Merge pull request #3378 from martin-frbg/issue3368-2

Rework generation of BFLOAT16 objects in CMAKE builds and fix missing CBLAS_XERBLA
This commit is contained in:
Martin Kroeker 2021-09-15 07:18:38 +02:00 committed by GitHub
commit 13182b2801
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
7 changed files with 110 additions and 26 deletions

View File

@ -132,7 +132,7 @@ endif ()
if (BUILD_BFLOAT16)
message(STATUS "Building Half Precision")
list(APPEND FLOAT_TYPES "BFLOAT16") # defines nothing
# list(APPEND FLOAT_TYPES "BFLOAT16") # defines nothing
endif ()
if (NOT DEFINED CORE OR "${CORE}" STREQUAL "UNKNOWN")

View File

@ -134,6 +134,8 @@ if (BUILD_BFLOAT16)
set(SHSWAPKERNEL ../arm/swap.c)
set(TOBF16KERNEL ../x86_64/tobf16.c)
set(BF16TOKERNEL ../x86_64/bf16to.c)
set(SBGEMVNKERNEL ../x86_64/sbgemv_n.c)
set(SBGEMVTKERNEL ../x86_64/sbgemv_t.c)
endif ()
endmacro ()

View File

@ -469,6 +469,9 @@ endif()
if (BUILD_COMPLEX16)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_COMPLEX16")
endif()
if (BUILD_BFLOAT16)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_BFLOAT16")
endif()
if(NOT MSVC)
set(CMAKE_ASM_FLAGS "${CMAKE_ASM_FLAGS} ${CCOMMON_OPT}")
endif()

View File

@ -81,6 +81,7 @@ foreach (float_type ${FLOAT_TYPES})
GenerateNamedObjects("gbmv_thread.c" "TRANSA" "gbmv_thread_t" false "" "" false ${float_type})
endif ()
# special defines for complex
if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
foreach (u_source ${U_SOURCES})
@ -197,6 +198,13 @@ foreach (float_type ${FLOAT_TYPES})
endif ()
endforeach ()
if (BUILD_BFLOAT16)
if (USE_THREAD)
GenerateNamedObjects("sbgemv_thread.c" "" "gemv_thread_n" false "" "" false "BFLOAT16")
GenerateNamedObjects("sbgemv_thread.c" "TRANSA" "gemv_thread_t" false "" "" false "BFLOAT16")
endif ()
endif ()
if ( BUILD_COMPLEX AND NOT BUILD_SINGLE)
if (USE_THREAD)
GenerateNamedObjects("gemv_thread.c" "" "gemv_thread_n" false "" "" false "SINGLE")

View File

@ -12,6 +12,12 @@ foreach (GEMM_DEFINE ${GEMM_DEFINES})
if (USE_THREAD AND NOT USE_SIMPLE_THREADED_LEVEL3)
GenerateNamedObjects("gemm.c" "${GEMM_DEFINE};THREADED_LEVEL3" "gemm_thread_${GEMM_DEFINE_LC}" 0)
endif ()
if (BUILD_BFLOAT16)
GenerateNamedObjects("gemm.c" "${GEMM_DEFINE}" "gemm_${GEMM_DEFINE_LC}" 0 "" "" false "BFLOAT16")
if (USE_THREAD AND NOT USE_SIMPLE_THREADED_LEVEL3)
GenerateNamedObjects("gemm.c" "${GEMM_DEFINE};THREADED_LEVEL3" "gemm_thread_${GEMM_DEFINE_LC}" 0 "" "" false "BFLOAT16")
endif ()
endif ()
endforeach ()
if ( BUILD_COMPLEX16 AND NOT BUILD_DOUBLE)

View File

@ -82,6 +82,7 @@ foreach (CBLAS_FLAG ${CBLAS_FLAGS})
GenerateNamedObjects("${BLAS3_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${DISABLE_COMPLEX})
GenerateNamedObjects("${BLAS3_MANGLED_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${MANGLE_COMPLEX})
GenerateNamedObjects("xerbla.c" "" "xerbla" ${CBLAS_FLAG} "" "" true)
#sdsdot, dsdot
if (BUILD_SINGLE OR BUILD_DOUBLE)
GenerateNamedObjects("sdsdot.c" "" "sdsdot" ${CBLAS_FLAG} "" "" true "SINGLE")
@ -104,6 +105,15 @@ endif ()
GenerateNamedObjects("imax.c" "USE_ABS;USE_MIN" "i*amin" ${CBLAS_FLAG})
GenerateNamedObjects("imax.c" "USE_MIN" "i*min" ${CBLAS_FLAG})
if (BUILD_BFLOAT16)
GenerateNamedObjects("bf16dot.c" "" "sbdot" ${CBLAS_FLAG} "" "" true "BFLOAT16")
GenerateNamedObjects("gemm.c" "" "sbgemm" ${CBLAS_FLAG} "" "" true "BFLOAT16")
GenerateNamedObjects("sbgemv.c" "" "sbgemv" ${CBLAS_FLAG} "" "" true "BFLOAT16")
GenerateNamedObjects("tobf16.c" "SINGLE_PREC" "sbstobf16" ${CBLAS_FLAG} "" "" true "BFLOAT16")
GenerateNamedObjects("tobf16.c" "DOUBLE_PREC" "sbdtobf16" ${CBLAS_FLAG} "" "" true "BFLOAT16")
GenerateNamedObjects("bf16to.c" "SINGLE_PREC" "sbf16tos" ${CBLAS_FLAG} "" "" true "BFLOAT16")
GenerateNamedObjects("bf16to.c" "DOUBLE_PREC" "dbf16tod" ${CBLAS_FLAG} "" "" true "BFLOAT16")
endif ()
# complex-specific sources
foreach (float_type ${FLOAT_TYPES})

