There is an existing question "Average of 3 long integers" that is specifically concerned with the efficient computation of the average of three signed integers.
The use of unsigned integers however allows for additional optimizations not applicable to the scenario covered in the previous question. This question is about the efficient computation of the average of three unsigned integers, where the average is rounded towards zero, i.e. in mathematical terms I want to compute ⌊ (a + b + c) / 3 ⌋.
A straightforward way to compute this average is
avg = a / 3 + b / 3 + c / 3 + (a % 3 + b % 3 + c % 3) / 3;
To first order, modern optimizing compilers will transform the divisions into multiplications with a reciprocal plus a shift, and the modulo operations into a back-multiply and a subtraction, where the back-multiply may use a scale_add idiom available on many architectures, e.g. lea on x86_64, add with lsl #n on ARM, iscadd on NVIDIA GPUs.
In trying to optimize the above in a generic fashion suitable for many common platforms, I observe that typically the cost of integer operations is in the relationship logical ≤ (add | sub) ≤ shift ≤ scale_add ≤ mul. Cost here refers to all of latency, throughput limitations, and power consumption. Any such differences become more pronounced when the integer type processed is wider than the native register width, e.g. when processing uint64_t data on a 32-bit processor.
My optimization strategy was therefore to minimize instruction count and replace "expensive" with "cheap" operations where possible, while not increasing register pressure and retaining exploitable parallelism for wide out-of-order processors.
The first observation is that we can reduce a sum of three operands into a sum of two operands by first applying a CSA (carry save adder) that produces a sum value and a carry value, where the carry value has twice the weight of the sum value. The cost of a software-based CSA is five logicals on most processors. Some processors, like NVIDIA GPUs, have a LOP3 instruction that can compute an arbitrary logical expression of three operands in one fell swoop, in which case CSA condenses to two LOP3s (note: I have yet convince the CUDA compiler to emit those two LOP3s; it currently produces four LOP3s!).
The second observation is that because we are computing the modulo of division by 3, we don't need a back-multiply to compute it. We can instead use dividend % 3 = ((dividend / 3) + dividend) & 3, reducing the modulo to an add plus a logical since we already have the division result. This is an instance of the general algorithm: dividend % (2n-1) = ((dividend / (2n-1) + dividend) & (2n-1).
Finally for the division by 3 in the correction term (a % 3 + b % 3 + c % 3) / 3 we don't need the code for generic division by 3. Since the dividend is very small, in [0, 6], we can simplify x / 3 into (3 * x) / 8 which requires just a scale_add plus a shift.
The code below shows my current work-in-progress. Using Compiler Explorer to check the code generated for various platforms shows the tight code I would expect (when compiled with -O3).
However, in timing the code on my Ivy Bridge x86_64 machine using the Intel 13.x compiler, a flaw became apparent: while my code improves latency (from 18 cycles to 15 cycles for uint64_t data) compared to the simple version, throughput worsens (from one result every 6.8 cycles to one result every 8.5 cycles for uint64_t data). Looking at the assembly code more closely it is quite apparent why that is: I basically managed to take the code down from roughly three-way parallelism to roughly two-way parallelism.
Is there a generically applicable optimization technique, beneficial on common processors in particular all flavors of x86 and ARM as well as GPUs, that preserves more parallelism? Alternatively, is there an optimization technique that further reduces overall operation count to make up for reduced parallelism? The computation of the correction term (tail in the code below) seems like a good target. The simplification (carry_mod_3 + sum_mod_3) / 2 looked enticing but delivers an incorrect result for one of the nine possible combinations.
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#define BENCHMARK (1)
#define SIMPLE_COMPUTATION (0)
#if BENCHMARK
#define T uint64_t
#else // !BENCHMARK
#define T uint8_t
#endif // BENCHMARK
T average_of_3 (T a, T b, T c)
{
T avg;
#if SIMPLE_COMPUTATION
avg = a / 3 + b / 3 + c / 3 + (a % 3 + b % 3 + c % 3) / 3;
#else // !SIMPLE_COMPUTATION
/* carry save adder */
T a_xor_b = a ^ b;
T sum = a_xor_b ^ c;
T carry = (a_xor_b & c) | (a & b);
/* here 2 * carry + sum = a + b + c */
T sum_div_3 = (sum / 3); // {MUL|MULHI}, SHR
T sum_mod_3 = (sum + sum_div_3) & 3; // ADD, AND
if (sizeof (size_t) == sizeof (T)) { // "native precision" (well, not always)
T two_carry_div_3 = (carry / 3) * 2; // MULHI, ANDN
T two_carry_mod_3 = (2 * carry + two_carry_div_3) & 6; // SCALE_ADD, AND
T head = two_carry_div_3 + sum_div_3; // ADD
T tail = (3 * (two_carry_mod_3 + sum_mod_3)) / 8; // ADD, SCALE_ADD, SHR
avg = head + tail; // ADD
} else {
T carry_div_3 = (carry / 3); // MUL, SHR
T carry_mod_3 = (carry + carry_div_3) & 3; // ADD, AND
T head = (2 * carry_div_3 + sum_div_3); // SCALE_ADD
T tail = (3 * (2 * carry_mod_3 + sum_mod_3)) / 8; // SCALE_ADD, SCALE_ADD, SHR
avg = head + tail; // ADD
}
#endif // SIMPLE_COMPUTATION
return avg;
}
#if !BENCHMARK
/* Test correctness on 8-bit data exhaustively. Should catch most errors */
int main (void)
{
T a, b, c, res, ref;
a = 0;
do {
b = 0;
do {
c = 0;
do {
res = average_of_3 (a, b, c);
ref = ((uint64_t)a + (uint64_t)b + (uint64_t)c) / 3;
if (res != ref) {
printf ("a=%08x b=%08x c=%08x res=%08x ref=%08x\n",
a, b, c, res, ref);
return EXIT_FAILURE;
}
c++;
} while (c);
b++;
} while (b);
a++;
} while (a);
return EXIT_SUCCESS;
}
#else // BENCHMARK
#include <math.h>
// A routine to give access to a high precision timer on most systems.
