Local variable alignment for ATI/AMD OpenCL - opencl

I've been having trouble with a misaligned structure. Here are the structures involved:
struct Ray
{
float4 origin;
float4 dir;
float len;
float dummy [3];
};
struct RayStack
{
struct Ray r [STACK_DEPTH];
int depth [STACK_DEPTH];
float refr [STACK_DEPTH];
int top;
float dummy [3];
};
Incidentally, STACK_DEPTH is a multiple of 4. I've been careful to make sure that all the structures are a multiple of 16 in size and that float4's within are on an aligned boundary.
The problem is when I use it as a local variable, the structure RayStack is unaligned:
struct RayStack stack;
printf("stack: %p\n", &stack);
The stack address ends up ending in 8 and not 0 as I would want for a 16-byte aligned structure. This causes a crash on ATI cards (although Intel and nVidia are not bothered by it). I've tried placing __attribute__((aligned(16))) in the structure (before and after), and in the local variable definition and that doesn't change anything. Actually, adding a printf statement fixed the problem although I have no idea how.
Is there away to ensure that the local variable stack is aligned on a 16 byte boundary and stop the crashing on ATI cards.
Thanks!

You do know that arrays in structures have to be aligned to a 16 byte boundary?
What's the "dummy" array for? Padding? If so, don't use arrays for padding.
From my experience with Nvidia, ATI and Intel, the following is the most safe method:
struct Ray
{
float4 origin;
float4 dir;
float len;
float padding1;
float padding2;
float padding3;
};
struct RayStack
{
struct Ray r[STACK_DEPTH];
int depth[STACK_DEPTH];
float refr[STACK_DEPTH];
int top;
float padding1;
float padding2;
float padding3;
};

Related

Static variable in OpenCL C

I'm writing a renderer from scratch using openCL and I have a little compilation problem on my kernel with the error :
CL_BUILD_PROGRAM : error: program scope variable must reside in constant address space static float* objects;
The problem is that this program compiles on my desktop (with nvidia drivers) and doesn't work on my laptop (with nvidia drivers), also I have the exact same kernel file in another project that works fine on both computers...
Does anyone have an idea what I could be doing wrong ?
As a clarification, I'm coding a raymarcher which's kernel takes a list of objects "encoded" in a float array that is needed a lot in the program and that's why I need it accessible to the hole kernel.
Here is the kernel code simplified :
float* objects;
float4 getDistCol(float3 position) {
int arr_length = objects[0];
float4 distCol = {INFINITY, 0, 0, 0};
int index = 1;
while (index < arr_length) {
float objType = objects[index];
if (compare(objType, SPHERE)) {
// Treats the part of the buffer as a sphere
index += SPHERE_ATR_LENGTH;
} else if (compare(objType, PLANE)) {
//Treats the part of the buffer as a plane
index += PLANE_ATR_LENGTH;
} else {
float4 errCol = {500, 1, 0, 0};
return errCol;
}
}
}
__kernel void mkernel(__global int *image, __constant int *dimension,
__constant float *position, __constant float *aimDir, __global float *objs) {
objects = objs;
// Gets ray direction and stuf
// ...
// ...
float4 distCol = RayMarch(ro, rd);
float3 impact = rd*distCol.x + ro;
col = distCol.yzw * GetLight(impact);
image[dimension[0]*dimension[1] - idx*dimension[1]+idy] = toInt(col);
Where getDistCol(float3 position) gets called a lot by a lot of functions and I would like to avoid having to pass my float buffer to every function that needs to call getDistCol()...
There is no "static" variables allowed in OpenCL C that you can declare outside of kernels and use across kernels. Some compilers might still tolerate this, others might not. Nvidia has recently changed their OpenCL compiler from LLVM 3.4 to NVVM 7 in a driver update, so you may have the 2 different compilers on your desktop/laptop GPUs.
In your case, the solution is to hand the global kernel parameter pointer over to the function:
float4 getDistCol(float3 position, __global float *objects) {
int arr_length = objects[0]; // access objects normally, as you would in the kernel
// ...
}
kernel void mkernel(__global int *image, __constant int *dimension, __constant float *position, __constant float *aimDir, __global float *objs) {
// ...
getDistCol(position, objs); // hand global objs pointer over to function
// ...
}
Lonely variables out in the wild are only allowed as constant memory space, which is useful for large tables. They are cached in L2$, so read-only access is potentially faster. Example
constant float objects[1234] = {
1.0f, 2.0f, ...
};

