skia2/tests/SkBlockAllocatorTest.cpp

665 lines
26 KiB
C++
Raw Normal View History

Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
/*
* Copyright 2020 Google LLC
*
* Use of this source code is governed by a BSD-style license that can be
* found in the LICENSE file.
*/
#include "src/core/SkBlockAllocator.h"
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
#include "tests/Test.h"
#include <cstring>
using Block = SkBlockAllocator::Block;
using GrowthPolicy = SkBlockAllocator::GrowthPolicy;
class BlockAllocatorTestAccess {
public:
template<size_t N>
static size_t ScratchBlockSize(SkSBlockAllocator<N>& pool) {
return (size_t) pool->scratchBlockSize();
}
};
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
// Helper functions for modifying the allocator in a controlled manner
template<size_t N>
static int block_count(const SkSBlockAllocator<N>& pool) {
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
int ct = 0;
for (const Block* b : pool->blocks()) {
(void) b;
ct++;
}
return ct;
}
template<size_t N>
static Block* get_block(SkSBlockAllocator<N>& pool, int blockIndex) {
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
Block* found = nullptr;
int i = 0;
for (Block* b: pool->blocks()) {
if (i == blockIndex) {
found = b;
break;
}
i++;
}
SkASSERT(found != nullptr);
return found;
}
// SkBlockAllocator holds on to the largest last-released block to reuse for new allocations,
// and this is still counted in its totalSize(). However, it's easier to reason about size - scratch
// in many of these tests.
template<size_t N>
static size_t total_size(SkSBlockAllocator<N>& pool) {
return pool->totalSize() - BlockAllocatorTestAccess::ScratchBlockSize(pool);
}
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
template<size_t N>
static size_t add_block(SkSBlockAllocator<N>& pool) {
size_t currentSize = total_size(pool);
SkBlockAllocator::Block* current = pool->currentBlock();
while(pool->currentBlock() == current) {
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
pool->template allocate<4>(pool->preallocSize() / 2);
}
return total_size(pool) - currentSize;
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
}
template<size_t N>
static void* alloc_byte(SkSBlockAllocator<N>& pool) {
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
auto br = pool->template allocate<1>(1);
return br.fBlock->ptr(br.fAlignedOffset);
}
DEF_TEST(SkBlockAllocatorPreallocSize, r) {
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
// Tests stack/member initialization, option #1 described in doc
SkBlockAllocator stack{GrowthPolicy::kFixed, 2048};
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
SkDEBUGCODE(stack.validate();)
REPORTER_ASSERT(r, stack.preallocSize() == sizeof(SkBlockAllocator));
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
REPORTER_ASSERT(r, stack.preallocUsableSpace() == (size_t) stack.currentBlock()->avail());
// Tests placement new initialization to increase head block size, option #2
void* mem = operator new(1024);
SkBlockAllocator* placement = new (mem) SkBlockAllocator(GrowthPolicy::kLinear, 1024,
1024 - sizeof(SkBlockAllocator));
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
REPORTER_ASSERT(r, placement->preallocSize() == 1024);
REPORTER_ASSERT(r, placement->preallocUsableSpace() < 1024 &&
placement->preallocUsableSpace() >= (1024 - sizeof(SkBlockAllocator)));
placement->~SkBlockAllocator();
operator delete(mem);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
// Tests inline increased preallocation, option #3
SkSBlockAllocator<2048> inlined{};
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
SkDEBUGCODE(inlined->validate();)
REPORTER_ASSERT(r, inlined->preallocSize() == 2048);
REPORTER_ASSERT(r, inlined->preallocUsableSpace() < 2048 &&
inlined->preallocUsableSpace() >= (2048 - sizeof(SkBlockAllocator)));
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
}
DEF_TEST(SkBlockAllocatorAlloc, r) {
SkSBlockAllocator<1024> pool{};
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
