector<size_t> &anCountIn) |
257 | 0 | { |
258 | 0 | CPLAssert(anStartIdxIn.size() == m_oInputArrayMetadata.anBlockSizes.size()); |
259 | 0 | CPLAssert(anStartIdxIn.size() == anCountIn.size()); |
260 | |
|
261 | 0 | if (!AllocateBuffer(abyBuffer, MultiplyElements(anCountIn))) |
262 | 0 | return false; |
263 | | |
264 | | // anStartIdxIn and anCountIn are expressed in the shape *before* encoding |
265 | | // We need to apply the potential transpositions before submitting them |
266 | | // to the decoder of the Array->Bytes decoder |
267 | 0 | std::vector<size_t> anStartIdx(anStartIdxIn); |
268 | 0 | std::vector<size_t> anCount(anCountIn); |
269 | 0 | for (auto &poCodec : m_apoCodecs) |
270 | 0 | { |
271 | 0 | poCodec->ChangeArrayShapeForward(anStartIdx, anCount); |
272 | 0 | } |
273 | |
|
274 | 0 | for (auto iter = m_apoCodecs.rbegin(); iter != m_apoCodecs.rend(); ++iter) |
275 | 0 | { |
276 | 0 | const auto &poCodec = *iter; |
277 | |
|
278 | 0 | if (!poCodec->DecodePartial(poFile, abyBuffer, m_abyTmp, anStartIdx, |
279 | 0 | anCount)) |
280 | 0 | return false; |
281 | 0 | std::swap(abyBuffer, m_abyTmp); |
282 | 0 | } |
283 | 0 | return true; |
284 | 0 | } |
285 | | |
286 | | /************************************************************************/ |
287 | | /* ZarrV3CodecSequence::BatchDecodePartial() */ |
288 | | /************************************************************************/ |
289 | | |
290 | | bool ZarrV3CodecSequence::BatchDecodePartial( |
291 | | VSIVirtualHandle *poFile, |
292 | | const std::vector<std::pair<std::vector<size_t>, std::vector<size_t>>> |
293 | | &anRequests, |
294 | | std::vector<ZarrByteVectorQuickResize> &aResults) |
295 | 0 | { |
296 | | // Only batch-decode when sharding is the sole codec. If other codecs |
297 | | // (e.g. transpose) precede it, indices and output need codec-specific |
298 | | // transformations that BatchDecodePartial does not handle. |
299 | 0 | if (m_apoCodecs.size() == 1) |
300 | 0 | { |
301 | 0 | auto *poSharding = dynamic_cast<ZarrV3CodecShardingIndexed *>( |
302 | 0 | m_apoCodecs.back().get()); |
303 | 0 | if (poSharding) |
304 | 0 | { |
305 | 0 | return poSharding->BatchDecodePartial(poFile, anRequests, aResults); |
306 | 0 | } |
307 | 0 | } |
308 | | |
309 | | // Fallback: sequential DecodePartial for non-sharding codec chains |
310 | 0 | aResults.resize(anRequests.size()); |
311 | 0 | for (size_t i = 0; i < anRequests.size(); ++i) |
312 | 0 | { |
313 | 0 | if (!DecodePartial(poFile, aResults[i], anRequests[i].first, |
314 | 0 | anRequests[i].second)) |
315 | 0 | return false; |
316 | 0 | } |
317 | 0 | return true; |
318 | 0 | } |
319 | | |
320 | | /************************************************************************/ |
321 | | /* ZarrV3CodecSequence::GetInnerMostBlockSize() */ |
322 | | /************************************************************************/ |
323 | | |
324 | | std::vector<size_t> ZarrV3CodecSequence::GetInnerMostBlockSize( |
325 | | const std::vector<size_t> &anOuterBlockSize) const |
326 | 0 | { |
327 | 0 | auto chunkSize = anOuterBlockSize; |
328 | 0 | for (auto iter = m_apoCodecs.rbegin(); iter != m_apoCodecs.rend(); ++iter) |
329 | 0 | { |
330 | 0 | const auto &poCodec = *iter; |
331 | 0 | if (m_bPartialDecodingPossible || |
332 | 0 | poCodec->GetName() != ZarrV3CodecShardingIndexed::NAME) |
333 | 0 | { |
334 | 0 | chunkSize = poCodec->GetInnerMostBlockSize(chunkSize); |
335 | 0 | } |
336 | 0 | } |
337 | 0 | return chunkSize; |
338 | 0 | } |