[Phi] Unify kernel build targets (#41091)

* unify_kernel_build_target

* fix dnn kernel failed

* fix dnn kernel loss target

* fix xpu compile failed

cmake/phi.cmake

@@ -146,17 +146,12 @@ function(kernel_library TARGET)
             endif()
         endif()
         if (WITH_XPU_KP)
-            if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/kps/${TARGET}.cu)
             # Change XPU2 file suffix
             # NOTE(chenweihang): If we can be sure that the *.kps suffix is no longer used, it can be copied directly to *.xpu
             file(COPY ${CMAKE_CURRENT_SOURCE_DIR}/kps/${TARGET}.cu DESTINATION ${CMAKE_CURRENT_BINARY_DIR}/kps)
             file(RENAME ${CMAKE_CURRENT_BINARY_DIR}/kps/${TARGET}.cu ${CMAKE_CURRENT_BINARY_DIR}/kps/${TARGET}.kps)
             list(APPEND kps_srcs ${CMAKE_CURRENT_BINARY_DIR}/kps/${TARGET}.kps)
         endif()
-            if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/cpu/${TARGET}.cc )
-                list(APPEND kps_srcs ${CMAKE_CURRENT_SOURCE_DIR}/cpu/${TARGET}.cc)
-            endif()
-        endif()
     else()
         # TODO(chenweihang): impl compile by source later
     endif()
@@ -219,71 +214,76 @@ function(kernel_library TARGET)
     list(LENGTH kps_srcs kps_srcs_len)
 
     # kernel source file level
-    # level 1: base device kernel
-    # - cpu_srcs / gpu_srcs / xpu_srcs / gpudnn_srcs / kps_srcs
+    # level 1: base device kernel (if any device or dnn kernel exists, the cpu_kernel must be exists!!!)
+    # - cpu_srcs / gpu_srcs / xpu_srcs / kps_srcs
+    # = dnn srcs: gpudnn_srcs
     # level 2: device-independent kernel
     # - common_srcs
-    set(base_device_kernels)
-    set(device_independent_kernel)
 
-    # 1. Base device kernel compile
-    if (${cpu_srcs_len} GREATER 0)
-        cc_library(${TARGET}_cpu${target_suffix} SRCS ${cpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
-        list(APPEND base_device_kernels ${TARGET}_cpu${target_suffix})
-    endif()
-    if (${gpu_srcs_len} GREATER 0)
-        if (WITH_GPU)
-            nv_library(${TARGET}_gpu${target_suffix} SRCS ${gpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
-        elseif (WITH_ROCM)
-            hip_library(${TARGET}_gpu${target_suffix} SRCS ${gpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
-        endif()
-        list(APPEND base_device_kernels ${TARGET}_gpu${target_suffix})
+    set(partial_build_flag 0)
+    set(base_build_flag 0)
+    if (${common_srcs_len} GREATER 0)
+        set(partial_build_flag 1)
     endif()
-    if (${xpu_srcs_len} GREATER 0)
-        cc_library(${TARGET}_xpu${target_suffix} SRCS ${xpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
-        list(APPEND base_device_kernels ${TARGET}_xpu${target_suffix})
+    if (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0 OR ${xpu_srcs_len} GREATER 0 OR ${kps_srcs_len} GREATER 0)
+        set(base_build_flag 1)
     endif()
+
+    # gpudnn or mkldnn needs to be compiled separately
+    set(dnn_kernels)
     if (${gpudnn_srcs_len} GREATER 0)
         if (WITH_GPU)
             nv_library(${TARGET}_gpudnn${target_suffix} SRCS ${gpudnn_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
         elseif (WITH_ROCM)
             hip_library(${TARGET}_gpudnn${target_suffix} SRCS ${gpudnn_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
         endif()
-        list(APPEND base_device_kernels ${TARGET}_gpudnn${target_suffix})
-    endif()
-    if (${kps_srcs_len} GREATER 0)
-        # only when WITH_XPU_KP, the kps_srcs_len can be > 0
-        xpu_library(${TARGET}_kps${target_suffix} SRCS ${kps_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
-        list(APPEND base_device_kernels ${TARGET}_kps${target_suffix})
+        list(APPEND dnn_kernels ${TARGET}_gpudnn${target_suffix})
     endif()
+    list(LENGTH dnn_kernels dnn_kernels_len)
 