SkDEBUGCODE(pool->validate();)
// Assumes the previous pointer was in the same block
auto validate_ptr = [&](int align, int size,
SkBlockAllocator::ByteRange br,
SkBlockAllocator::ByteRange* prevBR) {
uintptr_t pt = reinterpret_cast<uintptr_t>(br.fBlock->ptr(br.fAlignedOffset));
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
// Matches the requested align
REPORTER_ASSERT(r, pt % align == 0);
// And large enough
REPORTER_ASSERT(r, br.fEnd - br.fAlignedOffset >= size);
// And has enough padding for alignment
REPORTER_ASSERT(r, br.fAlignedOffset - br.fStart >= 0);
REPORTER_ASSERT(r, br.fAlignedOffset - br.fStart <= align - 1);
// And block of the returned struct is the current block of the allocator
REPORTER_ASSERT(r, pool->currentBlock() == br.fBlock);
// And make sure that we're past the required end of the previous allocation
if (prevBR) {
uintptr_t prevEnd =
reinterpret_cast<uintptr_t>(prevBR->fBlock->ptr(prevBR->fEnd - 1));
REPORTER_ASSERT(r, pt > prevEnd);
}
// And make sure that the entire byte range is safe to write into (excluding the dead space
// between "start" and "aligned offset," which is just padding and is left poisoned)
std::memset(br.fBlock->ptr(br.fAlignedOffset), 0xFF, br.fEnd - br.fAlignedOffset);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
};
auto p1 = pool->allocate<1>(14);
validate_ptr(1, 14, p1, nullptr);
auto p2 = pool->allocate<2>(24);
validate_ptr(2, 24, p2, &p1);
auto p4 = pool->allocate<4>(28);
validate_ptr(4, 28, p4, &p2);
auto p8 = pool->allocate<8>(40);
validate_ptr(8, 40, p8, &p4);
auto p16 = pool->allocate<16>(64);
validate_ptr(16, 64, p16, &p8);
auto p32 = pool->allocate<32>(96);
validate_ptr(32, 96, p32, &p16);
// All of these allocations should be in the head block
REPORTER_ASSERT(r, total_size(pool) == pool->preallocSize());
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
SkDEBUGCODE(pool->validate();)
// Requesting an allocation of avail() should not make a new block
size_t avail = pool->currentBlock()->avail<4>();
auto pAvail = pool->allocate<4>(avail);
validate_ptr(4, avail, pAvail, &p32);
// Remaining should be less than the alignment that was requested, and then
// the next allocation will make a new block
REPORTER_ASSERT(r, pool->currentBlock()->avail<4>() < 4);
auto pNextBlock = pool->allocate<4>(4);
validate_ptr(4, 4, pNextBlock, nullptr);
REPORTER_ASSERT(r, total_size(pool) > pool->preallocSize());
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
// Allocating more than avail() makes an another block
size_t currentSize = total_size(pool);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
size_t bigRequest = pool->currentBlock()->avail<4>() * 2;
auto pTooBig = pool->allocate<4>(bigRequest);
validate_ptr(4, bigRequest, pTooBig, nullptr);
REPORTER_ASSERT(r, total_size(pool) > currentSize);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
// Allocating more than the default growth policy (1024 in this case), will fulfill the request
REPORTER_ASSERT(r, total_size(pool) - currentSize < 4096);
currentSize = total_size(pool);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
auto pReallyTooBig = pool->allocate<4>(4096);
validate_ptr(4, 4096, pReallyTooBig, nullptr);
REPORTER_ASSERT(r, total_size(pool) >= currentSize + 4096);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
SkDEBUGCODE(pool->validate();)
}
DEF_TEST(SkBlockAllocatorResize, r) {
SkSBlockAllocator<1024> pool{};
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
SkDEBUGCODE(pool->validate();)
// Fixed resize from 16 to 32
SkBlockAllocator::ByteRange p = pool->allocate<4>(16);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
REPORTER_ASSERT(r, p.fBlock->avail<4>() > 16);
REPORTER_ASSERT(r, p.fBlock->resize(p.fStart, p.fEnd, 16));
p.fEnd += 16;
std::memset(p.fBlock->ptr(p.fAlignedOffset), 0x11, p.fEnd - p.fAlignedOffset);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
// Subsequent allocation is 32 bytes ahead of 'p' now, and 'p' cannot be resized further.
