I am trying to optimize a CUDA code with LLVM passes on a PowerPC system (RHEL 7.6 with no root access) equipped with V100 GPUs, CUDA 10.1, and LLVM 11 (built from source). Also, I tested clang, lli, and opt on a simple C++ code, and everything works just fine.
After days of searching, reading, and trials-and-errors, I managed to compile a simple CUDA source. The code is the famous axpy:
#include <iostream>
#define cudaCheckError() \
{ \
cudaError_t e = cudaGetLastError(); \
if (e != cudaSuccess) { \
printf("Cuda failure %s:%d: '%s'\n", __FILE__, __LINE__, \
cudaGetErrorString(e)); \
exit(EXIT_FAILURE); \
} \
}
__global__ void axpy(float a, float* x, float* y) {
y[threadIdx.x] = a * x[threadIdx.x];
}
int main(int argc, char* argv[]) {
const int kDataLen = 4;
float a = 2.0f;
float host_x[kDataLen] = {1.0f, 2.0f, 3.0f, 4.0f};
float host_y[kDataLen];
// Copy input data to device.
float* device_x;
float* device_y;
cudaMalloc(&device_x, kDataLen * sizeof(float));
cudaMalloc(&device_y, kDataLen * sizeof(float));
cudaMemcpy(device_x, host_x, kDataLen * sizeof(float),
cudaMemcpyHostToDevice);
// Launch the kernel.
axpy<<<1, kDataLen>>>(a, device_x, device_y);
cudaCheckError();
// Copy output data to host.
cudaDeviceSynchronize();
cudaMemcpy(host_y, device_y, kDataLen * sizeof(float),
cudaMemcpyDeviceToHost);
// Print the results.
for (int i = 0; i < kDataLen; ++i) {
std::cout << "y[" << i << "] = " << host_y[i] << "\n";
}
cudaDeviceReset();
return 0;
}
And I wrote the compilation steps in a Makefile like this (I know, its style can be improved!): I haven't added any LLVM passes yet.
BIN_FILE=axpy
SRC_FILE=$(BIN_FILE).cu
main: $(BIN_FILE)
$(BIN_FILE).ll: $(SRC_FILE)
clang++ -stdlib=libc++ -Wall $(BIN_FILE).cu --cuda-gpu-arch=sm_70 -S -c -emit-llvm
$(BIN_FILE)-cuda-nvptx64-nvidia-cuda-sm_70.ll: $(SRC_FILE)
clang++ -stdlib=libc++ -Wall $(BIN_FILE).cu --cuda-gpu-arch=sm_70 -S -c -emit-llvm
$(BIN_FILE).ptx: $(BIN_FILE)-cuda-nvptx64-nvidia-cuda-sm_70.ll
llc -march=nvptx64 $(BIN_FILE)-cuda-nvptx64-nvidia-cuda-sm_70.ll -o $(BIN_FILE).ptx
$(BIN_FILE)_dlink.o: $(BIN_FILE).ptx
ptxas -m64 --gpu-name=sm_70 $(BIN_FILE).ptx -o $(BIN_FILE).ptx.o
fatbinary --64 --create $(BIN_FILE).fatbin --image=profile=sm_70,file=$(BIN_FILE).ptx.o
nvcc $(BIN_FILE).fatbin -arch=sm_70 -dlink -o $(BIN_FILE)_dlink.o -rdc=true
# For the host code:
$(BIN_FILE).o: $(BIN_FILE).ll
llc -mcpu=ppc64 $(BIN_FILE).ll -o $(BIN_FILE).s
clang++ -c $(BIN_FILE).s -o $(BIN_FILE).o
# Link both object files together with a linker:
$(BIN_FILE): $(BIN_FILE).o $(BIN_FILE)_dlink.o
nvcc $(BIN_FILE).o $(BIN_FILE)_dlink.o -o $(BIN_FILE) -arch=sm_70 -lc++
clean:
rm *.ll *.s *.ptx *.ptx.o *.fatbin $(BIN_FILE) $(BIN_FILE).o $(BIN_FILE)_dlink.o
It seems all the steps run smoothly without any warning, but after running the executable file, I get the error:
Cuda failure axpy.cu:33: 'invalid device function'
I have also replaced the last linker command with the following, and it runs ok, but with the same error.
