5
\$\begingroup\$

I am attempting to optimize a piece of C code which aims to multiply a series of pairs of unsigned shorts and add the result. I am only concerned about the high 16 bits of the result, and I can guarantee that the sum of the multiples will fit in a 32-bit value. I initially coded this in C, and then rewrote it to use SSE2 intrinsics (slowest), and then rewrote it in SSE2 assembler (fastest). I am not an expert at x86 assembler and would appreciate any recommendations on how to speed this code up. Speed is the priority, this is a tight inner loop. It is OK to assume that input will be valid. In addition, portability is not a major concern, this code will only be used on computers with Intel i5 or i7 processors. Attend:

C

register int i; uint16_t* iw_ptr = n->iw; uint16_t* nw_ptr = n->nw; register uint32_t accvalue = 0; if (id < 2 * I_CNT) { for (i = 0; i < I_CNT; i++) { accvalue += (uint32_t) i_ptr[i] * (uint32_t) iw_ptr[i]; } } for (i = 0; i < id; i++) { accvalue += (uint32_t) n_ptr[i] * (uint32_t) nw_ptr[i]; } value = (uint16_t) (accvalue >> 16);

C SSE2

#define SSE2_I ((INPUT_COUNT+7)/8) #define SSE2_N ((NEURON_COUNT+7)/8) register int i; __m128i mm_sums; __m128i mm_arg1; __m128i mm_arg2; __m128i mm_accum = _mm_setzero_si128(); __m128i* mm_iptr = (__m128i*) i_ptr; __m128i* mm_nptr = (__m128i*) n_ptr; __m128i* mm_iwptr = (__m128i*) n->iw_ptr; __m128i* mm_nwptr = (__m128i*) n->nw_ptr; int id_8 = (id + 7)/8; // Round up to nearest multiple of 8 if (id < 2 * I_CNT) { for (i = 0; i < SSE2_I; i++) { mm_arg1 = _mm_loadu_si128(mm_iptr+i); mm_arg2 = _mm_loadu_si128(mm_iwptr+i); mm_sums = _mm_mulhi_epu16(mm_arg1, mm_arg2); mm_accum = _mm_adds_epu16(mm_accum, mm_sums); } } for (i = 0; i < id_8; i++) { mm_arg1 = _mm_loadu_si128(mm_nptr+i); mm_arg2 = _mm_loadu_si128(mm_nwptr+i); mm_sums = _mm_mulhi_epu16(mm_arg1, mm_arg2); mm_accum = _mm_adds_epu16(mm_accum, mm_sums); } _mm_storeu_si128(mm_accum_mem, mm_accum); for (i = 0; i < 8; i++) { value += *(((uint16_t*) mm_accum_mem) + i); }