auto pNext = pool->allocate<4>(16);
REPORTER_ASSERT(r, reinterpret_cast<uintptr_t>(pNext.fBlock->ptr(pNext.fAlignedOffset)) -
reinterpret_cast<uintptr_t>(pNext.fBlock->ptr(p.fAlignedOffset)) == 32);
REPORTER_ASSERT(r, p.fBlock == pNext.fBlock);
REPORTER_ASSERT(r, !p.fBlock->resize(p.fStart, p.fEnd, 48));
// Confirm that releasing pNext allows 'p' to be resized, and that it can be resized up to avail
REPORTER_ASSERT(r, p.fBlock->release(pNext.fStart, pNext.fEnd));
int fillBlock = p.fBlock->avail<4>();
REPORTER_ASSERT(r, p.fBlock->resize(p.fStart, p.fEnd, fillBlock));
p.fEnd += fillBlock;
std::memset(p.fBlock->ptr(p.fAlignedOffset), 0x22, p.fEnd - p.fAlignedOffset);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
// Confirm that resizing when there's not enough room fails
REPORTER_ASSERT(r, p.fBlock->avail<4>() < fillBlock);
REPORTER_ASSERT(r, !p.fBlock->resize(p.fStart, p.fEnd, fillBlock));
// Confirm that we can shrink 'p' back to 32 bytes and then further allocate again
int shrinkTo32 = p.fStart - p.fEnd + 32;
REPORTER_ASSERT(r, p.fBlock->resize(p.fStart, p.fEnd, shrinkTo32));
p.fEnd += shrinkTo32;
REPORTER_ASSERT(r, p.fEnd - p.fStart == 32);
std::memset(p.fBlock->ptr(p.fAlignedOffset), 0x33, p.fEnd - p.fAlignedOffset);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
pNext = pool->allocate<4>(16);
REPORTER_ASSERT(r, reinterpret_cast<uintptr_t>(pNext.fBlock->ptr(pNext.fAlignedOffset)) -
reinterpret_cast<uintptr_t>(pNext.fBlock->ptr(p.fAlignedOffset)) == 32);
SkDEBUGCODE(pool->validate();)
// Confirm that we can't shrink past the start of the allocation, but we can shrink it to 0
int shrinkTo0 = pNext.fStart - pNext.fEnd;
#ifndef SK_DEBUG
// Only test for false on release builds; a negative size should assert on debug builds
REPORTER_ASSERT(r, !pNext.fBlock->resize(pNext.fStart, pNext.fEnd, shrinkTo0 - 1));
#endif
REPORTER_ASSERT(r, pNext.fBlock->resize(pNext.fStart, pNext.fEnd, shrinkTo0));
}
DEF_TEST(SkBlockAllocatorRelease, r) {
SkSBlockAllocator<1024> pool{};
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
SkDEBUGCODE(pool->validate();)
// Successful allocate and release
auto p = pool->allocate<8>(32);
REPORTER_ASSERT(r, pool->currentBlock()->release(p.fStart, p.fEnd));
// Ensure the above release actually means the next allocation reuses the same space
auto p2 = pool->allocate<8>(32);
REPORTER_ASSERT(r, p.fStart == p2.fStart);
// Confirm that 'p2' cannot be released if another allocation came after it
auto p3 = pool->allocate<8>(64);
(void) p3;
REPORTER_ASSERT(r, !p2.fBlock->release(p2.fStart, p2.fEnd));
// Confirm that 'p4' can be released if 'p5' is released first, and confirm that 'p2' and 'p3'
// can be released simultaneously (equivalent to 'p3' then 'p2').
auto p4 = pool->allocate<8>(16);
auto p5 = pool->allocate<8>(96);
REPORTER_ASSERT(r, p5.fBlock->release(p5.fStart, p5.fEnd));
REPORTER_ASSERT(r, p4.fBlock->release(p4.fStart, p4.fEnd));
REPORTER_ASSERT(r, p2.fBlock->release(p2.fStart, p3.fEnd));
// And confirm that passing in the wrong size for the allocation fails
p = pool->allocate<8>(32);
REPORTER_ASSERT(r, !p.fBlock->release(p.fStart, p.fEnd - 16));
REPORTER_ASSERT(r, !p.fBlock->release(p.fStart, p.fEnd + 16));
REPORTER_ASSERT(r, p.fBlock->release(p.fStart, p.fEnd));
SkDEBUGCODE(pool->validate();)
}
DEF_TEST(SkBlockAllocatorRewind, r) {
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
// Confirm that a bunch of allocations and then releases in stack order fully goes back to the
// start of the block (i.e. unwinds the entire stack, and not just the last cursor position)
SkSBlockAllocator<1024> pool{};
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
SkDEBUGCODE(pool->validate();)
std::vector<SkBlockAllocator::ByteRange> ptrs;
ptrs.reserve(32); // silence clang-tidy performance warning
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