clang++ -stdlib=libc++ $(BIN_FILE).o $(BIN_FILE)_dlink.o -o $(BIN_FILE) -lcuda -lcudart -lcudadevrt \
-L/path-to-gcc-lib/
I really appreciate any help. It's worth mentioning that I need to utilize some of the modern features of CUDA, most of all cooperative groups, so I thought recent LLVM releases might be working. Wondering whether changing LLVM version might be helpful.
-- Edit:
The output of cuobjdump axpy.fatbin -ptx -sass
:
Fatbin elf code:
================
arch = sm_70
code version = [1,7]
producer = <unknown>
host = linux
compile_size = 64bit
code for sm_70
Function : _Z4axpyfPfS_
.headerflags @"EF_CUDA_SM70 EF_CUDA_PTX_SM(EF_CUDA_SM70)"
/*0000*/ MOV R1, c[0x0][0x28] ; /* 0x00000a0000017a02 */
/* 0x000fd00000000f00 */
/*0010*/ @!PT SHFL.IDX PT, RZ, RZ, RZ, RZ ; /* 0x000000fffffff389 */
/* 0x000fe200000e00ff */
/*0020*/ IADD3 R1, R1, -0x18, RZ ; /* 0xffffffe801017810 */
/* 0x000fe20007ffe0ff */
/*0030*/ IMAD.MOV.U32 R9, RZ, RZ, c[0x0][0x16c] ; /* 0x00005b00ff097624 */
/* 0x000fe200078e00ff */
/*0040*/ MOV R8, c[0x0][0x168] ; /* 0x00005a0000087a02 */
/* 0x000fe40000000f00 */
/*0050*/ IADD3 R2, P0, R1, c[0x0][0x20], RZ ; /* 0x0000080001027a10 */
/* 0x000fc80007f1e0ff */
/*0060*/ IADD3.X R3, RZ, c[0x0][0x24], RZ, P0, !PT ; /* 0x00000900ff037a10 */
/* 0x000fd000007fe4ff */
/*0070*/ ST.E.64.SYS [R2+0x8], R8 ; /* 0x0000000802007385 */
/* 0x0001e8000010eb08 */
/*0080*/ LD.E.64.SYS R4, [R2+0x8] ; /* 0x0000000802047980 */
/* 0x000ea2000010eb00 */
/*0090*/ IMAD.MOV.U32 R10, RZ, RZ, c[0x0][0x170] ; /* 0x00005c00ff0a7624 */
/* 0x000fe200078e00ff */
/*00a0*/ MOV R11, c[0x0][0x174] ; /* 0x00005d00000b7a02 */
/* 0x000fe20000000f00 */
/*00b0*/ IMAD.MOV.U32 R15, RZ, RZ, c[0x0][0x160] ; /* 0x00005800ff0f7624 */
/* 0x000fe200078e00ff */
/*00c0*/ S2R R13, SR_TID.X ; /* 0x00000000000d7919 */
/* 0x000eac0000002100 */
/*00d0*/ ST.E.64.SYS [R2+0x10], R10 ; /* 0x0000001002007385 */
/* 0x000fe8000010eb0a */
/*00e0*/ LD.E.64.SYS R6, [R2+0x10] ; /* 0x0000001002067980 */
/* 0x000ee8000010eb00 */
/*00f0*/ ST.E.SYS [R2], R15 ; /* 0x0000000002007385 */
/* 0x000fe8000010e90f */
/*0100*/ LD.E.SYS R0, [R2] ; /* 0x0000000002007980 */
/* 0x000e22000010e900 */
/*0110*/ IMAD.WIDE.U32 R4, R13, 0x4, R4 ; /* 0x000000040d047825 */
/* 0x004fd400078e0004 */
/*0120*/ LD.E.SYS R5, [R4] ; /* 0x0000000004057980 */
/* 0x000e22000010e900 */
/*0130*/ IMAD.WIDE.U32 R6, R13, 0x4, R6 ; /* 0x000000040d067825 */
/* 0x008fe400078e0006 */
/*0140*/ FMUL R9, R0, R5 ; /* 0x0000000500097220 */
/* 0x001fd00000400000 */
/*0150*/ ST.E.SYS [R6], R9 ; /* 0x0000000006007385 */
/* 0x000fe2000010e909 */
/*0160*/ EXIT ; /* 0x000000000000794d */
/* 0x000fea0003800000 */
/*0170*/ BRA 0x170; /* 0xfffffff000007947 */
/* 0x000fc0000383ffff */
.......................
