1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/gfx/skia/trunk/src/gpu/GrBufferAllocPool.cpp Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,488 @@
1.4 +
1.5 +/*
1.6 + * Copyright 2010 Google Inc.
1.7 + *
1.8 + * Use of this source code is governed by a BSD-style license that can be
1.9 + * found in the LICENSE file.
1.10 + */
1.11 +
1.12 +
1.13 +#include "GrBufferAllocPool.h"
1.14 +#include "GrDrawTargetCaps.h"
1.15 +#include "GrGpu.h"
1.16 +#include "GrIndexBuffer.h"
1.17 +#include "GrTypes.h"
1.18 +#include "GrVertexBuffer.h"
1.19 +
1.20 +#ifdef SK_DEBUG
1.21 + #define VALIDATE validate
1.22 +#else
1.23 + static void VALIDATE(bool = false) {}
1.24 +#endif
1.25 +
1.26 +// page size
1.27 +#define GrBufferAllocPool_MIN_BLOCK_SIZE ((size_t)1 << 12)
1.28 +
1.29 +GrBufferAllocPool::GrBufferAllocPool(GrGpu* gpu,
1.30 + BufferType bufferType,
1.31 + bool frequentResetHint,
1.32 + size_t blockSize,
1.33 + int preallocBufferCnt) :
1.34 + fBlocks(GrMax(8, 2*preallocBufferCnt)) {
1.35 +
1.36 + SkASSERT(NULL != gpu);
1.37 + fGpu = gpu;
1.38 + fGpu->ref();
1.39 + fGpuIsReffed = true;
1.40 +
1.41 + fBufferType = bufferType;
1.42 + fFrequentResetHint = frequentResetHint;
1.43 + fBufferPtr = NULL;
1.44 + fMinBlockSize = GrMax(GrBufferAllocPool_MIN_BLOCK_SIZE, blockSize);
1.45 +
1.46 + fBytesInUse = 0;
1.47 +
1.48 + fPreallocBuffersInUse = 0;
1.49 + fPreallocBufferStartIdx = 0;
1.50 + for (int i = 0; i < preallocBufferCnt; ++i) {
1.51 + GrGeometryBuffer* buffer = this->createBuffer(fMinBlockSize);
1.52 + if (NULL != buffer) {
1.53 + *fPreallocBuffers.append() = buffer;
1.54 + }
1.55 + }
1.56 +}
1.57 +
1.58 +GrBufferAllocPool::~GrBufferAllocPool() {
1.59 + VALIDATE();
1.60 + if (fBlocks.count()) {
1.61 + GrGeometryBuffer* buffer = fBlocks.back().fBuffer;
1.62 + if (buffer->isLocked()) {
1.63 + buffer->unlock();
1.64 + }
1.65 + }
1.66 + while (!fBlocks.empty()) {
1.67 + destroyBlock();
1.68 + }
1.69 + fPreallocBuffers.unrefAll();
1.70 + releaseGpuRef();
1.71 +}
1.72 +
1.73 +void GrBufferAllocPool::releaseGpuRef() {
1.74 + if (fGpuIsReffed) {
1.75 + fGpu->unref();
1.76 + fGpuIsReffed = false;
1.77 + }
1.78 +}
1.79 +
1.80 +void GrBufferAllocPool::reset() {
1.81 + VALIDATE();
1.82 + fBytesInUse = 0;
1.83 + if (fBlocks.count()) {
1.84 + GrGeometryBuffer* buffer = fBlocks.back().fBuffer;
1.85 + if (buffer->isLocked()) {
1.86 + buffer->unlock();
1.87 + }
1.88 + }
1.89 + // fPreallocBuffersInUse will be decremented down to zero in the while loop
1.90 + int preallocBuffersInUse = fPreallocBuffersInUse;
1.91 + while (!fBlocks.empty()) {
1.92 + this->destroyBlock();
1.93 + }
1.94 + if (fPreallocBuffers.count()) {
1.95 + // must set this after above loop.
1.96 + fPreallocBufferStartIdx = (fPreallocBufferStartIdx +
1.97 + preallocBuffersInUse) %
1.98 + fPreallocBuffers.count();
1.99 + }
1.100 + // we may have created a large cpu mirror of a large VB. Reset the size
1.101 + // to match our pre-allocated VBs.
