optimization - CUDA compiler produce unoptimal assembler

Question

I have compiled flowing simple test kernel (CUDA5, sm2.0):

__device__ void TestKernel(int *pdata)
    {
      int a0,b0,c0;

      a0 = pdata[0];
      b0 = pdata[1];

      c0 = a0 + b0;
      pdata[2] = c0;
    }

and expect something like flowing assembler:

LD R3,[R0]
LD R4,[R0+4]
IADD R4,R4,R3
ST [R0+8],R4

but from cuobjdump --dump-sass I see flowing much longer result:

/*0000*/     /*0x10001de428000000*/     MOV R0, R4;
/*0008*/     /*0x00001de428000000*/     MOV R0, R0;
/*0010*/     /*0x00001de428000000*/     MOV R0, R0;
/*0018*/     /*0x00001de428000000*/     MOV R0, R0;
/*0020*/     /*0x0000dc8580000000*/     LD R3, [R0];
/*0028*/     /*0x0c00dde428000000*/     MOV R3, R3;
/*0030*/     /*0x10011c034800c000*/     IADD R4, R0, 0x4;
/*0038*/     /*0x10011de428000000*/     MOV R4, R4;
/*0040*/     /*0x00411c8580000000*/     LD R4, [R4];
/*0048*/     /*0x10011de428000000*/     MOV R4, R4;
/*0050*/     /*0x1030dc0348000000*/     IADD R3, R3, R4;
/*0058*/     /*0x20001c034800c000*/     IADD R0, R0, 0x8;
/*0060*/     /*0x00001de428000000*/     MOV R0, R0;
/*0068*/     /*0x0000dc8590000000*/     ST [R0], R3;
/*0070*/     /*0x00001de790000000*/     RET;
/*0078*/     /*0x00001de780000000*/     EXIT;
/*0080*/     /*0x00001de780000000*/     EXIT;

Very strange to me MOVs instruction in addresses 8,10,18,28,38,60 also the immediate offset in load/store instruction doesn't used. So instead expected 4 (actually 6 including RET,EXIT) instruction I get 15 What is possible reason?

Answer 1

What you are seeing is almost certainly because you are compiling with debugging turned on. If I build your kernel I get this:

$ nvcc -arch=sm_30 -c asmprob.cu 
$ cuobjdump -sass asmprob.o

Fatbin elf code:
================
arch = sm_30
code version = [1,6]
producer = cuda
host = mac
compile_size = 32bit
identifier = asmprob.cu

    code for sm_30
        Function : _Z10TestKernelPi
    /*0008*/     /*0x10005de428004001*/     MOV R1, c [0x0] [0x44];
    /*0010*/     /*0x00009de428004005*/     MOV R2, c [0x0] [0x140];
    /*0018*/     /*0x10211c034800c000*/     IADD R4, R2, 0x4;
    /*0020*/     /*0x20209c034800c000*/     IADD R2, R2, 0x8;
    /*0028*/     /*0x0040dc8580000000*/     LD R3, [R4];
    /*0030*/     /*0xf0401c8583ffffff*/     LD R0, [R4+-0x4];
    /*0038*/     /*0x00301c0348000000*/     IADD R0, R3, R0;
    /*0048*/     /*0x00201c8590000000*/     ST [R2], R0;
    /*0050*/     /*0x00001de780000000*/     EXIT;
    /*0058*/     /*0xe0001de74003ffff*/     BRA 0x58;
    /*0060*/     /*0x00001de440000000*/     NOP CC.T;
    /*0068*/     /*0x00001de440000000*/     NOP CC.T;
    /*0070*/     /*0x00001de440000000*/     NOP CC.T;
    /*0078*/     /*0x00001de440000000*/     NOP CC.T;
        .................................

on the other hand, if I build it with debug settings, I get code just like you show:

$ nvcc -arch=sm_30 -G -c asmprob.cu 
$ cuobjdump -sass asmprob.o

Fatbin elf code:
================
arch = sm_30
code version = [1,6]
producer = cuda
host = mac
compile_size = 32bit
has debug info
compressed
identifier = asmprob.cu

