a few places in the gemm scheduler code were missing barriers;
the code likely worked OK due to heavy use of volatile / _Atomic
but there's no reason to get this incorrect
The use of _Atomic leads to really bad code generation in the compiler
(on x86, you get 2 "mfence" memory barriers around each access with gcc8, despite
x86 being ordered and cache coherent). But there's a fallback in the code that
just uses volatile which is more than plenty in practice.
If we're nervous about cross thread synchronization for these variables, we should
make the YIELD function be a compiler/memory barrier instead.
performance before (after last commit)
Matrix SGEMM cycles MPC DGEMM cycles MPC
48 x 48 10630.0 10.6 0.7% 18112.8 6.2 -0.7%
64 x 64 20374.8 13.0 1.9% 40487.0 6.5 0.4%
65 x 65 141955.2 1.9 -428.3% 146708.8 1.9 -179.2%
80 x 80 178921.1 2.9 -369.6% 186032.7 2.8 -156.6%
96 x 96 205436.2 4.3 -233.4% 224513.1 3.9 -97.0%
112 x 112 244408.2 5.8 -162.7% 262158.7 5.4 -47.1%
128 x 128 321334.5 6.5 -141.3% 333829.0 6.3 -29.2%
Performance with this patch (roughly a 2x improvement):
Matrix SGEMM cycles MPC DGEMM cycles MPC
48 x 48 10756.0 10.5 -0.5% 18296.7 6.1 -1.7%
64 x 64 20490.0 12.9 1.4% 40615.0 6.5 0.0%
65 x 65 83528.3 3.3 -210.9% 96319.0 2.9 -83.3%
80 x 80 101453.5 5.1 -166.3% 128021.7 4.0 -76.6%
96 x 96 149795.1 5.9 -143.1% 168059.4 5.3 -47.4%
112 x 112 191481.2 7.3 -105.8% 204165.0 6.9 -14.6%
128 x 128 265019.2 7.9 -99.0% 272006.4 7.7 -5.3%
The jobs array is getting initialized in O(compiled cpus^2) complexity.
Distros and people with bigger systems will use pretty high values
(128 or 256 or more) for this value, leading to interesting bubbles
in performance.
Baseline (single threaded performance) gets roughly 13 - 15 multiplications per cycle
in the interesting range (threading kicks in at 65x65 mult by 65x65).
The hardware is capable of 32 multiplications per cycle theoretically.
Matrix SGEMM cycles MPC DGEMM cycles MPC
48 x 48 10703.9 10.6 0.0% 17990.6 6.3 0.0%
64 x 64 20778.4 12.8 0.0% 40629.2 6.5 0.0%
65 x 65 26869.9 10.3 0.0% 52545.7 5.3 0.0%
80 x 80 38104.5 13.5 0.0% 72492.7 7.1 0.0%
96 x 96 61626.4 14.4 0.0% 113983.8 7.8 0.0%
112 x 112 91803.8 15.3 0.0% 180987.3 7.8 0.0%
128 x 128 133161.4 15.8 0.0% 258374.3 8.1 0.0%
When threading is turned on
TARGET=SKYLAKEX F_COMPILER=GFORTRAN SHARED=1 DYNAMIC_THREADS=1 USE_OPENMP=0 NUM_THREADS=128
Matrix SGEMM cycles MPC DGEMM cycles MPC
48 x 48 10725.9 10.5 -0.2% 18134.9 6.2 -0.8%
64 x 64 20500.6 12.9 1.3% 40929.1 6.5 -0.7%
65 x 65 2040832.1 0.1 -7495.2% 2097633.6 0.1 -3892.0%
80 x 80 2063129.1 0.2 -5314.4% 2119925.2 0.2 -2824.3%
96 x 96 2070374.5 0.4 -3259.6% 2173604.4 0.4 -1806.9%
112 x 112 2111721.5 0.7 -2169.6% 2263330.8 0.6 -1170.0%
128 x 128 2276181.5 0.9 -1609.3% 2377228.9 0.9 -820.1%
There is a deep deep cliff once you hit 65x65
With this patch
Matrix SGEMM cycles MPC DGEMM cycles MPC
48 x 48 10630.0 10.6 0.7% 18112.8 6.2 -0.7%
64 x 64 20374.8 13.0 1.9% 40487.0 6.5 0.4%
65 x 65 141955.2 1.9 -428.3% 146708.8 1.9 -179.2%
80 x 80 178921.1 2.9 -369.6% 186032.7 2.8 -156.6%
96 x 96 205436.2 4.3 -233.4% 224513.1 3.9 -97.0%
112 x 112 244408.2 5.8 -162.7% 262158.7 5.4 -47.1%
128 x 128 321334.5 6.5 -141.3% 333829.0 6.3 -29.2%
The cliff is very significantly reduced.
(more to follow)
When NUM_THREADS(MAX_CPU_NUNBERS) is very large ,e.g. 256.
typedef struct {
volatile BLASLONG working[MAX_CPU_NUMBER][CACHE_LINE_SIZE * DIVIDE_RATE];
} job_t;
job_t job[MAX_CPU_NUMBER];
The job array is equal 8MB.
Thus, We use malloc instead of stack allocation.
Martin Kroeker
Craig Donner
Arjan van de Ven
Zhiyong Dang
Andrew
Tim Moon
Werner Saar
Ralph Campbell
Timothy Gu
Zhang Xianyi