Repository navigation
Corrections to Chapter 1, The Mathematics of Parallelism.
1.7 Flynn's taxonomy - the four classes are now stated as SISD, SIMD, MISD and MIMD. MISD is noted as rare in practice. SIMT is presented as NVIDIA's extension of SIMD, not one of Flynn's classes. The summary bullet and exercise 9 are fixed.
1.4 Gustafson-Barsis - Amdahl's f (serial share of single-unit time) is now separated from Gustafson's s (serial share of the parallel run). In the scaled example f = 0.01/990.01 = 0.00001, which is why the two laws agree. A figure placeholder marks the Amdahl-squeeze / Gustafson-stretch pair beside Figure 1.2.
1.9 Cost of synchronisation - the 9% barrier share is described as lost efficiency per unit (1000/1100 = 0.91), not an Amdahl ceiling. The fixed-work case, where the barrier's share rises and the speedup flattens, is stated separately.
1.8.3 Doubling table - the serial-bound row now reads "double the number of units".
1.8.2 Animation 4 - the 16x16 tile raises the loop's bandwidth ceiling sixteenfold and the loop stays memory-bound at 4 FLOP/byte.
1.2 Speedup and efficiency - adds a note on superlinear speedup and the equal-work assumption behind S(p) <= p.
Figures - the roofline is now captioned Figure 1.6, which also fixes a duplicated zoom image, and the CPU/memory/I/O-bound figure is renumbered to Figure 1.7.