Reduce CI noise (#150)

* erand48() does not exist on windows (so much for <stdlib.h> ...)

* Deactivate fuzzing big sizes of matmul - pending #149

* Add setup and bench on GCC 8 as GCC10 OpenMP is broken

* size 2 also tends to stall?

* C++ compat shim

* For now skip the full GEMM "fuzz-like" tests.

benchmarks/matmul_gemm_blas/test_gemm_output.nim

@@ -79,7 +79,7 @@ proc testVsReference*(M, N, K: int) =
 when isMainModule:
   randomize(42) # For reproducibility
 
-  const sizes = [2,3,9,37,129,700]
+  const sizes = [2,3,9,37,129,700] # TODO: random syncScope stalls in CI https://github.com/mratsim/weave/issues/149
 
   init(Weave)
   for M in sizes:

demos/raytracing/README.md

@@ -27,30 +27,71 @@ compared to the single-threaded or the single parallel-for versions.
 Note: except for the nested parallelism which has RNG issue,
 the Nim and C++ versions are pixel equivalent.
 
+### Setup
+
 CPU: i9-9980XE, 18 cores, overclocked at 4.1GHz all-core turbo (from 3.0 nominal)
 The code was compiled with default flag, hence x86-64, hence SSE2.
 
 - Nim devel (1.3.5 2020-05-16) + GCC v10.1.0
   - `nim c --threads:off -d:danger`
   - `nim c --threads:on -d:danger`
 - GCC v10.1.0
-  - `-03`
+  - `-O3`
+  - `-O3 -fopenmp`
+- GCC v8.4.0
+  - `-O3`
   - `-O3 -fopenmp`
 - Clang v10.0.0
-  - `-03`
+  - `-O3`
   - `-O3 -fopenmp`
 
-| Bench            |         Nim | Clang C++ |     GCC C++ |
-|------------------|------------:|----------:|------------:|
-| Single-threaded  | 4min43.369s | 4m51.052s | 4min50.934s |
-| Multithreaded    |     13.211s |   14.428s | 2min14.616s |
-| Nested-parallel  |     12.981s |           |             |
-| Parallel speedup |      21.83x |    20.17x |       2.16x |
+### Commands
+
+
+```bash
+git clone https://github.com/mratsim/weave
+cd weave
+nimble install -y # install Weave dependencies, here synthesis, overwriting if asked.
+
+nim -v # Ensure you have nim 1.2.0 or more recent
+
+# Threads on (by default in this repo)
+nim c -d:danger -o:build/ray_threaded demos/raytracing/smallpt.nim
+
+# Threads off
+nim c -d:danger --threads:off -o:build/ray_single demos/raytracing/smallpt.nim
+
+g++ -O3 -o build/ray_gcc_single demos/raytracing/smallpt.cpp
+g++ -O3 -fopenmp -o build/ray_gcc_omp demos/raytracing/smallpt.cpp
+
+clang++ -O3 -o build/ray_clang_single demos/raytracing/smallpt.cpp
+clang++ -O3 -fopenmp -o build/ray_clang_single demos/raytracing/smallpt.cpp
+```
+
+### Results & Analysis
+
+GCC 10 has a significant OpenMP regression
+
+|      Bench       |     Nim     | Clang C++ OpenMP | GCC 10 C++ OpenMP | GCC 8 C++ OpenMP |
+| ---------------- | ----------: | ---------------: | ----------------: | ---------------: |
+| Single-threaded  | 4min43.369s |        4m51.052s |       4min50.934s |        4m50.648s |
+| Multithreaded    |     12.977s |          14.428s |       2min14.616s |          12.244s |
+| Nested-parallel  |     12.981s |                  |                   |                  |
+| Parallel speedup |      21.83x |           20.17x |             2.16x |           23.74x |
+
+Single-threaded Nim is 2.7% faster than Clang C++.
+Multithreaded Nim via Weave is 11.1% faster Clang C++.
+
+GCC 8 despite a simpler OpenMP design (usage of a global task queue instead of work-stealing)
+achieves a better speedup than both Weave and Clang.
+In that case, I expect it's because the tasks are so big that there is minimal contention
+on the task queue, furthermore the OpenMP schedule is "Dynamic" so we avoid the worst case scenario
+with static scheduling where a bunch of threads are assigned easy rays that never collide with a surface
+and a couple of threads are drowned in complex rays.
 
-Single-threaded Nim is 2.7% faster than Clang C++
-Multithreaded Nim via Weave is 11.1% faster Clang C++
+I have absolutely no idea of what happened to OpenMP in GCC 10.
 