for (int i = 0; i < 32; ++i) {
ptrs.push_back(pool->allocate<4>(16));
}
// Release everything in reverse order
SkDEBUGCODE(pool->validate();)
for (int i = 31; i >= 0; --i) {
auto br = ptrs[i];
REPORTER_ASSERT(r, br.fBlock->release(br.fStart, br.fEnd));
}
// If correct, we've rewound all the way back to the start of the block, so a new allocation
// will have the same location as ptrs[0]
SkDEBUGCODE(pool->validate();)
REPORTER_ASSERT(r, pool->allocate<4>(16).fStart == ptrs[0].fStart);
}
DEF_TEST(SkBlockAllocatorGrowthPolicy, r) {
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
static constexpr int kInitSize = 128;
static constexpr int kBlockCount = 5;
static constexpr size_t kExpectedSizes[SkBlockAllocator::kGrowthPolicyCount][kBlockCount] = {
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
// kFixed -> kInitSize per block
{ kInitSize, kInitSize, kInitSize, kInitSize, kInitSize },
// kLinear -> (block ct + 1) * kInitSize for next block
{ kInitSize, 2 * kInitSize, 3 * kInitSize, 4 * kInitSize, 5 * kInitSize },
// kFibonacci -> 1, 1, 2, 3, 5 * kInitSize for the blocks
{ kInitSize, kInitSize, 2 * kInitSize, 3 * kInitSize, 5 * kInitSize },
// kExponential -> 1, 2, 4, 8, 16 * kInitSize for the blocks
{ kInitSize, 2 * kInitSize, 4 * kInitSize, 8 * kInitSize, 16 * kInitSize },
};
for (int gp = 0; gp < SkBlockAllocator::kGrowthPolicyCount; ++gp) {
SkSBlockAllocator<kInitSize> pool{(GrowthPolicy) gp};
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
SkDEBUGCODE(pool->validate();)
REPORTER_ASSERT(r, kExpectedSizes[gp][0] == total_size(pool));
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
for (int i = 1; i < kBlockCount; ++i) {
REPORTER_ASSERT(r, kExpectedSizes[gp][i] == add_block(pool));
}
SkDEBUGCODE(pool->validate();)
}
}
DEF_TEST(SkBlockAllocatorReset, r) {
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
static constexpr int kBlockIncrement = 1024;
SkSBlockAllocator<kBlockIncrement> pool{GrowthPolicy::kLinear};
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
SkDEBUGCODE(pool->validate();)
void* firstAlloc = alloc_byte(pool);
// Add several blocks
add_block(pool);
add_block(pool);
add_block(pool);
SkDEBUGCODE(pool->validate();)
REPORTER_ASSERT(r, block_count(pool) == 4); // 3 added plus the implicit head
get_block(pool, 0)->setMetadata(2);
// Reset and confirm that there's only one block, a new allocation matches 'firstAlloc' again,
// and new blocks are sized based on a reset growth policy.
pool->reset();
SkDEBUGCODE(pool->validate();)
REPORTER_ASSERT(r,block_count(pool) == 1);
REPORTER_ASSERT(r, pool->preallocSize() == pool->totalSize());
REPORTER_ASSERT(r, get_block(pool, 0)->metadata() == 0);
REPORTER_ASSERT(r, firstAlloc == alloc_byte(pool));
REPORTER_ASSERT(r, 2 * kBlockIncrement == add_block(pool));
REPORTER_ASSERT(r, 3 * kBlockIncrement == add_block(pool));
SkDEBUGCODE(pool->validate();)
}
DEF_TEST(SkBlockAllocatorReleaseBlock, r) {
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
// This loops over all growth policies to make sure that the incremental releases update the
// sequence correctly for each policy.
for (int gp = 0; gp < SkBlockAllocator::kGrowthPolicyCount; ++gp) {
SkSBlockAllocator<1024> pool{(GrowthPolicy) gp};
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
SkDEBUGCODE(pool->validate();)
void* firstAlloc = alloc_byte(pool);
size_t b1Size = total_size(pool);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
size_t b2Size = add_block(pool);
size_t b3Size = add_block(pool);
size_t b4Size = add_block(pool);
SkDEBUGCODE(pool->validate();)
get_block(pool, 0)->setMetadata(1);
get_block(pool, 1)->setMetadata(2);
get_block(pool, 2)->setMetadata(3);
get_block(pool, 3)->setMetadata(4);
// Remove the 3 added blocks, but always remove the i = 1 to test intermediate removal (and
// on the last iteration, will test tail removal).
