Optimize half and float image conversion

1 files changed, 105 insertions, 1 deletions

diff --git a/core/io/image.cpp b/core/io/image.cpp
index 5498b448d7..6e638d2e99 100644
--- a/core/io/image.cpp
+++ b/core/io/image.cpp
@@ -508,6 +508,38 @@ static void _convert(int p_width, int p_height, const uint8_t *p_src, uint8_t *p
 	}
 }
 
+template <typename T, uint32_t read_channels, uint32_t write_channels, T def_zero, T def_one>
+static void _convert_fast(int p_width, int p_height, const T *p_src, T *p_dst) {
+	uint32_t dst_count = 0;
+	uint32_t src_count = 0;
+
+	const int resolution = p_width * p_height;
+
+	for (int i = 0; i < resolution; i++) {
+		memcpy(p_dst + dst_count, p_src + src_count, MIN(read_channels, write_channels) * sizeof(T));
+
+		if constexpr (write_channels > read_channels) {
+			const T def_value[4] = { def_zero, def_zero, def_zero, def_one };
+			memcpy(p_dst + dst_count + read_channels, &def_value[read_channels], (write_channels - read_channels) * sizeof(T));
+		}
+
+		dst_count += write_channels;
+		src_count += read_channels;
+	}
+}
+
+static bool _are_formats_compatible(Image::Format p_format0, Image::Format p_format1) {
+	if (p_format0 <= Image::FORMAT_RGBA8 && p_format1 <= Image::FORMAT_RGBA8) {
+		return true;
+	} else if (p_format0 <= Image::FORMAT_RGBAH && p_format0 >= Image::FORMAT_RH && p_format1 <= Image::FORMAT_RGBAH && p_format1 >= Image::FORMAT_RH) {
+		return true;
+	} else if (p_format0 <= Image::FORMAT_RGBAF && p_format0 >= Image::FORMAT_RF && p_format1 <= Image::FORMAT_RGBAF && p_format1 >= Image::FORMAT_RF) {
+		return true;
+	}
+
+	return false;
+}
+
 void Image::convert(Format p_new_format) {
 	ERR_FAIL_INDEX_MSG(p_new_format, FORMAT_MAX, "The Image format specified (" + itos(p_new_format) + ") is out of range. See Image's Format enum.");
 	if (data.size() == 0) {
@@ -524,7 +556,7 @@ void Image::convert(Format p_new_format) {
 	if (Image::is_format_compressed(format) || Image::is_format_compressed(p_new_format)) {
 		ERR_FAIL_MSG("Cannot convert to <-> from compressed formats. Use compress() and decompress() instead.");
 
-	} else if (format > FORMAT_RGBA8 || p_new_format > FORMAT_RGBA8) {
+	} else if (!_are_formats_compatible(format, p_new_format)) {
 		//use put/set pixel which is slower but works with non byte formats
 		Image new_img(width, height, mipmaps, p_new_format);
 
