SPIR-V Dialect to LLVM Dialect conversion manual

This manual describes the conversion from SPIR-V Dialect to LLVM Dialect. It assumes familiarity with both, and describes the design choices behind the modelling of SPIR-V concepts in LLVM Dialect. The conversion is an ongoing work, and is expected to grow as more features are implemented.

Conversion can be performed by invoking an appropriate conversion pass:

mlir-opt -convert-spirv-to-llvm <filename.mlir>

This pass performs type and operation conversions for SPIR-V operations as described in this document.

Type Conversion

This section describes how SPIR-V Dialect types are mapped to LLVM Dialect.

Scalar types

SPIR-V DialectLLVM Dialect
i<bitwidth>!llvm.i<bitwidth>
si<bitwidth>!llvm.i<bitwidth>
ui<bitwidth>!llvm.i<bitwidth>
f16f16
f32f32
f64f64

Vector types

SPIR-V DialectLLVM Dialect
vector<<count> x <scalar-type>>vector<<count> x <scalar-type>>

Pointer types

A SPIR-V pointer also takes a Storage Class. At the moment, conversion does not take it into account.

SPIR-V DialectLLVM Dialect
!spv.ptr< <element-type>, <storage-class> >!llvm.ptr<<element-type>>

Array types

SPIR-V distinguishes between array type and run-time array type, the length of which is not known at compile time. In LLVM, it is possible to index beyond the end of the array. Therefore, runtime array can be implemented as a zero length array type.

Moreover, SPIR-V supports the notion of array stride. Currently only natural strides (based on VulkanLayoutUtils) are supported. They are also mapped to LLVM array.

SPIR-V DialectLLVM Dialect
!spv.array<<count> x <element-type>>!llvm.array<<count> x <element-type>>
!spv.rtarray< <element-type> >!llvm.array<0 x <element-type>>

Struct types

Members of SPIR-V struct types may have decorations and offset information. Currently, there is no support of member decorations conversion for structs. For more information see section on Decorations.

Usually we expect that each struct member has a natural size and alignment. However, there are cases (e.g. in graphics) where one would place struct members explicitly at particular offsets. This case is not supported at the moment. Hence, we adhere to the following mapping:

  • Structs with no offset are modelled as LLVM packed structures.

  • Structs with natural offset (i.e. offset that equals to cumulative size of the previous struct elements or is a natural alignment) are mapped to naturally padded structs.

  • Structs with unnatural offset (i.e. offset that is not equal to cumulative size of the previous struct elements) are not supported. In this case, offsets can be emulated with padding fields (e.g. integers). However, such a design would require index recalculation in the conversion of ops that involve memory addressing.

Examples of SPIR-V struct conversion are: ```mlir !spv.struct<i8, i32> => !llvm.struct<packed (i8, i32)> !spv.struct<i8 [0], i32 [4]> => !llvm.struct<(i8, i32)>

// error !spv.struct<i8 [0], i32 [8]> ```

Not implemented types

The rest of the types not mentioned explicitly above are not supported by the conversion. This includes ImageType and MatrixType.

Operation Conversion

This section describes how SPIR-V Dialect operations are converted to LLVM Dialect. It lists already working conversion patterns, as well as those that are an ongoing work.

There are also multiple ops for which there is no clear mapping in LLVM. Conversion for those have to be discussed within the community on the case-by-case basis.

Arithmetic ops

SPIR-V arithmetic ops mostly have a direct equivalent in LLVM Dialect. Such exceptions as spv.SMod and spv.FMod are rare.

SPIR-V Dialect opLLVM Dialect op
spv.FAddllvm.fadd
spv.FDivllvm.fdiv
spv.FNegatellvm.fneg
spv.FMulllvm.fmul
spv.FRemllvm.frem
spv.FSubllvm.fsub
spv.IAddllvm.add
spv.IMulllvm.mul
spv.ISubllvm.sub
spv.SDivllvm.sdiv
spv.SRemllvm.srem
spv.UDivllvm.udiv
spv.UModllvm.urem

Bitwise ops

SPIR-V has a range of bit ops that are mapped to LLVM dialect ops, intrinsics or may have a specific conversion pattern.