1.102 + fCpuData.reset(fMinBlockSize);
1.103 + SkASSERT(0 == fPreallocBuffersInUse);
1.104 + VALIDATE();
1.105 +}
1.106 +
1.107 +void GrBufferAllocPool::unlock() {
1.108 + VALIDATE();
1.109 +
1.110 + if (NULL != fBufferPtr) {
1.111 + BufferBlock& block = fBlocks.back();
1.112 + if (block.fBuffer->isLocked()) {
1.113 + block.fBuffer->unlock();
1.114 + } else {
1.115 + size_t flushSize = block.fBuffer->sizeInBytes() - block.fBytesFree;
1.116 + flushCpuData(fBlocks.back().fBuffer, flushSize);
1.117 + }
1.118 + fBufferPtr = NULL;
1.119 + }
1.120 + VALIDATE();
1.121 +}
1.122 +
1.123 +#ifdef SK_DEBUG
1.124 +void GrBufferAllocPool::validate(bool unusedBlockAllowed) const {
1.125 + if (NULL != fBufferPtr) {
1.126 + SkASSERT(!fBlocks.empty());
1.127 + if (fBlocks.back().fBuffer->isLocked()) {
1.128 + GrGeometryBuffer* buf = fBlocks.back().fBuffer;
1.129 + SkASSERT(buf->lockPtr() == fBufferPtr);
1.130 + } else {
1.131 + SkASSERT(fCpuData.get() == fBufferPtr);
1.132 + }
1.133 + } else {
1.134 + SkASSERT(fBlocks.empty() || !fBlocks.back().fBuffer->isLocked());
1.135 + }
1.136 + size_t bytesInUse = 0;
1.137 + for (int i = 0; i < fBlocks.count() - 1; ++i) {
1.138 + SkASSERT(!fBlocks[i].fBuffer->isLocked());
1.139 + }
1.140 + for (int i = 0; i < fBlocks.count(); ++i) {
1.141 + size_t bytes = fBlocks[i].fBuffer->sizeInBytes() - fBlocks[i].fBytesFree;
1.142 + bytesInUse += bytes;
1.143 + SkASSERT(bytes || unusedBlockAllowed);
1.144 + }
1.145 +
1.146 + SkASSERT(bytesInUse == fBytesInUse);
1.147 + if (unusedBlockAllowed) {
1.148 + SkASSERT((fBytesInUse && !fBlocks.empty()) ||
1.149 + (!fBytesInUse && (fBlocks.count() < 2)));
1.150 + } else {
1.151 + SkASSERT((0 == fBytesInUse) == fBlocks.empty());
1.152 + }
1.153 +}
1.154 +#endif
1.155 +
1.156 +void* GrBufferAllocPool::makeSpace(size_t size,
1.157 + size_t alignment,
1.158 + const GrGeometryBuffer** buffer,
1.159 + size_t* offset) {
1.160 + VALIDATE();
1.161 +
1.162 + SkASSERT(NULL != buffer);
1.163 + SkASSERT(NULL != offset);
1.164 +
1.165 + if (NULL != fBufferPtr) {
1.166 + BufferBlock& back = fBlocks.back();
1.167 + size_t usedBytes = back.fBuffer->sizeInBytes() - back.fBytesFree;
1.168 + size_t pad = GrSizeAlignUpPad(usedBytes,
1.169 + alignment);
1.170 + if ((size + pad) <= back.fBytesFree) {
1.171 + usedBytes += pad;
1.172 + *offset = usedBytes;
1.173 + *buffer = back.fBuffer;
1.174 + back.fBytesFree -= size + pad;
1.175 + fBytesInUse += size + pad;
1.176 + VALIDATE();
1.177 + return (void*)(reinterpret_cast<intptr_t>(fBufferPtr) + usedBytes);
1.178 + }
1.179 + }
1.180 +
1.181 + // We could honor the space request using by a partial update of the current
1.182 + // VB (if there is room). But we don't currently use draw calls to GL that
1.183 + // allow the driver to know that previously issued draws won't read from
1.184 + // the part of the buffer we update. Also, the GL buffer implementation
1.185 + // may be cheating on the actual buffer size by shrinking the buffer on
1.186 + // updateData() if the amount of data passed is less than the full buffer
1.187 + // size.
