d3dx9具有一个函数D3DXFloat16To32Array可以实现这个转换。
开源引擎ogre也有类似模块OgreBitwise.h中也有转换函数。这个float16格式是1－5－10；

typedef enum _D3DDECLTYPE
{
    D3DDECLTYPE_FLOAT1    =  0,  // 1D float expanded to (value, 0., 0., 1.)
    D3DDECLTYPE_FLOAT2    =  1,  // 2D float expanded to (value, value, 0., 1.)
    D3DDECLTYPE_FLOAT3    =  2,  // 3D float expanded to (value, value, value, 1.)
    D3DDECLTYPE_FLOAT4    =  3,  // 4D float
    D3DDECLTYPE_D3DCOLOR  =  4,  // 4D packed unsigned bytes mapped to 0. to 1. range
                                 // Input is in D3DCOLOR format (ARGB) expanded to (R, G, B, A)
    D3DDECLTYPE_UBYTE4    =  5,  // 4D unsigned byte
    D3DDECLTYPE_SHORT2    =  6,  // 2D signed short expanded to (value, value, 0., 1.)
    D3DDECLTYPE_SHORT4    =  7,  // 4D signed short

// The following types are valid only with vertex shaders >= 2.0

    D3DDECLTYPE_UBYTE4N   =  8,  // Each of 4 bytes is normalized by dividing to 255.0
    D3DDECLTYPE_SHORT2N   =  9,  // 2D signed short normalized (v[0]/32767.0,v[1]/32767.0,0,1)
    D3DDECLTYPE_SHORT4N   = 10,  // 4D signed short normalized (v[0]/32767.0,v[1]/32767.0,v[2]/32767.0,v[3]/32767.0)
    D3DDECLTYPE_USHORT2N  = 11,  // 2D unsigned short normalized (v[0]/65535.0,v[1]/65535.0,0,1)
    D3DDECLTYPE_USHORT4N  = 12,  // 4D unsigned short normalized (v[0]/65535.0,v[1]/65535.0,v[2]/65535.0,v[3]/65535.0)
    D3DDECLTYPE_UDEC3     = 13,  // 3D unsigned 10 10 10 format expanded to (value, value, value, 1)
    D3DDECLTYPE_DEC3N     = 14,  // 3D signed 10 10 10 format normalized and expanded to (v[0]/511.0, v[1]/511.0, v[2]/511.0, 1)
    D3DDECLTYPE_FLOAT16_2 = 15,  // Two 16-bit floating point values, expanded to (value, value, 0, 1)
    D3DDECLTYPE_FLOAT16_4 = 16,  // Four 16-bit floating point values
    D3DDECLTYPE_UNUSED    = 17,  // When the type field in a decl is unused.
} D3DDECLTYPE;
可能大部分都很少用到。
最近发现老外常用D3DDECLTYPE_FLOAT16_2，D3DDECLTYPE_UBYTE4N，
应该可以节约几何带宽，比较数据小了很多，相比较float4而言。
其中这个UBYTE4N，就是你的app提供byte类型，在提交给shader处理时，驱动自动帮你除255.0f转换成浮点数。蛮好，蛮好。

关于Vertex Declaration和vs输入相关联的问题：
app的定义的顶点数据格式，和输入给vertex shader处理顶点，是怎么关联的呢，一开始误以为是按照内存分布，就如c＋＋中的结构体，其实不是，app的vertex是通过Usage Semantics 跟shader function对应来关联，也就是相同的sematic相关联。

// A structure for our custom vertex type. We added texture coordinates
struct CUSTOMVERTEX
{
    D3DXVECTOR3 position; // The position
    D3DCOLOR    color;    // The color
    FLOAT       tu, tv;   // The texture coordinates
};
//
D3DVERTEXELEMENT9 decl[] =
 {
  {0 ,0  ,D3DDECLTYPE_FLOAT3 ,  D3DDECLMETHOD_DEFAULT ,D3DDECLUSAGE_POSITION ,0},
  {0 ,12 ,D3DDECLTYPE_D3DCOLOR ,D3DDECLMETHOD_DEFAULT ,D3DDECLUSAGE_COLOR ,   0},
  {0 ,16 ,D3DDECLTYPE_FLOAT2 ,  D3DDECLMETHOD_DEFAULT ,D3DDECLUSAGE_TEXCOORD ,0},
  D3DDECL_END()
 };
//在shader中
struct VertexShaderInput
{
 float2 Texcoord  :TEXCOORD0;
 float4 VertexColor      :COLOR0;
 float4 Position   :POSITION;
};
struct VertexShaderInput
{
 float2 Texcoord  :TEXCOORD0;
 float4 Position   :POSITION;
 float4 VertexColor      :COLOR0;
};

上述三个成员可以随便交换，重要的sematic的修饰。 row_major float4x4 mWorldViewProj;
//
float3 pos=mul(mWorldViewProj,vPosition);
对应指令为：

dp3 oPos.x, c0, v0  
dp3 oPos.y, c1, v0  
dp3 oPos.z, c2, v0  











column_major float4x4 mWorldViewProj;
//
float3 pos=mul(vPosition，mWorldViewProj);
对应指令为：

dp3 oPos.x, v0, c0  
dp3 oPos.y, v0, c1  
dp3 oPos.z, v0, c2   
原因是：
对于mul指令
Performs matrix multiplication between x and y. 
If x is a vector, it treated as a row vector. 
If y is a vector, it is treated as a column vector.