Function decodeBytesSimd

SHTest/Externals/meshoptimizer/vertexcodec.cpp:929–973

#if defined(SIMD_SSE) || defined(SIMD_AVX) || defined(SIMD_NEON) || defined(SIMD_WASM)
SIMD_TARGET
static const unsigned char* decodeBytesSimd(const unsigned char* data, const unsigned char* data_end, unsigned char* buffer, size_t buffer_size)
{
	assert(buffer_size % kByteGroupSize == 0);
	assert(kByteGroupSize == 16);

	const unsigned char* header = data;

	// round number of groups to 4 to get number of header bytes
	size_t header_size = (buffer_size / kByteGroupSize + 3) / 4;

	if (size_t(data_end - data) < header_size)
		return NULL;

	data += header_size;

	size_t i = 0;

	// fast-path: process 4 groups at a time, do a shared bounds check - each group reads <=24b
	for (; i + kByteGroupSize * 4 <= buffer_size && size_t(data_end - data) >= kByteGroupDecodeLimit * 4; i += kByteGroupSize * 4)
	{
		size_t header_offset = i / kByteGroupSize;
		unsigned char header_byte = header[header_offset / 4];

		data = decodeBytesGroupSimd(data, buffer + i + kByteGroupSize * 0, (header_byte >> 0) & 3);
		data = decodeBytesGroupSimd(data, buffer + i + kByteGroupSize * 1, (header_byte >> 2) & 3);
		data = decodeBytesGroupSimd(data, buffer + i + kByteGroupSize * 2, (header_byte >> 4) & 3);
		data = decodeBytesGroupSimd(data, buffer + i + kByteGroupSize * 3, (header_byte >> 6) & 3);
	}

	// slow-path: process remaining groups
	for (; i < buffer_size; i += kByteGroupSize)
	{
		if (size_t(data_end - data) < kByteGroupDecodeLimit)
			return NULL;

		size_t header_offset = i / kByteGroupSize;

		int bitslog2 = (header[header_offset / 4] >> ((header_offset % 4) * 2)) & 3;

		data = decodeBytesGroupSimd(data, buffer + i, bitslog2);
	}

	return data;
}

SIMD_TARGET
static const unsigned char* decodeVertexBlockSimd(const unsigned char* data, const unsigned char* data_end, unsigned char* vertex_data, size_t vertex_count, size_t vertex_size, unsigned char last_vertex[256])