Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
未验证 提交 3a77d027 编写于 作者: C Chen Weihang 提交者： GitHub
浏览文件
操作

[Phi] Remove gpudnn suffix & polish cmake (#40239) 

* remove gpudnn suffix & polish cmake

* fix typo
上级 9aa6bfc7
隐藏空白更改
内联 并排
Showing with 59 addition and 63 deletion
cmake/phi.cmake
浏览文件 @ 3a77d027
...@@ -134,8 +134,8 @@ function(kernel_library TARGET) ...@@ -134,8 +134,8 @@ function(kernel_library TARGET)
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/kps/${TARGET}.cu) if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/kps/${TARGET}.cu)
list(APPEND gpu_srcs ${CMAKE_CURRENT_SOURCE_DIR}/kps/${TARGET}.cu) list(APPEND gpu_srcs ${CMAKE_CURRENT_SOURCE_DIR}/kps/${TARGET}.cu)
endif() endif()
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/gpudnn/${TARGET}_gpudnn.cu) if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/gpudnn/${TARGET}.cu)
list(APPEND gpudnn_srcs ${CMAKE_CURRENT_SOURCE_DIR}/gpudnn/${TARGET}_gpudnn.cu) list(APPEND gpudnn_srcs ${CMAKE_CURRENT_SOURCE_DIR}/gpudnn/${TARGET}.cu)
endif() endif()
endif() endif()
if (WITH_XPU) if (WITH_XPU)
...@@ -197,92 +197,88 @@ function(kernel_library TARGET) ...@@ -197,92 +197,88 @@ function(kernel_library TARGET)
# kernel source file level # kernel source file level
# level 1: base device kernel # level 1: base device kernel
# - cpu_srcs / gpu_srcs / xpu_srcs / kps_srcs # - cpu_srcs / gpu_srcs / xpu_srcs / gpudnn_srcs / kps_srcs
# level 2: device-independent kernel # level 2: device-independent kernel
# - common_srcs # - common_srcs
# level 3: Kernel implemented by reusing device-independent kernel # level 3: Kernel implemented by reusing device-independent kernel
# - selected_rows_srcs # - selected_rows_srcs
set(base_device_kernels)
set(device_independent_kernel)
set(high_level_kernels)
# Build Target according different src organization # 1. Base device kernel compile
if((${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0 OR if (${cpu_srcs_len} GREATER 0)
${xpu_srcs_len} GREATER 0 OR ${gpudnn_srcs_len} GREATER 0 OR ${kps_srcs_len} GREATER 0) AND cc_library(${TARGET}_cpu SRCS ${cpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
(${common_srcs_len} GREATER 0 OR ${selected_rows_srcs_len} GREATER 0)) list(APPEND base_device_kernels ${TARGET}_cpu)
# If the common_srcs/selected_rows_srcs depends on specific device srcs, build target using this rule. endif()
if (${gpu_srcs_len} GREATER 0)
if (WITH_GPU) if (WITH_GPU)
if (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0 OR ${gpudnn_srcs_len} GREATER 0) nv_library(${TARGET}_gpu SRCS ${gpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
nv_library(${TARGET}_part SRCS ${cpu_srcs} ${gpu_srcs} ${gpudnn_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
nv_library(${TARGET} SRCS ${common_srcs} ${selected_rows_srcs} DEPS ${TARGET}_part)
endif()
elseif (WITH_ROCM) elseif (WITH_ROCM)
if (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0 OR ${gpudnn_srcs_len} GREATER 0) hip_library(${TARGET}_gpu SRCS ${gpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
hip_library(${TARGET}_part SRCS ${cpu_srcs} ${gpu_srcs} ${gpudnn_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
hip_library(${TARGET} SRCS ${common_srcs} ${selected_rows_srcs} DEPS ${TARGET}_part)
endif()
elseif (WITH_XPU_KP)
if (${cpu_srcs_len} GREATER 0 OR ${xpu_srcs_len} GREATER 0 OR ${kps_srcs_len} GREATER 0)
xpu_library(${TARGET}_part SRCS ${cpu_srcs} ${xpu_srcs} ${kps_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
xpu_library(${TARGET} SRCS ${common_srcs} ${selected_rows_srcs} DEPS ${TARGET}_part)
endif()
else()
if (${cpu_srcs_len} GREATER 0 OR ${xpu_srcs_len} GREATER 0)
cc_library(${TARGET}_part SRCS ${cpu_srcs} ${xpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
cc_library(${TARGET} SRCS ${common_srcs} ${selected_rows_srcs} DEPS ${TARGET}_part)
endif()
endif() endif()
# If there are only specific device srcs, build target using this rule. list(APPEND base_device_kernels ${TARGET}_gpu)
elseif (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0 OR ${xpu_srcs_len} GREATER 0 OR ${gpudnn_srcs_len} GREATER 0 OR ${kps_srcs_len} GREATER 0) endif()
if (${xpu_srcs_len} GREATER 0)
cc_library(${TARGET}_xpu SRCS ${xpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
list(APPEND base_device_kernels ${TARGET}_xpu)
endif()
if (${gpudnn_srcs_len} GREATER 0)
if (WITH_GPU) if (WITH_GPU)
if (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0 OR ${gpudnn_srcs_len} GREATER 0) nv_library(${TARGET}_gpudnn SRCS ${gpudnn_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
nv_library(${TARGET} SRCS ${cpu_srcs} ${gpu_srcs} ${gpudnn_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
endif()
elseif (WITH_ROCM) elseif (WITH_ROCM)
if (${cpu_srcs_len} GREATER 0 OR ${gpu_srcs_len} GREATER 0 OR ${gpudnn_srcs_len} GREATER 0) hip_library(${TARGET}_gpudnn SRCS ${gpudnn_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
hip_library(${TARGET} SRCS ${cpu_srcs} ${gpu_srcs} ${gpudnn_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
