Merge branch 'develop' into arm64_cmake_small_matrix_opt

martin-frbg · Oct 3, 2024 · b4495a8 · b4495a8
2 parents 7087b0a + 68eefe6
commit b4495a8
Show file tree

Hide file tree

Showing 26 changed files with 602 additions and 146 deletions.
diff --git a/.github/workflows/loongarch64.yml b/.github/workflows/loongarch64.yml
@@ -23,6 +23,15 @@ jobs:
           - target: LOONGSON2K1000
             triple: loongarch64-unknown-linux-gnu
             opts: NO_SHARED=1 DYNAMIC_ARCH=1 TARGET=LOONGSON2K1000
+          - target: LA64_GENERIC
+            triple:  loongarch64-unknown-linux-gnu
+            opts: NO_SHARED=1 DYNAMIC_ARCH=1 TARGET=LA64_GENERIC
+          - target: LA464
+            triple: loongarch64-unknown-linux-gnu
+            opts: NO_SHARED=1 DYNAMIC_ARCH=1 TARGET=LA464
+          - target: LA264
+            triple: loongarch64-unknown-linux-gnu
+            opts: NO_SHARED=1 DYNAMIC_ARCH=1 TARGET=LA264
           - target: DYNAMIC_ARCH
             triple: loongarch64-unknown-linux-gnu
             opts: NO_SHARED=1 DYNAMIC_ARCH=1 TARGET=GENERIC

diff --git a/.github/workflows/loongarch64_clang.yml b/.github/workflows/loongarch64_clang.yml
@@ -20,6 +20,12 @@ jobs:
             opts: NO_SHARED=1 DYNAMIC_ARCH=1 TARGET=LOONGSON3R5
           - target: LOONGSON2K1000
             opts: NO_SHARED=1 DYNAMIC_ARCH=1 TARGET=LOONGSON2K1000
+          - target: LA64_GENERIC
+            opts: NO_SHARED=1 DYNAMIC_ARCH=1 TARGET=LA64_GENERIC
+          - target: LA464
+            opts: NO_SHARED=1 DYNAMIC_ARCH=1 TARGET=LA464
+          - target: LA264
+            opts: NO_SHARED=1 DYNAMIC_ARCH=1 TARGET=LA264
           - target: DYNAMIC_ARCH
             opts: NO_SHARED=1 DYNAMIC_ARCH=1 TARGET=GENERIC
 

diff --git a/Makefile.install b/Makefile.install
@@ -14,6 +14,9 @@ endif
 ifeq ($(INTERFACE64),1)
 USE_64BITINT=1
 endif
+ifeq ($(USE_OPENMP),1)
+	FOMP_OPT:= -fopenmp
+endif
 
 PREFIX ?= /opt/OpenBLAS
 
@@ -178,6 +181,7 @@ endif
 	@echo 'libnamesuffix='$(LIBNAMESUFFIX) >> "$(PKGFILE)"
 	@echo 'libsuffix='$(SYMBOLSUFFIX) >> "$(PKGFILE)"
 	@echo 'includedir='$(OPENBLAS_INCLUDE_DIR) >> "$(PKGFILE)"
+	@echo 'omp_opt='$(FOMP_OPT) >> "$(PKGFILE)"
 	@echo 'openblas_config= USE_64BITINT='$(INTERFACE64) 'DYNAMIC_ARCH='$(DYNAMIC_ARCH) 'DYNAMIC_OLDER='$(DYNAMIC_OLDER) 'NO_CBLAS='$(NO_CBLAS) 'NO_LAPACK='$(NO_LAPACK) 'NO_LAPACKE='$(NO_LAPACKE) 'NO_AFFINITY='$(NO_AFFINITY) 'USE_OPENMP='$(USE_OPENMP) $(TARGET) 'MAX_THREADS='$(NUM_THREADS)>> "$(PKGFILE)"
 	@echo 'version='$(VERSION) >> "$(PKGFILE)"
 	@echo 'extralib='$(PKG_EXTRALIB) >> "$(PKGFILE)"

diff --git a/Makefile.system b/Makefile.system
@@ -727,7 +727,7 @@ endif
 endif
 
 ifeq ($(ARCH), loongarch64)
-DYNAMIC_CORE = LOONGSON3R5 LOONGSON2K1000 LOONGSONGENERIC
+DYNAMIC_CORE = LA64_GENERIC LA264 LA464
 endif
 