REPORTER_ASSERT(r, total_size(pool) == b1Size + b2Size + b3Size + b4Size);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
pool->releaseBlock(get_block(pool, 1));
REPORTER_ASSERT(r, block_count(pool) == 3);
REPORTER_ASSERT(r, get_block(pool, 1)->metadata() == 3);
REPORTER_ASSERT(r, total_size(pool) == b1Size + b3Size + b4Size);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
pool->releaseBlock(get_block(pool, 1));
REPORTER_ASSERT(r, block_count(pool) == 2);
REPORTER_ASSERT(r, get_block(pool, 1)->metadata() == 4);
REPORTER_ASSERT(r, total_size(pool) == b1Size + b4Size);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
pool->releaseBlock(get_block(pool, 1));
REPORTER_ASSERT(r, block_count(pool) == 1);
REPORTER_ASSERT(r, total_size(pool) == b1Size);
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
// Since we're back to just the head block, if we add a new block, the growth policy should
// match the original sequence instead of continuing with "b5Size'"
pool->resetScratchSpace();
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
size_t size = add_block(pool);
REPORTER_ASSERT(r, size == b2Size);
pool->releaseBlock(get_block(pool, 1));
// Explicitly release the head block and confirm it's reset
pool->releaseBlock(get_block(pool, 0));
REPORTER_ASSERT(r, total_size(pool) == pool->preallocSize());
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
REPORTER_ASSERT(r, block_count(pool) == 1);
REPORTER_ASSERT(r, firstAlloc == alloc_byte(pool));
REPORTER_ASSERT(r, get_block(pool, 0)->metadata() == 0); // metadata reset too
// Confirm that if we have > 1 block, but release the head block we can still access the
// others
add_block(pool);
add_block(pool);
pool->releaseBlock(get_block(pool, 0));
REPORTER_ASSERT(r, block_count(pool) == 3);
SkDEBUGCODE(pool->validate();)
}
}
DEF_TEST(SkBlockAllocatorIterateAndRelease, r) {
SkSBlockAllocator<256> pool;
pool->headBlock()->setMetadata(1);
add_block(pool);
add_block(pool);
add_block(pool);
// Loop forward and release the blocks
int releaseCount = 0;
for (auto* b : pool->blocks()) {
pool->releaseBlock(b);
releaseCount++;
}
REPORTER_ASSERT(r, releaseCount == 4);
// pool should have just the head block, but was reset
REPORTER_ASSERT(r, pool->headBlock()->metadata() == 0);
REPORTER_ASSERT(r, block_count(pool) == 1);
// Add more blocks
pool->headBlock()->setMetadata(1);
add_block(pool);
add_block(pool);
add_block(pool);
// Loop in reverse and release the blocks
releaseCount = 0;
for (auto* b : pool->rblocks()) {
pool->releaseBlock(b);
releaseCount++;
}
REPORTER_ASSERT(r, releaseCount == 4);
// pool should have just the head block, but was reset
REPORTER_ASSERT(r, pool->headBlock()->metadata() == 0);
REPORTER_ASSERT(r, block_count(pool) == 1);
}
DEF_TEST(SkBlockAllocatorScratchBlockReserve, r) {
SkSBlockAllocator<256> pool;
size_t added = add_block(pool);
REPORTER_ASSERT(r, BlockAllocatorTestAccess::ScratchBlockSize(pool) == 0);
size_t total = pool->totalSize();
pool->releaseBlock(pool->currentBlock());
// Total size shouldn't have changed, the released block should become scratch
REPORTER_ASSERT(r, pool->totalSize() == total);
REPORTER_ASSERT(r, BlockAllocatorTestAccess::ScratchBlockSize(pool) == added);
// But a reset definitely deletes any scratch block
pool->reset();
REPORTER_ASSERT(r, BlockAllocatorTestAccess::ScratchBlockSize(pool) == 0);
// Reserving more than what's available adds a scratch block, and current block remains avail.
size_t avail = pool->currentBlock()->avail();
size_t reserve = avail + 1;
pool->reserve(reserve);
REPORTER_ASSERT(r, (size_t) pool->currentBlock()->avail() == avail);
// And rounds up to the fixed size of this pool's growth policy
REPORTER_ASSERT(r, BlockAllocatorTestAccess::ScratchBlockSize(pool) >= reserve &&
BlockAllocatorTestAccess::ScratchBlockSize(pool) % 256 == 0);
// Allocating more than avail activates the scratch block (so totalSize doesn't change)
size_t preAllocTotalSize = pool->totalSize();
pool->allocate<1>(avail + 1);
REPORTER_ASSERT(r, BlockAllocatorTestAccess::ScratchBlockSize(pool) == 0);
REPORTER_ASSERT(r, pool->totalSize() == preAllocTotalSize);
// When reserving less than what's still available in the current block, no scratch block is
// added.
pool->reserve(pool->currentBlock()->avail());
REPORTER_ASSERT(r, BlockAllocatorTestAccess::ScratchBlockSize(pool) == 0);
// Unless checking available bytes is disabled
pool->reserve(pool->currentBlock()->avail(), SkBlockAllocator::kIgnoreExistingBytes_Flag);
REPORTER_ASSERT(r, BlockAllocatorTestAccess::ScratchBlockSize(pool) > 0);
// If kIgnoreGrowthPolicy is specified, the new scratch block should not have been updated to
// follow the size (which in this case is a fixed 256 bytes per block).
pool->resetScratchSpace();
pool->reserve(32, SkBlockAllocator::kIgnoreGrowthPolicy_Flag);
REPORTER_ASSERT(r, BlockAllocatorTestAccess::ScratchBlockSize(pool) > 0 &&
BlockAllocatorTestAccess::ScratchBlockSize(pool) < 256);
// When requesting an allocation larger than the current block and the scratch block, a new
// block is added, and the scratch block remains scratch.
