Add direct copy when output and input vertex formats are matching.

cpp/render/vertex_format_opengles.hpp

@@ -71,6 +71,20 @@ struct alignas(4) rk_output_row {
     }
 };
 
+template<typename _output, unsigned _cols, bool _signed, bool _normalized>
+struct alignas(4) rk_output_row<_output, _output, _cols, _signed, _normalized> {
+    rk_output<_output, _output, _signed, _normalized> output_col[_cols];
+
+    [[RK_FAST]]
+    inline void convert(
+        rk_input_row<_output, _cols> const & __restrict src) {
+        static_assert(sizeof(output_col) == sizeof(src.input_col));
+        rk_output<_output, _output, _signed, _normalized> const * const input_col =
+            reinterpret_cast<rk_output<_output, _output, _signed, _normalized> const *>(src.input_col);
+        *output_col = *input_col;
+    }
+};
+
 template<typename _input, unsigned _cols, bool _signed, bool _normalized>
 struct alignas(4) rk_output_row<rk_packed<_signed, _cols>, _input, _cols, _signed, _normalized> {
     rk_output<rk_packed<_signed, _cols>, _input, _signed, _normalized> output_cols;
@@ -100,6 +114,20 @@ struct alignas(4) rk_output_format {
     }
 };
 
+template<typename _output, unsigned _cols, unsigned _rows, bool _signed, bool _normalized>
+struct alignas(4) rk_output_format<_output, _output, _cols, _rows, _signed, _normalized> {
+    rk_output_row<_output, _output, _cols, _signed, _normalized> output_row[_rows];
+
+    [[RK_FAST]]
+    inline void convert(
+        rk_input_format<_output, _cols, _rows> const & __restrict src) {
+        static_assert(sizeof(output_row) == sizeof(src.input_row));
+        rk_output_row<_output, _output, _cols, _signed, _normalized> const * const input_row =
+            reinterpret_cast<rk_output_row<_output, _output, _cols, _signed, _normalized> const *>(src.input_row);
+        *output_row = *input_row;
+    }
+};
+
 template<typename _input, bool _signed, bool _normalized>
 struct alignas(alignof(rk_packed<_signed, 3>)) rk_output<rk_packed<_signed, 3>, _input, _signed, _normalized> {
     rk_packed<_signed, 3> output;
@@ -221,8 +249,8 @@ typedef rk_vertex::rk_format<rk_short, rk_float,           3, 1, true, false> rk
 typedef rk_vertex::rk_format<rk_short,            rk_float, 3, 1, true,  true>  rk_vec3_short_norm;
 typedef rk_vertex::rk_format<rk_packed<true,  3>, rk_float, 3, 1, true,  false> rk_vec3_int10;
 typedef rk_vertex::rk_format<rk_packed<true,  3>, rk_float, 3, 1, true,  true>  rk_vec3_int10_norm;
-typedef rk_vertex::rk_format<rk_packed<true, 3>, rk_float, 3, 1, true, false> rk_vec3_uint10;
+typedef rk_vertex::rk_format<rk_packed<false, 3>, rk_float, 3, 1, false, false> rk_vec3_uint10;
-typedef rk_vertex::rk_format<rk_packed<true, 3>, rk_float, 3, 1, true, true>  rk_vec3_uint10_norm;
+typedef rk_vertex::rk_format<rk_packed<false, 3>, rk_float, 3, 1, false, true>  rk_vec3_uint10_norm;
 typedef rk_vertex::rk_format<rk_float,            rk_float, 3, 3, true,  false> rk_mat3_float;
 typedef rk_vertex::rk_format<rk_packed<true,  3>, rk_float, 3, 3, true,  false> rk_mat3_int10;
 typedef rk_vertex::rk_format<rk_packed<true,  3>, rk_float, 3, 3, true,  true>  rk_mat3_int10_norm;