Merge branch 'demo131' into Issue/862

xmake/iluvatar.lua

@@ -42,10 +42,14 @@ target("infiniop-iluvatar")
     add_links("cudart", "cublas", "cudnn")
 
     set_warnings("all", "error")
-    add_cuflags("-Wno-error=unused-private-field")
+    add_cuflags("-Wno-error=unused-private-field", "-Wno-error=unused-variable", "-Wno-unused-variable")
     add_cuflags("-fPIC", "-x", "ivcore", "-std=c++17", {force = true})
+    if has_config("ivcore-20") then
+        add_cuflags("--cuda-gpu-arch=ivcore20", {force = true})
+    end
     add_culdflags("-fPIC")
-    add_cxflags("-fPIC")
+    add_cxflags("-fPIC", "-Wno-error=unused-variable", "-Wno-unused-variable")
+    add_cxxflags("-fPIC", "-Wno-error=unused-variable", "-Wno-unused-variable")
 
     -- set_languages("cxx17") 天数似乎不能用这个配置
     add_files("../src/infiniop/devices/nvidia/*.cu", "../src/infiniop/ops/*/nvidia/*.cu")
@@ -54,7 +58,7 @@ target("infiniop-iluvatar")
     add_files("../src/infiniop/ops/dequantize_awq/iluvatar/*.cu")
 
     if has_config("ninetoothed") then
-        add_files("../build/ninetoothed/*.c", {cxflags = {"-Wno-return-type"}})
+        add_files("../build/ninetoothed/*.c", "../build/ninetoothed/*.cpp", {cxxflags = {"-Wno-return-type"}})
     end
 target_end()
 
@@ -73,6 +77,7 @@ target("infinirt-iluvatar")
     add_cuflags("-fPIC", "-x", "ivcore", "-std=c++17", {force = true})
     add_culdflags("-fPIC")
     add_cxflags("-fPIC")
+    add_cxxflags("-fPIC")
 
     -- set_languages("cxx17") 天数似乎不能用这个配置
     add_files("../src/infinirt/cuda/*.cu")
@@ -94,6 +99,7 @@ target("infiniccl-iluvatar")
         add_cuflags("-fPIC", "-x", "ivcore", "-std=c++17", {force = true})
         add_culdflags("-fPIC")
         add_cxflags("-fPIC")
+        add_cxxflags("-fPIC")
 
         local nccl_root = os.getenv("NCCL_ROOT")
         if nccl_root then

xmake/kunlun.lua

@@ -75,6 +75,7 @@ target("infiniop-kunlun")
     on_install(function (target) end)
 
     add_cxflags("-lstdc++ -fPIC -Wno-error=unused-function")
+    add_cxxflags("-lstdc++ -fPIC -Wno-error=unused-function")
     set_warnings("all", "error")
 
     set_languages("cxx17")
@@ -102,6 +103,7 @@ target("infinirt-kunlun")
     -- Add include dirs
     add_files("$(projectdir)/src/infinirt/kunlun/*.cc")
     add_cxflags("-lstdc++ -Wall -Werror -fPIC")
+    add_cxxflags("-lstdc++ -Wall -Werror -fPIC")
 target_end()
 
 target("infiniccl-kunlun")
@@ -117,5 +119,6 @@ target("infiniccl-kunlun")
         add_links("bkcl")
         add_files("$(projectdir)/src/infiniccl/kunlun/*.cc")
         add_cxflags("-lstdc++ -fPIC")
+        add_cxxflags("-lstdc++ -fPIC")
     end
 target_end()

xmake/metax.lua

@@ -48,11 +48,20 @@ target("infiniop-metax")
     set_languages("cxx17")
     set_warnings("all", "error")
     add_cxflags("-lstdc++", "-fPIC", "-Wno-defaulted-function-deleted", "-Wno-strict-aliasing", {force = true})
+    add_cxxflags("-lstdc++", "-fPIC", "-Wno-defaulted-function-deleted", "-Wno-strict-aliasing", {force = true})
     add_files("../src/infiniop/devices/metax/*.cc", "../src/infiniop/ops/*/metax/*.cc")
     add_files("../src/infiniop/ops/*/metax/*.maca", {rule = "maca"})
 
     if has_config("ninetoothed") then
-        add_files("../build/ninetoothed/*.c", {cxflags = {"-include stdlib.h", "-Wno-return-type"}})
+        add_includedirs(MACA_ROOT .. "/include/mcr")
+        add_files("../build/ninetoothed/*.c", "../build/ninetoothed/*.cpp", {
+            cxflags = {
+                "-include stdlib.h", 
+                "-Wno-return-type", 
+                "-Wno-implicit-function-declaration",
+                "-Wno-builtin-declaration-mismatch"
+            }
+        })
     end
 target_end()
 