SkBlockAllocator::Block* oldTail = pool->currentBlock();
avail = oldTail->avail();
size_t scratchAvail = 2 * avail;
pool->reserve(scratchAvail);
REPORTER_ASSERT(r, BlockAllocatorTestAccess::ScratchBlockSize(pool) >= scratchAvail);
// This allocation request is higher than oldTail's available, and the scratch size so we
// should add a new block and scratch size should stay the same.
scratchAvail = BlockAllocatorTestAccess::ScratchBlockSize(pool);
pool->allocate<1>(scratchAvail + 1);
REPORTER_ASSERT(r, pool->currentBlock() != oldTail);
REPORTER_ASSERT(r, BlockAllocatorTestAccess::ScratchBlockSize(pool) == scratchAvail);
}
DEF_TEST(SkBlockAllocatorStealBlocks, r) {
SkSBlockAllocator<256> poolA;
SkSBlockAllocator<128> poolB;
Reland "Support moving blocks from one allocator to another" This reverts commit 423dd689df15c2ae00a05fc70d86b558d44b9686. Reason for revert: had missed corner case when tail block was empty Original change's description: > Revert "Support moving blocks from one allocator to another" > > This reverts commit 0f064041bfd939c65ae63b84b3584c97db2736a0. > > Reason for revert: unit test failures on some windows bots > > Original change's description: > > Support moving blocks from one allocator to another > > > > Uses this new capability and the previously implemented reserve() > > feature to implement efficient concatenation of GrTAllocators. > > > > Change-Id: I2aff42eaf75e74e3b595d3069b6a271fa7329465 > > Reviewed-on: https://skia-review.googlesource.com/c/skia/+/303665 > > Commit-Queue: Michael Ludwig <michaelludwig@google.com> > > Reviewed-by: Robert Phillips <robertphillips@google.com> > > TBR=robertphillips@google.com,csmartdalton@google.com,michaelludwig@google.com > > Change-Id: I931f2462ecf6e04d40a671336d0de7d80efd313d > No-Presubmit: true > No-Tree-Checks: true > No-Try: true > Reviewed-on: https://skia-review.googlesource.com/c/skia/+/304604 > Reviewed-by: Michael Ludwig <michaelludwig@google.com> > Commit-Queue: Michael Ludwig <michaelludwig@google.com> TBR=robertphillips@google.com,csmartdalton@google.com,michaelludwig@google.com # Not skipping CQ checks because this is a reland. Change-Id: Ia793c2f0e7ab2f3fd437871fa7fb4f56979f9ceb Reviewed-on: https://skia-review.googlesource.com/c/skia/+/304739 Auto-Submit: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com> Reviewed-by: Michael Ludwig <michaelludwig@google.com> Commit-Queue: Michael Ludwig <michaelludwig@google.com>
2020-07-21 17:04:42 +00:00
add_block(poolA);
add_block(poolA);
add_block(poolA);
add_block(poolB);
add_block(poolB);
char* bAlloc = (char*) alloc_byte(poolB);
*bAlloc = 't';
const SkBlockAllocator::Block* allocOwner = poolB->findOwningBlock(bAlloc);
Reland "Support moving blocks from one allocator to another" This reverts commit 423dd689df15c2ae00a05fc70d86b558d44b9686. Reason for revert: had missed corner case when tail block was empty Original change's description: > Revert "Support moving blocks from one allocator to another" > > This reverts commit 0f064041bfd939c65ae63b84b3584c97db2736a0. > > Reason for revert: unit test failures on some windows bots > > Original change's description: > > Support moving blocks from one allocator to another > > > > Uses this new capability and the previously implemented reserve() > > feature to implement efficient concatenation of GrTAllocators. > > > > Change-Id: I2aff42eaf75e74e3b595d3069b6a271fa7329465 > > Reviewed-on: https://skia-review.googlesource.com/c/skia/+/303665 > > Commit-Queue: Michael Ludwig <michaelludwig@google.com> > > Reviewed-by: Robert Phillips <robertphillips@google.com> > > TBR=robertphillips@google.com,csmartdalton@google.com,michaelludwig@google.com > > Change-Id: I931f2462ecf6e04d40a671336d0de7d80efd313d > No-Presubmit: true > No-Tree-Checks: true > No-Try: true > Reviewed-on: https://skia-review.googlesource.com/c/skia/+/304604 > Reviewed-by: Michael Ludwig <michaelludwig@google.com> > Commit-Queue: Michael Ludwig <michaelludwig@google.com> TBR=robertphillips@google.com,csmartdalton@google.com,michaelludwig@google.com # Not skipping CQ checks because this is a reland. Change-Id: Ia793c2f0e7ab2f3fd437871fa7fb4f56979f9ceb Reviewed-on: https://skia-review.googlesource.com/c/skia/+/304739 Auto-Submit: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com> Reviewed-by: Michael Ludwig <michaelludwig@google.com> Commit-Queue: Michael Ludwig <michaelludwig@google.com>
2020-07-21 17:04:42 +00:00
REPORTER_ASSERT(r, block_count(poolA) == 4);
REPORTER_ASSERT(r, block_count(poolB) == 3);
size_t aSize = poolA->totalSize();
size_t bSize = poolB->totalSize();
size_t theftSize = bSize - poolB->preallocSize();
// This steal should move B's 2 heap blocks to A, bringing A to 6 and B to just its head
poolA->stealHeapBlocks(poolB.allocator());
REPORTER_ASSERT(r, block_count(poolA) == 6);
REPORTER_ASSERT(r, block_count(poolB) == 1);
REPORTER_ASSERT(r, poolB->preallocSize() == poolB->totalSize());
REPORTER_ASSERT(r, poolA->totalSize() == aSize + theftSize);
REPORTER_ASSERT(r, *bAlloc == 't');
REPORTER_ASSERT(r, (uintptr_t) poolA->findOwningBlock(bAlloc) == (uintptr_t) allocOwner);
REPORTER_ASSERT(r, !poolB->findOwningBlock(bAlloc));
// Redoing the steal now that B is just a head block should be a no-op
poolA->stealHeapBlocks(poolB.allocator());
REPORTER_ASSERT(r, block_count(poolA) == 6);
REPORTER_ASSERT(r, block_count(poolB) == 1);
}
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
// These tests ensure that the allocation padding mechanism works as intended
struct TestMeta {
int fX1;
int fX2;
};
struct alignas(32) TestMetaBig {
int fX1;
int fX2;
};
DEF_TEST(SkBlockAllocatorMetadata, r) {
SkSBlockAllocator<1024> pool{};
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
SkDEBUGCODE(pool->validate();)
// Allocation where alignment of user data > alignment of metadata
SkASSERT(alignof(TestMeta) < 16);
auto p1 = pool->allocate<16, sizeof(TestMeta)>(16);
SkDEBUGCODE(pool->validate();)
REPORTER_ASSERT(r, p1.fAlignedOffset - p1.fStart >= (int) sizeof(TestMeta));
TestMeta* meta = static_cast<TestMeta*>(p1.fBlock->ptr(p1.fAlignedOffset - sizeof(TestMeta)));
// Confirm alignment for both pointers
REPORTER_ASSERT(r, reinterpret_cast<uintptr_t>(meta) % alignof(TestMeta) == 0);
REPORTER_ASSERT(r, reinterpret_cast<uintptr_t>(p1.fBlock->ptr(p1.fAlignedOffset)) % 16 == 0);
// Access fields to make sure 'meta' matches compilers expectations...
meta->fX1 = 2;
meta->fX2 = 5;
// Repeat, but for metadata that has a larger alignment than the allocation
SkASSERT(alignof(TestMetaBig) == 32);
auto p2 = pool->allocate<alignof(TestMetaBig), sizeof(TestMetaBig)>(16);
SkDEBUGCODE(pool->validate();)
REPORTER_ASSERT(r, p2.fAlignedOffset - p2.fStart >= (int) sizeof(TestMetaBig));
TestMetaBig* metaBig = static_cast<TestMetaBig*>(
p2.fBlock->ptr(p2.fAlignedOffset - sizeof(TestMetaBig)));
// Confirm alignment for both pointers
REPORTER_ASSERT(r, reinterpret_cast<uintptr_t>(metaBig) % alignof(TestMetaBig) == 0);
REPORTER_ASSERT(r, reinterpret_cast<uintptr_t>(p2.fBlock->ptr(p2.fAlignedOffset)) % 16 == 0);
// Access fields
metaBig->fX1 = 3;
metaBig->fX2 = 6;
// Ensure metadata values persist after allocations
REPORTER_ASSERT(r, meta->fX1 == 2 && meta->fX2 == 5);
REPORTER_ASSERT(r, metaBig->fX1 == 3 && metaBig->fX2 == 6);
}
DEF_TEST(SkBlockAllocatorAllocatorMetadata, r) {
SkSBlockAllocator<256> pool{};
SkDEBUGCODE(pool->validate();)
REPORTER_ASSERT(r, pool->metadata() == 0); // initial value
pool->setMetadata(4);
REPORTER_ASSERT(r, pool->metadata() == 4);
// Releasing the head block doesn't change the allocator's metadata (even though that's where
// it is stored).
