Octahedral Normal/Tangent Compression

Implementation of Octahedral normal compression into Godot 4.0

core/math/vector3.cpp

@@ -117,6 +117,22 @@ Vector3 Vector3::octahedron_decode(const Vector2 &p_oct) {
 	return n.normalized();
 }
 
+Vector2 Vector3::octahedron_tangent_encode(const float sign) const {
+	Vector2 res = this->octahedron_encode();
+	res.y = res.y * 0.5f + 0.5f;
+	res.y = sign >= 0.0f ? res.y : 1 - res.y;
+	return res;
+}
+
+Vector3 Vector3::octahedron_tangent_decode(const Vector2 &p_oct, float *sign) {
+	Vector2 oct_compressed = p_oct;
+	oct_compressed.y = oct_compressed.y * 2 - 1;
+	*sign = oct_compressed.y >= 0.0f ? 1.0f : -1.0f;
+	oct_compressed.y = Math::abs(oct_compressed.y);
+	Vector3 res = Vector3::octahedron_decode(oct_compressed);
+	return res;
+}
+
 Basis Vector3::outer(const Vector3 &p_with) const {
 	Vector3 row0(x * p_with.x, x * p_with.y, x * p_with.z);
 	Vector3 row1(y * p_with.x, y * p_with.y, y * p_with.z);

core/math/vector3.h

@@ -111,6 +111,8 @@ struct _NO_DISCARD_ Vector3 {
 
 	Vector2 octahedron_encode() const;
 	static Vector3 octahedron_decode(const Vector2 &p_oct);
+	Vector2 octahedron_tangent_encode(const float sign) const;
+	static Vector3 octahedron_tangent_decode(const Vector2 &p_oct, float *sign);
 
 	_FORCE_INLINE_ Vector3 cross(const Vector3 &p_with) const;
 	_FORCE_INLINE_ real_t dot(const Vector3 &p_with) const;

scene/resources/mesh.cpp

@@ -863,27 +863,6 @@ static Mesh::PrimitiveType _old_primitives[7] = {
 };
 #endif // DISABLE_DEPRECATED
 