View File

@ -91,6 +91,15 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
GenerateNamedObjects("${KERNELDIR}/${DSDOTKERNEL}" "DSDOT" "d*dot_k" false "" "" false "SINGLE")
GenerateNamedObjects("${KERNELDIR}/${DSDOTKERNEL}" "DSDOT" "dsdot_k" false "" "" false "SINGLE")
# sbdot
if (BUILD_BFLOAT16)
GenerateNamedObjects("${KERNELDIR}/${SBDOTKERNEL}" "SBDOT" "dot_k" false "" "" false "BFLOAT16")
GenerateNamedObjects("${KERNELDIR}/${BF16TOKERNEL}" "SINGLE" "f16tos_k" false "" "" false "BFLOAT16")
GenerateNamedObjects("${KERNELDIR}/${BF16TOKERNEL}" "DOUBLE" "bf16tod_k" false "" "" false "DOUBLE")
GenerateNamedObjects("${KERNELDIR}/${TOBF16KERNEL}" "SINGLE" "stobf16_k" false "" "" false "BFLOAT16")
GenerateNamedObjects("${KERNELDIR}/${TOBF16KERNEL}" "DOUBLE" "dtobf16_k" false "" "" false "BFLOAT16")
endif()
if ((BUILD_COMPLEX OR BUILD_DOUBLE) AND NOT BUILD_SINGLE)
GenerateNamedObjects("${KERNELDIR}/${SAMAXKERNEL}" "USE_ABS" "amax_k" false "" "" false "SINGLE")
GenerateNamedObjects("${KERNELDIR}/${SAMINKERNEL}" "USE_ABS;USE_MIN" "amin_k" false "" "" false "SINGLE")
@ -149,9 +158,6 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
GenerateNamedObjects("generic/ger.c" "" "ger_k" false "" "" "" 3)
foreach (float_type ${FLOAT_TYPES})
string(SUBSTRING ${float_type} 0 1 float_char)
if (${float_type} STREQUAL "BFLOAT16")
set (float_char "SB")
endif ()
if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
GenerateNamedObjects("${KERNELDIR}/${${float_char}GERUKERNEL}" "" "geru_k" false "" "" false ${float_type})
GenerateNamedObjects("${KERNELDIR}/${${float_char}GERCKERNEL}" "CONJ" "gerc_k" false "" "" false ${float_type})
@ -185,6 +191,10 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
GenerateNamedObjects("${KERNELDIR}/${SGEMVNKERNEL}" "" "gemv_n" false "" "" false "SINGLE")
GenerateNamedObjects("${KERNELDIR}/${SGEMVTKERNEL}" "TRANS" "gemv_t" false "" "" false "SINGLE")
endif ()
if (BUILD_BFLOAT16)
GenerateNamedObjects("${KERNELDIR}/${SBGEMVNKERNEL}" "" "gemv_n" false "" "" false "BFLOAT16")
GenerateNamedObjects("${KERNELDIR}/${SBGEMVTKERNEL}" "" "gemv_t" false "" "" false "BFLOAT16")
endif ()
# Makefile.L3
set(USE_TRMM false)
string(TOUPPER ${TARGET_CORE} UC_TARGET_CORE)
@ -209,15 +219,8 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
GenerateNamedObjects("${KERNELDIR}/${SGEMMDIRECTPERFORMANT}" "" "gemm_direct_performant" false "" "" false SINGLE)
endif()
foreach (float_type SINGLE DOUBLE BFLOAT16)
foreach (float_type SINGLE DOUBLE)
string(SUBSTRING ${float_type} 0 1 float_char)
if (${float_type} STREQUAL "BFLOAT16")
if (NOT ${BUILD_BFLOAT16})
continue ()
else ()
set (float_char "SB")
endif ()
endif ()
GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMKERNEL}" "" "gemm_kernel" false "" "" false ${float_type})
endforeach()
if (BUILD_COMPLEX16 AND NOT BUILD_DOUBLE)
@ -253,11 +256,24 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
