1# Uniform Block to StructuredBuffer Translation 2 3## Background 4In ANGLE's D3D11 backend, we normally translate GLSL uniform blocks to 5HLSL constant buffers. We run into a compile performance issue with 6[fxc](https://docs.microsoft.com/en-us/windows/win32/direct3dtools/fxc) 7and dynamic constant buffer indexing, 8[anglebug.com/40096608](https://bugs.chromium.org/p/angleproject/issues/detail?id=3682) 9 10## Solution 11We translate a uniform block into a StructuredBuffer when the following three 12conditions are satisfied: 13* The uniform block has only one array member, and the array size is larger than or 14equal to 50; 15* In the shader, all the accesses of the member are through indexing operator; 16* The type of the array member must be any of the following: 17 * a scalar or vector type. 18 * a mat2x4, mat3x4 or mat4x4 matrix type in column major layout, a mat4x2, mat4x2 or 19 mat4x3 matrix type in row major layout. 20 * a structure type with no array or structure members, where all of the structure's 21 fields satisify the prior type conditions. 22 23## Analysis 24A typical use case for uniform block to StructuredBuffer translation is for shaders with 25one large array member in a uniform block. For example: 26``` 27// GLSL code 28uniform buffer { 29 TYPE buf [100]; 30}; 31``` 32Will be translated into 33``` 34// HLSL code 35StructuredBuffer <TYPETRANSLATED> bufTranslated: register(tN); 36``` 37 38However, even with the above limitation, there are still many shaders where we cannot 39apply our translation. They are divided into two classes. The first case is when the 40shader accesses a "whole entity" uniform block array member element. The second is when 41the shader uses the std140 layout. 42 43### Operate uniform block array member as whole entity 44According to ESSL spec 3.0, 5.7 Structure and Array Operations, the following operators 45are allowed to operate on arrays as whole entities: 46 47| Operator Name | Operator | 48| :---------------- | :------: | 49| field or method selector | . | 50| assignment | == != | 51| Ternary operator | ?: | 52| Sequence operator | , | 53| indexing | [] | 54 55However, after translating to StructuredBuffer, the uniform array member cannot be used as 56a whole entity since its type has been changed. The member is no longer an array. 57After the change, we only support the indexed operation since it is the most common use case. 58Other operator usage is unsupported. Example unsupported usages: 59 60| Operator On the Uniform Array Member | examples | 61| :------ | :------ | 62| method selector | buf.length(); // Angle don’t support it, too. | 63| equality == != | TYPE var[NUMBER] = {…}; <br> if (var == buf); | 64| assignment = | TYPE var[NUMBER] = {…}; <br> var = buf; | 65| Ternary operator ?: | // Angle don’t support it, too. | 66| Sequence operator , | TYPE var1[NUMBER] = {…}; <br> TYPE var2[NUMBER] = (var1, buf); | 67| Function arguments | void func(TYPE a[NUMBER]); <br> func(buf); | 68| Function return type | TYPE[NUMBER] func() { return buf;} <br> TYPE var[NUMBER] = func(); | 69 70### Std140 limitation 71GLSL uniform blocks follow std140 layout packing rules. StructuredBufer has a different set 72of packing rules. So we may need to explicitly pad the type `TYPETRANSLATED` to follow std140 73rules. Alternately, we can just simply only support those types which don't need to be padded. 74These are the supported translation types which do not require translation emulation: 75 76| GLSL TYPE | TRANSLATED HLSL TYPE | 77| :------ | :------ | 78| vec4/ivec4/uvec4/bvec4 | float4/int4/uint4/bool4 | 79| mat2x4 (column_major) | float2x4 (row_major) | 80| mat3x4 (column_major) | float3x4 (row_major) | 81| mat4x4 (column_major) | float4x4 (row_major) | 82| mat4x2 (row_major) | float4x3 (column_major) | 83| mat4x3 (row_major) | float4x3 (column_major) | 84| mat4x4 (row_major) | float4x4 (column_major) | 85 86These are the supported translation types which require some basic translation emulation: 87 88| GLSL TYPE | TRANSLATED HLSL TYPE | examples | 89| :------ | :------ | :------ | 90|float/int/uint/bool |float4/int4/uint4/bool4|GLSL: float var = buf[0]; <br> HLSL: float var = buf[0].x; | 91|vec2/ivec2/uvec2/bvec2|float4/int4/uint4/bool4|GLSL: vec2 var = buf[0]; <br> HLSL: float2 var = buf[0].xy; | 92|vec3/ivec3/uvec3/bvec3|float4/int4/uint4/bool4|GLSL: vec3 var = buf[0]; <br> HLSL: float3 var = buf[0].xyz;| 93 94These are the unsupported translation types which require more complex translation emulation: 95 96| GLSL TYPE | TRANSLATED HLSL TYPE | 97| :------ | :------ | 98| mat2x2 (column_major) | float2x4 (row_major) | 99| mat2x3 (column_major) | float2x4 (row_major) | 100| mat3x2 (column_major) | float3x4 (row_major) | 101| mat3x3 (column_major) | float3x4 (row_major) | 102| mat4x2 (column_major) | float4x4 (row_major) | 103| mat4x3 (column_major) | float4x4 (row_major) | 104| mat2x2 (row_major) | float4x2 (column_major) | 105| mat2x3 (row_major) | float4x3 (column_major) | 106| mat2x4 (row_major) | float4x4 (column_major) | 107| mat3x2 (row_major) | float4x2 (column_major) | 108| mat3x3 (row_major) | float4x3 (column_major) | 109| mat3x4 (row_major) | float4x4 (column_major) | 110 111 112Take mat3x2(column_major) for an example, the uniform buffer's memory layout is as shown below. 113 114|index|0 |1 |... | 115|:--- | :------ | :------ | :--- | 116|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|... | 117 118 119And the declaration of the uniform block in vertex shader may be as shown below. 120``` 121layout(std140) uniform buffer { 122 mat3x2 buf [100]; 123}; 124 125void main(void) { 126 ... 127 vec2 var = buf[0][2] 128 ... 129} 130``` 131Will be translated to 132 133``` 134#pragma pack_matrix(row_major) 135StructuredBuffer<float3x4> bufTranslated: register(t0); 136 137float3x2 GetFloat3x2FromFloat3x4Rowmajor(float3x4 mat) 138{ 139 float3x2 res = { 0.0 }; 140 res[0] = mat[0].xy; 141 res[1] = mat[1].xy; 142 res[2] = mat[2].xy; 143 return res; 144} 145 146VS_OUTPUT main(VS_INPUT input) { 147 ... 148 float3x2 var = GetFloat3x2FromFloat3x4Rowmajor(bufTranslated[0]) 149} 150``` 151 152When accessing the element of the `buf` variable, we would need to extract a float3x2 from 153a float3x4 for every element.