Merge pull request #3456 from martin-frbg/issue3444

Add/restore a GENERIC target for MIPS32 and support MIPS32 cross-compilation using CMAKE

cmake/arch.cmake

@@ -109,7 +109,7 @@ if (${ARCH} STREQUAL "ia64")
   endif ()
 endif ()
 
-if (MIPS64)
+if (MIPS32 OR MIPS64)
   set(NO_BINARY_MODE 1)
 endif ()
 

cmake/cc.cmake

@@ -15,6 +15,11 @@ if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU" OR ${CMAKE_C_COMPILER_ID} STREQUAL "LS
 
   if (NO_BINARY_MODE)
 
+    if (MIPS32)
+        set(CCOMMON_OPT "${CCOMMON_OPT} -mabi=32")
+      set(BINARY_DEFINED 1)
+    endif ()
+
     if (MIPS64)
       if (BINARY64)
         set(CCOMMON_OPT "${CCOMMON_OPT} -mabi=64")

cmake/prebuild.cmake

@@ -416,7 +416,7 @@ endif ()
     set(ZGEMM_UNROLL_M 4)
     set(ZGEMM_UNROLL_N 4)
     set(SYMV_P 16)
-elseif ("${TCORE}" STREQUAL "VORTEX")
+  elseif ("${TCORE}" STREQUAL "VORTEX")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define ARMV8\n"
       "#define L1_CODE_SIZE\t32768\n"
@@ -439,6 +439,34 @@ elseif ("${TCORE}" STREQUAL "VORTEX")
     set(ZGEMM_UNROLL_M 4)
     set(ZGEMM_UNROLL_N 4)
     set(SYMV_P 16)
+  elseif ("${TCORE}" STREQUAL "P5600")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L2_SIZE 1048576\n"
+      "#define DTB_SIZE 4096\n"
+      "#define DTB_DEFAULT_ENTRIES 64\n")
+    set(SGEMM_UNROLL_M 2)
+    set(SGEMM_UNROLL_N 2)
+    set(DGEMM_UNROLL_M 2)
+    set(DGEMM_UNROLL_N 2)
+    set(CGEMM_UNROLL_M 2)
+    set(CGEMM_UNROLL_N 2)
+    set(ZGEMM_UNROLL_M 2)
+    set(ZGEMM_UNROLL_N 2)
+    set(SYMV_P 16)
+  elseif ("${TCORE}" MATCHES "MIPS")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L2_SIZE 262144\n"
+      "#define DTB_SIZE 4096\n"
+      "#define DTB_DEFAULT_ENTRIES 64\n")
+    set(SGEMM_UNROLL_M 2)
+    set(SGEMM_UNROLL_N 2)
+    set(DGEMM_UNROLL_M 2)
+    set(DGEMM_UNROLL_N 2)
+    set(CGEMM_UNROLL_M 2)
+    set(CGEMM_UNROLL_N 2)
+    set(ZGEMM_UNROLL_M 2)
+    set(ZGEMM_UNROLL_N 2)
+    set(SYMV_P 16)
   elseif ("${TCORE}" STREQUAL "POWER6")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L1_DATA_SIZE 32768\n"

cmake/system_check.cmake

@@ -73,6 +73,8 @@ elseif (${CMAKE_CROSSCOMPILING})
     else ()
        set(X86 1)
     endif()
+   elseif (${TARGET} STREQUAL "P5600" OR ${TARGET} MATCHES "MIPS.*")
+       set(MIPS32 1)
    elseif (${TARGET} STREQUAL "ARMV7")
        set(ARM 1)
    else()
@@ -88,6 +90,10 @@ elseif(X86)
   set(ARCH "x86")
 elseif(PPC)
   set(ARCH "power")
+elseif(MIPS32)
+  set(ARCH "mips")
+elseif(MIPS64)
+  set(ARCH "mips64")
 elseif(ARM)
   set(ARCH "arm")
 elseif(ARM64)