GenerateNamedObjects("${KERNELDIR}/${SGEMM_BETA}" "" "gemm_beta" false "" "" false "SINGLE")
endif ()
if (BUILD_BFLOAT16)
if (SBGEMMINCOPY)
GenerateNamedObjects("${KERNELDIR}/${SBGEMMINCOPY}" "" "${SBGEMMINCOPYOBJ}" false "" "" true "BFLOAT16")
endif ()
if (SBGEMMITCOPY)
GenerateNamedObjects("${KERNELDIR}/${SBGEMMITCOPY}" "" "${SBGEMMITCOPYOBJ}" false "" "" true "BFLOAT16")
endif ()
if (SBGEMMONCOPY)
GenerateNamedObjects("${KERNELDIR}/${SBGEMMONCOPY}" "" "${SBGEMMONCOPYOBJ}" false "" "" true "BFLOAT16")
endif ()
if (SBGEMMOTCOPY)
GenerateNamedObjects("${KERNELDIR}/${SBGEMMOTCOPY}" "" "${SBGEMMOTCOPYOBJ}" false "" "" true "BFLOAT16")
endif ()
GenerateNamedObjects("${KERNELDIR}/${SBGEMMKERNEL}" "" "gemm_kernel" false "" "" false "BFLOAT16")
GenerateNamedObjects("${KERNELDIR}/${SBGEMM_BETA}" "" "gemm_beta" false "" "" false "BFLOAT16")
endif ()
foreach (float_type ${FLOAT_TYPES})
string(SUBSTRING ${float_type} 0 1 float_char)
if (${float_type} STREQUAL "BFLOAT16")
set (float_char "SB")
endif ()
if (${float_char}GEMMINCOPY)
GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMINCOPY}" "${float_type}" "${${float_char}GEMMINCOPYOBJ}" false "" "" true ${float_type})
endif ()
@ -568,6 +584,44 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMM_SMALL_K_B0_TN}" "B0" "gemm_small_kernel_b0_tn" false "" "" false ${float_type})
GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMM_SMALL_K_B0_NT}" "B0" "gemm_small_kernel_b0_tt" false "" "" false ${float_type})
endif ()
if (BUILD_BFLOAT16)
if (NOT DEFINED SBGEMM_SMALL_M_PERMIT)
set(SBGEMM_SMALL_M_PERMIT ../generic/gemm_small_matrix_permit.c)
endif ()
if (NOT DEFINED SBGEMM_SMALL_K_NN)
set(SBGEMM_SMALL_K_NN ../generic/gemm_small_matrix_kernel_nn.c)
endif ()
if (NOT DEFINED SBGEMM_SMALL_K_NT)
set(SBGEMM_SMALL_K_NT ../generic/gemm_small_matrix_kernel_nt.c)
endif ()
if (NOT DEFINED SBGEMM_SMALL_K_TN)
set(SBGEMM_SMALL_K_TN ../generic/gemm_small_matrix_kernel_tn.c)
endif ()
if (NOT DEFINED SBGEMM_SMALL_K_TT)
set(SBGEMM_SMALL_K_TT ../generic/gemm_small_matrix_kernel_tt.c)
endif ()
if (NOT DEFINED SBGEMM_SMALL_K_B0_NN)
set(SBGEMM_SMALL_K_B0_NN ../generic/gemm_small_matrix_kernel_nn.c)
endif ()
if (NOT DEFINED SBGEMM_SMALL_K_B0_NT)
set(SBGEMM_SMALL_K_B0_NT ../generic/gemm_small_matrix_kernel_nt.c)
endif ()
if (NOT DEFINED SBGEMM_SMALL_K_B0_TN)
set(SBGEMM_SMALL_K_B0_TN ../generic/gemm_small_matrix_kernel_tn.c)
endif ()
if (NOT DEFINED SBGEMM_SMALL_K_B0_TT)
set($SBGEMM_SMALL_K_B0_TT ../generic/gemm_small_matrix_kernel_tt.c)
endif ()
GenerateNamedObjects("${KERNELDIR}/${SBGEMM_SMALL_M_PERMIT}" "" "gemm_small_matrix_permit" false "" "" false "BFLOAT16")
GenerateNamedObjects("${KERNELDIR}/${SBGEMM_SMALL_K_NN}" "" "gemm_small_kernel_nn" false "" "" false "BFLOAT16")
GenerateNamedObjects("${KERNELDIR}/${SBGEMM_SMALL_K_NT}" "" "gemm_small_kernel_nt" false "" "" false "BFLOAT16")
GenerateNamedObjects("${KERNELDIR}/${SBGEMM_SMALL_K_TN}" "" "gemm_small_kernel_tn" false "" "" false "BFLOAT16")