1.188 +
1.189 + if (!createBlock(size)) {
1.190 + return NULL;
1.191 + }
1.192 + SkASSERT(NULL != fBufferPtr);
1.193 +
1.194 + *offset = 0;
1.195 + BufferBlock& back = fBlocks.back();
1.196 + *buffer = back.fBuffer;
1.197 + back.fBytesFree -= size;
1.198 + fBytesInUse += size;
1.199 + VALIDATE();
1.200 + return fBufferPtr;
1.201 +}
1.202 +
1.203 +int GrBufferAllocPool::currentBufferItems(size_t itemSize) const {
1.204 + VALIDATE();
1.205 + if (NULL != fBufferPtr) {
1.206 + const BufferBlock& back = fBlocks.back();
1.207 + size_t usedBytes = back.fBuffer->sizeInBytes() - back.fBytesFree;
1.208 + size_t pad = GrSizeAlignUpPad(usedBytes, itemSize);
1.209 + return static_cast<int>((back.fBytesFree - pad) / itemSize);
1.210 + } else if (fPreallocBuffersInUse < fPreallocBuffers.count()) {
1.211 + return static_cast<int>(fMinBlockSize / itemSize);
1.212 + }
1.213 + return 0;
1.214 +}
1.215 +
1.216 +int GrBufferAllocPool::preallocatedBuffersRemaining() const {
1.217 + return fPreallocBuffers.count() - fPreallocBuffersInUse;
1.218 +}
1.219 +
1.220 +int GrBufferAllocPool::preallocatedBufferCount() const {
1.221 + return fPreallocBuffers.count();
1.222 +}
1.223 +
1.224 +void GrBufferAllocPool::putBack(size_t bytes) {
1.225 + VALIDATE();
1.226 +
1.227 + // if the putBack unwinds all the preallocated buffers then we will
1.228 + // advance the starting index. As blocks are destroyed fPreallocBuffersInUse
1.229 + // will be decremented. I will reach zero if all blocks using preallocated
1.230 + // buffers are released.
1.231 + int preallocBuffersInUse = fPreallocBuffersInUse;
1.232 +
1.233 + while (bytes) {
1.234 + // caller shouldnt try to put back more than they've taken
1.235 + SkASSERT(!fBlocks.empty());
1.236 + BufferBlock& block = fBlocks.back();
1.237 + size_t bytesUsed = block.fBuffer->sizeInBytes() - block.fBytesFree;
1.238 + if (bytes >= bytesUsed) {
1.239 + bytes -= bytesUsed;
1.240 + fBytesInUse -= bytesUsed;
1.241 + // if we locked a vb to satisfy the make space and we're releasing
1.242 + // beyond it, then unlock it.
1.243 + if (block.fBuffer->isLocked()) {
1.244 + block.fBuffer->unlock();
1.245 + }
1.246 + this->destroyBlock();
1.247 + } else {
1.248 + block.fBytesFree += bytes;
1.249 + fBytesInUse -= bytes;
1.250 + bytes = 0;
1.251 + break;
1.252 + }
1.253 + }
1.254 + if (!fPreallocBuffersInUse && fPreallocBuffers.count()) {
1.255 + fPreallocBufferStartIdx = (fPreallocBufferStartIdx +
1.256 + preallocBuffersInUse) %
1.257 + fPreallocBuffers.count();
1.258 + }
1.259 + VALIDATE();
1.260 +}
1.261 +
1.262 +bool GrBufferAllocPool::createBlock(size_t requestSize) {
1.263 +
1.264 + size_t size = GrMax(requestSize, fMinBlockSize);
1.265 + SkASSERT(size >= GrBufferAllocPool_MIN_BLOCK_SIZE);
1.266 +
1.267 + VALIDATE();
1.268 +
1.269 + BufferBlock& block = fBlocks.push_back();
1.270 +
1.271 + if (size == fMinBlockSize &&
1.272 + fPreallocBuffersInUse < fPreallocBuffers.count()) {
1.273 +
1.274 + uint32_t nextBuffer = (fPreallocBuffersInUse +
1.275 + fPreallocBufferStartIdx) %
1.276 + fPreallocBuffers.count();
1.277 + block.fBuffer = fPreallocBuffers[nextBuffer];
1.278 + block.fBuffer->ref();
1.279 + ++fPreallocBuffersInUse;
1.280 + } else {
1.281 + block.fBuffer = this->createBuffer(size);
1.282 + if (NULL == block.fBuffer) {
1.283 + fBlocks.pop_back();
1.284 + return false;
1.285 + }
1.286 + }
1.287 +
1.288 + block.fBytesFree = size;
1.289 + if (NULL != fBufferPtr) {
1.290 + SkASSERT(fBlocks.count() > 1);
1.291 + BufferBlock& prev = fBlocks.fromBack(1);
1.292 + if (prev.fBuffer->isLocked()) {
1.293 + prev.fBuffer->unlock();
1.294 + } else {
1.295 + flushCpuData(prev.fBuffer,
1.296 + prev.fBuffer->sizeInBytes() - prev.fBytesFree);
1.297 + }
1.298 + fBufferPtr = NULL;
1.299 + }
1.300 +
1.301 + SkASSERT(NULL == fBufferPtr);
1.302 +
1.303 + // If the buffer is CPU-backed we lock it because it is free to do so and saves a copy.