-    # 2. Device-independent kernel compile
-    if (${common_srcs_len} GREATER 0)
+    if (${partial_build_flag} EQUAL 0 AND ${base_build_flag} EQUAL 1)
         if (WITH_GPU)
-            nv_library(${TARGET}_common${target_suffix} SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels})
+            if (${dnn_kernels_len} GREATER 0)
+                nv_library(${TARGET}_base${target_suffix} SRCS ${cpu_srcs} ${gpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
+                nv_library(${TARGET}${target_suffix} DEPS ${TARGET}_base${target_suffix} ${dnn_kernels})
+            else()
+                nv_library(${TARGET}${target_suffix} SRCS ${cpu_srcs} ${gpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
+            endif()
         elseif (WITH_ROCM)
-            hip_library(${TARGET}_common${target_suffix} SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels})
+            if (${dnn_kernels_len} GREATER 0)
+                hip_library(${TARGET}_base${target_suffix} SRCS ${cpu_srcs} ${gpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
+                hip_library(${TARGET}${target_suffix} DEPS ${TARGET}_base${target_suffix} ${dnn_kernels})
+            else()
+                hip_library(${TARGET}${target_suffix} SRCS ${cpu_srcs} ${gpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
+            endif()
         elseif (WITH_XPU_KP)
-            xpu_library(${TARGET}_common${target_suffix} SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels})
+            xpu_library(${TARGET}${target_suffix} SRCS ${cpu_srcs} ${kps_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
         else()
-            cc_library(${TARGET}_common${target_suffix} SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels})
+            cc_library(${TARGET}${target_suffix} SRCS ${cpu_srcs} ${xpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
         endif()
-        list(APPEND device_independent_kernel ${TARGET}_common${target_suffix})
+    elseif (${partial_build_flag} EQUAL 1 AND ${base_build_flag} EQUAL 1)
+        if (WITH_GPU)
+            nv_library(${TARGET}_base${target_suffix} SRCS ${cpu_srcs} ${gpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
+            nv_library(${TARGET}${target_suffix} SRCS ${common_srcs} DEPS ${TARGET}_base${target_suffix} ${dnn_kernels})
+        elseif (WITH_ROCM)
+            hip_library(${TARGET}_base${target_suffix} SRCS ${cpu_srcs} ${gpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
+            hip_library(${TARGET}${target_suffix} SRCS ${common_srcs} DEPS ${TARGET}_base${target_suffix} ${dnn_kernels})
+        elseif (WITH_XPU_KP)
+            xpu_library(${TARGET}_base${target_suffix} SRCS ${cpu_srcs} ${kps_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
+            xpu_library(${TARGET}${target_suffix} SRCS ${common_srcs} DEPS ${TARGET}_base${target_suffix})
+        else()
+            cc_library(${TARGET}_base${target_suffix} SRCS ${cpu_srcs} ${xpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
+            cc_library(${TARGET}${target_suffix} SRCS ${common_srcs} DEPS ${TARGET}_base${target_suffix})
         endif()
-
-
-    # 3. Unify target compile
-    list(LENGTH base_device_kernels base_device_kernels_len)
-    list(LENGTH device_independent_kernel device_independent_kernel_len)
-    if (${base_device_kernels_len} GREATER 0 OR ${device_independent_kernel_len} GREATER 0)
+    elseif (${partial_build_flag} EQUAL 1 AND ${base_build_flag} EQUAL 0)
         if (WITH_GPU)
-            nv_library(${TARGET}${target_suffix} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels} ${device_independent_kernel})
+            nv_library(${TARGET}${target_suffix} SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
         elseif (WITH_ROCM)
-            hip_library(${TARGET}${target_suffix} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels} ${device_independent_kernel})
+            hip_library(${TARGET}${target_suffix} SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
         elseif (WITH_XPU_KP)
-            xpu_library(${TARGET}${target_suffix} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels} ${device_independent_kernel})
+            xpu_library(${TARGET}${target_suffix} SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
         else()
-            cc_library(${TARGET}${target_suffix} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels} ${device_independent_kernel})
+            cc_library(${TARGET}${target_suffix} SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
         endif()
     else()
         set(target_build_flag 0)