-// Convert Octahedron-mapped normalized vector back to Cartesian
-// Assumes normalized format (elements of v within range [-1, 1])
-Vector3 _oct_to_norm(const Vector2 v) {
-	Vector3 res(v.x, v.y, 1 - (Math::absf(v.x) + Math::absf(v.y)));
-	float t = MAX(-res.z, 0.0f);
-	res.x += t * -SIGN(res.x);
-	res.y += t * -SIGN(res.y);
-	return res.normalized();
-}
-
-// Convert Octahedron-mapped normalized tangent vector back to Cartesian
-// out_sign provides the direction for the original cartesian tangent
-// Assumes normalized format (elements of v within range [-1, 1])
-Vector3 _oct_to_tangent(const Vector2 v, float *out_sign) {
-	Vector2 v_decompressed = v;
-	v_decompressed.y = Math::absf(v_decompressed.y) * 2 - 1;
-	Vector3 res = _oct_to_norm(v_decompressed);
-	*out_sign = SIGN(v[1]);
-	return res;
-}
-
 void _fix_array_compatibility(const Vector<uint8_t> &p_src, uint32_t p_old_format, uint32_t p_new_format, uint32_t p_elements, Vector<uint8_t> &vertex_data, Vector<uint8_t> &attribute_data, Vector<uint8_t> &skin_data) {
 	uint32_t dst_vertex_stride;
 	uint32_t dst_attribute_stride;
@@ -954,127 +933,93 @@ void _fix_array_compatibility(const Vector<uint8_t> &p_src, uint32_t p_old_forma
 					if ((p_old_format & OLD_ARRAY_COMPRESS_NORMAL) && (p_old_format & OLD_ARRAY_FORMAT_TANGENT) && (p_old_format & OLD_ARRAY_COMPRESS_TANGENT)) {
 						for (uint32_t i = 0; i < p_elements; i++) {
 							const int8_t *src = (const int8_t *)&src_vertex_ptr[i * src_vertex_stride + src_offset];
-							uint32_t *dst = (uint32_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_NORMAL]];
-							const Vector2 src_vec(src[0] / 127.0f, src[1] / 127.0f);
-
-							const Vector3 res = _oct_to_norm(src_vec) * Vector3(0.5, 0.5, 0.5) + Vector3(0.5, 0.5, 0.5);
-							*dst = 0;
-							*dst |= CLAMP(int(res.x * 1023.0f), 0, 1023);
-							*dst |= CLAMP(int(res.y * 1023.0f), 0, 1023) << 10;
-							*dst |= CLAMP(int(res.z * 1023.0f), 0, 1023) << 20;
+							int16_t *dst = (int16_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_NORMAL]];
+
+							dst[0] = (int16_t)CLAMP(src[0] / 127.0f * 32767, -32768, 32767);
+							dst[1] = (int16_t)CLAMP(src[1] / 127.0f * 32767, -32768, 32767);
 						}
-						src_offset += sizeof(int8_t) * 2;
+						src_offset += sizeof(int16_t) * 2;
 					} else {
 						for (uint32_t i = 0; i < p_elements; i++) {
 							const int16_t *src = (const int16_t *)&src_vertex_ptr[i * src_vertex_stride + src_offset];
-							uint32_t *dst = (uint32_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_NORMAL]];
-							const Vector2 src_vec(src[0] / 32767.0f, src[1] / 32767.0f);
-
-							const Vector3 res = _oct_to_norm(src_vec) * Vector3(0.5, 0.5, 0.5) + Vector3(0.5, 0.5, 0.5);
-							*dst = 0;
-							*dst |= CLAMP(int(res.x * 1023.0f), 0, 1023);
-							*dst |= CLAMP(int(res.y * 1023.0f), 0, 1023) << 10;
-							*dst |= CLAMP(int(res.z * 1023.0f), 0, 1023) << 20;
+							int16_t *dst = (int16_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_NORMAL]];
+
+							dst[0] = src[0];
+							dst[1] = src[1];
 						}
 						src_offset += sizeof(int16_t) * 2;
 					}
 				} else { // No Octahedral compression
 					if (p_old_format & OLD_ARRAY_COMPRESS_NORMAL) {
-						const float multiplier = 1.f / 127.f * 1023.0f;
-
 						for (uint32_t i = 0; i < p_elements; i++) {
 							const int8_t *src = (const int8_t *)&src_vertex_ptr[i * src_vertex_stride + src_offset];
-							uint32_t *dst = (uint32_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_NORMAL]];
+							const Vector3 original_normal(src[0], src[1], src[2]);
+							Vector2 res = original_normal.octahedron_encode();
 
-							*dst = 0;
-							*dst |= CLAMP(int(src[0] * multiplier), 0, 1023);
-							*dst |= CLAMP(int(src[1] * multiplier), 0, 1023) << 10;
-							*dst |= CLAMP(int(src[2] * multiplier), 0, 1023) << 20;
+							uint16_t *dst = (uint16_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_NORMAL]];
+							dst[0] = (uint16_t)CLAMP(res.x * 65535, 0, 65535);
+							dst[1] = (uint16_t)CLAMP(res.y * 65535, 0, 65535);
 						}
-						src_offset += sizeof(uint32_t);
+						src_offset += sizeof(uint16_t) * 2;
 					} else {
 						for (uint32_t i = 0; i < p_elements; i++) {
 							const float *src = (const float *)&src_vertex_ptr[i * src_vertex_stride + src_offset];
-							uint32_t *dst = (uint32_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_NORMAL]];
+							const Vector3 original_normal(src[0], src[1], src[2]);
+							Vector2 res = original_normal.octahedron_encode();
 
-							*dst = 0;
-							*dst |= CLAMP(int(src[0] * 1023.0), 0, 1023);
-							*dst |= CLAMP(int(src[1] * 1023.0), 0, 1023) << 10;
-							*dst |= CLAMP(int(src[2] * 1023.0), 0, 1023) << 20;
+							uint16_t *dst = (uint16_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_NORMAL]];
+							dst[0] = (uint16_t)CLAMP(res.x * 65535, 0, 65535);
+							dst[1] = (uint16_t)CLAMP(res.y * 65535, 0, 65535);
 						}
-						src_offset += sizeof(float) * 3;
+						src_offset += sizeof(uint16_t) * 2;
 					}
 				}
 