1.304 + // Otherwise when buffer locking is supported:
1.305 + // a) If the frequently reset hint is set we only lock when the requested size meets a
1.306 + // threshold (since we don't expect it is likely that we will see more vertex data)
1.307 + // b) If the hint is not set we lock if the buffer size is greater than the threshold.
1.308 + bool attemptLock = block.fBuffer->isCPUBacked();
1.309 + if (!attemptLock && fGpu->caps()->bufferLockSupport()) {
1.310 + if (fFrequentResetHint) {
1.311 + attemptLock = requestSize > GR_GEOM_BUFFER_LOCK_THRESHOLD;
1.312 + } else {
1.313 + attemptLock = size > GR_GEOM_BUFFER_LOCK_THRESHOLD;
1.314 + }
1.315 + }
1.316 +
1.317 + if (attemptLock) {
1.318 + fBufferPtr = block.fBuffer->lock();
1.319 + }
1.320 +
1.321 + if (NULL == fBufferPtr) {
1.322 + fBufferPtr = fCpuData.reset(size);
1.323 + }
1.324 +
1.325 + VALIDATE(true);
1.326 +
1.327 + return true;
1.328 +}
1.329 +
1.330 +void GrBufferAllocPool::destroyBlock() {
1.331 + SkASSERT(!fBlocks.empty());
1.332 +
1.333 + BufferBlock& block = fBlocks.back();
1.334 + if (fPreallocBuffersInUse > 0) {
1.335 + uint32_t prevPreallocBuffer = (fPreallocBuffersInUse +
1.336 + fPreallocBufferStartIdx +
1.337 + (fPreallocBuffers.count() - 1)) %
1.338 + fPreallocBuffers.count();
1.339 + if (block.fBuffer == fPreallocBuffers[prevPreallocBuffer]) {
1.340 + --fPreallocBuffersInUse;
1.341 + }
1.342 + }
1.343 + SkASSERT(!block.fBuffer->isLocked());
1.344 + block.fBuffer->unref();
1.345 + fBlocks.pop_back();
1.346 + fBufferPtr = NULL;
1.347 +}
1.348 +
1.349 +void GrBufferAllocPool::flushCpuData(GrGeometryBuffer* buffer,
1.350 + size_t flushSize) {
1.351 + SkASSERT(NULL != buffer);
1.352 + SkASSERT(!buffer->isLocked());
1.353 + SkASSERT(fCpuData.get() == fBufferPtr);
1.354 + SkASSERT(flushSize <= buffer->sizeInBytes());
1.355 + VALIDATE(true);
1.356 +
1.357 + if (fGpu->caps()->bufferLockSupport() &&
1.358 + flushSize > GR_GEOM_BUFFER_LOCK_THRESHOLD) {
1.359 + void* data = buffer->lock();
1.360 + if (NULL != data) {
1.361 + memcpy(data, fBufferPtr, flushSize);
1.362 + buffer->unlock();
1.363 + return;
1.364 + }
1.365 + }
1.366 + buffer->updateData(fBufferPtr, flushSize);
1.367 + VALIDATE(true);
1.368 +}
1.369 +
1.370 +GrGeometryBuffer* GrBufferAllocPool::createBuffer(size_t size) {
1.371 + if (kIndex_BufferType == fBufferType) {
1.372 + return fGpu->createIndexBuffer(size, true);
1.373 + } else {
1.374 + SkASSERT(kVertex_BufferType == fBufferType);
1.375 + return fGpu->createVertexBuffer(size, true);
1.376 + }
1.377 +}
1.378 +
1.379 +////////////////////////////////////////////////////////////////////////////////
1.380 +
1.381 +GrVertexBufferAllocPool::GrVertexBufferAllocPool(GrGpu* gpu,
1.382 + bool frequentResetHint,
1.383 + size_t bufferSize,
1.384 + int preallocBufferCnt)
1.385 +: GrBufferAllocPool(gpu,
1.386 + kVertex_BufferType,
1.387 + frequentResetHint,
1.388 + bufferSize,
1.389 + preallocBufferCnt) {
1.390 +}
1.391 +
1.392 +void* GrVertexBufferAllocPool::makeSpace(size_t vertexSize,
1.393 + int vertexCount,
1.394 + const GrVertexBuffer** buffer,
1.395 + int* startVertex) {
1.396 +
1.397 + SkASSERT(vertexCount >= 0);
1.398 + SkASSERT(NULL != buffer);
