mirror of https://github.com/alibaba/MNN.git
91 lines
3.5 KiB
GLSL
91 lines
3.5 KiB
GLSL
layout(std430) buffer;
|
|
layout(FORMAT, binding=0) writeonly uniform PRECISION image3D uOutput;
|
|
layout(location=1) uniform mediump sampler3D uInput;
|
|
layout(location=2) uniform mediump sampler3D uKernel;
|
|
|
|
layout(binding=3) readonly buffer bias{
|
|
vec4 data[];
|
|
} uBias;
|
|
|
|
layout(location=4) uniform ivec2 uPad;
|
|
layout(location=5) uniform ivec2 uKernelSize;
|
|
layout(location=6) uniform ivec2 uStride;
|
|
layout(location=7) uniform ivec2 uDilate;
|
|
layout(location=8) uniform int uUnroll;
|
|
|
|
layout(location=10) uniform ivec3 uOutputSize;
|
|
layout(location=11) uniform ivec3 uInputSize;
|
|
|
|
#define UP_DIV(x, y) (((x)+(y)-1)/(y))
|
|
|
|
//weight : oc ic h w -> oc/4, ic/4, ky kx ic4 oc4
|
|
layout (local_size_x = XLOCAL, local_size_y = YLOCAL, local_size_z = ZLOCAL) in;
|
|
|
|
void main()
|
|
{
|
|
if (all(lessThan(ivec3(gl_GlobalInvocationID), uOutputSize)))
|
|
{
|
|
ivec3 pos = ivec3(gl_GlobalInvocationID)*ivec3(uUnroll, 1, 1);
|
|
int kernelX = uKernelSize.x;
|
|
ivec3 inputSize = uInputSize;
|
|
ivec2 s0 = pos.xy*uStride-uPad;
|
|
int fx, fy, fz;
|
|
ivec2 sfxy = max(ivec2(0), (UP_DIV(-s0, uDilate)));
|
|
ivec2 efxy = min(uKernelSize, UP_DIV(inputSize.xy-s0, uDilate));
|
|
vec4 color = uBias.data[pos.z];
|
|
vec4 color2 = color;
|
|
vec4 color3 = color;
|
|
vec4 color4 = color;
|
|
int kernelY = pos.z;
|
|
for (fy=sfxy.y; fy<efxy.y; ++fy)
|
|
{
|
|
int sy = fy*uDilate.y + s0.y;
|
|
for (fx=0; fx<kernelX; ++fx)
|
|
{
|
|
int kernelZ = fx + fy*kernelX;
|
|
int sx1 = fx*uDilate.x + s0.x;
|
|
int sx2 = sx1 + uStride.x;
|
|
int sx3 = sx1 + uStride.x * 2;
|
|
int sx4 = sx1 + uStride.x * 3;
|
|
float m1 = sx1 >= 0&& sx1 < inputSize.x ? 1.0 : 0.0;
|
|
float m2 = sx2 >= 0&& sx2 < inputSize.x ? 1.0 : 0.0;
|
|
float m3 = sx3 >= 0&& sx3 < inputSize.x ? 1.0 : 0.0;
|
|
float m4 = sx4 >= 0&& sx4 < inputSize.x ? 1.0 : 0.0;
|
|
fz = 0;
|
|
for (; fz<inputSize.z; ++fz)
|
|
{
|
|
int kernelX = 4*fz;
|
|
vec4 k0 = texelFetch(uKernel, ivec3(kernelX+0, kernelY, kernelZ), 0);
|
|
vec4 k1 = texelFetch(uKernel, ivec3(kernelX+1, kernelY, kernelZ), 0);
|
|
vec4 k2 = texelFetch(uKernel, ivec3(kernelX+2, kernelY, kernelZ), 0);
|
|
vec4 k3 = texelFetch(uKernel, ivec3(kernelX+3, kernelY, kernelZ), 0);
|
|
|
|
mat4 k = mat4(k0, k1, k2, k3);
|
|
|
|
color += k*texelFetch(uInput, ivec3(sx1, sy, fz), 0) * m1;
|
|
color2 += k*texelFetch(uInput, ivec3(sx2, sy, fz), 0) * m2;
|
|
color3 += k*texelFetch(uInput, ivec3(sx3, sy, fz), 0) * m3;
|
|
color4 += k*texelFetch(uInput, ivec3(sx4, sy, fz), 0) * m4;
|
|
}
|
|
}
|
|
}
|
|
#ifdef RELU
|
|
color = max(color, vec4(0));
|
|
color2 = max(color2, vec4(0));
|
|
color3 = max(color3, vec4(0));
|
|
color4 = max(color4, vec4(0));
|
|
#endif
|
|
#ifdef RELU6
|
|
color = clamp(color, vec4(0), vec4(6));
|
|
color2 = clamp(color2, vec4(0), vec4(6));
|
|
color3 = clamp(color3, vec4(0), vec4(6));
|
|
color4 = clamp(color4, vec4(0), vec4(6));
|
|
#endif
|
|
imageStore(uOutput, ivec3(pos.x+0, pos.y, pos.z), color);
|
|
imageStore(uOutput, ivec3(pos.x+1, pos.y, pos.z), color2);
|
|
imageStore(uOutput, ivec3(pos.x+2, pos.y, pos.z), color3);
|
|
imageStore(uOutput, ivec3(pos.x+3, pos.y, pos.z), color4);
|
|
}
|
|
|
|
}
|