 			} break;
 			case OLD_ARRAY_TANGENT: {
 				if (p_old_format & OLD_ARRAY_FLAG_USE_OCTAHEDRAL_COMPRESSION) {
-					if (p_old_format & OLD_ARRAY_COMPRESS_TANGENT) { // int8
+					if (p_old_format & OLD_ARRAY_COMPRESS_TANGENT) { // int8 SNORM -> uint16 UNORM
 						for (uint32_t i = 0; i < p_elements; i++) {
 							const int8_t *src = (const int8_t *)&src_vertex_ptr[i * src_vertex_stride + src_offset];
-							uint32_t *dst = (uint32_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_TANGENT]];
-							const Vector2 src_vec(src[0] / 127.0f, src[1] / 127.0f);
-							float out_sign;
-							const Vector3 res = _oct_to_tangent(src_vec, &out_sign) * Vector3(0.5, 0.5, 0.5) + Vector3(0.5, 0.5, 0.5);
-
-							*dst = 0;
-							*dst |= CLAMP(int(res.x * 1023.0), 0, 1023);
-							*dst |= CLAMP(int(res.y * 1023.0), 0, 1023) << 10;
-							*dst |= CLAMP(int(res.z * 1023.0), 0, 1023) << 20;
-							if (out_sign > 0) {
-								*dst |= 3 << 30;
-							}
+							uint16_t *dst = (uint16_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_TANGENT]];
+
+							dst[0] = (uint16_t)CLAMP((src[0] / 127.0f * .5f + .5f) * 65535, 0, 65535);
+							dst[1] = (uint16_t)CLAMP((src[1] / 127.0f * .5f + .5f) * 65535, 0, 65535);
 						}
-						src_offset += sizeof(int8_t) * 2;
-					} else { // int16
+						src_offset += sizeof(uint16_t) * 2;
+					} else { // int16 SNORM -> uint16 UNORM
 						for (uint32_t i = 0; i < p_elements; i++) {
 							const int16_t *src = (const int16_t *)&src_vertex_ptr[i * src_vertex_stride + src_offset];
-							uint32_t *dst = (uint32_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_TANGENT]];
-							const Vector2 src_vec(src[0] / 32767.0f, src[1] / 32767.0f);
-							float out_sign;
-							Vector3 res = _oct_to_tangent(src_vec, &out_sign) * Vector3(0.5, 0.5, 0.5) + Vector3(0.5, 0.5, 0.5);
-
-							*dst = 0;
-							*dst |= CLAMP(int(res.x * 1023.0), 0, 1023);
-							*dst |= CLAMP(int(res.y * 1023.0), 0, 1023) << 10;
-							*dst |= CLAMP(int(res.z * 1023.0), 0, 1023) << 20;
-							if (out_sign > 0) {
-								*dst |= 3 << 30;
-							}
+							uint16_t *dst = (uint16_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_TANGENT]];
+
+							dst[0] = (uint16_t)CLAMP((src[0] / 32767.0f * .5f + .5f) * 65535, 0, 65535);
+							dst[1] = (uint16_t)CLAMP((src[1] / 32767.0f * .5f + .5f) * 65535, 0, 65535);
 						}
-						src_offset += sizeof(int16_t) * 2;
+						src_offset += sizeof(uint16_t) * 2;
 					}
 				} else { // No Octahedral compression
 					if (p_old_format & OLD_ARRAY_COMPRESS_TANGENT) {
-						const float multiplier = 1.f / 127.f * 1023.0f;
-
 						for (uint32_t i = 0; i < p_elements; i++) {
 							const int8_t *src = (const int8_t *)&src_vertex_ptr[i * src_vertex_stride + src_offset];
-							uint32_t *dst = (uint32_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_TANGENT]];
-
-							*dst = 0;
-							*dst |= CLAMP(int(src[0] * multiplier), 0, 1023);
-							*dst |= CLAMP(int(src[1] * multiplier), 0, 1023) << 10;
-							*dst |= CLAMP(int(src[2] * multiplier), 0, 1023) << 20;
-							if (src[3] > 0) {
-								*dst |= 3 << 30;
-							}
+							const Vector3 original_tangent(src[0], src[1], src[2]);
+							Vector2 res = original_tangent.octahedron_tangent_encode(src[3]);
+
+							uint16_t *dst = (uint16_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_NORMAL]];
+							dst[0] = (uint16_t)CLAMP(res.x * 65535, 0, 65535);
+							dst[1] = (uint16_t)CLAMP(res.y * 65535, 0, 65535);
 						}
-						src_offset += sizeof(uint32_t);
+						src_offset += sizeof(uint16_t) * 2;
 					} else {
 						for (uint32_t i = 0; i < p_elements; i++) {
 							const float *src = (const float *)&src_vertex_ptr[i * src_vertex_stride + src_offset];
-							uint32_t *dst = (uint32_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_TANGENT]];
-
-							*dst = 0;
-							*dst |= CLAMP(int(src[0] * 1023.0), 0, 1023);
-							*dst |= CLAMP(int(src[1] * 1023.0), 0, 1023) << 10;
-							*dst |= CLAMP(int(src[2] * 1023.0), 0, 1023) << 20;
-							if (src[3] > 0) {
-								*dst |= 3 << 30;
-							}
+							const Vector3 original_tangent(src[0], src[1], src[2]);
+							Vector2 res = original_tangent.octahedron_tangent_encode(src[3]);
+
+							uint16_t *dst = (uint16_t *)&dst_vertex_ptr[i * dst_vertex_stride + dst_offsets[Mesh::ARRAY_NORMAL]];
+							dst[0] = (uint16_t)CLAMP(res.x * 65535, 0, 65535);
+							dst[1] = (uint16_t)CLAMP(res.y * 65535, 0, 65535);
 						}
-						src_offset += sizeof(float) * 4;
+						src_offset += sizeof(uint16_t) * 2;
 					}
 				}
 			} break;