-Note: I only have 18 cores but we observe over 18x speedup
+Note: I only have 18 cores but we observe speedups in the 20x
 with Weave and LLVM. This is probably due to 2 factors:
 - Raytracing is pure compute, in particular contrary to high-performance computing
   and machine learning workloads which are also very memory-intensive (matrices and tensors with thousands to millions of elements)
@@ -65,7 +106,7 @@ with Weave and LLVM. This is probably due to 2 factors:
 
 ## License
 
-Kevin beason code is licensed under (mail redacted to avoid spam)
+Kevin Beason code is licensed under (mail redacted to avoid spam)
 
 ```
 LICENSE

demos/raytracing/smallpt.nim

@@ -101,7 +101,15 @@ func intersect(r: Ray, t: var float64, id: var int32): bool =
       id = i.int32
   return t < inf
 
-proc erand48(xi: var array[3, cushort]): cdouble {.importc, header:"<stdlib.h>", sideeffect.} # Need the same RNG for comparison
+when defined(cpp):
+  # Seems like Nim codegen for mutable arrays is slightly different from the C++ API
+  # and needs a compatibility shim
+  proc erand48(xi: ptr cushort): cdouble {.importc, header:"<stdlib.h>", sideeffect.}
+  proc erand48(xi: var array[3, cushort]): float64 {.inline.} =
+    erand48(xi[0].addr)
+else:
+  # Need the same RNG for comparison
+  proc erand48(xi: var array[3, cushort]): cdouble {.importc, header:"<stdlib.h>", sideeffect.}
 
 proc radiance(r: Ray, depth: int32, xi: var array[3, cushort]): Vec =
   var t: float64                      # distance to intersection

weave.nimble

@@ -55,7 +55,8 @@ task test, "Run Weave tests":
 
   test "-d:WV_LazyFlowvar", "tests/test_background_jobs.nim"
 
-  test "", "demos/raytracing/smallpt.nim"
+  when not defined(windows): # Does not support erand48
+    test "", "demos/raytracing/smallpt.nim"
 
   test "", "benchmarks/dfs/weave_dfs.nim"
   test "", "benchmarks/fibonacci/weave_fib.nim"
@@ -90,9 +91,10 @@ task test, "Run Weave tests":
   # - spawn
   # - spawnDelayed by pledges
   # - syncScope
-  when not defined(windows) and (defined(i386) or defined(amd64)):
-    if not existsEnv"TEST_LANG" or getEnv"TEST_LANG" != "cpp":
-      test "-d:danger", "benchmarks/matmul_gemm_blas/test_gemm_output.nim"
+  when false: # TODO, not sure why this stalls why the gemm_weave_nestable don't - https://github.com/mratsim/weave/pull/150
+    when not defined(windows) and (defined(i386) or defined(amd64)):
+      if not existsEnv"TEST_LANG" or getEnv"TEST_LANG" != "cpp":
+        test "-d:danger", "benchmarks/matmul_gemm_blas/test_gemm_output.nim"
 
 task test_gc_arc, "Run Weave tests with --gc:arc":
   test "--gc:arc", "weave/cross_thread_com/channels_spsc_single.nim"
@@ -123,7 +125,8 @@ task test_gc_arc, "Run Weave tests with --gc:arc":
 
   test "--gc:arc -d:WV_LazyFlowvar", "tests/test_background_jobs.nim"
 
-  test "--gc:arc", "demos/raytracing/smallpt.nim"
+  when not defined(windows):
+    test "--gc:arc", "demos/raytracing/smallpt.nim"
 
   test "--gc:arc", "benchmarks/dfs/weave_dfs.nim"
   test "--gc:arc", "benchmarks/fibonacci/weave_fib.nim"
@@ -158,9 +161,10 @@ task test_gc_arc, "Run Weave tests with --gc:arc":
   # - spawn
   # - spawnDelayed by pledges
   # - syncScope
-  when not defined(windows) and (defined(i386) or defined(amd64)):
-    if not existsEnv"TEST_LANG" or getEnv"TEST_LANG" != "cpp":
-      test "-d:danger", "benchmarks/matmul_gemm_blas/test_gemm_output.nim"
+  when false: # TODO, not sure why this stalls why the gemm_weave_nestable don't - https://github.com/mratsim/weave/pull/150
+    when not defined(windows) and (defined(i386) or defined(amd64)):
+      if not existsEnv"TEST_LANG" or getEnv"TEST_LANG" != "cpp":
+        test "-d:danger", "benchmarks/matmul_gemm_blas/test_gemm_output.nim"
 
 task gen_book, "Generate Weave documentation":
   exec "mdbook build docs"