    code for sm_30
        Function : _Z10TestKernelPi
    /*0000*/     /*0x10005de428004001*/     MOV R1, c [0x0] [0x44];
    /*0008*/     /*0x00001de218000005*/     MOV32I R0, 0x140;
    /*0010*/     /*0x00001c8614000000*/     LDC R0, c [0x0] [R0];
    /*0018*/     /*0x00001de428000000*/     MOV R0, R0;
    /*0020*/     /*0x00009c8580000000*/     LD R2, [R0];
    /*0028*/     /*0x08009de428000000*/     MOV R2, R2;
    /*0030*/     /*0x1000dc034800c000*/     IADD R3, R0, 0x4;
    /*0038*/     /*0x0c00dde428000000*/     MOV R3, R3;
    /*0040*/     /*0x0030dc8580000000*/     LD R3, [R3];
    /*0048*/     /*0x0c00dde428000000*/     MOV R3, R3;
    /*0050*/     /*0x0c209c0348000000*/     IADD R2, R2, R3;
    /*0058*/     /*0x20001c034800c000*/     IADD R0, R0, 0x8;
    /*0060*/     /*0x00001de428000000*/     MOV R0, R0;
    /*0068*/     /*0x00009c8590000000*/     ST [R0], R2;
    /*0070*/     /*0x40001de740000000*/     BRA 0x88;
    /*0078*/     /*0x00001de780000000*/     EXIT;
    /*0080*/     /*0x00001de780000000*/     EXIT;
    /*0088*/     /*0x00001de780000000*/     EXIT;
    /*0090*/     /*0x00001de780000000*/     EXIT;
    /*0098*/     /*0xe0001de74003ffff*/     BRA 0x98;
    /*00a0*/     /*0x00001de440000000*/     NOP CC.T;
    /*00a8*/     /*0x00001de440000000*/     NOP CC.T;
    /*00b0*/     /*0x00001de440000000*/     NOP CC.T;
    /*00b8*/     /*0x00001de440000000*/     NOP CC.T;
        .................................

Which makes me think that your question is "why doesn't the compiler produce optimal code when I disable optimisations and compile for the debugger?", which is something of a rhetorical question, methinks....

---

EDIT:

And lest there be any doubts that enabling GPU debugging disables compiler optimisation, consider the following output from ′nvcc′:

