KernelOp

enum class MayaFlux::Portal::Graphics::KernelOp : uint8_t

strong

Named operation the SPIR-V emitter knows how to lower to opcodes.

Each value maps to a fixed, unambiguous instruction sequence. No string parsing or expression compilation is involved.

Naming convention: operands are listed in application order. PC fields are consumed in declaration order from ShaderSpec::pc_fields. SSBO operands are the InOut/Output bindings in declaration order.

Single-SSBO elementwise — arithmetic (sig[i] = f(sig[i], pc...)): Scale sig[i] = sig[i] * pc[0] ScaleOffset sig[i] = sig[i] * pc[0] + pc[1] Fma sig[i] = sig[i] * pc[0] + pc[1] (fused, single instruction) Offset sig[i] = sig[i] + pc[0] Clip sig[i] = clamp(sig[i], pc[0], pc[1]) Abs sig[i] = abs(sig[i]) Negate sig[i] = -sig[i] Floor sig[i] = floor(sig[i]) Ceil sig[i] = ceil(sig[i]) Round sig[i] = round(sig[i]) Trunc sig[i] = trunc(sig[i]) Fract sig[i] = fract(sig[i]) Sqrt sig[i] = sqrt(sig[i]) InverseSqrt sig[i] = 1/sqrt(sig[i])

Single-SSBO elementwise — transcendental (sig[i] = f(sig[i])): Sin sig[i] = sin(sig[i]) Cos sig[i] = cos(sig[i]) Tan sig[i] = tan(sig[i]) Asin sig[i] = asin(sig[i]) Acos sig[i] = acos(sig[i]) Atan sig[i] = atan(sig[i]) Sinh sig[i] = sinh(sig[i]) Cosh sig[i] = cosh(sig[i]) Tanh sig[i] = tanh(sig[i]) Exp sig[i] = e^sig[i] Exp2 sig[i] = 2^sig[i] Log sig[i] = ln(sig[i]) Log2 sig[i] = log2(sig[i])

Two-SSBO elementwise (out[i] = f(a[i], b[i])): Add out[i] = a[i] + b[i] Multiply out[i] = a[i] * b[i] Mix out[i] = a[i] + (b[i] - a[i]) * pc[0] Sub out[i] = a[i] - b[i] Pow out[i] = pow(a[i], b[i]) Atan2 out[i] = atan(a[i], b[i]) Min out[i] = min(a[i], b[i]) MaxTwo out[i] = max(a[i], b[i]) Step out[i] = step(a[i], b[i]) edge=a, x=b

Two-SSBO + one PC: SmoothStep out[i] = smoothstep(a[i], b[i], pc[0])

Image body elementwise (out = f(pixel, pc)): CompareGE out[ch] = pixel[ch] >= pc[0] ? 1.0 : 0.0 CompareGEPreserve out[ch] = pixel[ch] >= pc[0] ? pc[1] : pixel[ch] ChannelDot out = dot(pixel.rgba, pc[0..3]) broadcast to all channels ChannelReplicate out = pixel[pc_channel_index].xxxx (single channel to all)

Reduction operations (one InOut SSBO, shared memory): Sum accumulate + into shared[lid], tree reduce Max accumulate max into shared[lid], tree reduce IndexScale sig[i] = float(i) * sig[i]

Enumerator
Scale
ScaleOffset
Fma
Offset
Clip
Abs
Negate
Floor
Ceil
Round
Trunc
Fract
Sqrt
InverseSqrt
Sin
Cos
Tan
Asin
Acos
Atan
Sinh
Cosh
Tanh
Exp
Exp2
Log
Log2
Add
Multiply
Mix
Sub
Pow
Atan2
Min
MaxTwo
Step
SmoothStep
Sum
Max
ScanSum	Inclusive prefix sum: shared[i] = sum(x[0..i])
MaxIndex	Reduction variant: finds max value AND its index. First InOut SSBO holds the values (overwritten with the max in slot 0). Second InOut SSBO (UINT32) receives the winning index in slot 0.
IndexScale	out[i] = float(i) * a[i]. The only op that uses the invocation index itself as an arithmetic operand rather than purely for addressing.
CompareGE	out[ch] = pixel[ch] >= pc[0] ? 1.0 : 0.0
CompareGEPreserve	out[ch] = pixel[ch] >= pc[0] ? pc[1] : pixel[ch]
ChannelDot	out = dot(pixel.rgba, pc[0..3]) broadcast to all channels
ChannelReplicate	out = pixel[pc_channel_index].xxxx (single channel to all)

Definition at line 227 of file ShaderSpec.hpp.

                    : uint8_t {
    // arithmetic
    Scale,
    ScaleOffset,
    Fma,
    Offset,
    Clip,
    Abs,
    Negate,
    Floor,
    Ceil,
    Round,
    Trunc,
    Fract,
    Sqrt,
    InverseSqrt,
    // transcendental
    Sin,
    Cos,
    Tan,
    Asin,
    Acos,
    Atan,
    Sinh,
    Cosh,
    Tanh,
    Exp,
    Exp2,
    Log,
    Log2,
    // two-SSBO
    Add,
    Multiply,
    Mix,
    Sub,
    Pow,
    Atan2,
    Min,
    MaxTwo,
    Step,
    SmoothStep,
    // reduction
    Sum,
    Max,
    ScanSum, ///< Inclusive prefix sum: shared[i] = sum(x[0..i])
    MaxIndex, ///< Reduction variant: finds max value AND its index.
    ///< First InOut SSBO holds the values (overwritten with the
    ///< max in slot 0). Second InOut SSBO (UINT32) receives the
    ///< winning index in slot 0.
    IndexScale, ///< out[i] = float(i) * a[i]. The only op that uses the
    ///< invocation index itself as an arithmetic operand rather
    ///< than purely for addressing.
 
    CompareGE, ///< out[ch] = pixel[ch] >= pc[0] ? 1.0 : 0.0
    CompareGEPreserve, ///< out[ch] = pixel[ch] >= pc[0] ? pc[1] : pixel[ch]
    ChannelDot, ///< out = dot(pixel.rgba, pc[0..3]) broadcast to all channels
    ChannelReplicate, ///< out = pixel[pc_channel_index].xxxx (single channel to all)
};