servers/rendering/renderer_rd/shaders/scene_forward_clustered.glsl

@@ -15,11 +15,11 @@ layout(location = 0) in vec3 vertex_attrib;
 //only for pure render depth when normal is not used
 
 #ifdef NORMAL_USED
-layout(location = 1) in vec3 normal_attrib;
+layout(location = 1) in vec2 normal_attrib;
 #endif
 
 #if defined(TANGENT_USED) || defined(NORMAL_MAP_USED) || defined(LIGHT_ANISOTROPY_USED)
-layout(location = 2) in vec4 tangent_attrib;
+layout(location = 2) in vec2 tangent_attrib;
 #endif
 
 #if defined(COLOR_USED)
@@ -58,6 +58,13 @@ layout(location = 10) in uvec4 bone_attrib;
 layout(location = 11) in vec4 weight_attrib;
 #endif
 
+vec3 oct_to_vec3(vec2 e) {
+	vec3 v = vec3(e.xy, 1.0 - abs(e.x) - abs(e.y));
+	float t = max(-v.z, 0.0);
+	v.xy += t * -sign(v.xy);
+	return v;
+}
+
 /* Varyings */
 
 layout(location = 0) out vec3 vertex_interp;
@@ -231,12 +238,13 @@ void vertex_shader(in uint instance_index, in bool is_multimesh, in SceneData sc
 
 	vec3 vertex = vertex_attrib;
 #ifdef NORMAL_USED
-	vec3 normal = normal_attrib * 2.0 - 1.0;
+	vec3 normal = oct_to_vec3(normal_attrib * 2.0 - 1.0);
 #endif
 
 #if defined(TANGENT_USED) || defined(NORMAL_MAP_USED) || defined(LIGHT_ANISOTROPY_USED)
-	vec3 tangent = tangent_attrib.xyz * 2.0 - 1.0;
-	float binormalf = tangent_attrib.a * 2.0 - 1.0;
+	vec2 signed_tangent_attrib = tangent_attrib * 2.0 - 1.0;
+	vec3 tangent = oct_to_vec3(vec2(signed_tangent_attrib.x, abs(signed_tangent_attrib.y) * 2.0 - 1.0));
+	float binormalf = sign(signed_tangent_attrib.y);
 	vec3 binormal = normalize(cross(normal, tangent) * binormalf);
 #endif
 

servers/rendering/renderer_rd/shaders/scene_forward_mobile.glsl

@@ -16,11 +16,11 @@ layout(location = 0) in vec3 vertex_attrib;
 //only for pure render depth when normal is not used
 