Direct conversion

As with arithmetic ops, most of bitwise ops have a semantically equivalent op in LLVM:

SPIR-V Dialect opLLVM Dialect op
spv.BitwiseAndllvm.and
spv.BitwiseOrllvm.or
spv.BitwiseXorllvm.xor

Also, some of bitwise ops can be modelled with LLVM intrinsics:

SPIR-V Dialect opLLVM Dialect intrinsic
spv.BitCountllvm.intr.ctpop
spv.BitReversellvm.intr.bitreverse

spv.Not

spv.Not is modelled with a xor operation with a mask with all bits set.

                            %mask = llvm.mlir.constant(-1 : i32) : i32
%0 = spv.Not %op : i32  =>  %0  = llvm.xor %op, %mask : i32

Bitfield ops

SPIR-V dialect has three bitfield ops: spv.BitFieldInsert, spv.BitFieldSExtract and spv.BitFieldUExtract. This section will first outline the general design of conversion patterns for this ops, and then describe each of them.

All of these ops take base, offset and count (insert for spv.BitFieldInsert) as arguments. There are two important things to note:

  • offset and count are always scalar. This means that we can have the following case:

    %0 = spv.BitFieldSExtract %base, %offset, %count : vector<2xi32>, i8, i8
    

    To be able to proceed with conversion algorithms described below, all operands have to be of the same type and bitwidth. This requires broadcasting of offset and count to vectors, for example for the case above it gives:

    // Broadcasting offset
    %offset0 = llvm.mlir.undef : vector<2xi8>
    %zero = llvm.mlir.constant(0 : i32) : i32
    %offset1 = llvm.insertelement %offset, %offset0[%zero : i32] : vector<2xi8>
    %one = llvm.mlir.constant(1 : i32) : i32
    %vec_offset = llvm.insertelement  %offset, %offset1[%one : i32] : vector<2xi8>
    
    // Broadcasting count
    // ...
    
  • offset and count may have different bitwidths from base. In this case, both of these operands have to be zero extended (since they are treated as unsigned by the specification) or truncated. For the above example it would be:

    // Zero extending offset after broadcasting
    %res_offset = llvm.zext %vec_offset: vector<2xi8> to vector<2xi32>
    

    Also, note that if the bitwidth of offset or count is greater than the bitwidth of base, truncation is still permitted. This is because the ops have a defined behaviour with offset and count being less than the size of base. It creates a natural upper bound on what values offset and count can take, which is 64. This can be expressed in less than 8 bits.

Now, having these two cases in mind, we can proceed with conversion for the ops and their operands.

spv.BitFieldInsert

This operation is implemented as a series of LLVM Dialect operations. First step would be to create a mask with bits set outside [offset, offset + count - 1]. Then, unchanged bits are extracted from base that are outside of [offset, offset + count - 1]. The result is ored with shifted insert.

// Create mask
// %minus_one = llvm.mlir.constant(-1 : i32) : i32
// %t0        = llvm.shl %minus_one, %count : i32
// %t1        = llvm.xor %t0, %minus_one : i32
// %t2        = llvm.shl %t1, %offset : i32
// %mask      = llvm.xor %t2, %minus_one : i32

// Extract unchanged bits from the Base
// %new_base  = llvm.and %base, %mask : i32

// Insert new bits
// %sh_insert = llvm.shl %insert, %offset : i32
// %res       = llvm.or %new_base, %sh_insert : i32
%res = spv.BitFieldInsert %base, %insert, %offset, %count : i32, i32, i32
spv.BitFieldSExtract

To implement spv.BitFieldSExtract, base is shifted left by [sizeof(base) - (count + offset)], so that the bit at offset + count - 1 is the most significant bit. After, the result is shifted right, filling the bits with the sign bit.

