Release

[pedroac]

No description provided.

[gemini-code-assist]

Code Review

This pull request updates the sprat-cli suite to version 0.5.1, introducing sprite deduplication, artifact reduction via dilation, and GPU-native texture compression. It refactors the transform system for better metadata control and optimizes core packing algorithms. Feedback focuses on reducing memory overhead during dilation, adopting modern C++20 idioms, ensuring arithmetic safety against overflows, and improving maintainability by replacing magic numbers and manual hash implementations with named constants and shared utilities.

[gemini-code-assist]

std::vector<unsigned char> read_buf = atlas; creates a full copy of the atlas buffer. For large atlases (e.g., 16k x 16k), this can consume a significant amount of memory (1GB for RGBA). Consider using a smaller temporary buffer only for the region being dilated or optimizing the two-buffer approach to avoid a full atlas copy.

[gemini-code-assist]

Since the project is using C++20 features (like std::ranges::sort), it is better to use std::popcount from the <bit> header instead of manual compiler-specific intrinsics. This improves portability and readability.

Suggested change

	#ifdef _MSC_VER
	static inline int popcount64(unsigned long long x) { return (int)__popcnt64(x); }
	#else
	static inline int popcount64(unsigned long long x) { return __builtin_popcountll(x); }
	#endif
	static inline int popcount64(unsigned long long x) { return std::popcount(x); }

[gemini-code-assist]

Please include the <bit> header to support the use of std::popcount.

Suggested change

	#include <unordered_map>
	#include <unordered_set>
	#include <optional>
	#include <numeric>
	#include <cstring>
	#include <unordered_map>
	#include <unordered_set>
	#include <optional>
	#include <numeric>
	#include <cstring>
	#include <bit>

[gemini-code-assist]

This manual compaction loop can be replaced with std::erase_if (available since C++20), which is more idiomatic and concise.

Suggested change

	size_t write = 0;
	for (size_t ri = 0; ri < free_rects.size(); ++ri) {
	if (free_rects[ri].w > 0 && free_rects[ri].h > 0) {
	if (write != ri) {
	free_rects[write] = free_rects[ri];
	}
	++write;
	}
	}
	prune_free_rects(free_rects);
	free_rects.resize(write);
	std::erase_if(free_rects, [](const Rect& r) { return r.w <= 0 || r.h <= 0; });

[gemini-code-assist]

The PR adds src/core/fnv1a.h which provides a reusable fnv1a_hash function. It should be used here instead of re-implementing the algorithm manually with magic numbers.

Suggested change

	uint64_t fnv = 14695981039346656037ULL;
	for (size_t bi = 0; bi < nbytes; ++bi) {
	fnv ^= px[bi];
	fnv *= 1099511628211ULL;
	}
	entry_content_hash = fnv;
	entry_content_hash = sprat::core::fnv1a_hash(px, nbytes);

[gemini-code-assist]

The perceptual deduplication uses an $O(N^2)$ pairwise comparison. While it is gated by sprite dimensions, it could still be a performance bottleneck for large asset sets with many sprites of the same size. Consider using a more efficient spatial indexing structure (like a BK-tree or VP-tree) if performance becomes an issue for large projects.

[gemini-code-assist]

std::function has a small overhead due to type erasure. Since the lambda find is only used locally and is implemented iteratively (not recursively), using auto is more efficient.

Suggested change

	std::function<size_t(size_t)> find = [&](size_t x) -> size_t {
	auto find = [&](size_t x) -> size_t {

[gemini-code-assist]

The calculation of compressed_bytes uses int for width and height. While unlikely given typical atlas limits, width * height could overflow a 32-bit signed integer. It is safer to cast to size_t before multiplication.

Suggested change

	size_t compressed_bytes = (format == "dxt1" || format == "DXT1")
	? (width * height) / 2
	: width * height;
	size_t compressed_bytes = (format == "dxt1" || format == "DXT1")
	? (static_cast<size_t>(width) * height) / 2
	: static_cast<size_t>(width) * height;

[gemini-code-assist]

The DDS header flags use magic numbers. Using named constants (e.g., DDSD_CAPS, DDSD_HEIGHT, etc.) would improve readability and maintainability.