 ifeq ($(ARCH), riscv64)
@@ -1720,8 +1720,8 @@ LAPACK_FFLAGS := $(filter-out -msse3 -mssse3 -msse4.1 -mavx -mavx2 -mskylake-avx
 override FFLAGS := $(filter-out -msse3 -mssse3 -msse4.1 -mavx -mavx2 -mskylake-avx512 ,$(FFLAGS))
 endif
 ifeq ($(F_COMPILER),FLANGNEW)
-LAPACK_FFLAGS := $(filter-out -m32 -m64 -msse3 -mssse3 -msse4.1 -mavx -mavx2 -mskylake-avx512 ,$(FFLAGS))
-override FFLAGS := $(filter-out -m32 -m64 -msse3 -mssse3 -msse4.1 -mavx -mavx2 -mskylake-avx512 ,$(FFLAGS))
+LAPACK_FFLAGS := $(filter-out -m32 -m64 -msse3 -mssse3 -msse4.1 -mavx -mavx2 -mskylake-avx512 -mtune=% -mabi=% ,$(FFLAGS))
+override FFLAGS := $(filter-out -m32 -m64 -msse3 -mssse3 -msse4.1 -mavx -mavx2 -mskylake-avx512 -mtune=% -mabi=% ,$(FFLAGS))
 endif
 
 LAPACK_CFLAGS = $(CFLAGS)

diff --git a/TargetList.txt b/TargetList.txt
@@ -126,9 +126,17 @@ x280
 RISCV64_ZVL256B
 
 11.LOONGARCH64:
+// LOONGSONGENERIC/LOONGSON2K1000/LOONGSON3R5 are legacy names,
+// and it is recommended to use the more standardized naming conventions
+// LA64_GENERIC/LA264/LA464. You can still specify TARGET as
+// LOONGSONGENERIC/LOONGSON2K1000/LOONGSON3R5 during compilation or runtime,
+// and they will be internally relocated to LA64_GENERIC/LA264/LA464.
 LOONGSONGENERIC
-LOONGSON3R5
 LOONGSON2K1000
+LOONGSON3R5
+LA64_GENERIC
+LA264
+LA464
 
 12. Elbrus E2000:
 E2K

diff --git a/cblas.h b/cblas.h
@@ -407,13 +407,13 @@ void cblas_cimatcopy(OPENBLAS_CONST enum CBLAS_ORDER CORDER, OPENBLAS_CONST enum
 void cblas_zimatcopy(OPENBLAS_CONST enum CBLAS_ORDER CORDER, OPENBLAS_CONST enum CBLAS_TRANSPOSE CTRANS, OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST double* calpha, double* a, 
 		     OPENBLAS_CONST blasint clda, OPENBLAS_CONST blasint cldb); 
 
-void cblas_sgeadd(OPENBLAS_CONST enum CBLAS_ORDER CORDER,OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST float calpha, float *a, OPENBLAS_CONST blasint clda, OPENBLAS_CONST float cbeta, 
+void cblas_sgeadd(OPENBLAS_CONST enum CBLAS_ORDER CORDER,OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST float calpha, OPENBLAS_CONST float *a, OPENBLAS_CONST blasint clda, OPENBLAS_CONST float cbeta, 
 		  float *c, OPENBLAS_CONST blasint cldc); 
-void cblas_dgeadd(OPENBLAS_CONST enum CBLAS_ORDER CORDER,OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST double calpha, double *a, OPENBLAS_CONST blasint clda, OPENBLAS_CONST double cbeta, 
+void cblas_dgeadd(OPENBLAS_CONST enum CBLAS_ORDER CORDER,OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST double calpha, OPENBLAS_CONST double *a, OPENBLAS_CONST blasint clda, OPENBLAS_CONST double cbeta, 
 		  double *c, OPENBLAS_CONST blasint cldc); 
-void cblas_cgeadd(OPENBLAS_CONST enum CBLAS_ORDER CORDER,OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST float *calpha, float *a, OPENBLAS_CONST blasint clda, OPENBLAS_CONST float *cbeta, 
+void cblas_cgeadd(OPENBLAS_CONST enum CBLAS_ORDER CORDER,OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST float *calpha, OPENBLAS_CONST float *a, OPENBLAS_CONST blasint clda, OPENBLAS_CONST float *cbeta, 
 		  float *c, OPENBLAS_CONST blasint cldc); 
-void cblas_zgeadd(OPENBLAS_CONST enum CBLAS_ORDER CORDER,OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST double *calpha, double *a, OPENBLAS_CONST blasint clda, OPENBLAS_CONST double *cbeta, 
+void cblas_zgeadd(OPENBLAS_CONST enum CBLAS_ORDER CORDER,OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST double *calpha, OPENBLAS_CONST double *a, OPENBLAS_CONST blasint clda, OPENBLAS_CONST double *cbeta, 
 		  double *c, OPENBLAS_CONST blasint cldc); 
 
 void cblas_sgemm_batch(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_TRANSPOSE * TransA_array, OPENBLAS_CONST enum CBLAS_TRANSPOSE * TransB_array, OPENBLAS_CONST blasint * M_array, OPENBLAS_CONST blasint * N_array, OPENBLAS_CONST blasint * K_array,

diff --git a/cmake/arch.cmake b/cmake/arch.cmake
@@ -94,6 +94,10 @@ if (DYNAMIC_ARCH)
     endif ()
   endif ()
 