// Calculate the amount to shift left.
// %size    = llvm.mlir.constant(32 : i32) : i32
// %t0      = llvm.add %count, %offset : i32
// %t1      = llvm.sub %size, %t0 : i32

// Shift left and then right to extract the bits
// %sh_left = llvm.shl %base, %t1 : i32
// %t2      = llvm.add %offset, %t1 : i32
// %res     = llvm.ashr %sh_left, %t2 : i32
%res = spv.BitFieldSExtract %base, %offset, %count : i32, i32, i32
spv.BitFieldUExtract

For this op a similar pattern as for spv.BitFieldInsert is used. First, a mask with bits set at [0, count - 1] is created. Then base is shifted by offset and the mask is applied.

// Create a mask
// %minus_one = llvm.mlir.constant(-1 : i32) : i32
// %t0        = llvm.shl %minus_one, %count : i32
// mask       = llvm.xor  %t0, %minus_one : i32

// Shift Base and apply mask
// %sh_base   = llvm.lshr %base, %offset : i32
// %res       = llvm.and %sh_base, %mask : i32
%res = spv.BitFieldUExtract %base, %offset, %count : i32, i32, i32

Cast ops

Direct conversions

SPIR-V Dialect opLLVM Dialect op
spv.ConvertFToSllvm.fptosi
spv.ConvertFToUllvm.fptoui
spv.ConvertSToFllvm.sitofp
spv.ConvertUToFllvm.uitofp

spv.Bitcast

This operation has a direct counterpart in LLVM: llvm.bitcast. It is treated separately since it also supports pointer to pointer bit pattern-preserving type conversion, apart from regular scalar or vector of numerical type.

Special cases

Special cases include spv.FConvert, spv.SConvert and spv.UConvert. These operations are either a truncate or extend. Let's denote the operand component width as A, and result component width as R. Then, the following mappings are used:

spv.FConvert
CaseLLVM Dialect op
A < Rllvm.fpext
A > Rllvm.fptrunc
spv.SConvert
CaseLLVM Dialect op
A < Rllvm.sext
A > Rllvm.trunc
spv.UConvert
CaseLLVM Dialect op
A < Rllvm.zext
A > Rllvm.trunc

The case when A = R is not possible, based on SPIR-V Dialect specification:

The component width cannot equal the component width in Result Type.

Comparison ops

SPIR-V comparison ops are mapped to LLVM icmp and fcmp operations.

SPIR-V Dialect opLLVM Dialect op
spv.IEqualllvm.icmp "eq"
spv.INotEqualllvm.icmp "ne"
spv.FOrdEqualllvm.fcmp "oeq"
spv.FOrdGreaterThanllvm.fcmp "ogt"
spv.FOrdGreaterThanEqualllvm.fcmp "oge"
spv.FOrdLessThanllvm.fcmp "olt"
spv.FOrdLessThanEqualllvm.fcmp "ole"
spv.FOrdNotEqualllvm.fcmp "one"
spv.FUnordEqualllvm.fcmp "ueq"
spv.FUnordGreaterThanllvm.fcmp "ugt"
spv.FUnordGreaterThanEqualllvm.fcmp "uge"
spv.FUnordLessThanllvm.fcmp "ult"
spv.FUnordLessThanEqualllvm.fcmp "ule"
spv.FUnordNotEqualllvm.fcmp "une"
spv.SGreaterThanllvm.icmp "sgt"
spv.SGreaterThanEqualllvm.icmp "sge"
spv.SLessThanllvm.icmp "slt"
spv.SLessThanEqualllvm.icmp "sle"
spv.UGreaterThanllvm.icmp "ugt"
spv.UGreaterThanEqualllvm.icmp "uge"
spv.ULessThanllvm.icmp "ult"
spv.ULessThanEqualllvm.icmp "ule"

Composite ops

Currently, conversion supports rewrite patterns for spv.CompositeExtract and spv.CompositeInsert. We distinguish two cases for these operations: when the composite object is a vector, and when the composite object is of a non-vector type (i.e. struct, array or runtime array).