SSE2 and x86 Assembler

#define SSE2_I ((I_CNT+7)/8) #define SSE2_N ((N_CNT+7)/8) __m128i* mm_iptr = (__m128i*) i_ptr; __m128i* mm_nptr = (__m128i*) n_ptr; __m128i* mm_iwptr = (__m128i*) n->iw_ptr; __m128i* mm_nwptr = (__m128i*) n->nw_ptr; int id_8 = (id + 7)/8; // Divide by 8, round up asm( "MOVL %5, %%eax \n\t" // Current ID "MOVL %6, %%ebx \n\t" // 2*I_CNT "CMP %%eax, %%ebx \n\t" // If ID >= 2 * input count "JGE Nstart1 \n\t" // Skip the first step "MOVL %1, %%eax \n\t" "MOVL %3, %%ebx \n\t" "MOVDQA (%%eax), %%xmm0 \n\t" "MOVDQA (%%ebx), %%xmm1 \n\t" "PMULHUW %%xmm1, %%xmm0 \n\t" #if SSE2_I > 1 "MOVDQA 0x10(%%eax), %%xmm1 \n\t" "MOVDQA 0x10(%%ebx), %%xmm2 \n\t" "PMULHUW %%xmm1, %%xmm2 \n\t" "PADDUSW %%xmm2, %%xmm0 \n\t" #endif #if SSE2_I > 2 "MOVDQA 0x20(%%eax), %%xmm1 \n\t" "MOVDQA 0x20(%%ebx), %%xmm2 \n\t" "PMULHUW %%xmm1, %%xmm2 \n\t" "PADDUSW %%xmm2, %%xmm0 \n\t" #endif #if SSE2_I > 3 "MOVDQA 0x30(%%eax), %%xmm1 \n\t" "MOVDQA 0x30(%%ebx), %%xmm2 \n\t" "PMULHUW %%xmm1, %%xmm2 \n\t" "PADDUSW %%xmm2, %%xmm0 \n\t" #endif "JMP Nstart2 \n\t" "Nstart1: \n\t" // This is our first multiplication "MOVL %2, %%edi \n\t" "MOVL %4, %%esi \n\t" "MOVDQA (%%edi), %%xmm0 \n\t" "MOVDQA (%%esi), %%xmm1 \n\t" "PMULHUW %%xmm1, %%xmm0 \n\t" "JMP Nstart3 \n\t" "Nstart2: \n\t" // This is not our first multiplication "MOVL %2, %%edi \n\t" "MOVL %4, %%esi \n\t" "MOVDQA (%%edi), %%xmm1 \n\t" "MOVDQA (%%esi), %%xmm2 \n\t" "PMULHUW %%xmm1, %%xmm2 \n\t" "PADDUSW %%xmm2, %%xmm0 \n\t" "Nstart3: \n\t" "MOVL %7, %%ebx \n\t" // Current ID, divided by 8, rounded up. // The number of rounds we have to do "DEC %%ebx \n\t" // If it is now 0 or -1 "JLE Endloop \n\t" // We don't have to do any more rounds "MOVL $0x10, %%eax \n\t" // The offset "Loop: \n\t" "MOVDQA (%%edi, %%eax), %%xmm1 \n\t" "MOVDQA (%%esi, %%eax), %%xmm2 \n\t" "PMULHUW %%xmm1, %%xmm2 \n\t" "PADDUSW %%xmm2, %%xmm0 \n\t" "ADDL $0x10, %%eax \n\t" "DEC %%ebx \n\t" // The round we just did "JNE Loop \n\t" // If not zero, do it again "Endloop: \n\t" /** * PREPARE FOR THE ADDING OF THE WORDS * xmm0 3 2 1 0 * xmm1 0 0 3 2 0b00001110 * xmm0 X X 1+3 0+2 PADDUSW * xmm1 X X X 1+3 0b00000001 * xmm0 X X X sum PADDUSW * This however is still two words - we were shuffling doublewords... * But that's not all we can shuffle! * Shuffle lowwords 0b00000001 * PADDUSW * Extract to register as a 16-bit word */ "PSHUFD $0x0E, %%xmm0, %%xmm1 \n\t" "PADDUSW %%xmm1, %%xmm0 \n\t" "PSHUFD $0x01, %%xmm0, %%xmm1 \n\t" "PADDUSW %%xmm1, %%xmm0 \n\t" "PSHUFLW $0x01, %%xmm0, %%xmm1 \n\t" "PADDUSW %%xmm1, %%xmm0 \n\t" "PEXTRW $0x01, %%xmm0, %%eax \n\t" : "=d" (value) // Outputs : "g" (mm_iptr), "g" (mm_nptr), "g" (mm_iwptr), "g" (mm_nwptr), "g" (id), "g" (2*INPUT_COUNT), "g" (id_8) // Inputs : "%xmm0", "%xmm1", "%xmm2", "%eax", "%ebx", "%edi", "%esi" // Clobbered );

I have tried to:

Use the prefetchnta instruction on a few of the pointers (like %2 and %4), this has only increased the execution time.

Use more of the XMM registers by grouping the MOVDQA, MOVDQA and PMULHUW, PADDUSW to up to three iterations at a time. I expected this to be at least a little useful by grouping the reads from memory together and slightly increasing the number of instructions between a read and the corresponding use of that data. This provided no speedup.

I would like to be able to skip the PMULHUW, PADDUSW sequence if either of the operands is all-zeroes. However, MOVDQA does not set the zero register and I can't see any easy way to test that. Is it possible to jump if an XMM register is all zeroes?

Did I screw up any of the assembly loop code?

Any other ways to get one of these running faster?

\$\endgroup\$

1 Answer 1

Reset to default
5
\$\begingroup\$

Some suggestions:

  • forget asm (for now at least) and stick with SSE intrinsics - you can concentrate on optimisation and let the compiler worry about the implementation details like register allocation, instruction scheduling and loop unrolling

  • use a decent compiler - e.g. Intel ICC typically generates better code than gcc in most cases, and Visual Studio typically generates worse code (this is just based on empirical evidence - try as many different compilers as is practical to see which gives best results)

  • use x86-64 if you can - this gives you twice as many SSE registers to play with (16 versus 8) which enables more loop unrolling etc

  • if your data contains a lot of zeroes such that it's worth the cost of testing and branching to avoid redundant multiplies then you can use _mm_testz_si128 (PTEST) to handle this case - make sure you time with/without this optimisation as its potential benefit is highly data-dependent

  • you are doing very little computation relative to the number of loads and stores so you may hit an optimisation brick wall due to finite memory bandwidth (you may have hit this already) - if possible try combining some other operations from before/after this loop so that you can mitigate the cost of loads and stores

  • try and ensure that your data is 16 byte aligned and use only aligned loads/stores if you possibly can - even though Core i7 supposedly has zero performance penalty for misaligned loads in practice there is still an overhead (probably due to increased cache footprint)

  • don't mess with prefetch instructions (for now at least) - it's hard to beat the automatic prefetch in Core i7 and you may well end up reducing performance rather then improving it

\$\endgroup\$
1
  • 1
    \$\begingroup\$ Thanks for the advice. I tried with the SSE intrinsics but they were actually slower, and I had heard that GCC was not great at optimizing them. (Here's some proof: paddusw %xmm1, %xmm0; movdqa %xmm0, %xmm1.) I'm downloading ICC right now (very slowly) and I'll try it. \$\endgroup\$ Commented Jan 6, 2012 at 2:09

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.