1.399 + SkASSERT(NULL != startVertex);
1.400 +
1.401 + size_t offset = 0; // assign to suppress warning
1.402 + const GrGeometryBuffer* geomBuffer = NULL; // assign to suppress warning
1.403 + void* ptr = INHERITED::makeSpace(vertexSize * vertexCount,
1.404 + vertexSize,
1.405 + &geomBuffer,
1.406 + &offset);
1.407 +
1.408 + *buffer = (const GrVertexBuffer*) geomBuffer;
1.409 + SkASSERT(0 == offset % vertexSize);
1.410 + *startVertex = static_cast<int>(offset / vertexSize);
1.411 + return ptr;
1.412 +}
1.413 +
1.414 +bool GrVertexBufferAllocPool::appendVertices(size_t vertexSize,
1.415 + int vertexCount,
1.416 + const void* vertices,
1.417 + const GrVertexBuffer** buffer,
1.418 + int* startVertex) {
1.419 + void* space = makeSpace(vertexSize, vertexCount, buffer, startVertex);
1.420 + if (NULL != space) {
1.421 + memcpy(space,
1.422 + vertices,
1.423 + vertexSize * vertexCount);
1.424 + return true;
1.425 + } else {
1.426 + return false;
1.427 + }
1.428 +}
1.429 +
1.430 +int GrVertexBufferAllocPool::preallocatedBufferVertices(size_t vertexSize) const {
1.431 + return static_cast<int>(INHERITED::preallocatedBufferSize() / vertexSize);
1.432 +}
1.433 +
1.434 +int GrVertexBufferAllocPool::currentBufferVertices(size_t vertexSize) const {
1.435 + return currentBufferItems(vertexSize);
1.436 +}
1.437 +
1.438 +////////////////////////////////////////////////////////////////////////////////
1.439 +
1.440 +GrIndexBufferAllocPool::GrIndexBufferAllocPool(GrGpu* gpu,
1.441 + bool frequentResetHint,
1.442 + size_t bufferSize,
1.443 + int preallocBufferCnt)
1.444 +: GrBufferAllocPool(gpu,
1.445 + kIndex_BufferType,
1.446 + frequentResetHint,
1.447 + bufferSize,
1.448 + preallocBufferCnt) {
1.449 +}
1.450 +
1.451 +void* GrIndexBufferAllocPool::makeSpace(int indexCount,
1.452 + const GrIndexBuffer** buffer,
1.453 + int* startIndex) {
1.454 +
1.455 + SkASSERT(indexCount >= 0);
1.456 + SkASSERT(NULL != buffer);
1.457 + SkASSERT(NULL != startIndex);
1.458 +
1.459 + size_t offset = 0; // assign to suppress warning
1.460 + const GrGeometryBuffer* geomBuffer = NULL; // assign to suppress warning
1.461 + void* ptr = INHERITED::makeSpace(indexCount * sizeof(uint16_t),
1.462 + sizeof(uint16_t),
1.463 + &geomBuffer,
1.464 + &offset);
1.465 +
1.466 + *buffer = (const GrIndexBuffer*) geomBuffer;
1.467 + SkASSERT(0 == offset % sizeof(uint16_t));
1.468 + *startIndex = static_cast<int>(offset / sizeof(uint16_t));
1.469 + return ptr;
1.470 +}
1.471 +
1.472 +bool GrIndexBufferAllocPool::appendIndices(int indexCount,
1.473 + const void* indices,
1.474 + const GrIndexBuffer** buffer,
1.475 + int* startIndex) {
1.476 + void* space = makeSpace(indexCount, buffer, startIndex);
1.477 + if (NULL != space) {
1.478 + memcpy(space, indices, sizeof(uint16_t) * indexCount);
1.479 + return true;
1.480 + } else {
1.481 + return false;
1.482 + }
1.483 +}
1.484 +
1.485 +int GrIndexBufferAllocPool::preallocatedBufferIndices() const {
1.486 + return static_cast<int>(INHERITED::preallocatedBufferSize() / sizeof(uint16_t));
1.487 +}
1.488 +
1.489 +int GrIndexBufferAllocPool::currentBufferIndices() const {
1.490 + return currentBufferItems(sizeof(uint16_t));
1.491 +}