OpenCL void pointer arithmetic - strange behavior

I have wrote an OpenCL kernel that is using the opencl-opengl interoperability to read vertices and indices, but probably this is not even important because I am just doing simple pointer addition in order to get a specific vertex by index.
uint pos = (index + base)*stride;
Here i am calculating the absolute position in bytes, in my example pos is 28,643,328 with a stride of 28, index = 0 and base = 1,022,976. Well, that seems correct.
Unfortunately, I cant use vload3 directly because the offset parameter isn't calculated as an absolute address in bytes. So I just add pos to the pointer void* vertices_gl
void* new_addr = vertices_gl+pos;
new_addr is in my example = 0x2f90000 and this is where the strange part begins,
vertices_gl = 0x303f000
The result (new_addr) should be 0x4B90000 (0x303f000 + 28,643,328)
I dont understand why the address vertices_gl is getting decreased by 716,800 (0xAF000)
I'm targeting the GPU: AMD Radeon HD5830
Ps: for those wondering, I am using a printf to get these values :) ( couldn't get CodeXL working)
There is no pointer arithmetic for void* pointers. Use char* pointers to perform byte-wise pointer computations.
Or a lot better than that: Use the real type the pointer is pointing to, and don't multiply offsets. Simply write vertex[index+base] assuming vertex points to your type containing 28 bytes of data.
Performance consideration: Align your vertex attributes to a power of two for coalesced memory access. This means, add 4 bytes of padding after each vertex entry. To automatically do this, use float8 as the vertex type if your attributes are all floating point values. I assume you work with position and normal data or something similar, so it might be a good idea to write a custom struct which encapsulates both vectors in a convenient and self-explaining way:
// Defining a type for the vertex data. This is 32 bytes large.
// You can share this code in a header for inclusion in both OpenCL and C / C++!
typedef struct {
float4 pos;
float4 normal;
} VertexData;
// Example kernel
__kernel void computeNormalKernel(__global VertexData *vertex, uint base) {
uint index = get_global_id(0);
VertexData thisVertex = vertex[index+base]; // It can't be simpler!
thisVertex.normal = computeNormal(...); // Like you'd do it in C / C++!
vertex[index+base] = thisVertex; // Of couse also when writing
}
Note: This code doesn't work with your stride of 28 if you just change one of the float4s to a float3, since float3 also consumes 4 floats of memory. But you can write it like this, which will not add padding (but note that this will penalize memory access bandwidth):
typedef struct {
float pos[4];
float normal[3]; // Assuming you want 3 floats here
} VertexData;

Does CUDA support pointer-aliasing?

The reason why I ask this is because there is some strange bug in my code and I suspect it could be some aliasing problem:
__shared__ float x[32];
__shared__ unsigned int xsum[32];
int idx=threadIdx.x;
unsigned char * xchar=(unsigned char *)x;
//...do something
if (threadIdx.x<32)
{
xchar[4*idx]&=somestring[0];
xchar[4*idx+1]&=somestring[1];
xchar[4*idx+2]&=somestring[2];
xchar[4*idx+3]&=somestring[3];
xsum[idx]+=*((unsigned int *)(x+idx));//<-Looks like the compiler sometimes fail to recongize this as the aliasing of xchar;
};
The compiler only needs to honour aliasing between compatible types. Since char and float are not compatible, the compiler is free to assume the pointers never alias.
If you want to do bitwise operations on float, firstly convert (via __float_as_int()) to unsigned integer, then operate on that, and finally convert back to float (using __int_as_float()).
I think you have a race condition here. But I don't know what is somestring. If it is the same for all threads you can do like this:
__shared__ float x[32];
unsigned char * xchar=(unsigned char *)x;
//...do something
if(threadIdx.x<4) {
xchar[threadIdx.x]&=somestring[threadIdx.x];
}
__syncthreads();
unsigned int xsum+=*((unsigned int *)x);
It means that every thread shares the same array and therefore, xsum is the same between all threads. If you want that each thread has its own array, you have to allocate an array of 32*number_of_threads_in_block and use an offset.
PS: the code above works only in 1D block. In 2D or 3D you have to compute you own threadID and be sure that only 4 threads execute the code.