endif()
elseif (WITH_XPU_KP)
if (${cpu_srcs_len} GREATER 0 OR ${xpu_srcs_len} GREATER 0 OR ${kps_srcs_len} GREATER 0)
xpu_library(${TARGET} SRCS ${cpu_srcs} ${xpu_srcs} ${kps_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
endif()
else()
if (${cpu_srcs_len} GREATER 0 OR ${xpu_srcs_len} GREATER 0)
cc_library(${TARGET} SRCS ${cpu_srcs} ${xpu_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
endif()
endif() endif()
# If the selected_rows_srcs depends on common_srcs, build target using this rule. list(APPEND base_device_kernels ${TARGET}_gpudnn)
elseif (${common_srcs_len} GREATER 0 AND ${selected_rows_srcs_len} GREATER 0) endif()
if (${kps_srcs_len} GREATER 0)
# only when WITH_XPU_KP, the kps_srcs_len can be > 0
xpu_library(${TARGET}_kps SRCS ${kps_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps})
list(APPEND base_device_kernels ${TARGET}_kps)
endif()
# 2. Device-independent kernel compile
if (${common_srcs_len} GREATER 0)
if (WITH_GPU) if (WITH_GPU)
nv_library(${TARGET}_part SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps}) nv_library(${TARGET}_common SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels})
nv_library(${TARGET} SRCS ${selected_rows_srcs} DEPS ${TARGET}_part)
elseif (WITH_ROCM) elseif (WITH_ROCM)
hip_library(${TARGET}_part SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps}) hip_library(${TARGET}_common SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels})
hip_library(${TARGET} SRCS ${selected_rows_srcs} DEPS ${TARGET}_part)
elseif (WITH_XPU_KP) elseif (WITH_XPU_KP)
xpu_library(${TARGET}_part SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps}) xpu_library(${TARGET}_common SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels})
xpu_library(${TARGET} SRCS ${selected_rows_srcs} DEPS ${TARGET}_part)
else() else()
cc_library(${TARGET}_part SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps}) cc_library(${TARGET}_common SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels})
cc_library(${TARGET} SRCS ${selected_rows_srcs} DEPS ${TARGET}_part)
endif() endif()
# If there are only common_srcs or selected_rows_srcs, build target using below rules. list(APPEND device_independent_kernel ${TARGET}_common)
elseif (${common_srcs_len} GREATER 0) endif()
# 3. Reusing kernel compile
if (${selected_rows_srcs_len} GREATER 0)
if (WITH_GPU) if (WITH_GPU)
nv_library(${TARGET} SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps}) nv_library(${TARGET}_sr SRCS ${selected_rows_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels} ${device_independent_kernel})
elseif (WITH_ROCM) elseif (WITH_ROCM)
hip_library(${TARGET} SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps}) hip_library(${TARGET}_sr SRCS ${selected_rows_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels} ${device_independent_kernel})
elseif (WITH_XPU_KP) elseif (WITH_XPU_KP)
xpu_library(${TARGET} SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps}) xpu_library(${TARGET}_sr SRCS ${selected_rows_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels} ${device_independent_kernel})
else() else()
cc_library(${TARGET} SRCS ${common_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps}) cc_library(${TARGET}_sr SRCS ${selected_rows_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels} ${device_independent_kernel})
endif() endif()
elseif (${selected_rows_srcs_len} GREATER 0) list(APPEND high_level_kernels ${TARGET}_sr)
endif()
# 4. Unify target compile
list(LENGTH base_device_kernels base_device_kernels_len)
list(LENGTH device_independent_kernel device_independent_kernel_len)
list(LENGTH high_level_kernels high_level_kernels_len)
if (${base_device_kernels_len} GREATER 0 OR ${device_independent_kernel_len} GREATER 0 OR
${high_level_kernels_len} GREATER 0)
if (WITH_GPU) if (WITH_GPU)
nv_library(${TARGET} SRCS ${selected_rows_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps}) nv_library(${TARGET} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels} ${device_independent_kernel} ${high_level_kernels})
elseif (WITH_ROCM) elseif (WITH_ROCM)
hip_library(${TARGET} SRCS ${selected_rows_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps}) hip_library(${TARGET} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels} ${device_independent_kernel} ${high_level_kernels})
elseif (WITH_XPU_KP) elseif (WITH_XPU_KP)
xpu_library(${TARGET} SRCS ${selected_rows_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps}) xpu_library(${TARGET} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels} ${device_independent_kernel} ${high_level_kernels})
else() else()
cc_library(${TARGET} SRCS ${selected_rows_srcs} DEPS ${kernel_library_DEPS} ${kernel_deps}) cc_library(${TARGET} DEPS ${kernel_library_DEPS} ${kernel_deps} ${base_device_kernels} ${device_independent_kernel} ${high_level_kernels})
endif() endif()
else() else()
set(target_build_flag 0) set(target_build_flag 0)
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或