Composite typeSPIR-V Dialect opLLVM Dialect op
vectorspv.CompositeExtractllvm.extractelement
vectorspv.CompositeInsertllvm.insertelement
non-vectorspv.CompositeExtractllvm.extractvalue
non-vectorspv.CompositeInsertllvm.insertvalue

spv.EntryPoint and spv.ExecutionMode

First of all, it is important to note that there is no direct representation of entry points in LLVM. At the moment, we use the following approach:

  • spv.EntryPoint is simply removed.

  • In contrast, spv.ExecutionMode may contain important information about the entry point. For example, LocalSize provides information about the work-group size that can be reused.

    In order to preserve this information, spv.ExecutionMode is converted to a struct global variable that stores the execution mode id and any variables associated with it. In C, the struct has the structure shown below.

    // No values are associated      // There are values that are associated
    // with this entry point.        // with this entry point.
    struct {                         struct {
      int32_t executionMode;             int32_t executionMode;
    };                                   int32_t values[];
                                     };
    
    // spv.ExecutionMode @empty "ContractionOff"
    llvm.mlir.global external constant @{{.*}}() : !llvm.struct<(i32)> {
      %0   = llvm.mlir.undef : !llvm.struct<(i32)>
      %1   = llvm.mlir.constant(31 : i32) : i32
      %ret = llvm.insertvalue %1, %0[0 : i32] : !llvm.struct<(i32)>
      llvm.return %ret : !llvm.struct<(i32)>
    }
    

Logical ops

Logical ops follow a similar pattern as bitwise ops, with the difference that they operate on i1 or vector of i1 values. The following mapping is used to emulate SPIR-V ops behaviour:

SPIR-V Dialect opLLVM Dialect op
spv.LogicalAndllvm.and
spv.LogicalOrllvm.or
spv.LogicalEqualllvm.icmp "eq"
spv.LogicalNotEqualllvm.icmp "ne"

spv.LogicalNot has the same conversion pattern as bitwise spv.Not. It is modelled with xor operation with a mask with all bits set.

                                  %mask = llvm.mlir.constant(-1 : i1) : i1
%0 = spv.LogicalNot %op : i1  =>  %0    = llvm.xor %op, %mask : i1

Memory ops

This section describes the conversion patterns for SPIR-V dialect operations that concern memory.

spv.AccessChain

spv.AccessChain is mapped to llvm.getelementptr op. In order to create a valid LLVM op, we also add a 0 index to the spv.AccessChain's indices list in order to go through the pointer.

// Access the 1st element of the array
%i   = spv.Constant 1: i32
%var = spv.Variable : !spv.ptr<!spv.struct<f32, !spv.array<4xf32>>, Function>
%el  = spv.AccessChain %var[%i, %i] : !spv.ptr<!spv.struct<f32, !spv.array<4xf32>>, Function>, i32, i32

// Corresponding LLVM dialect code
%i   = ...
%var = ...
%0   = llvm.mlir.constant(0 : i32) : i32
%el  = llvm.getelementptr %var[%0, %i, %i] : (!llvm.ptr<struct<packed (f32, array<4 x f32>)>>, i32, i32, i32)

spv.Load and spv.Store

These ops are converted to their LLVM counterparts: llvm.load and llvm.store. If the op has a memory access attribute, then there are the following cases, based on the value of the attribute:

  • Aligned: alignment is passed on to LLVM op builder, for example: mlir // llvm.store %ptr, %val {alignment = 4 : i64} : !llvm.ptr<f32> spv.Store "Function" %ptr, %val ["Aligned", 4] : f32

  • None: same case as if there is no memory access attribute.

  • Nontemporal: set nontemporal flag, for example: mlir // %res = llvm.load %ptr {nontemporal} : !llvm.ptr<f32> %res = spv.Load "Function" %ptr ["Nontemporal"] : f32

  • Volatile: mark the op as volatile, for example: mlir // %res = llvm.load volatile %ptr : !llvm.ptr<f32> %res = spv.Load "Function" %ptr ["Volatile"] : f32 Otherwise the conversion fails as other cases (MakePointerAvailable, MakePointerVisible, NonPrivatePointer) are not supported yet.

spv.GlobalVariable and spv.mlir.addressof

spv.GlobalVariable is modelled with llvm.mlir.global op. However, there is a difference that has to be pointed out.