GenerateNamedObjects("${KERNELDIR}/${SBGEMM_SMALL_K_NT}" "" "gemm_small_kernel_tt" false "" "" false "BFLOAT16")
GenerateNamedObjects("${KERNELDIR}/${SBGEMM_SMALL_K_B0_NN}" "B0" "gemm_small_kernel_b0_nn" false "" "" false "BFLOAT16")
GenerateNamedObjects("${KERNELDIR}/${SBGEMM_SMALL_K_B0_NT}" "B0" "gemm_small_kernel_b0_nt" false "" "" false "BFLOAT16")
GenerateNamedObjects("${KERNELDIR}/${SBGEMM_SMALL_K_B0_TN}" "B0" "gemm_small_kernel_b0_tn" false "" "" false "BFLOAT16")
GenerateNamedObjects("${KERNELDIR}/${SBGEMM_SMALL_K_B0_NT}" "B0" "gemm_small_kernel_b0_tt" false "" "" false "BFLOAT16")
endif ()
endif ()
if (NOT DEFINED ${float_char}OMATCOPY_CN)
@ -702,6 +756,7 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
#geadd
GenerateNamedObjects("${KERNELDIR}/${${float_char}GEADD_KERNEL}" "" "geadd_k" false "" "" false ${float_type})
endforeach ()
if (BUILD_DOUBLE AND NOT BUILD_SINGLE)
GenerateNamedObjects("${KERNELDIR}/${STRSMKERNEL_LN}" "UPPER;LN;TRSMKERNEL" "trsm_kernel_LN" false "" "" false "SINGLE")
GenerateNamedObjects("${KERNELDIR}/${STRSMKERNEL_LT}" "LT;TRSMKERNEL" "trsm_kernel_LT" false "" "" false "SINGLE")
@ -840,22 +895,22 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
GenerateNamedObjects("generic/trsm_ltcopy_${SGEMM_UNROLL_N}.c" "OUTER;LOWER" "trsm_oltncopy" false "" ${TSUFFIX} false "SINGLE")
if (SGEMMINCOPY)
GenerateNamedObjects("${KERNELDIR}/${SGEMMINCOPY}" "SINGLE" "${SGEMMINCOPYOBJ}" false "" "" true "SINGLE")
GenerateNamedObjects("${KERNELDIR}/${SGEMMINCOPY}" "SINGLE" "${SGEMMINCOPYOBJ}" false "" "" true "SINGLE")
endif ()
if (SGEMMITCOPY)
GenerateNamedObjects("${KERNELDIR}/${SGEMMITCOPY}" "SINGLE" "${SGEMMITCOPYOBJ}" false "" "" true "SINGLE")
endif ()
if (SGEMMONCOPY)
GenerateNamedObjects("${KERNELDIR}/${SGEMMONCOPY}" "SINGLE" "${SGEMMONCOPYOBJ}" false "" "" true "SINGLE")
endif ()
if (SGEMMOTCOPY)
GenerateNamedObjects("${KERNELDIR}/${SGEMMOTCOPY}" "SINGLE" "${SGEMMOTCOPYOBJ}" false "" "" true "SINGLE")
if (SGEMMITCOPY)
GenerateNamedObjects("${KERNELDIR}/${SGEMMITCOPY}" "SINGLE" "${SGEMMITCOPYOBJ}" false "" "" true "SINGLE")
endif ()
if (SGEMMONCOPY)
GenerateNamedObjects("${KERNELDIR}/${SGEMMONCOPY}" "SINGLE" "${SGEMMONCOPYOBJ}" false "" "" true "SINGLE")
endif ()
if (SGEMMOTCOPY)
GenerateNamedObjects("${KERNELDIR}/${SGEMMOTCOPY}" "SINGLE" "${SGEMMOTCOPYOBJ}" false "" "" true "SINGLE")
endif ()
GenerateNamedObjects("${KERNELDIR}/${SGEMVNKERNEL}" "" "gemv_n" false "" "" false "SINGLE")
GenerateNamedObjects("${KERNELDIR}/${SGEMVTKERNEL}" "TRANS" "gemv_t" false "" "" false "SINGLE")
endif ()
if (BUILD_COMPLEX16 AND NOT BUILD_DOUBLE)
if (BUILD_COMPLEX16 AND NOT BUILD_DOUBLE)
GenerateNamedObjects("generic/neg_tcopy_${DGEMM_UNROLL_M}.c" "" "neg_tcopy" false "" ${TSUFFIX} false "DOUBLE")
GenerateNamedObjects("generic/laswp_ncopy_${DGEMM_UNROLL_N}.c" "" "laswp_ncopy" false "" ${TSUFFIX} false "DOUBLE")
endif ()