+  if (LOONGARCH64)
+    set(DYNAMIC_CORE LOONGSONGENERIC LOONGSON2K1000 LOONGSON3R5)
+  endif ()
+
   if (EXISTS ${PROJECT_SOURCE_DIR}/config_kernel.h)
 	  message (FATAL_ERROR "Your build directory contains a file config_kernel.h, probably from a previous compilation with make. This will conflict with the cmake compilation and cause strange compiler errors - please remove the file before trying again")
   endif ()

diff --git a/cmake/fc.cmake b/cmake/fc.cmake
@@ -61,21 +61,25 @@ if (${F_COMPILER} STREQUAL "GFORTRAN" OR ${F_COMPILER} STREQUAL "F95" OR CMAKE_F
     endif ()
     if (LOONGARCH64)
       if (BINARY64)
-	CHECK_C_COMPILER_FLAG("-mabi=lp64d" COMPILER_SUPPORT_LP64D_ABI)
-        if(COMPILER_SUPPORT_LP64D_ABI)
-	  set(FCOMMON_OPT "${FCOMMON_OPT} -mabi=lp64d")
-	else()
-	  set(FCOMMON_OPT "${FCOMMON_OPT} -mabi=lp64")
-	endif ()
+        if (NOT CMAKE_Fortran_COMPILER_ID MATCHES "LLVMFlang.*")
+	  CHECK_C_COMPILER_FLAG("-mabi=lp64d" COMPILER_SUPPORT_LP64D_ABI)
+          if(COMPILER_SUPPORT_LP64D_ABI)
+	    set(FCOMMON_OPT "${FCOMMON_OPT} -mabi=lp64d")
+	  else()
+	    set(FCOMMON_OPT "${FCOMMON_OPT} -mabi=lp64")
+	  endif ()
+        endif ()
         if (INTERFACE64) 
           set(FCOMMON_OPT "${FCOMMON_OPT} -fdefault-integer-8") 
         endif () 
       else ()
-	CHECK_C_COMPILER_FLAG("-mabi=ilp32d" COMPILER_SUPPORT_ILP32D_ABI)
-	if(COMPILER_SUPPORT_ILP32D_ABI)
-	  set(FCOMMON_OPT "${FCOMMON_OPT} -mabi=ilp32d")
-	else()
-	  set(FCOMMON_OPT "${FCOMMON_OPT} -mabi=lp32")
+        if (NOT CMAKE_Fortran_COMPILER_ID MATCHES "LLVMFlang.*")
+	  CHECK_C_COMPILER_FLAG("-mabi=ilp32d" COMPILER_SUPPORT_ILP32D_ABI)
+	  if(COMPILER_SUPPORT_ILP32D_ABI)
+	    set(FCOMMON_OPT "${FCOMMON_OPT} -mabi=ilp32d")
+	  else()
+	    set(FCOMMON_OPT "${FCOMMON_OPT} -mabi=lp32")
+          endif ()
 	endif ()
       endif ()
     endif ()