 #ifdef NORMAL_USED
-layout(location = 1) in vec3 normal_attrib;
+layout(location = 1) in vec2 normal_attrib;
 #endif
 
 #if defined(TANGENT_USED) || defined(NORMAL_MAP_USED) || defined(LIGHT_ANISOTROPY_USED)
-layout(location = 2) in vec4 tangent_attrib;
+layout(location = 2) in vec2 tangent_attrib;
 #endif
 
 #if defined(COLOR_USED)
@@ -59,6 +59,13 @@ layout(location = 10) in uvec4 bone_attrib;
 layout(location = 11) in vec4 weight_attrib;
 #endif
 
+vec3 oct_to_vec3(vec2 e) {
+	vec3 v = vec3(e.xy, 1.0 - abs(e.x) - abs(e.y));
+	float t = max(-v.z, 0.0);
+	v.xy += t * -sign(v.xy);
+	return v;
+}
+
 /* Varyings */
 
 layout(location = 0) highp out vec3 vertex_interp;
@@ -229,12 +236,13 @@ void main() {
 
 	vec3 vertex = vertex_attrib;
 #ifdef NORMAL_USED
-	vec3 normal = normal_attrib * 2.0 - 1.0;
+	vec3 normal = oct_to_vec3(normal_attrib * 2.0 - 1.0);
 #endif
 
 #if defined(TANGENT_USED) || defined(NORMAL_MAP_USED) || defined(LIGHT_ANISOTROPY_USED)
-	vec3 tangent = tangent_attrib.xyz * 2.0 - 1.0;
-	float binormalf = tangent_attrib.a * 2.0 - 1.0;
+	vec3 signed_tangent_attrib = tangent_attrib * 2.0 - 1.0;
+	vec3 tangent = oct_to_vec3(vec2(signed_tangent_attrib.x, abs(signed_tangent_attrib.y) * 2.0 - 1.0));
+	float binormalf = sign(signed_tangent_attrib.y);
 	vec3 binormal = normalize(cross(normal, tangent) * binormalf);
 #endif
 

servers/rendering/renderer_rd/storage_rd/mesh_storage.cpp

@@ -1057,10 +1057,9 @@ void MeshStorage::_mesh_surface_generate_version_for_input_mask(Mesh::Surface::V
 				} break;
 				case RS::ARRAY_NORMAL: {
 					vd.offset = stride;
+					vd.format = RD::DATA_FORMAT_R16G16_UNORM;
+					stride += sizeof(uint16_t) * 2;
 
-					vd.format = RD::DATA_FORMAT_A2B10G10R10_UNORM_PACK32;
-
-					stride += sizeof(uint32_t);
 					if (mis) {
 						buffer = mis->vertex_buffer;
 					} else {
@@ -1069,9 +1068,9 @@ void MeshStorage::_mesh_surface_generate_version_for_input_mask(Mesh::Surface::V
 				} break;
 				case RS::ARRAY_TANGENT: {
 					vd.offset = stride;
+					vd.format = RD::DATA_FORMAT_R16G16_UNORM;
+					stride += sizeof(uint16_t) * 2;
 
-					vd.format = RD::DATA_FORMAT_A2B10G10R10_UNORM_PACK32;
-					stride += sizeof(uint32_t);
 					if (mis) {
 						buffer = mis->vertex_buffer;
 					} else {

servers/rendering_server.cpp

@@ -398,16 +398,14 @@ Error RenderingServer::_surface_set_data(Array p_arrays, uint32_t p_format, uint
 
 				const Vector3 *src = array.ptr();
 				for (int i = 0; i < p_vertex_array_len; i++) {
-					Vector3 n = src[i] * Vector3(0.5, 0.5, 0.5) + Vector3(0.5, 0.5, 0.5);
-
-					uint32_t value = 0;
-					value |= CLAMP(int(n.x * 1023.0), 0, 1023);
-					value |= CLAMP(int(n.y * 1023.0), 0, 1023) << 10;
-					value |= CLAMP(int(n.z * 1023.0), 0, 1023) << 20;
+					Vector2 res = src[i].octahedron_encode();
+					int16_t vector[2] = {
+						(int16_t)CLAMP(res.x * 65535, 0, 65535),
+						(int16_t)CLAMP(res.y * 65535, 0, 65535),
+					};
 
