commit 1.2

source/tricks.cxx

@@ -0,0 +1,64 @@
+module;
+
+#include <simd>
+
+module tricks;
+
+import std;
+
+namespace brc {
+
+std::size_t find_delim(std::string_view string, char target) {
+	constexpr auto width = std::simd::vec<std::uint8_t>::size();
+	const auto needle = std::simd::vec<std::uint8_t>(static_cast<std::uint8_t>(target));
+
+	auto i = std::size_t(0);
+	for (; i + width <= string.size(); i += width) {
+		const auto chunk = std::simd::vec<std::uint8_t>(
+			std::span<const std::uint8_t, width>(reinterpret_cast<const std::uint8_t*>(string.data() + i), width));
+
+		const auto mask = chunk == needle;
+		if (const auto bits = mask.to_bitset(); bits.any()) {
+			return i + std::countr_zero(bits.to_ullong());
+		}
+	}
+
+	for (; i < string.size(); ++i) {
+		if (string[i] == target) return i;
+	}
+	return string.size();
+}
+
+std::inplace_vector<std::size_t, 4> find_delim_batch(std::string_view string, char target) {
+	constexpr auto BatchSize = 4;
+	auto sizes = std::inplace_vector<std::size_t, BatchSize>();
+
+	constexpr auto width = std::simd::vec<std::uint8_t>::size();
+	const auto needle = std::simd::vec<std::uint8_t>(static_cast<std::uint8_t>(target));
+
+	std::size_t i = 0;
+
+	for (; i + width <= string.size() && sizes.size() < BatchSize; i += width) {
+		const auto chunk = std::simd::vec<std::uint8_t>(
+			std::span<const std::uint8_t, width>(reinterpret_cast<const std::uint8_t*>(string.data() + i), width));
+
+		const auto mask = chunk == needle;
+		auto bits = mask.to_bitset().to_ullong();
+
+		while (bits != 0 && sizes.size() < BatchSize) {
+			const auto bit_pos = std::countr_zero(bits);
+			sizes.push_back(i + bit_pos);
+			bits &= bits - 1;
+		}
+	}
+
+	for (; i < string.size() && sizes.size() < BatchSize; ++i) {
+		if (string[i] == target) {
+			sizes.push_back(i);
+		}
+	}
+
+	return sizes;
+}
+
+}  // namespace brc