$ nvcc -arch=sm_30 -G -c --dryrun asmprob.cu 
#$ _SPACE_= 
#$ _CUDART_=cudart
#$ _HERE_=/usr/local/cuda/bin
#$ _THERE_=/usr/local/cuda/bin
#$ _TARGET_SIZE_=
#$ TOP=/usr/local/cuda/bin/..
#$ PATH=/usr/local/cuda/bin/../open64/bin:/usr/local/cuda/bin/../nvvm:/usr/local/cuda/bin:/opt/local/bin:/opt/local/sbin:/Library/Frameworks/Python.framework/Versions/Current/bin:/usr/bin:/bin:/usr/sbin:/sbin:/usr/local/bin:/usr/local/git/bin:/usr/texbin:/usr/X11/bin:/usr/NX/bin:/usr/local/bin:/Users/talonmies/bin:/usr/local/cuda/bin
#$ INCLUDES="-I/usr/local/cuda/bin/../include"  
#$ LIBRARIES=  "-L/usr/local/cuda/bin/../lib" -lcudart
#$ CUDAFE_FLAGS=
#$ OPENCC_FLAGS=
#$ PTXAS_FLAGS=
#$ gcc -D__CUDA_ARCH__=300 -E -x c++  -DCUDA_DOUBLE_MATH_FUNCTIONS  -D__CUDACC__ -D__NVCC__  "-I/usr/local/cuda/bin/../include"   -include "cuda_runtime.h" -m32 -malign-double -o "/tmp/tmpxft_00005ceb_00000000-6_asmprob.cpp1.ii" "asmprob.cu" 
#$ cudafe --m32 --gnu_version=40201 -tused --no_remove_unneeded_entities --debug_mode  --gen_c_file_name "/tmp/tmpxft_00005ceb_00000000-3_asmprob.cudafe1.c" --stub_file_name "/tmp/tmpxft_00005ceb_00000000-3_asmprob.cudafe1.stub.c" --gen_device_file_name "/tmp/tmpxft_00005ceb_00000000-3_asmprob.cudafe1.gpu" --nv_arch "compute_30" --gen_module_id_file --module_id_file_name "/tmp/tmpxft_00005ceb_00000000-2_asmprob.module_id" --include_file_name "tmpxft_00005ceb_00000000-1_asmprob.fatbin.c" "/tmp/tmpxft_00005ceb_00000000-6_asmprob.cpp1.ii" 
#$ gcc -D__CUDA_ARCH__=300 -E -x c  -DCUDA_DOUBLE_MATH_FUNCTIONS  -D__CUDACC__ -D__NVCC__ -D__CUDANVVM__  -D__CUDA_PREC_DIV -D__CUDA_PREC_SQRT "-I/usr/local/cuda/bin/../include"   -m32 -malign-double -o "/tmp/tmpxft_00005ceb_00000000-7_asmprob.cpp2.i" "/tmp/tmpxft_00005ceb_00000000-3_asmprob.cudafe1.gpu" 
#$ cudafe -w --m32 --gnu_version=40201 --c --debug_mode  --gen_c_file_name "/tmp/tmpxft_00005ceb_00000000-8_asmprob.cudafe2.c" --stub_file_name "/tmp/tmpxft_00005ceb_00000000-8_asmprob.cudafe2.stub.c" --gen_device_file_name "/tmp/tmpxft_00005ceb_00000000-8_asmprob.cudafe2.gpu" --nv_arch "compute_30" --module_id_file_name "/tmp/tmpxft_00005ceb_00000000-2_asmprob.module_id" --include_file_name "tmpxft_00005ceb_00000000-1_asmprob.fatbin.c" "/tmp/tmpxft_00005ceb_00000000-7_asmprob.cpp2.i" 
#$ gcc -D__CUDA_ARCH__=300 -E -x c  -DCUDA_DOUBLE_MATH_FUNCTIONS  -D__CUDABE__ -D__CUDANVVM__  -D__CUDA_PREC_DIV -D__CUDA_PREC_SQRT "-I/usr/local/cuda/bin/../include"   -m32 -malign-double -o "/tmp/tmpxft_00005ceb_00000000-9_asmprob.cpp3.i" "/tmp/tmpxft_00005ceb_00000000-8_asmprob.cudafe2.gpu" 
#$ filehash -s " -g --dont-merge-basicblocks --return-at-end " "/tmp/tmpxft_00005ceb_00000000-9_asmprob.cpp3.i" > "/tmp/tmpxft_00005ceb_00000000-10_asmprob.hash"
#$ gcc -E -x c++ -D__CUDACC__ -D__NVCC__  "-I/usr/local/cuda/bin/../include"   -include "cuda_runtime.h" -m32 -malign-double -o "/tmp/tmpxft_00005ceb_00000000-4_asmprob.cpp4.ii" "asmprob.cu" 
#$ cudafe++ --m32 --gnu_version=40201 --parse_templates --debug_mode  --gen_c_file_name "/tmp/tmpxft_00005ceb_00000000-3_asmprob.cudafe1.cpp" --stub_file_name "tmpxft_00005ceb_00000000-3_asmprob.cudafe1.stub.c" --module_id_file_name "/tmp/tmpxft_00005ceb_00000000-2_asmprob.module_id" "/tmp/tmpxft_00005ceb_00000000-4_asmprob.cpp4.ii" 
#$ cicc  -arch compute_30 -m32 -ftz=0 -prec_div=1 -prec_sqrt=1 -fmad=1 -g -O0 "/tmp/tmpxft_00005ceb_00000000-11_asmprob" "/tmp/tmpxft_00005ceb_00000000-9_asmprob.cpp3.i"  -o "/tmp/tmpxft_00005ceb_00000000-5_asmprob.ptx"
#$ ptxas  -arch=sm_30 -m32  -g --dont-merge-basicblocks --return-at-end "/tmp/tmpxft_00005ceb_00000000-5_asmprob.ptx"  -o "/tmp/tmpxft_00005ceb_00000000-12_asmprob.sm_30.cubin" 
#$ fatbinary --create="/tmp/tmpxft_00005ceb_00000000-1_asmprob.fatbin" -32 --key="xxxxxxxxxx" --ident="asmprob.cu" --cmdline=" -g --dont-merge-basicblocks --return-at-end " -g "--image=profile=sm_30,file=/tmp/tmpxft_00005ceb_00000000-12_asmprob.sm_30.cubin" "--image=profile=compute_30,file=/tmp/tmpxft_00005ceb_00000000-5_asmprob.ptx" --embedded-fatbin="/tmp/tmpxft_00005ceb_00000000-1_asmprob.fatbin.c" --cuda
#$ rm /tmp/tmpxft_00005ceb_00000000-1_asmprob.fatbin
#$ gcc -D__CUDA_ARCH__=300 -E -x c++  -DCUDA_DOUBLE_MATH_FUNCTIONS   -D__CUDA_PREC_DIV -D__CUDA_PREC_SQRT "-I/usr/local/cuda/bin/../include"   -m32 -malign-double -o "/tmp/tmpxft_00005ceb_00000000-13_asmprob.ii" "/tmp/tmpxft_00005ceb_00000000-3_asmprob.cudafe1.cpp" 
#$ gcc -c -x c++ "-I/usr/local/cuda/bin/../include"   -fpreprocessed -m32 -malign-double -o "asmprob.o" "/tmp/tmpxft_00005ceb_00000000-13_asmprob.ii" 

Note the device code compilation phase command:

cicc  -arch compute_30 -m32 -ftz=0 -prec_div=1 -prec_sqrt=1 -fmad=1 -g -O0 <----

ie. debugging builds compile with optimisation set to 0.