fix: get_tuple NIF helper + compilation configs and speedup (#1479)

exla/Makefile

@@ -73,7 +73,7 @@ ifeq ($(NVCC_TEST),nvcc)
   NVCC := nvcc
 	NVCCFLAGS += -DCUDA_ENABLED
 else
-  NVCC := g++
+  NVCC := $(CXX)
 	NVCCFLAGS = $(CFLAGS)
 endif
 

exla/README.md

@@ -59,6 +59,8 @@ mix deps.get
 mix test
 ```
 
+By default, EXLA passes `["-jN"]` as a Make argument, where `N` is `System.schedulers_online() - 2`, capped at `1`. `config :exla, :make_args, ...` can be used to override this default setting.
+
 In order to run tests on a specific device, use the `EXLA_TARGET` environment variable, which is a dev-only variable for this project (it has no effect when using EXLA as a dependency). For example, `EXLA_TARGET=cuda` or `EXLA_TARGET=rocm`. Make sure to also specify `XLA_TARGET` to fetch or compile a proper version of the XLA binary.
 
 ### Building with Docker

exla/c_src/exla/exla_nif_util.cc

@@ -1,10 +1,10 @@
 #include "exla_nif_util.h"
 
+#include "mlir/IR/Builders.h"
 #include "mlir/IR/BuiltinTypes.h"
+#include "stablehlo/dialect/StablehloOps.h"
 #include "xla/primitive_util.h"
 #include "xla/shape_util.h"
-#include "mlir/IR/Builders.h"
-#include "stablehlo/dialect/StablehloOps.h"
 
 namespace exla {
 namespace nif {
@@ -190,7 +190,7 @@ int get_tuple(ErlNifEnv* env, ERL_NIF_TERM tuple, std::vector<int64>& var) {
   var.reserve(length);
 
   for (int i = 0; i < length; i++) {
-    int data;
+    int64 data;
     if (!get(env, terms[i], &data)) return 0;
     var.push_back(data);
   }

exla/mix.exs

@@ -5,6 +5,9 @@ defmodule EXLA.MixProject do
   @version "0.7.1"
 
   def project do
+    make_args =
+      Application.get_env(:exla, :make_args) || ["-j#{max(System.schedulers_online() - 2, 1)}"]
+
     [
       app: :exla,
       version: @version,
@@ -34,7 +37,8 @@ defmodule EXLA.MixProject do
           "MIX_BUILD_EMBEDDED" => "#{Mix.Project.config()[:build_embedded]}",
           "CWD_RELATIVE_TO_PRIV_PATH" => cwd_relative_to_priv
         }
-      end
+      end,
+      make_args: make_args
     ]
   end