#if defined(_WIN32)
#if !defined(WIN32_LEAN_AND_MEAN)
#define WIN32_LEAN_AND_MEAN
#endif
#include <windows.h>
double second (void)
{
LARGE_INTEGER t;
static double oofreq;
static int checkedForHighResTimer;
static BOOL hasHighResTimer;
if (!checkedForHighResTimer) {
hasHighResTimer = QueryPerformanceFrequency (&t);
oofreq = 1.0 / (double)t.QuadPart;
checkedForHighResTimer = 1;
}
if (hasHighResTimer) {
QueryPerformanceCounter (&t);
return (double)t.QuadPart * oofreq;
} else {
return (double)GetTickCount() * 1.0e-3;
}
}
#elif defined(__linux__) || defined(__APPLE__)
#include <stddef.h>
#include <sys/time.h>
double second (void)
{
struct timeval tv;
gettimeofday(&tv, NULL);
return (double)tv.tv_sec + (double)tv.tv_usec * 1.0e-6;
}
#else
#error unsupported platform
#endif
#define N (3000000)
int main (void)
{
double start, stop, elapsed = INFINITY;
int i, k;
T a, b;
T avg0 = 0xffffffff, avg1 = 0xfffffffe;
T avg2 = 0xfffffffd, avg3 = 0xfffffffc;
T avg4 = 0xfffffffb, avg5 = 0xfffffffa;
T avg6 = 0xfffffff9, avg7 = 0xfffffff8;
T avg8 = 0xfffffff7, avg9 = 0xfffffff6;
T avg10 = 0xfffffff5, avg11 = 0xfffffff4;
T avg12 = 0xfffffff2, avg13 = 0xfffffff2;
T avg14 = 0xfffffff1, avg15 = 0xfffffff0;
a = 0x31415926;
b = 0x27182818;
avg0 = average_of_3 (a, b, avg0);
for (k = 0; k < 5; k++) {
start = second();
for (i = 0; i < N; i++) {
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
avg0 = average_of_3 (a, b, avg0);
b = (b + avg0) ^ a;
a = (a ^ b) + avg0;
}
stop = second();
elapsed = fmin (stop - start, elapsed);
}
printf ("a=%016llx b=%016llx avg=%016llx",
(uint64_t)a, (uint64_t)b, (uint64_t)avg0);
printf ("\rlatency: each average_of_3() took %.6e seconds\n",
elapsed / 16 / N);
a = 0x31415926;
b = 0x27182818;
avg0 = average_of_3 (a, b, avg0);
for (k = 0; k < 5; k++) {
start = second();
for (i = 0; i < N; i++) {
avg0 = average_of_3 (a, b, avg0);
avg1 = average_of_3 (a, b, avg1);
avg2 = average_of_3 (a, b, avg2);
avg3 = average_of_3 (a, b, avg3);
avg4 = average_of_3 (a, b, avg4);
avg5 = average_of_3 (a, b, avg5);
avg6 = average_of_3 (a, b, avg6);
avg7 = average_of_3 (a, b, avg7);
avg8 = average_of_3 (a, b, avg8);
avg9 = average_of_3 (a, b, avg9);
avg10 = average_of_3 (a, b, avg10);
avg11 = average_of_3 (a, b, avg11);
avg12 = average_of_3 (a, b, avg12);
avg13 = average_of_3 (a, b, avg13);
avg14 = average_of_3 (a, b, avg14);
avg15 = average_of_3 (a, b, avg15);
b = (b + avg0) ^ a;
a = (a ^ b) + avg0;
}
stop = second();
elapsed = fmin (stop - start, elapsed);
}
printf ("a=%016llx b=%016llx avg=%016llx", (uint64_t)a, (uint64_t)b,
(uint64_t)(avg0 + avg1 + avg2 + avg3 + avg4 + avg5 + avg6 + avg7 +
avg8 + avg9 +avg10 +avg11 +avg12 +avg13 +avg14 +avg15));
printf ("\rthroughput: each average_of_3() took %.6e seconds\n",
elapsed / 16 / N);
return EXIT_SUCCESS;
}
#endif // BENCHMARK
Let me throw my hat in the ring. Not doing anything too tricky here, I think.
#include <stdint.h>
uint64_t average_of_three(uint64_t a, uint64_t b, uint64_t c) {
uint64_t hi = (a >> 32) + (b >> 32) + (c >> 32);
uint64_t lo = hi + (a & 0xffffffff) + (b & 0xffffffff) + (c & 0xffffffff);
return 0x55555555 * hi + lo / 3;
}
Following discussion below about different splits, here's a version that saves a multiply at the expense of three bitwise-ANDs:
T hi = (a >> 2) + (b >> 2) + (c >> 2);
T lo = (a & 3) + (b & 3) + (c & 3);
avg = hi + (hi + lo) / 3;