257 lines
6.3 KiB
Plaintext
257 lines
6.3 KiB
Plaintext
#pragma clang diagnostic ignored "-Wmissing-prototypes"
|
|
|
|
#include <metal_stdlib>
|
|
#include <simd/simd.h>
|
|
|
|
using namespace metal;
|
|
|
|
template<typename T, uint N>
|
|
inline void spvArrayCopyFromConstantToStack(thread T (&dst)[N], constant T (&src)[N])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
dst[i] = src[i];
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N>
|
|
inline void spvArrayCopyFromConstantToThreadGroup(threadgroup T (&dst)[N], constant T (&src)[N])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
dst[i] = src[i];
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N>
|
|
inline void spvArrayCopyFromStackToStack(thread T (&dst)[N], thread const T (&src)[N])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
dst[i] = src[i];
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N>
|
|
inline void spvArrayCopyFromStackToThreadGroup(threadgroup T (&dst)[N], thread const T (&src)[N])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
dst[i] = src[i];
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N>
|
|
inline void spvArrayCopyFromThreadGroupToStack(thread T (&dst)[N], threadgroup const T (&src)[N])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
dst[i] = src[i];
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N>
|
|
inline void spvArrayCopyFromThreadGroupToThreadGroup(threadgroup T (&dst)[N], threadgroup const T (&src)[N])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
dst[i] = src[i];
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N>
|
|
inline void spvArrayCopyFromDeviceToDevice(device T (&dst)[N], device const T (&src)[N])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
dst[i] = src[i];
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N>
|
|
inline void spvArrayCopyFromConstantToDevice(device T (&dst)[N], constant T (&src)[N])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
dst[i] = src[i];
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N>
|
|
inline void spvArrayCopyFromStackToDevice(device T (&dst)[N], thread const T (&src)[N])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
dst[i] = src[i];
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N>
|
|
inline void spvArrayCopyFromThreadGroupToDevice(device T (&dst)[N], threadgroup const T (&src)[N])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
dst[i] = src[i];
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N>
|
|
inline void spvArrayCopyFromDeviceToStack(thread T (&dst)[N], device const T (&src)[N])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
dst[i] = src[i];
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N>
|
|
inline void spvArrayCopyFromDeviceToThreadGroup(threadgroup T (&dst)[N], device const T (&src)[N])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
dst[i] = src[i];
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N, uint M>
|
|
inline void spvArrayCopyFromConstantToStack(thread T (&dst)[N][M], constant T (&src)[N][M])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
spvArrayCopyFromConstantToStack(dst[i], src[i]);
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N, uint M>
|
|
inline void spvArrayCopyFromConstantToThreadGroup(threadgroup T (&dst)[N][M], constant T (&src)[N][M])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
spvArrayCopyFromConstantToThreadGroup(dst[i], src[i]);
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N, uint M>
|
|
inline void spvArrayCopyFromStackToStack(thread T (&dst)[N][M], thread const T (&src)[N][M])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
spvArrayCopyFromStackToStack(dst[i], src[i]);
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N, uint M>
|
|
inline void spvArrayCopyFromStackToThreadGroup(threadgroup T (&dst)[N][M], thread const T (&src)[N][M])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
spvArrayCopyFromStackToThreadGroup(dst[i], src[i]);
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N, uint M>
|
|
inline void spvArrayCopyFromThreadGroupToStack(thread T (&dst)[N][M], threadgroup const T (&src)[N][M])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
spvArrayCopyFromThreadGroupToStack(dst[i], src[i]);
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N, uint M>
|
|
inline void spvArrayCopyFromThreadGroupToThreadGroup(threadgroup T (&dst)[N][M], threadgroup const T (&src)[N][M])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
spvArrayCopyFromThreadGroupToThreadGroup(dst[i], src[i]);
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N, uint M>
|
|
inline void spvArrayCopyFromDeviceToDevice(device T (&dst)[N][M], device const T (&src)[N][M])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
spvArrayCopyFromDeviceToDevice(dst[i], src[i]);
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N, uint M>
|
|
inline void spvArrayCopyFromConstantToDevice(device T (&dst)[N][M], constant T (&src)[N][M])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
spvArrayCopyFromConstantToDevice(dst[i], src[i]);
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N, uint M>
|
|
inline void spvArrayCopyFromStackToDevice(device T (&dst)[N][M], thread const T (&src)[N][M])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
spvArrayCopyFromStackToDevice(dst[i], src[i]);
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N, uint M>
|
|
inline void spvArrayCopyFromThreadGroupToDevice(device T (&dst)[N][M], threadgroup const T (&src)[N][M])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
spvArrayCopyFromThreadGroupToDevice(dst[i], src[i]);
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N, uint M>
|
|
inline void spvArrayCopyFromDeviceToStack(thread T (&dst)[N][M], device const T (&src)[N][M])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
spvArrayCopyFromDeviceToStack(dst[i], src[i]);
|
|
}
|
|
}
|
|
|
|
template<typename T, uint N, uint M>
|
|
inline void spvArrayCopyFromDeviceToThreadGroup(threadgroup T (&dst)[N][M], device const T (&src)[N][M])
|
|
{
|
|
for (uint i = 0; i < N; i++)
|
|
{
|
|
spvArrayCopyFromDeviceToThreadGroup(dst[i], src[i]);
|
|
}
|
|
}
|
|
|
|
struct BUF
|
|
{
|
|
int a;
|
|
float b;
|
|
float c;
|
|
};
|
|
|
|
constant uint3 gl_WorkGroupSize [[maybe_unused]] = uint3(1u);
|
|
|
|
constant float _16[2] = { 1.0, 2.0 };
|
|
constant float _19[2] = { 3.0, 4.0 };
|
|
constant float _20[2][2] = { { 1.0, 2.0 }, { 3.0, 4.0 } };
|
|
constant float _21[2][2][2] = { { { 1.0, 2.0 }, { 3.0, 4.0 } }, { { 1.0, 2.0 }, { 3.0, 4.0 } } };
|
|
|
|
kernel void main0(device BUF& o [[buffer(0)]])
|
|
{
|
|
float c[2][2][2];
|
|
spvArrayCopyFromConstantToStack(c, _21);
|
|
o.a = int(c[1][1][1]);
|
|
float _43[2] = { o.b, o.c };
|
|
float _48[2] = { o.b, o.b };
|
|
float _49[2][2] = { { _43[0], _43[1] }, { _48[0], _48[1] } };
|
|
float _54[2] = { o.c, o.c };
|
|
float _59[2] = { o.c, o.b };
|
|
float _60[2][2] = { { _54[0], _54[1] }, { _59[0], _59[1] } };
|
|
float _61[2][2][2] = { { { _49[0][0], _49[0][1] }, { _49[1][0], _49[1][1] } }, { { _60[0][0], _60[0][1] }, { _60[1][0], _60[1][1] } } };
|
|
float d[2][2][2];
|
|
spvArrayCopyFromStackToStack(d, _61);
|
|
float e[2][2][2];
|
|
spvArrayCopyFromStackToStack(e, d);
|
|
o.b = e[1][0][1];
|
|
}
|
|
|