| 
 | 1 | +find_package(OpenCL REQUIRED)  | 
 | 2 | +find_package(Python3 REQUIRED)  | 
 | 3 | + | 
 | 4 | +set(TARGET_NAME ggml-opencl)  | 
 | 5 | + | 
 | 6 | +ggml_add_backend_library(${TARGET_NAME}  | 
 | 7 | +                         ggml-opencl.cpp  | 
 | 8 | +                         ../../include/ggml-opencl.h)  | 
 | 9 | +target_link_libraries(${TARGET_NAME} PRIVATE ${OpenCL_LIBRARIES})  | 
 | 10 | +target_include_directories(${TARGET_NAME} PRIVATE ${OpenCL_INCLUDE_DIRS})  | 
 | 11 | + | 
 | 12 | +if (GGML_OPENCL_PROFILING)  | 
 | 13 | +    message(STATUS "OpenCL profiling enabled (increases CPU overhead)")  | 
 | 14 | +    add_compile_definitions(GGML_OPENCL_PROFILING)  | 
 | 15 | +endif ()  | 
 | 16 | + | 
 | 17 | +add_compile_definitions(GGML_OPENCL_SOA_Q)  | 
 | 18 | + | 
 | 19 | +if (GGML_OPENCL_USE_ADRENO_KERNELS)  | 
 | 20 | +    message(STATUS "OpenCL will use matmul kernels optimized for Adreno")  | 
 | 21 | +    add_compile_definitions(GGML_OPENCL_USE_ADRENO_KERNELS)  | 
 | 22 | +endif ()  | 
 | 23 | + | 
 | 24 | +if (GGML_OPENCL_EMBED_KERNELS)  | 
 | 25 | +    add_compile_definitions(GGML_OPENCL_EMBED_KERNELS)  | 
 | 26 | + | 
 | 27 | +    set(OPENCL_CL_SOURCE_EMBED         "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl.cl.h")  | 
 | 28 | +    set(OPENCL_MM_CL_SOURCE_EMBED      "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_mm.cl.h")  | 
 | 29 | +    set(OPENCL_CVT_CL_SOURCE_EMBED     "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_cvt.cl.h")  | 
 | 30 | + | 
 | 31 | +    set(OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED             "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_gemv_noshuffle.cl.h")  | 
 | 32 | +    set(OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED     "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_gemv_noshuffle_general.cl.h")  | 
 | 33 | +    set(OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED          "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_mul_mat_Ab_Bi_8x4.cl.h")  | 
 | 34 | +    set(OPENCL_TRANSPOSE_16_SOURCE_EMBED               "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_16.cl.h")  | 
 | 35 | +    set(OPENCL_TRANSPOSE_32_SOURCE_EMBED               "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_32.cl.h")  | 
 | 36 | +    set(OPENCL_TRANSPOSE_32_16_SOURCE_EMBED            "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_32_16.cl.h")  | 
 | 37 | + | 
 | 38 | +    set(EMBED_KERNEL_SCRIPT             "${CMAKE_CURRENT_SOURCE_DIR}/kernels/embed_kernel.py")  | 
 | 39 | +    file(MAKE_DIRECTORY                 "${CMAKE_BINARY_DIR}/autogenerated")  | 
 | 40 | + | 
 | 41 | +    include_directories("${CMAKE_BINARY_DIR}/autogenerated")  | 
 | 42 | + | 
 | 43 | +    # Python must be accessible from command line  | 
 | 44 | +    add_custom_command(  | 
 | 45 | +        OUTPUT ${OPENCL_CL_SOURCE_EMBED}  | 
 | 46 | +        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}  | 
 | 47 | +            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl.cl  | 
 | 48 | +            ${OPENCL_CL_SOURCE_EMBED}  | 
 | 49 | +        DEPENDS kernels/ggml-opencl.cl ${EMBED_KERNEL_SCRIPT}  | 
 | 50 | +        COMMENT "Generate ggml-opencl.cl.h"  | 
 | 51 | +    )  | 
 | 52 | + | 
 | 53 | +    add_custom_command(  | 
 | 54 | +        OUTPUT ${OPENCL_MM_CL_SOURCE_EMBED}  | 
 | 55 | +        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}  | 
 | 56 | +            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_mm.cl  | 
 | 57 | +            ${OPENCL_MM_CL_SOURCE_EMBED}  | 
 | 58 | +        DEPENDS kernels/ggml-opencl_mm.cl ${EMBED_KERNEL_SCRIPT}  | 
 | 59 | +        COMMENT "Generate ggml-opencl_mm.cl.h"  | 
 | 60 | +    )  | 
 | 61 | + | 
 | 62 | +    add_custom_command(  | 
 | 63 | +        OUTPUT ${OPENCL_CVT_CL_SOURCE_EMBED}  | 
 | 64 | +        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}  | 
 | 65 | +            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_cvt.cl  | 
 | 66 | +            ${OPENCL_CVT_CL_SOURCE_EMBED}  | 
 | 67 | +        DEPENDS kernels/ggml-opencl_cvt.cl ${EMBED_KERNEL_SCRIPT}  | 
 | 68 | +        COMMENT "Generate ggml-opencl_cvt.cl.h"  | 
 | 69 | +    )  | 
 | 70 | + | 
 | 71 | +    add_custom_command(  | 
 | 72 | +        OUTPUT ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED}  | 
 | 73 | +        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}  | 
 | 74 | +            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_gemv_noshuffle.cl  | 
 | 75 | +            ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED}  | 
 | 76 | +        DEPENDS kernels/ggml-opencl_gemv_noshuffle.cl ${EMBED_KERNEL_SCRIPT}  | 
 | 77 | +        COMMENT "Generate ggml-opencl_gemv_noshuffle.cl.h"  | 
 | 78 | +    )  | 
 | 79 | + | 
 | 80 | +    add_custom_command(  | 