The output ends with dots.
The problem was not related to PowerPC architecture. I needed to pass the fatbin
file to the host-side compilation command with -Xclang -fcuda-include-gpubinary -Xclang axpy.fatbin
to replicate the whole compilation behavior.
Here is the corrected Makefile:
BIN_FILE=axpy
SRC_FILE=$(BIN_FILE).cu
main: $(BIN_FILE)
# Host Side
$(BIN_FILE).ll: $(SRC_FILE) $(BIN_FILE).fatbin
clang++ -stdlib=libc++ -Wall -Werror $(BIN_FILE).cu -march=ppc64le --cuda-host-only -relocatable-pch \
-Xclang -fcuda-include-gpubinary -Xclang $(BIN_FILE).fatbin -S -g -c -emit-llvm
$(BIN_FILE).o: $(BIN_FILE).ll
llc -march=ppc64le $(BIN_FILE).ll -o $(BIN_FILE).s
clang++ -c -Wall $(BIN_FILE).s -o $(BIN_FILE).o
# GPU Side
$(BIN_FILE)-cuda-nvptx64-nvidia-cuda-sm_70.ll: $(SRC_FILE)
clang++ -x cuda -stdlib=libc++ -Wall -Werror $(BIN_FILE).cu --cuda-device-only \
--cuda-gpu-arch=sm_70 -S -g -emit-llvm
$(BIN_FILE).ptx: $(BIN_FILE)-cuda-nvptx64-nvidia-cuda-sm_70.ll
llc -march=nvptx64 -mcpu=sm_70 -mattr=+ptx64 $(BIN_FILE)-cuda-nvptx64-nvidia-cuda-sm_70.ll -o $(BIN_FILE).ptx
$(BIN_FILE).ptx.o: $(BIN_FILE).ptx
ptxas -m64 --gpu-name=sm_70 $(BIN_FILE).ptx -o $(BIN_FILE).ptx.o
$(BIN_FILE).fatbin: $(BIN_FILE).ptx.o
fatbinary --64 --create $(BIN_FILE).fatbin --image=profile=sm_70,file=$(BIN_FILE).ptx.o \
--image=profile=compute_70,file=$(BIN_FILE).ptx -link
$(BIN_FILE)_dlink.o: $(BIN_FILE).fatbin
nvcc $(BIN_FILE).fatbin -gencode arch=compute_70,code=sm_70 \
-dlink -o $(BIN_FILE)_dlink.o -lcudart -lcudart_static -lcudadevrt
# Link both object files together (either nvcc or clang works here):
$(BIN_FILE): $(BIN_FILE).o $(BIN_FILE)_dlink.o
nvcc $(BIN_FILE).o $(BIN_FILE)_dlink.o -o $(BIN_FILE) -arch=sm_70 -lc++
Figure 1 in this link includes the creation steps of the fatbinary file.