Uniform Block to StructuredBuffer Translation

Background

In ANGLE's D3D11 backend, we normally translate GLSL uniform blocks to HLSL constant buffers. We run into a compile performance issue with fxc and dynamic constant buffer indexing, anglebug.com/40096608

Solution

We translate a uniform block into a StructuredBuffer when the following three conditions are satisfied:

• The uniform block has only one array member, and the array size is larger than or equal to 50;
• In the shader, all the accesses of the member are through indexing operator;
• The type of the array member must be any of the following:
  • a scalar or vector type.
  • a mat2x4, mat3x4 or mat4x4 matrix type in column major layout, a mat4x2, mat4x2 or mat4x3 matrix type in row major layout.
  • a structure type with no array or structure members, where all of the structure's fields satisify the prior type conditions.

Analysis

A typical use case for uniform block to StructuredBuffer translation is for shaders with one large array member in a uniform block. For example:

// GLSL code
uniform buffer {
    TYPE buf [100];
};

Will be translated into

// HLSL code
StructuredBuffer <TYPETRANSLATED>  bufTranslated: register(tN);

However, even with the above limitation, there are still many shaders where we cannot apply our translation. They are divided into two classes. The first case is when the shader accesses a “whole entity” uniform block array member element. The second is when the shader uses the std140 layout.

Operate uniform block array member as whole entity

According to ESSL spec 3.0, 5.7 Structure and Array Operations, the following operators are allowed to operate on arrays as whole entities:

Operator Name	Operator
field or method selector	.
assignment	== !=
Ternary operator	?:
Sequence operator	,
indexing	[]

However, after translating to StructuredBuffer, the uniform array member cannot be used as a whole entity since its type has been changed. The member is no longer an array. After the change, we only support the indexed operation since it is the most common use case. Other operator usage is unsupported. Example unsupported usages:

Operator On the Uniform Array Member	examples
method selector	buf.length(); // Angle don’t support it, too.
equality == !=	TYPE var[NUMBER] = {…}; if (var == buf);
assignment =	TYPE var[NUMBER] = {…}; var = buf;
Ternary operator ?:	// Angle don’t support it, too.
Sequence operator ,	TYPE var1[NUMBER] = {…}; TYPE var2[NUMBER] = (var1, buf);
Function arguments	void func(TYPE a[NUMBER]); func(buf);
Function return type	TYPE[NUMBER] func() { return buf;} TYPE var[NUMBER] = func();

Std140 limitation

GLSL uniform blocks follow std140 layout packing rules. StructuredBufer has a different set of packing rules. So we may need to explicitly pad the type TYPETRANSLATED to follow std140 rules. Alternately, we can just simply only support those types which don't need to be padded. These are the supported translation types which do not require translation emulation:

GLSL TYPE	TRANSLATED HLSL TYPE
vec4/ivec4/uvec4/bvec4	float4/int4/uint4/bool4
mat2x4 (column_major)	float2x4 (row_major)
mat3x4 (column_major)	float3x4 (row_major)
mat4x4 (column_major)	float4x4 (row_major)
mat4x2 (row_major)	float4x3 (column_major)
mat4x3 (row_major)	float4x3 (column_major)
mat4x4 (row_major)	float4x4 (column_major)

These are the supported translation types which require some basic translation emulation:

GLSL TYPE	TRANSLATED HLSL TYPE	examples
float/int/uint/bool	float4/int4/uint4/bool4	GLSL: float var = buf[0]; HLSL: float var = buf[0].x;
vec2/ivec2/uvec2/bvec2	float4/int4/uint4/bool4	GLSL: vec2 var = buf[0]; HLSL: float2 var = buf[0].xy;
vec3/ivec3/uvec3/bvec3	float4/int4/uint4/bool4	GLSL: vec3 var = buf[0]; HLSL: float3 var = buf[0].xyz;

These are the unsupported translation types which require more complex translation emulation:

GLSL TYPE	TRANSLATED HLSL TYPE
mat2x2 (column_major)	float2x4 (row_major)
mat2x3 (column_major)	float2x4 (row_major)
mat3x2 (column_major)	float3x4 (row_major)
mat3x3 (column_major)	float3x4 (row_major)
mat4x2 (column_major)	float4x4 (row_major)
mat4x3 (column_major)	float4x4 (row_major)
mat2x2 (row_major)	float4x2 (column_major)
mat2x3 (row_major)	float4x3 (column_major)
mat2x4 (row_major)	float4x4 (column_major)
mat3x2 (row_major)	float4x2 (column_major)
mat3x3 (row_major)	float4x3 (column_major)
mat3x4 (row_major)	float4x4 (column_major)

Take mat3x2(column_major) for an example, the uniform buffer's memory layout is as shown below.

index	0	1	...
data	1 2 x x 3 4 x x 5 6 x x	7 8 x x 9 10 x x 11 12 x x	...

And the declaration of the uniform block in vertex shader may be as shown below.

layout(std140) uniform buffer {
    mat3x2 buf [100];
};

void main(void) {
    ...
    vec2 var = buf[0][2]
    ...
}

Will be translated to

#pragma pack_matrix(row_major)
StructuredBuffer<float3x4> bufTranslated: register(t0);

float3x2 GetFloat3x2FromFloat3x4Rowmajor(float3x4 mat)
{
    float3x2 res = { 0.0 };
    res[0] = mat[0].xy;
    res[1] = mat[1].xy;
    res[2] = mat[2].xy;
    return res;
}

VS_OUTPUT main(VS_INPUT input) {
    ...
    float3x2 var = GetFloat3x2FromFloat3x4Rowmajor(bufTranslated[0])
}

When accessing the element of the buf variable, we would need to extract a float3x2 from a float3x4 for every element.