kernel/mips/KERNEL.generic

@@ -0,0 +1,160 @@
+SGEMM_BETA = ../generic/gemm_beta.c
+DGEMM_BETA = ../generic/gemm_beta.c
+CGEMM_BETA = ../generic/zgemm_beta.c
+ZGEMM_BETA = ../generic/zgemm_beta.c
+
+STRMMKERNEL	= ../generic/trmmkernel_2x2.c
+DTRMMKERNEL	= ../generic/trmmkernel_2x2.c
+CTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
+ZTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
+
+SGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
+SGEMMONCOPY    =  ../generic/gemm_ncopy_2.c
+SGEMMOTCOPY    =  ../generic/gemm_tcopy_2.c
+SGEMMONCOPYOBJ =  sgemm_oncopy.o
+SGEMMOTCOPYOBJ =  sgemm_otcopy.o
+
+DGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
+DGEMMONCOPY    = ../generic/gemm_ncopy_2.c
+DGEMMOTCOPY    = ../generic/gemm_tcopy_2.c
+DGEMMONCOPYOBJ = dgemm_oncopy.o
+DGEMMOTCOPYOBJ = dgemm_otcopy.o
+
+CGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
+CGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
+CGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
+CGEMMONCOPYOBJ =  cgemm_oncopy.o
+CGEMMOTCOPYOBJ =  cgemm_otcopy.o
+
+ZGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
+ZGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
+ZGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
+ZGEMMONCOPYOBJ =  zgemm_oncopy.o
+ZGEMMOTCOPYOBJ =  zgemm_otcopy.o
+
+STRSMKERNEL_LN	=  ../generic/trsm_kernel_LN.c
+STRSMKERNEL_LT	=  ../generic/trsm_kernel_LT.c
+STRSMKERNEL_RN	=  ../generic/trsm_kernel_RN.c
+STRSMKERNEL_RT	=  ../generic/trsm_kernel_RT.c
+
+DTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+DTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+DTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+CTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+CTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+CTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+CTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+ZTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+ZTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+ZTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+#Pure C for other kernels
+SAMAXKERNEL  = ../mips/amax.c
+DAMAXKERNEL  = ../mips/amax.c
+CAMAXKERNEL  = ../mips/zamax.c
+ZAMAXKERNEL  = ../mips/zamax.c
+
+SAMINKERNEL  = ../mips/amin.c
+DAMINKERNEL  = ../mips/amin.c
+CAMINKERNEL  = ../mips/zamin.c
+ZAMINKERNEL  = ../mips/zamin.c
+
+SMAXKERNEL   = ../mips/max.c
+DMAXKERNEL   = ../mips/max.c
+
+SMINKERNEL   = ../mips/min.c
+DMINKERNEL   = ../mips/min.c
+
+ISAMAXKERNEL = ../mips/iamax.c
+IDAMAXKERNEL = ../mips/iamax.c
+ICAMAXKERNEL = ../mips/izamax.c
+IZAMAXKERNEL = ../mips/izamax.c
+
+ISAMINKERNEL = ../mips/iamin.c
+IDAMINKERNEL = ../mips/iamin.c
+ICAMINKERNEL = ../mips/izamin.c
+IZAMINKERNEL = ../mips/izamin.c
+
+ISMAXKERNEL  = ../mips/imax.c
+IDMAXKERNEL  = ../mips/imax.c
+
+ISMINKERNEL  = ../mips/imin.c
+IDMINKERNEL  = ../mips/imin.c
+
+SASUMKERNEL  = ../mips/asum.c
+DASUMKERNEL  = ../mips/asum.c
+CASUMKERNEL  = ../mips/zasum.c
+ZASUMKERNEL  = ../mips/zasum.c
+
+SSUMKERNEL  = ../mips/sum.c
+DSUMKERNEL  = ../mips/sum.c
+CSUMKERNEL  = ../mips/zsum.c
+ZSUMKERNEL  = ../mips/zsum.c
+
+SAXPYKERNEL  = ../mips/axpy.c
+DAXPYKERNEL  = ../mips/axpy.c
+CAXPYKERNEL  = ../mips/zaxpy.c
+ZAXPYKERNEL  = ../mips/zaxpy.c
+
+SCOPYKERNEL  = ../mips/copy.c
+DCOPYKERNEL  = ../mips/copy.c
+CCOPYKERNEL  = ../mips/zcopy.c
+ZCOPYKERNEL  = ../mips/zcopy.c
+
+SDOTKERNEL   = ../mips/dot.c
+DDOTKERNEL   = ../mips/dot.c
+CDOTKERNEL   = ../mips/zdot.c
+ZDOTKERNEL   = ../mips/zdot.c
+
+SNRM2KERNEL  = ../mips/nrm2.c
+DNRM2KERNEL  = ../mips/nrm2.c
+CNRM2KERNEL  = ../mips/znrm2.c
+ZNRM2KERNEL  = ../mips/znrm2.c
+
+SROTKERNEL   = ../mips/rot.c
+DROTKERNEL   = ../mips/rot.c
+CROTKERNEL   = ../mips/zrot.c
+ZROTKERNEL   = ../mips/zrot.c
+
+SSCALKERNEL  = ../mips/scal.c
+DSCALKERNEL  = ../mips/scal.c
+CSCALKERNEL  = ../mips/zscal.c
+ZSCALKERNEL  = ../mips/zscal.c
+
+SSWAPKERNEL  = ../mips/swap.c
+DSWAPKERNEL  = ../mips/swap.c
+CSWAPKERNEL  = ../mips/zswap.c
+ZSWAPKERNEL  = ../mips/zswap.c
+
+SGEMVNKERNEL = ../mips/gemv_n.c
+DGEMVNKERNEL = ../mips/gemv_n.c
+CGEMVNKERNEL = ../mips/zgemv_n.c
+ZGEMVNKERNEL = ../mips/zgemv_n.c
+
+SGEMVTKERNEL = ../mips/gemv_t.c
+DGEMVTKERNEL = ../mips/gemv_t.c
+CGEMVTKERNEL = ../mips/zgemv_t.c
+ZGEMVTKERNEL = ../mips/zgemv_t.c
+
+SSYMV_U_KERNEL =  ../generic/symv_k.c
+SSYMV_L_KERNEL =  ../generic/symv_k.c
+DSYMV_U_KERNEL =  ../generic/symv_k.c
+DSYMV_L_KERNEL =  ../generic/symv_k.c
+QSYMV_U_KERNEL =  ../generic/symv_k.c
+QSYMV_L_KERNEL =  ../generic/symv_k.c
+CSYMV_U_KERNEL =  ../generic/zsymv_k.c
+CSYMV_L_KERNEL =  ../generic/zsymv_k.c
+ZSYMV_U_KERNEL =  ../generic/zsymv_k.c
+ZSYMV_L_KERNEL =  ../generic/zsymv_k.c
+XSYMV_U_KERNEL =  ../generic/zsymv_k.c
+XSYMV_L_KERNEL =  ../generic/zsymv_k.c
+
+ZHEMV_U_KERNEL =  ../generic/zhemv_k.c
+ZHEMV_L_KERNEL =  ../generic/zhemv_k.c
+
+CGEMM3MKERNEL    = ../generic/zgemm3mkernel_dump.c
+ZGEMM3MKERNEL    = ../generic/zgemm3mkernel_dump.c