In SPIR-V dialect, the global variable returns a pointer, whereas in LLVM dialect the global holds an actual value. This difference is handled by spv.mlir.addressof and llvm.mlir.addressof ops that both return a pointer and are used to reference the global.

// Original SPIR-V module
spv.module Logical GLSL450 {
  spv.GlobalVariable @struct : !spv.ptr<!spv.struct<f32, !spv.array<10xf32>>, Private>
  spv.func @func() -> () "None" {
    %0 = spv.mlir.addressof @struct : !spv.ptr<!spv.struct<f32, !spv.array<10xf32>>, Private>
    spv.Return
  }
}

// Converted result
module {
  llvm.mlir.global private @struct() : !llvm.struct<packed (f32, [10 x f32])>
  llvm.func @func() {
    %0 = llvm.mlir.addressof @struct : !llvm.ptr<struct<packed (f32, [10 x f32])>>
    llvm.return
  }
}

The SPIR-V to LLVM conversion does not involve modelling of workgroups. Hence, we say that only current invocation is in conversion's scope. This means that global variables with pointers of Input, Output, and Private storage classes are supported. Also, StorageBuffer storage class is allowed for executing mlir-spirv-cpu-runner.

Moreover, bind that specifies the descriptor set and the binding number and built_in that specifies SPIR-V BuiltIn decoration have no conversion into LLVM dialect.

Currently llvm.mlir.globals are created with private linkage for Private storage class and External for other storage classes, based on SPIR-V spec:

By default, functions and global variables are private to a module and cannot be accessed by other modules. However, a module may be written to export or import functions and global (module scope) variables.

If the global variable's pointer has Input storage class, then a constant flag is added to LLVM op:

spv.GlobalVariable @var : !spv.ptr<f32, Input>    =>    llvm.mlir.global external constant @var() : f32

spv.Variable

Per SPIR-V dialect spec, spv.Variable allocates an object in memory, resulting in a pointer to it, which can be used with spv.Load and spv.Store. It is also a function-level variable.

spv.Variable is modelled as llvm.alloca op. If initialized, an additional store instruction is used. Note that there is no initialization for arrays and structs since constants of these types are not supported in LLVM dialect (TODO). Also, at the moment initialization is only possible via spv.Constant.

// Conversion of VariableOp without initialization
                                                               %size = llvm.mlir.constant(1 : i32) : i32
%res = spv.Variable : !spv.ptr<vector<3xf32>, Function>   =>   %res  = llvm.alloca  %size x vector<3xf32> : (i32) -> !llvm.ptr<vec<3 x f32>>

// Conversion of VariableOp with initialization
                                                               %c    = llvm.mlir.constant(0 : i64) : i64
%c   = spv.Constant 0 : i64                                    %size = llvm.mlir.constant(1 : i32) : i32
%res = spv.Variable init(%c) : !spv.ptr<i64, Function>    =>   %res  = llvm.alloca %[[SIZE]] x i64 : (i32) -> !llvm.ptr<i64>
                                                               llvm.store %c, %res : !llvm.ptr<i64>

Note that simple conversion to alloca may not be sufficient if the code has some scoping. For example, if converting ops executed in a loop into allocas, a stack overflow may occur. For this case, stacksave/stackrestore pair can be used (TODO).

Miscellaneous ops with direct conversions

There are multiple SPIR-V ops that do not fit in a particular group but can be converted directly to LLVM dialect. Their conversion is addressed in this section.

SPIR-V Dialect opLLVM Dialect op
spv.Selectllvm.select
spv.Undefllvm.mlir.undef

Shift ops

Shift operates on two operands: shift and base.