@@ -655,6 +687,78 @@ void Image::convert(Format p_new_format) {
 			case FORMAT_RGBA8 | (FORMAT_RGB8 << 8):
 				_convert<3, true, 3, false, false, false>(mip_width, mip_height, rptr, wptr);
 				break;
+			case FORMAT_RH | (FORMAT_RGH << 8):
+				_convert_fast<uint16_t, 1, 2, 0x0000, 0x3C00>(mip_width, mip_height, (const uint16_t *)rptr, (uint16_t *)wptr);
+				break;
+			case FORMAT_RH | (FORMAT_RGBH << 8):
+				_convert_fast<uint16_t, 1, 3, 0x0000, 0x3C00>(mip_width, mip_height, (const uint16_t *)rptr, (uint16_t *)wptr);
+				break;
+			case FORMAT_RH | (FORMAT_RGBAH << 8):
+				_convert_fast<uint16_t, 1, 4, 0x0000, 0x3C00>(mip_width, mip_height, (const uint16_t *)rptr, (uint16_t *)wptr);
+				break;
+			case FORMAT_RGH | (FORMAT_RH << 8):
+				_convert_fast<uint16_t, 2, 1, 0x0000, 0x3C00>(mip_width, mip_height, (const uint16_t *)rptr, (uint16_t *)wptr);
+				break;
+			case FORMAT_RGH | (FORMAT_RGBH << 8):
+				_convert_fast<uint16_t, 2, 3, 0x0000, 0x3C00>(mip_width, mip_height, (const uint16_t *)rptr, (uint16_t *)wptr);
+				break;
+			case FORMAT_RGH | (FORMAT_RGBAH << 8):
+				_convert_fast<uint16_t, 2, 4, 0x0000, 0x3C00>(mip_width, mip_height, (const uint16_t *)rptr, (uint16_t *)wptr);
+				break;
+			case FORMAT_RGBH | (FORMAT_RH << 8):
+				_convert_fast<uint16_t, 3, 1, 0x0000, 0x3C00>(mip_width, mip_height, (const uint16_t *)rptr, (uint16_t *)wptr);
+				break;
+			case FORMAT_RGBH | (FORMAT_RGH << 8):
+				_convert_fast<uint16_t, 3, 2, 0x0000, 0x3C00>(mip_width, mip_height, (const uint16_t *)rptr, (uint16_t *)wptr);
+				break;
+			case FORMAT_RGBH | (FORMAT_RGBAH << 8):
+				_convert_fast<uint16_t, 3, 4, 0x0000, 0x3C00>(mip_width, mip_height, (const uint16_t *)rptr, (uint16_t *)wptr);
+				break;
+			case FORMAT_RGBAH | (FORMAT_RH << 8):
+				_convert_fast<uint16_t, 4, 1, 0x0000, 0x3C00>(mip_width, mip_height, (const uint16_t *)rptr, (uint16_t *)wptr);
+				break;
+			case FORMAT_RGBAH | (FORMAT_RGH << 8):
+				_convert_fast<uint16_t, 4, 2, 0x0000, 0x3C00>(mip_width, mip_height, (const uint16_t *)rptr, (uint16_t *)wptr);
+				break;
+			case FORMAT_RGBAH | (FORMAT_RGBH << 8):
+				_convert_fast<uint16_t, 4, 3, 0x0000, 0x3C00>(mip_width, mip_height, (const uint16_t *)rptr, (uint16_t *)wptr);
+				break;
+			case FORMAT_RF | (FORMAT_RGF << 8):
+				_convert_fast<uint32_t, 1, 2, 0x00000000, 0x3F800000>(mip_width, mip_height, (const uint32_t *)rptr, (uint32_t *)wptr);
+				break;
+			case FORMAT_RF | (FORMAT_RGBF << 8):
+				_convert_fast<uint32_t, 1, 3, 0x00000000, 0x3F800000>(mip_width, mip_height, (const uint32_t *)rptr, (uint32_t *)wptr);
+				break;
+			case FORMAT_RF | (FORMAT_RGBAF << 8):
+				_convert_fast<uint32_t, 1, 4, 0x00000000, 0x3F800000>(mip_width, mip_height, (const uint32_t *)rptr, (uint32_t *)wptr);
+				break;
+			case FORMAT_RGF | (FORMAT_RF << 8):
+				_convert_fast<uint32_t, 2, 1, 0x00000000, 0x3F800000>(mip_width, mip_height, (const uint32_t *)rptr, (uint32_t *)wptr);
+				break;
+			case FORMAT_RGF | (FORMAT_RGBF << 8):
+				_convert_fast<uint32_t, 2, 3, 0x00000000, 0x3F800000>(mip_width, mip_height, (const uint32_t *)rptr, (uint32_t *)wptr);
+				break;
+			case FORMAT_RGF | (FORMAT_RGBAF << 8):
+				_convert_fast<uint32_t, 2, 4, 0x00000000, 0x3F800000>(mip_width, mip_height, (const uint32_t *)rptr, (uint32_t *)wptr);
+				break;
+			case FORMAT_RGBF | (FORMAT_RF << 8):
+				_convert_fast<uint32_t, 3, 1, 0x00000000, 0x3F800000>(mip_width, mip_height, (const uint32_t *)rptr, (uint32_t *)wptr);
+				break;
+			case FORMAT_RGBF | (FORMAT_RGF << 8):
+				_convert_fast<uint32_t, 3, 2, 0x00000000, 0x3F800000>(mip_width, mip_height, (const uint32_t *)rptr, (uint32_t *)wptr);
+				break;
+			case FORMAT_RGBF | (FORMAT_RGBAF << 8):
+				_convert_fast<uint32_t, 3, 4, 0x00000000, 0x3F800000>(mip_width, mip_height, (const uint32_t *)rptr, (uint32_t *)wptr);
+				break;
+			case FORMAT_RGBAF | (FORMAT_RF << 8):
+				_convert_fast<uint32_t, 4, 1, 0x00000000, 0x3F800000>(mip_width, mip_height, (const uint32_t *)rptr, (uint32_t *)wptr);
+				break;
+			case FORMAT_RGBAF | (FORMAT_RGF << 8):
+				_convert_fast<uint32_t, 4, 2, 0x00000000, 0x3F800000>(mip_width, mip_height, (const uint32_t *)rptr, (uint32_t *)wptr);
+				break;
+			case FORMAT_RGBAF | (FORMAT_RGBF << 8):
+				_convert_fast<uint32_t, 4, 3, 0x00000000, 0x3F800000>(mip_width, mip_height, (const uint32_t *)rptr, (uint32_t *)wptr);
+				break;
 		}
 	}