 | 81 | +        OUTPUT ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED}  | 
 | 82 | +        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}  | 
 | 83 | +            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_gemv_noshuffle_general.cl  | 
 | 84 | +            ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED}  | 
 | 85 | +        DEPENDS kernels/ggml-opencl_gemv_noshuffle_general.cl ${EMBED_KERNEL_SCRIPT}  | 
 | 86 | +        COMMENT "Generate ggml-opencl_gemv_noshuffle_general.cl.h"  | 
 | 87 | +    )  | 
 | 88 | + | 
 | 89 | +    add_custom_command(  | 
 | 90 | +        OUTPUT ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED}  | 
 | 91 | +        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}  | 
 | 92 | +            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl  | 
 | 93 | +            ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED}  | 
 | 94 | +        DEPENDS kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl ${EMBED_KERNEL_SCRIPT}  | 
 | 95 | +        COMMENT "Generate ggml-opencl_mul_mat_Ab_Bi_8x4.cl.cl.h"  | 
 | 96 | +    )  | 
 | 97 | + | 
 | 98 | +    add_custom_command(  | 
 | 99 | +        OUTPUT ${OPENCL_TRANSPOSE_16_SOURCE_EMBED}  | 
 | 100 | +        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}  | 
 | 101 | +            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_16.cl  | 
 | 102 | +            ${OPENCL_TRANSPOSE_16_SOURCE_EMBED}  | 
 | 103 | +        DEPENDS kernels/ggml-opencl_transpose_16.cl ${EMBED_KERNEL_SCRIPT}  | 
 | 104 | +        COMMENT "Generate ggml-opencl_transpose_16.cl.h"  | 
 | 105 | +    )  | 
 | 106 | + | 
 | 107 | +    add_custom_command(  | 
 | 108 | +        OUTPUT ${OPENCL_TRANSPOSE_32_SOURCE_EMBED}  | 
 | 109 | +        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}  | 
 | 110 | +            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_32.cl  | 
 | 111 | +            ${OPENCL_TRANSPOSE_32_SOURCE_EMBED}  | 
 | 112 | +        DEPENDS kernels/ggml-opencl_transpose_32.cl ${EMBED_KERNEL_SCRIPT}  | 
 | 113 | +        COMMENT "Generate ggml-opencl_transpose_32.cl.h"  | 
 | 114 | +    )  | 
 | 115 | + | 
 | 116 | +    add_custom_command(  | 
 | 117 | +        OUTPUT ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED}  | 
 | 118 | +        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}  | 
 | 119 | +            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_32_16.cl  | 
 | 120 | +            ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED}  | 
 | 121 | +        DEPENDS kernels/ggml-opencl_transpose_32_16.cl ${EMBED_KERNEL_SCRIPT}  | 
 | 122 | +        COMMENT "Generate ggml-opencl_transpose_32_16.cl.h"  | 
 | 123 | +    )  | 
 | 124 | + | 
 | 125 | +    target_sources(${TARGET_NAME} PRIVATE  | 
 | 126 | +                   ${OPENCL_CL_SOURCE_EMBED}  | 
 | 127 | +                   ${OPENCL_MM_CL_SOURCE_EMBED}  | 
 | 128 | +                   ${OPENCL_CVT_CL_SOURCE_EMBED}  | 
 | 129 | +                   ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED}  | 
 | 130 | +                   ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED}  | 
 | 131 | +                   ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED}  | 
 | 132 | +                   ${OPENCL_TRANSPOSE_16_SOURCE_EMBED}  | 
 | 133 | +                   ${OPENCL_TRANSPOSE_32_SOURCE_EMBED}  | 
 | 134 | +                   ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED})  | 
 | 135 | +else ()  | 
 | 136 | +    # copy ggml-opencl.cl to bin directory  | 
 | 137 | +    configure_file(kernels/ggml-opencl.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl.cl COPYONLY)  | 
 | 138 | +    configure_file(kernels/ggml-opencl_mm.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_mm.cl COPYONLY)  | 
 | 139 | +    configure_file(kernels/ggml-opencl_cvt.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_cvt.cl COPYONLY)  | 
 | 140 | + | 
 | 141 | +    configure_file(kernels/ggml-opencl_gemv_noshuffle.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_gemv_noshuffle.cl COPYONLY)  | 
 | 142 | +    configure_file(kernels/ggml-opencl_gemv_noshuffle_general.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_gemv_noshuffle_general.cl COPYONLY)  | 
 | 143 | +    configure_file(kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_mul_mat_Ab_Bi_8x4.cl COPYONLY)  | 
 | 144 | +    configure_file(kernels/ggml-opencl_transpose_16.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_16.cl COPYONLY)  | 
 | 145 | +    configure_file(kernels/ggml-opencl_transpose_32.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_32.cl COPYONLY)  | 
 | 146 | +    configure_file(kernels/ggml-opencl_transpose_32_16.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_32_16.cl COPYONLY)  | 
 | 147 | +endif ()  | 
0 commit comments