In SPIR-V dialect, shift and base may have different bit width. On the contrary, in LLVM Dialect both base and shift have to be of the same bitwidth. This leads to the following conversions:

  • if base has the same bitwidth as shift, the conversion is straightforward.

  • if base has a greater bit width than shift, shift is sign or zero extended first. Then the extended value is passed to the shift.

  • otherwise, the conversion is considered to be illegal.

// Shift without extension
%res0 = spv.ShiftRightArithmetic %0, %2 : i32, i32  =>  %res0 = llvm.ashr %0, %2 : i32

// Shift with extension
                                                        %ext  = llvm.sext %1 : i16 to i32
%res1 = spv.ShiftRightArithmetic %0, %1 : i32, i16  =>  %res1 = llvm.ashr %0, %ext: i32

spv.Constant

At the moment spv.Constant conversion supports scalar and vector constants only.

Mapping

spv.Constant is mapped to llvm.mlir.constant. This is a straightforward conversion pattern with a special case when the argument is signed or unsigned.

Special case

SPIR-V constant can be a signed or unsigned integer. Since LLVM Dialect does not have signedness semantics, this case should be handled separately.

The conversion casts constant value attribute to a signless integer or a vector of signless integers. This is correct because in SPIR-V, like in LLVM, how to interpret an integer number is also dictated by the opcode. However, in reality hardware implementation might show unexpected behavior. Therefore, it is better to handle it case-by-case, given that the purpose of the conversion is not to cover all possible corner cases.

// %0 = llvm.mlir.constant(0 : i8) : i8
%0 = spv.Constant  0 : i8

// %1 = llvm.mlir.constant(dense<[2, 3, 4]> : vector<3xi32>) : vector<3xi32>
%1 = spv.Constant dense<[2, 3, 4]> : vector<3xui32>

Not implemented ops

There is no support of the following ops:

  • All atomic ops
  • All group ops
  • All matrix ops
  • All OCL ops

As well as:

  • spv.CompositeConstruct
  • spv.ControlBarrier
  • spv.CopyMemory
  • spv.FMod
  • spv.GLSL.Acos
  • spv.GLSL.Asin
  • spv.GLSL.Atan
  • spv.GLSL.Cosh
  • spv.GLSL.FSign
  • spv.GLSL.SAbs
  • spv.GLSL.Sinh
  • spv.GLSL.SSign
  • spv.MemoryBarrier
  • spv.mlir.referenceof
  • spv.SMod
  • spv.SpecConstant
  • spv.Unreachable
  • spv.VectorExtractDynamic

Control flow conversion

Branch ops

spv.Branch and spv.BranchConditional are mapped to llvm.br and llvm.cond_br. Branch weights for spv.BranchConditional are mapped to corresponding branch_weights attribute of llvm.cond_br. When translated to proper LLVM, branch_weights are converted into LLVM metadata associated with the conditional branch.

spv.FunctionCall

spv.FunctionCall maps to llvm.call. For example:

%0 = spv.FunctionCall @foo() : () -> i32    =>    %0 = llvm.call @foo() : () -> f32
spv.FunctionCall @bar(%0) : (i32) -> ()     =>    llvm.call @bar(%0) : (f32) -> ()

spv.mlir.selection and spv.mlir.loop

Control flow within spv.mlir.selection and spv.mlir.loop is lowered directly to LLVM via branch ops. The conversion can only be applied to selection or loop with all blocks being reachable. Moreover, selection and loop control attributes (such as Flatten or Unroll) are not supported at the moment.