-					memcpy(&vw[p_offsets[ai] + i * p_vertex_stride], &value, 4);
+					memcpy(&vw[p_offsets[ai] + i * p_vertex_stride], vector, 4);
 				}
-
 			} break;
 
 			case RS::ARRAY_TANGENT: {
@@ -416,33 +414,32 @@ Error RenderingServer::_surface_set_data(Array p_arrays, uint32_t p_format, uint
 				if (type == Variant::PACKED_FLOAT32_ARRAY) {
 					Vector<float> array = p_arrays[ai];
 					ERR_FAIL_COND_V(array.size() != p_vertex_array_len * 4, ERR_INVALID_PARAMETER);
-					const float *src = array.ptr();
+					const float *src_ptr = array.ptr();
 
 					for (int i = 0; i < p_vertex_array_len; i++) {
-						uint32_t value = 0;
-						value |= CLAMP(int((src[i * 4 + 0] * 0.5 + 0.5) * 1023.0), 0, 1023);
-						value |= CLAMP(int((src[i * 4 + 1] * 0.5 + 0.5) * 1023.0), 0, 1023) << 10;
-						value |= CLAMP(int((src[i * 4 + 2] * 0.5 + 0.5) * 1023.0), 0, 1023) << 20;
-						if (src[i * 4 + 3] > 0) {
-							value |= 3UL << 30;
-						}
-
-						memcpy(&vw[p_offsets[ai] + i * p_vertex_stride], &value, 4);
+						const Vector3 src(src_ptr[i * 4 + 0], src_ptr[i * 4 + 1], src_ptr[i * 4 + 2]);
+						Vector2 res = src.octahedron_tangent_encode(src_ptr[i * 4 + 3]);
+						int16_t vector[2] = {
+							(int16_t)CLAMP(res.x * 65535, 0, 65535),
+							(int16_t)CLAMP(res.y * 65535, 0, 65535),
+						};
+
+						memcpy(&vw[p_offsets[ai] + i * p_vertex_stride], vector, 4);
 					}
 				} else { // PACKED_FLOAT64_ARRAY
 					Vector<double> array = p_arrays[ai];
 					ERR_FAIL_COND_V(array.size() != p_vertex_array_len * 4, ERR_INVALID_PARAMETER);
-					const double *src = array.ptr();
+					const double *src_ptr = array.ptr();
 
 					for (int i = 0; i < p_vertex_array_len; i++) {
-						uint32_t value = 0;
-						value |= CLAMP(int((src[i * 4 + 0] * 0.5 + 0.5) * 1023.0), 0, 1023);
-						value |= CLAMP(int((src[i * 4 + 1] * 0.5 + 0.5) * 1023.0), 0, 1023) << 10;
-						value |= CLAMP(int((src[i * 4 + 2] * 0.5 + 0.5) * 1023.0), 0, 1023) << 20;
-						if (src[i * 4 + 3] > 0) {
-							value |= 3UL << 30;
-						}
-						memcpy(&vw[p_offsets[ai] + i * p_vertex_stride], &value, 4);
+						const Vector3 src(src_ptr[i * 4 + 0], src_ptr[i * 4 + 1], src_ptr[i * 4 + 2]);
+						Vector2 res = src.octahedron_tangent_encode(src_ptr[i * 4 + 3]);
+						int16_t vector[2] = {
+							(int16_t)CLAMP(res.x * 65535, 0, 65535),
+							(int16_t)CLAMP(res.y * 65535, 0, 65535),
+						};
+
+						memcpy(&vw[p_offsets[ai] + i * p_vertex_stride], vector, 4);
 					}
 				}
 			} break;