@@ -63,6 +72,7 @@ target("infinirt-metax")
     add_deps("infini-utils")
     set_warnings("all", "error")
     add_cxflags("-lstdc++ -fPIC")
+    add_cxxflags("-lstdc++ -fPIC")
     add_files("../src/infinirt/metax/*.cc")
 target_end()
 
@@ -73,6 +83,7 @@ target("infiniccl-metax")
     set_warnings("all", "error")
     if not is_plat("windows") then
         add_cxflags("-fPIC")
+        add_cxxflags("-fPIC")
     end
     if has_config("ccl") then
         if has_config("use-mc") then

xmake/moore.lua

@@ -42,6 +42,7 @@ target("infiniop-moore")
     set_languages("cxx17")
     set_warnings("all", "error")
     add_cxflags("-lstdc++", "-fPIC", "-Wno-comment")
+    add_cxxflags("-lstdc++", "-fPIC", "-Wno-comment")
     add_files("../src/infiniop/devices/moore/*.cc")
     add_files("../src/infiniop/ops/*/moore/*.mu", {rule = "mu"})
 
@@ -56,6 +57,7 @@ target("infinirt-moore")
     add_deps("infini-utils")
     set_warnings("all", "error")
     add_cxflags("-lstdc++", "-fPIC")
+    add_cxxflags("-lstdc++", "-fPIC")
     add_files("../src/infinirt/moore/*.cc")
 target_end()
 
@@ -66,6 +68,7 @@ target("infiniccl-moore")
     set_warnings("all", "error")
     if not is_plat("windows") then
         add_cxflags("-fPIC")
+        add_cxxflags("-fPIC")
     end
     if has_config("ccl") then
         add_links("libmccl.so")

xmake/nvidia.lua

@@ -4,11 +4,9 @@ if CUDNN_ROOT ~= nil then
 end
 
 local CUTLASS_ROOT = os.getenv("CUTLASS_ROOT") or os.getenv("CUTLASS_HOME") or os.getenv("CUTLASS_PATH")
-local CUTE_ROOT = os.getenv("CUTE_ROOT") or os.getenv("CUTE_HOME") or os.getenv("CUTE_PATH")
 
 if CUTLASS_ROOT ~= nil then
     add_includedirs(CUTLASS_ROOT)
-    add_includedirs(CUTE_ROOT)
 end
 
 target("infiniop-nvidia")
@@ -22,7 +20,6 @@ target("infiniop-nvidia")
     if has_config("cudnn") then
         add_links("cudnn")
     end
-    add_cugencodes("native")
 
     on_load(function (target)
         import("lib.detect.find_tool")
@@ -36,11 +33,6 @@ target("infiniop-nvidia")
 
             target:add("linkdirs", path.directory(path.directory(nvcc_path)) .. "/lib64/stubs")
             target:add("links", "cuda")
-
-            local cuda_arch = get_config("cuda_arch")
-            if cuda_arch ~= nil then
-                target:add("cu-cxxflags", "-arch=", cuda_arch)
-            end
         end
     end)
 
@@ -56,20 +48,33 @@ target("infiniop-nvidia")
         add_cuflags("-Xcompiler=-fPIC")
         add_cuflags("--extended-lambda")
         add_culdflags("-Xcompiler=-fPIC")
+        add_cxflags("-fPIC")
         add_cxxflags("-fPIC")
+        add_cflags("-fPIC")
         add_cuflags("--expt-relaxed-constexpr")
         if CUDNN_ROOT ~= nil then
             add_linkdirs(CUDNN_ROOT .. "/lib")
         end
     end
 