diff --git a/cmake/openblas.pc.in b/cmake/openblas.pc.in
@@ -9,5 +9,5 @@ Name: OpenBLAS
 Description: OpenBLAS is an optimized BLAS library based on GotoBLAS2 1.13 BSD version
 Version: @OpenBLAS_VERSION@
 URL: https://github.com/OpenMathLib/OpenBLAS
-Libs: @OpenMP_C_FLAGS@ -L${libdir} -l${libnameprefix}openblas${libnamesuffix}${libsuffix} 
-Cflags: -I${includedir}
+Libs: -L${libdir} -l${libnameprefix}openblas${libnamesuffix}${libsuffix} 
+Cflags: -I${includedir} @OpenMP_C_FLAGS@ 
diff --git a/cmake/prebuild.cmake b/cmake/prebuild.cmake
@@ -1349,6 +1349,54 @@ endif ()
       "#define DTB_DEFAULT_ENTRIES 128\n"
       "#define DTB_SIZE 4096\n"
       "#define L2_ASSOCIATIVE 4\n")
+  elseif ("${TCORE}" STREQUAL "LOONGSONGENERIC")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define DTB_DEFAULT_ENTRIES 64\n")
+      set(SGEMM_UNROLL_M 2)
+      set(SGEMM_UNROLL_N 8)
+      set(DGEMM_UNROLL_M 2)
+      set(DGEMM_UNROLL_N 8)
+      set(CGEMM_UNROLL_M 1)
+      set(CGEMM_UNROLL_N 4)
+      set(ZGEMM_UNROLL_M 1)
+      set(ZGEMM_UNROLL_N 4)
+      set(CGEMM3M_UNROLL_M 2)
+      set(CGEMM3M_UNROLL_N 8)
+      set(ZGEMM3M_UNROLL_M 2)
+      set(ZGEMM3M_UNROLL_N 8)
+  elseif ("${TCORE}" STREQUAL "LOONGSON2K1000")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define DTB_DEFAULT_ENTRIES 64\n")
+      set(HAVE_LSX  1)
+      set(SGEMM_UNROLL_M 2)
+      set(SGEMM_UNROLL_N 8)
+      set(DGEMM_UNROLL_M 8)
+      set(DGEMM_UNROLL_N 4)
+      set(CGEMM_UNROLL_M 8)
+      set(CGEMM_UNROLL_N 4)
+      set(ZGEMM_UNROLL_M 4)
+      set(ZGEMM_UNROLL_N 4)
+      set(CGEMM3M_UNROLL_M 2)
+      set(CGEMM3M_UNROLL_N 8)
+      set(ZGEMM3M_UNROLL_M 8)
+      set(ZGEMM3M_UNROLL_N 4)
+  elseif ("${TCORE}" STREQUAL "LOONGSON3R5")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define DTB_DEFAULT_ENTRIES 64\n")
+      set(HAVE_LASX 1)
+      set(HAVE_LSX  1)
+      set(SGEMM_UNROLL_M 16)
+      set(SGEMM_UNROLL_N 8)
+      set(DGEMM_UNROLL_M 16)
+      set(DGEMM_UNROLL_N 6)
+      set(CGEMM_UNROLL_M 16)
+      set(CGEMM_UNROLL_N 4)
+      set(ZGEMM_UNROLL_M 8)
+      set(ZGEMM_UNROLL_N 4)
+      set(CGEMM3M_UNROLL_M 16)
+      set(CGEMM3M_UNROLL_N 8)
+      set(ZGEMM3M_UNROLL_M 16)
+      set(ZGEMM3M_UNROLL_N 6)
   endif()
   set(SBGEMM_UNROLL_M 8)
   set(SBGEMM_UNROLL_N 4)

diff --git a/cmake/system.cmake b/cmake/system.cmake
@@ -388,7 +388,7 @@ if (NEED_PIC)
   endif()
 endif ()
 
-if (X86_64 OR ${CORE} STREQUAL POWER10 OR ARM64)
+if (X86_64 OR ${CORE} STREQUAL POWER10 OR ARM64 OR LOONGARCH64)
   set(SMALL_MATRIX_OPT TRUE)
 endif ()
 if (ARM64)
@@ -403,7 +403,7 @@ if (SMALL_MATRIX_OPT)
 endif ()
 
 if (DYNAMIC_ARCH)
-  if (X86 OR X86_64 OR ARM64 OR POWER OR RISCV64)
+  if (X86 OR X86_64 OR ARM64 OR POWER OR RISCV64 OR LOONGARCH64)
     set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_ARCH")
     if (DYNAMIC_OLDER)
       set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_OLDER")

diff --git a/cmake/system_check.cmake b/cmake/system_check.cmake
@@ -104,6 +104,8 @@ elseif(ARM)
   set(ARCH "arm")
 elseif(ARM64)
   set(ARCH "arm64")
+elseif(LOONGARCH64)
+  set(ARCH "loongarch64")
 else()
   set(ARCH ${CMAKE_SYSTEM_PROCESSOR} CACHE STRING "Target Architecture")
 endif ()

diff --git a/common_loongarch64.h b/common_loongarch64.h
@@ -281,9 +281,13 @@ REALNAME: ;\
 #define GNUSTACK
 #endif /* defined(__linux__) && defined(__ELF__) */
 
+#ifdef __clang__
+#define EPILOGUE .end
+#else
 #define EPILOGUE      \
     .end    REALNAME ;\
     GNUSTACK
+#endif
 
 #define PROFCODE