代码改变世界

unity shader矩阵相关优化

2020-04-18 17:07  kk20161206  阅读(532)  评论(0编辑  收藏  举报

https://chengkehan.github.io/MulOptimizing.html

总结:推荐用o.vertex = UnityObjectToClipPos(v.vertex); 而不是o.vertex = mul(UNITY_MATRIX_MVP, v.vertex);

因为后者会拆分为:

当开启 GPU Instancing 时,UNITY_MATRIX_MVP 会被解释为 mul(UNITY_MATRIX_VP, unity_ObjectToWorld),而 unity_ObjectToWorld 又会被解释为 unity_ObjectToWorldArray[unity_InstanceID]。这样 UNITY_MATRIX_MVP 不再是一个 uniform 常量,而变成了两个矩阵相乘,所以就需要像最开始的代码那样进行优化。

因为:

定义两个矩阵:

uniform float4x4 mat0;
uniform float4x4 mat1;

利用这两个的累积变换,对一个点进行变换操作,可以是类似下面这样:

o.vertex = mul(mat0, mul(mat1, v.vertex));

也可以是这样:

o.vertex = mul(mul(mat0, mat1), v.vertex);

这两种方式得到的结果是一样的,但是计算量却相差很大。最简单的验证方式就是手工计算一遍,就能体会到了。另一种更直观的方式,可以看一下对应的 glsl。

第一种:

// cg
o.vertex = mul(mat0, mul(mat1, v.vertex));
// glsl
u_xlat0 = in_POSITION0.yyyy * hlslcc_mtx4x4mat1[1];
u_xlat0 = hlslcc_mtx4x4mat1[0] * in_POSITION0.xxxx + u_xlat0;
u_xlat0 = hlslcc_mtx4x4mat1[2] * in_POSITION0.zzzz + u_xlat0;
u_xlat0 = hlslcc_mtx4x4mat1[3] * in_POSITION0.wwww + u_xlat0;
u_xlat1 = u_xlat0.yyyy * hlslcc_mtx4x4mat0[1];
u_xlat1 = hlslcc_mtx4x4mat0[0] * u_xlat0.xxxx + u_xlat1;
u_xlat1 = hlslcc_mtx4x4mat0[2] * u_xlat0.zzzz + u_xlat1;
gl_Position = hlslcc_mtx4x4mat0[3] * u_xlat0.wwww + u_xlat1;

第二种:

// cg
o.vertex = mul(mul(mat0, mat1), v.vertex);
// glsl
u_xlat0 = hlslcc_mtx4x4mat0[1] * hlslcc_mtx4x4mat1[1].yyyy;
u_xlat0 = hlslcc_mtx4x4mat0[0] * hlslcc_mtx4x4mat1[1].xxxx + u_xlat0;
u_xlat0 = hlslcc_mtx4x4mat0[2] * hlslcc_mtx4x4mat1[1].zzzz + u_xlat0;
u_xlat0 = hlslcc_mtx4x4mat0[3] * hlslcc_mtx4x4mat1[1].wwww + u_xlat0;
u_xlat0 = u_xlat0 * in_POSITION0.yyyy;
u_xlat1 = hlslcc_mtx4x4mat0[1] * hlslcc_mtx4x4mat1[0].yyyy;
u_xlat1 = hlslcc_mtx4x4mat0[0] * hlslcc_mtx4x4mat1[0].xxxx + u_xlat1;
u_xlat1 = hlslcc_mtx4x4mat0[2] * hlslcc_mtx4x4mat1[0].zzzz + u_xlat1;
u_xlat1 = hlslcc_mtx4x4mat0[3] * hlslcc_mtx4x4mat1[0].wwww + u_xlat1;
u_xlat0 = u_xlat1 * in_POSITION0.xxxx + u_xlat0;
u_xlat1 = hlslcc_mtx4x4mat0[1] * hlslcc_mtx4x4mat1[2].yyyy;
u_xlat1 = hlslcc_mtx4x4mat0[0] * hlslcc_mtx4x4mat1[2].xxxx + u_xlat1;
u_xlat1 = hlslcc_mtx4x4mat0[2] * hlslcc_mtx4x4mat1[2].zzzz + u_xlat1;
u_xlat1 = hlslcc_mtx4x4mat0[3] * hlslcc_mtx4x4mat1[2].wwww + u_xlat1;
u_xlat0 = u_xlat1 * in_POSITION0.zzzz + u_xlat0;
u_xlat1 = hlslcc_mtx4x4mat0[1] * hlslcc_mtx4x4mat1[3].yyyy;
u_xlat1 = hlslcc_mtx4x4mat0[0] * hlslcc_mtx4x4mat1[3].xxxx + u_xlat1;
u_xlat1 = hlslcc_mtx4x4mat0[2] * hlslcc_mtx4x4mat1[3].zzzz + u_xlat1;
u_xlat1 = hlslcc_mtx4x4mat0[3] * hlslcc_mtx4x4mat1[3].wwww + u_xlat1;
gl_Position = u_xlat1 * in_POSITION0.wwww + u_xlat0;

这样很直观的可以看到指令数量上的差别,指令越多也就意味着计算量越大,GPU 的压力也越大,但是两种方法得到的结果是一样的,所以这一点在优化时需要注意了。