Does CUDA (latest Tesla, Quadro) accelerate integer summations when atomicAdd is used on same shared memory address for all threads of block?

Accelerated Computing CUDA CUDA Programming and Performance
1

In some new CUDA architectures, does an SM unit use a dedicated prefix-sum (sum-scan) hardware for computing atomicAdd(s_sameAddress, 1) quicker than a warp-primitive based reduction/scan algorithm? Even when all warps of block target the same shared memory address?

2

I’ve not heard of any such hardware. There are whitepapers published for many chip architectures, did you read about it somewhere?

There is a compiler optimization that can identify warp-wide atomic activity, and aggregate it using warp shuffle into an atomic performed by a single thread, thus reducing “atomic pressure”. You can read about the basics here

3

I was experimenting warp-primitive based implementation versus atomicAdd based implementation for reduction and atomicAdd version was a bit faster even though I had optimized warp-primitive version. I don’t remember exact algorithm, there was warp-level-aggregation but not thread-level-aggregation.

4

Adding some detail about comparison, in histogram algorithm:

#pragma unroll 4
for (int j = 0; j < 4; j++) {
   uint32_t k = ka[j];
   uint32_t mask = __match_any_sync(0xFFFFFFFF, k);
   const int count = __popc(mask);
   const int leaderLane = __ffs(mask) - 1;
   if (warpLane == leaderLane) {
        atomicAdd(&s_histogram[k], count);
   }
}

this is 30% slower than this:

#pragma unroll 4
for (int j = 0; j < 4; j++) {
   uint32_t k = ka[j];
   atomicAdd(&s_histogram[k], 1);
}

when all inputs are duplicates.

Despite having many times more atomicAdds, its faster.

5

studying the sass may yield insight. First of all the compiler may already be doing warp-aggregation in the 2nd case. Second, it may have found a simpler instruction sequence to aggregate the count across the warp.

6

It produced this part:

        /*02f0*/               @P3 BRA 0x390 ;                                     /* 0x0000009000003947 */
                                                                                   /* 0x000fea0003800000 */
        /*0300*/                   LDG.E.128.CONSTANT R4, [R2.64+0x280000] ;       /* 0x2800000402047981 */
                                                                                   /* 0x002ea4000c1e9d00 */
        /*0310*/                   F2I.U32.TRUNC.NTZ R4, R4 ;                      /* 0x0000000400047305 */
                                                                                   /* 0x004e70000020f000 */
        /*0320*/                   F2I.U32.TRUNC.NTZ R5, R5 ;                      /* 0x0000000500057305 */
                                                                                   /* 0x000eb0000020f000 */
        /*0330*/                   F2I.U32.TRUNC.NTZ R6, R6 ;                      /* 0x0000000600067305 */
                                                                                   /* 0x000ee2000020f000 */
        /*0340*/                   ATOMS.POPC.INC.32 RZ, [R4.X4+URZ] ;             /* 0x0000000004ff7f8c */
                                                                                   /* 0x0023ee000d00403f */
        /*0350*/                   F2I.U32.TRUNC.NTZ R7, R7 ;                      /* 0x0000000700077305 */
                                                                                   /* 0x000f22000020f000 */
        /*0360*/                   ATOMS.POPC.INC.32 RZ, [R5.X4+URZ] ;             /* 0x0000000005ff7f8c */
                                                                                   /* 0x0043e8000d00403f */
        /*0370*/                   ATOMS.POPC.INC.32 RZ, [R6.X4+URZ] ;             /* 0x0000000006ff7f8c */
                                                                                   /* 0x0083e8000d00403f */
        /*0380*/                   ATOMS.POPC.INC.32 RZ, [R7.X4+URZ] ;             /* 0x0000000007ff7f8c */
                                                                                   /* 0x0103e4000d00403f */
        /*0390*/                   BSYNC B0 ;                                      /* 0x0000000000007941 */

ATOMS.POPC.INC instruction looks like atomic increment but also about population count. Maybe a fused instruction like FMA?

7

You could try replacing the 1 with another number to prevent .INC variants. Just for testing.