pool->releaseBlock(pool->headBlock());
REPORTER_ASSERT(r, pool->metadata() == 4);
// But resetting the whole allocator brings things back to as if it were newly constructed
pool->reset();
REPORTER_ASSERT(r, pool->metadata() == 0);
}
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
template<size_t Align, size_t Padding>
static void run_owning_block_test(skiatest::Reporter* r, SkBlockAllocator* pool) {
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
auto br = pool->allocate<Align, Padding>(1);
void* userPtr = br.fBlock->ptr(br.fAlignedOffset);
void* metaPtr = br.fBlock->ptr(br.fAlignedOffset - Padding);
Block* block = pool->owningBlock<Align, Padding>(userPtr, br.fStart);
REPORTER_ASSERT(r, block == br.fBlock);
block = pool->owningBlock<Align>(metaPtr, br.fStart);
REPORTER_ASSERT(r, block == br.fBlock);
block = reinterpret_cast<Block*>(reinterpret_cast<uintptr_t>(userPtr) - br.fAlignedOffset);
REPORTER_ASSERT(r, block == br.fBlock);
}
template<size_t Padding>
static void run_owning_block_tests(skiatest::Reporter* r, SkBlockAllocator* pool) {
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
run_owning_block_test<1, Padding>(r, pool);
run_owning_block_test<2, Padding>(r, pool);
run_owning_block_test<4, Padding>(r, pool);
run_owning_block_test<8, Padding>(r, pool);
run_owning_block_test<16, Padding>(r, pool);
run_owning_block_test<32, Padding>(r, pool);
run_owning_block_test<64, Padding>(r, pool);
run_owning_block_test<128, Padding>(r, pool);
}
DEF_TEST(SkBlockAllocatorOwningBlock, r) {
SkSBlockAllocator<1024> pool{};
Refactor GrMemoryPool into reusable GrBlockAllocator This moves the byte block linked list structure outside of GrMemoryPool into a new type, GrBlockAllocator. This new type is solely responsible for managing the byte blocks, tracking where the next allocation occurs, and creating/destroying the byte blocks. It also tries to encapsulate all of/most alignment related math, while making it convenient for clients to add per-allocation padding/metadata. It has added functionality compared to the original block linked list that was embedded in GrMemoryPool: - Supports resetting the entire allocator - Supports resizing the last allocation - Is able to rewind an entire stack of allocations, instead of just the last allocation. - Supports multiple block growth policies instead of just a fixed size. - Supports templated alignment, and variable alignment within a single GrBlockAllocator - Query the amount of available memory - Performs as much math as possible in 32-bit ints with static asserts that ensure overflow won't happen. Some of this flexibility was added so that the GrBlockAllocator can be used to implement an arena allocator similar to SkArenaAlloc, or to replace the use of SkTArray in GrQuadBuffer. It is also likely possible that GrAllocator can be written on top of GrBlockAllocator. I will try to perform these consolidations in later CLs. Change-Id: Ia6c8709642369b66b88eb1bc46264fb2aa9b62ab Reviewed-on: https://skia-review.googlesource.com/c/skia/+/262216 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Robert Phillips <robertphillips@google.com>
2020-03-17 14:14:48 +00:00
SkDEBUGCODE(pool->validate();)
run_owning_block_tests<1>(r, pool.allocator());
run_owning_block_tests<2>(r, pool.allocator());
run_owning_block_tests<4>(r, pool.allocator());
run_owning_block_tests<8>(r, pool.allocator());
run_owning_block_tests<16>(r, pool.allocator());
run_owning_block_tests<32>(r, pool.allocator());
// And some weird numbers
run_owning_block_tests<3>(r, pool.allocator());
run_owning_block_tests<9>(r, pool.allocator());
run_owning_block_tests<17>(r, pool.allocator());
}