param.h

@@ -3602,6 +3602,20 @@ is a big desktop or server with abundant cache rather than a phone or embedded d
 #define XGEMM_DEFAULT_UNROLL_M 1
 #endif
 
+#ifdef ARCH_MIPS
+#define SGEMM_DEFAULT_P  128
+#define DGEMM_DEFAULT_P  128
+#define CGEMM_DEFAULT_P  96
+#define ZGEMM_DEFAULT_P  64
+#define SGEMM_DEFAULT_Q  240
+#define DGEMM_DEFAULT_Q  120
+#define CGEMM_DEFAULT_Q  120
+#define ZGEMM_DEFAULT_Q  120
+#define SGEMM_DEFAULT_R  12288
+#define DGEMM_DEFAULT_R  8192
+#define CGEMM_DEFAULT_R  4096
+#define ZGEMM_DEFAULT_R  4096
+#else
 #define SGEMM_DEFAULT_P sgemm_p
 #define DGEMM_DEFAULT_P dgemm_p
 #define QGEMM_DEFAULT_P qgemm_p
@@ -3622,6 +3636,7 @@ is a big desktop or server with abundant cache rather than a phone or embedded d
 #define CGEMM_DEFAULT_Q 128
 #define ZGEMM_DEFAULT_Q 128
 #define XGEMM_DEFAULT_Q 128
+#endif
 
 #define SYMV_P	16