Avoiding data alignment in OpenCL

I need to pass a complex data type to OpenCL as a buffer and I want (if possible) to avoid the buffer alignment.
In OpenCL I need to use two structures to differentiate the data passed in the buffer casting to them:
typedef struct
{
char a;
float2 position;
} s1;
typedef struct
{
char a;
float2 position;
char b;
} s2;
I define the kernel in this way:
__kernel void
Foo(
__global const void* bufferData,
const int amountElements // in the buffer
)
{
// Now I cast to one of the structs depending on an extra value
__global s1* x = (__global s1*)bufferData;
}
And it works well only when I align the data passed in the buffer.
The question is: Is there a way to use _attribute_ ((packed)) or _attribute_((aligned(1))) to avoid the alignment in data passed in the buffer?
If padding the smaller structure is not an option, I suggest passing another parameter to let your kernel function know what the type is - maybe just the size of the elements.
Since you have data types that are 9 and 10 bytes, it may be worth a try padding them both out to 12 bytes depending on how many of them you read within your kernel.
Something else you may be interested in is the extension: cl_khr_byte_addressable_store
http://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/cl_khr_byte_addressable_store.html
update:
I didn't realize you were passing a mixed array, I thought It was uniform in type. If you want to track the type on a per-element basis, you should pass a list of the types (or codes). Using float2 on its own in bufferData would probably be faster as well.
__kernel void
Foo(
__global const float2* bufferData,
__global const char* bufferTypes,
const int amountElements // in the buffer
)

CUDA device pointer manipulation

I've used:
float *devptr;
//...
cudaMalloc(&devptr, sizeofarray);
cudaMemcpy(devptr, hostptr, sizeofarray, cudaMemcpyHostToDevice);
in CUDA C to allocate and populate an array.
Now I'm trying to run a cuda kernel, e.g.:
__global__ void kernelname(float *ptr)
{
//...
}
in that array but with an offset value.
In C/C++ it would be someting like this:
kernelname<<<dimGrid, dimBlock>>>(devptr+offset);
However, this doesn't seem to work.
Is there a way to do this without sending the offset value to the kernel in a separate argument and use that offset in the kernel code?
Any ideas on how to do this?
Pointer arithmetic does work just fine in CUDA. You can add an offset to a CUDA pointer in host code and it will work correctly (remembering the offset isn't a byte offset, it is a plain word or element offset).
EDIT: A simple working example:
#include <cstdio>
int main(void)
{
const int na = 5, nb = 4;
float a[na] = { 1.2, 3.4, 5.6, 7.8, 9.0 };
float *_a, b[nb];
size_t sza = size_t(na) * sizeof(float);
size_t szb = size_t(nb) * sizeof(float);
cudaFree(0);
cudaMalloc((void **)&_a, sza );
cudaMemcpy( _a, a, sza, cudaMemcpyHostToDevice);
cudaMemcpy( b, _a+1, szb, cudaMemcpyDeviceToHost);
for(int i=0; i<nb; i++)
printf("%d %f\n", i, b[i]);
cudaThreadExit();
}
Here, you can see a word/element offset has been applied to the device pointer in the second cudaMemcpy call to start the copy from the second word, not the first.
Pointer arithmetic does work on host side code, it's used fairly often in the example code provided by nvidia.
"Linear memory exists on the device in a 40-bit address space, so separately allocated entities can reference one another via pointers, for example, in a binary tree."
Read more at: http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#ixzz4KialMz00
And from the performance primitives (npp) documentation, a perfect example of pointer arithmetic.
"4.5.1 Select-Channel Source-Image Pointer
This is a pointer to the channel-of-interest within the first pixel of the source image. E.g. if pSrc is the
pointer to the first pixel inside the ROI of a three channel image. Using the appropriate select-channel copy
primitive one could copy the second channel of this source image into the first channel of a destination
image given by pDst by offsetting the pointer by one:
nppiCopy_8u_C3CR(pSrc + 1, nSrcStep, pDst, nDstStep, oSizeROI);"
*Note: this works without multiplying by the number of bytes per data element because the compiler is aware of the data type of the pointer, and calculates the address accordingly.
In C and C++, pointer arithmetic can be accomplished as above or by the notation &ptr[offset] (to return device memory address of data instead of value, value will not work on device memory from host side code). When using either notation the size of the data type is automatically handled, and the offset is specified as a number of data elements rather than bytes.

Resources