PTX fma and other flags (#585)

* PTX fma and other flags

* we can keep cuda debug info now, that's cool

* version bump

* Update XLA.jl

* Update src/XLA.jl

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

---------

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

Author: wsmoses

Project.toml

@@ -1,7 +1,7 @@
 name = "Reactant"
 uuid = "3c362404-f566-11ee-1572-e11a4b42c853"
 authors = ["William Moses <[email protected]>", "Valentin Churavy <[email protected]>", "Sergio Sánchez Ramírez <[email protected]>", "Paul Berg <[email protected]>", "Avik Pal <[email protected]>"]
-version = "0.2.20"
+version = "0.2.21"
 
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
@@ -70,7 +70,7 @@ PythonCall = "0.9"
 Random = "1.10"
 Random123 = "1.7"
 ReactantCore = "0.1.4"
-Reactant_jll = "0.0.45"
+Reactant_jll = "0.0.46"
 Scratch = "1.2"
 Sockets = "1.10"
 SpecialFunctions = "2.4"

deps/ReactantExtra/API.cpp

@@ -446,6 +446,12 @@ std::vector<int64_t> col_major(int64_t dim) {
   return minor_to_major;
 }
 
+extern "C" void ReactantLLVMParseCommandLineOptions(int argc, const char *const *argv,
+                                 const char *Overview) {
+    llvm::cl::ParseCommandLineOptions(argc, argv, StringRef(Overview),
+                                    &llvm::nulls());
+}
+
 std::vector<int64_t> row_major(int64_t dim) {
   std::vector<int64_t> minor_to_major;
   for (int i = 0; i < dim; i++) {

deps/ReactantExtra/BUILD

@@ -450,7 +450,8 @@ cc_library(
 "-Wl,-exported_symbol,_ProfilerActivityStart",
 "-Wl,-exported_symbol,_ProfilerActivityEnd",
 "-Wl,-exported_symbol,_ReactantFuncSetArgAttr",
-"-Wl,-exported_symbol,_ReactantCudaDriverGetVersion"
+"-Wl,-exported_symbol,_ReactantCudaDriverGetVersion",
+"-Wl,-exported_symbol,_ReactantLLVMParseCommandLineOptions"
     ]}),
     deps = [
                 "@enzyme//:EnzymeMLIR",

deps/ReactantExtra/WORKSPACE

@@ -136,7 +136,7 @@ http_archive(
 )
 
 # load("@jax//third_party/xla:workspace.bzl", "XLA_COMMIT", "XLA_SHA256")
-XLA_COMMIT = "281c11225c4a0bb7b710a290610a06d71194febd"
+XLA_COMMIT = "e0c92850a41cf5208744d8a919b969fa3506863c"
 XLA_SHA256 = ""
 
 http_archive(

ext/ReactantCUDAExt.jl

@@ -400,11 +400,10 @@ function compile(job)
         if !isempty(errors)
             throw(GPUCompiler.InvalidIRError(job, errors))
         end
-        LLVM.strip_debuginfo!(mod)
+        # LLVM.strip_debuginfo!(mod)
         modstr = string(mod)
         # This is a bit weird since we're taking a module from julia's llvm into reactant's llvm version
         # it is probably safer to reparse a string using the right llvm module api, so we will do that.
-
         mmod = MLIR.IR.Module(
             @ccall MLIR.API.mlir_c.ConvertLLVMStrToMLIR(
                 modstr::Cstring, MLIR.IR.context()::MLIR.API.MlirContext

src/XLA.jl

@@ -2,6 +2,13 @@ module XLA
 
 import ...MLIR
 
+function LLVMclopts(opts...)
+    args = ["", opts...]
+    @ccall MLIR.API.mlir_c.ReactantLLVMParseCommandLineOptions(
+        length(args)::Cint, args::Ptr{Cstring}, C_NULL::Ptr{Cvoid}
+    )::Cvoid
+end
+
 mutable struct Client
     client::Ptr{Cvoid}
 
@@ -50,6 +57,7 @@ function CPUClient(asynchronous=false, node_id=0, num_nodes=1; checkcount=true)
     end
     f = Libdl.dlsym(Reactant_jll.libReactantExtra_handle, "MakeCPUClient")
     client = ccall(f, Ptr{Cvoid}, (UInt, Cint, Cint), asynchronous, node_id, num_nodes)
+    LLVMclopts("-nvptx-fma-level=1")
     #client = @ccall MLIR.API.mlir_c.MakeCPUClient(asynchronous::UInt8, node_id::Cint, num_nodes::Cint)::Ptr{Cvoid}
     return Client(client)
 end
@@ -73,6 +81,7 @@ function GPUClient(node_id=0, num_nodes=1, platform="gpu")
     if client == C_NULL
         throw(AssertionError(unsafe_string(refstr[])))
     end
+    LLVMclopts("-nvptx-fma-level=1")
     return Client(client)
 end
 
@@ -83,6 +92,7 @@ function TPUClient(tpu_path::String)
     if client == C_NULL
         throw(AssertionError(unsafe_string(refstr[])))
     end
+    LLVMclopts("-nvptx-fma-level=1")
     return Client(client)
 end