// Conversion of selection
%cond = spv.Constant true                               %cond = llvm.mlir.constant(true) : i1
spv.mlir.selection {
  spv.BranchConditional %cond, ^true, ^false            llvm.cond_br %cond, ^true, ^false

^true:                                                                                              ^true:
  // True block code                                    // True block code
  spv.Branch ^merge                             =>      llvm.br ^merge

^false:                                               ^false:
  // False block code                                   // False block code
  spv.Branch ^merge                                     llvm.br ^merge

^merge:                                               ^merge:
  spv.mlir.merge                                            llvm.br ^continue
}
// Remaining code                                                                           ^continue:
                                                        // Remaining code
// Conversion of loop
%cond = spv.Constant true                               %cond = llvm.mlir.constant(true) : i1
spv.mlir.loop {
  spv.Branch ^header                                    llvm.br ^header

^header:                                              ^header:
  // Header code                                        // Header code
  spv.BranchConditional %cond, ^body, ^merge    =>      llvm.cond_br %cond, ^body, ^merge

^body:                                                ^body:
  // Body code                                          // Body code
  spv.Branch ^continue                                  llvm.br ^continue

^continue:                                            ^continue:
  // Continue code                                      // Continue code
  spv.Branch ^header                                    llvm.br ^header

^merge:                                               ^merge:
  spv.mlir.merge                                            llvm.br ^remaining
}
// Remaining code                                     ^remaining:
                                                        // Remaining code

Decorations conversion

Note: these conversions have not been implemented yet

GLSL extended instruction set

This section describes how SPIR-V ops from GLSL extended instructions set are mapped to LLVM Dialect.

Direct conversions

SPIR-V Dialect opLLVM Dialect op
spv.GLSL.Ceilllvm.intr.ceil
spv.GLSL.Cosllvm.intr.cos
spv.GLSL.Expllvm.intr.exp
spv.GLSL.FAbsllvm.intr.fabs
spv.GLSL.Floorllvm.intr.floor
spv.GLSL.FMaxllvm.intr.maxnum
spv.GLSL.FMinllvm.intr.minnum
spv.GLSL.Logllvm.intr.log
spv.GLSL.Sinllvm.intr.sin
spv.GLSL.Sqrtllvm.intr.sqrt
spv.GLSL.SMaxllvm.intr.smax
spv.GLSL.SMinllvm.intr.smin

Special cases

spv.InverseSqrt is mapped to:

                                           %one  = llvm.mlir.constant(1.0 : f32) : f32
%res = spv.InverseSqrt %arg : f32    =>    %sqrt = "llvm.intr.sqrt"(%arg) : (f32) -> f32
                                           %res  = fdiv %one, %sqrt : f32

spv.Tan is mapped to:

                                   %sin = "llvm.intr.sin"(%arg) : (f32) -> f32
%res = spv.Tan %arg : f32    =>    %cos = "llvm.intr.cos"(%arg) : (f32) -> f32
                                   %res = fdiv %sin, %cos : f32

spv.Tanh is modelled using the equality tanh(x) = {exp(2x) - 1}/{exp(2x) + 1}:

                                     %two   = llvm.mlir.constant(2.0: f32) : f32
                                     %2xArg = llvm.fmul %two, %arg : f32
                                     %exp   = "llvm.intr.exp"(%2xArg) : (f32) -> f32
%res = spv.Tanh %arg : f32     =>    %one   = llvm.mlir.constant(1.0 : f32) : f32
                                     %num   = llvm.fsub %exp, %one : f32
                                     %den   = llvm.fadd %exp, %one : f32
                                     %res   = llvm.fdiv %num, %den : f32

Function conversion and related ops

This section describes the conversion of function-related operations from SPIR-V to LLVM dialect.

spv.func

This op declares or defines a SPIR-V function and it is converted to llvm.func. This conversion handles signature conversion, and function control attributes remapping to LLVM dialect function passthrough attribute.

The following mapping is used to map SPIR-V function control to LLVM function attributes:

SPIR-V Function Control AttributesLLVM Function Attributes
NoneNo function attributes passed
Inlinealwaysinline
DontInlinenoinline
Purereadonly
Constreadnone

spv.Return and spv.ReturnValue

In LLVM IR, functions may return either 1 or 0 value. Hence, we map both ops to llvm.return with or without a return value.