-    add_cuflags("-Xcompiler=-Wno-error=deprecated-declarations")
+    add_cuflags("-Xcompiler=-Wno-error=deprecated-declarations", "-Xcompiler=-Wno-error=unused-function")
+
+    local arch_opt = get_config("cuda_arch")
+    if arch_opt and type(arch_opt) == "string" then
+        for _, arch in ipairs(arch_opt:split(",")) do
+            arch = arch:trim()
+            local compute = arch:gsub("sm_", "compute_")
+            add_cuflags("-gencode=arch=" .. compute .. ",code=" .. arch)
+        end
+    else
+        add_cugencodes("native")
+    end
 
     set_languages("cxx17")
-    add_files("../src/infiniop/devices/nvidia/*.cu", "../src/infiniop/ops/*/nvidia/*.cu")
+    add_files("../src/infiniop/devices/nvidia/*.cu", "../src/infiniop/ops/*/nvidia/*.cu", "../src/infiniop/ops/*/*/nvidia/*.cu")
 
     if has_config("ninetoothed") then
-        add_files("../build/ninetoothed/*.c")
+        add_files("../build/ninetoothed/*.c", "../build/ninetoothed/*.cpp")
     end
 target_end()
 
@@ -89,6 +94,7 @@ target("infinirt-nvidia")
         add_cuflags("-Xcompiler=-fPIC")
         add_culdflags("-Xcompiler=-fPIC")
         add_cxflags("-fPIC")
+        add_cxxflags("-fPIC")
     end
 
     set_languages("cxx17")
@@ -108,6 +114,7 @@ target("infiniccl-nvidia")
             add_cuflags("-Xcompiler=-fPIC")
             add_culdflags("-Xcompiler=-fPIC")
             add_cxflags("-fPIC")
+            add_cxxflags("-fPIC")
 
             local nccl_root = os.getenv("NCCL_ROOT")
             if nccl_root then

xmake/qy.lua

@@ -88,6 +88,7 @@ target("infiniop-qy")
         add_cuflags("-Xcompiler=-fPIC")
         add_cuflags("--extended-lambda")
         add_culdflags("-Xcompiler=-fPIC")
+        add_cxflags("-fPIC")
         add_cxxflags("-fPIC")
         add_cuflags("--expt-relaxed-constexpr")
         if CUDNN_ROOT ~= nil then
@@ -98,10 +99,10 @@ target("infiniop-qy")
     add_cuflags("-Xcompiler=-Wno-error=deprecated-declarations")
 
     set_languages("cxx17")
-    add_files("../src/infiniop/devices/nvidia/*.cu", "../src/infiniop/ops/*/nvidia/*.cu")
+    add_files("../src/infiniop/devices/nvidia/*.cu", "../src/infiniop/ops/*/nvidia/*.cu", "../src/infiniop/ops/*/*/nvidia/*.cu")
 
     if has_config("ninetoothed") then
-        add_files("../build/ninetoothed/*.c")
+        add_files("../build/ninetoothed/*.c", "../build/ninetoothed/*.cpp")
     end
 target_end()
 
@@ -117,6 +118,7 @@ target("infinirt-qy")
         add_cuflags("-Xcompiler=-fPIC")
         add_culdflags("-Xcompiler=-fPIC")
         add_cxflags("-fPIC")
+        add_cxxflags("-fPIC")
     end
 
     set_languages("cxx17")
@@ -133,6 +135,7 @@ target("infiniccl-qy")
             add_cuflags("-Xcompiler=-fPIC")
             add_culdflags("-Xcompiler=-fPIC")
             add_cxflags("-fPIC")
+            add_cxxflags("-fPIC")
 
             local nccl_root = os.getenv("NCCL_ROOT")
             if nccl_root then

xmake/test.lua

@@ -24,7 +24,7 @@ target("infiniop-test")
     add_links("infiniop", "infinirt")
 
     if has_config("omp") then
-        add_cxflags("-fopenmp")
+        add_cxxflags("-fopenmp")
         add_ldflags("-fopenmp")
     end
 
@@ -88,8 +88,6 @@ target("infinicore-test")
     add_files(os.projectdir().."/src/infinicore/ops/*/*.cc")
     add_files(os.projectdir().."/src/infinicore/nn/*.cc")
 
-    add_files(os.projectdir().."/src/infinicore-test/*.cc")
-    add_files(os.projectdir().."/src/infinicore-test/*/*.cc")
-
+    add_files(os.projectdir().."/src/infinicore-test/**.cc")
     set_installdir(INFINI_ROOT)
 target_end()