Module ops

Module in SPIR-V has one region that contains one block. It is defined via spv.module op that also takes a range of attributes:

  • Addressing model
  • Memory model
  • Version-Capability-Extension attribute

spv.module is converted into ModuleOp. This plays a role of enclosing scope to LLVM ops. At the moment, SPIR-V module attributes are ignored.

mlir-spirv-cpu-runner

mlir-spirv-cpu-runner allows to execute gpu dialect kernel on the CPU via SPIR-V to LLVM dialect conversion. Currently, only single-threaded kernel is supported.

To build the runner, add the following option to cmake: bash -DMLIR_ENABLE_SPIRV_CPU_RUNNER=1

Pipeline

The gpu module with the kernel and the host code undergo the following transformations:

  • Convert the gpu module into SPIR-V dialect, lower ABI attributes and update version, capability and extension.

  • Emulate the kernel call by converting the launching operation into a normal function call. The data from the host side to the device is passed via copying to global variables. These are created in both the host and the kernel code and later linked when nested modules are folded.

  • Convert SPIR-V dialect kernel to LLVM dialect via the new conversion path.

After these passes, the IR transforms into a nested LLVM module - a main module representing the host code and a kernel module. These modules are linked and executed using ExecutionEngine.

Walk-through

This section gives a detailed overview of the IR changes while running mlir-spirv-cpu-runner. First, consider that we have the following IR. (For simplicity some type annotations and function implementations have been omitted).

gpu.module @foo {
  gpu.func @bar(%arg: memref<8xi32>) {
    // Kernel code.
    gpu.return
  }
}

func @main() {
  // Fill the buffer with some data
  %buffer = memref.alloc : memref<8xi32>
  %data = ...
  call fillBuffer(%buffer, %data)

  "gpu.launch_func"(/*grid dimensions*/, %buffer) {
    kernel = @foo::bar
  }
}

Lowering gpu dialect to SPIR-V dialect results in

spv.module @__spv__foo /*VCE triple and other metadata here*/ {
  spv.GlobalVariable @__spv__foo_arg bind(0,0) : ...
  spv.func @bar() {
    // Kernel code.
  }
  spv.EntryPoint @bar, ...
}

func @main() {
  // Fill the buffer with some data.
  %buffer = memref.alloc : memref<8xi32>
  %data = ...
  call fillBuffer(%buffer, %data)

  "gpu.launch_func"(/*grid dimensions*/, %buffer) {
    kernel = @foo::bar
  }
}

Then, the lowering from standard dialect to LLVM dialect is applied to the host code.

spv.module @__spv__foo /*VCE triple and other metadata here*/ {
  spv.GlobalVariable @__spv__foo_arg bind(0,0) : ...
  spv.func @bar() {
    // Kernel code.
  }
  spv.EntryPoint @bar, ...
}

// Kernel function declaration.
llvm.func @__spv__foo_bar() : ...

llvm.func @main() {
  // Fill the buffer with some data.
  llvm.call fillBuffer(%buffer, %data)

  // Copy data to the global variable, call kernel, and copy the data back.
  %addr = llvm.mlir.addressof @__spv__foo_arg_descriptor_set0_binding0 : ...
  "llvm.intr.memcpy"(%addr, %buffer) : ...
  llvm.call @__spv__foo_bar()
  "llvm.intr.memcpy"(%buffer, %addr) : ...

  llvm.return
}

Finally, SPIR-V module is converted to LLVM and the symbol names are resolved for the linkage.

module @__spv__foo {
  llvm.mlir.global @__spv__foo_arg_descriptor_set0_binding0 : ...
  llvm.func @__spv__foo_bar() {
    // Kernel code.
  }
}

// Kernel function declaration.
llvm.func @__spv__foo_bar() : ...

llvm.func @main() {
  // Fill the buffer with some data.
  llvm.call fillBuffer(%buffer, %data)

  // Copy data to the global variable, call kernel, and copy the data back.
  %addr = llvm.mlir.addressof @__spv__foo_arg_descriptor_set0_binding0 : ...
  "llvm.intr.memcpy"(%addr, %buffer) : ...
  llvm.call @__spv__foo_bar()
  "llvm.intr.memcpy"(%buffer, %addr) : ...

  llvm.return
}