| @@ -59,7 +59,8 @@ ifeq ($(CORE), Z14) | |||
| USE_TRMM = 1 | |||
| endif | |||
| #ifndef SHGEMMKERNEL | |||
| ifeq ($(BUILD_HALF), 1) | |||
| ifndef SHGEMMKERNEL | |||
| SHGEMM_BETA = ../generic/gemm_beta.c | |||
| SHGEMMKERNEL = ../generic/gemmkernel_2x2.c | |||
| SHGEMMINCOPY = ../generic/gemm_ncopy_2.c | |||
| @@ -70,12 +71,13 @@ SHGEMMINCOPYOBJ = shgemm_incopy$(TSUFFIX).$(SUFFIX) | |||
| SHGEMMITCOPYOBJ = shgemm_itcopy$(TSUFFIX).$(SUFFIX) | |||
| SHGEMMONCOPYOBJ = shgemm_oncopy$(TSUFFIX).$(SUFFIX) | |||
| SHGEMMOTCOPYOBJ = shgemm_otcopy$(TSUFFIX).$(SUFFIX) | |||
| #endif | |||
| endif | |||
| SHKERNELOBJS += \ | |||
| shgemm_kernel$(TSUFFIX).$(SUFFIX) \ | |||
| $(SHGEMMINCOPYOBJ) $(SHGEMMITCOPYOBJ) \ | |||
| $(SHGEMMONCOPYOBJ) $(SHGEMMOTCOPYOBJ) | |||
| endif | |||
| SKERNELOBJS += \ | |||
| sgemm_kernel$(TSUFFIX).$(SUFFIX) \ | |||
| @@ -110,7 +112,9 @@ XKERNELOBJS += \ | |||
| $(XGEMMINCOPYOBJ) $(XGEMMITCOPYOBJ) \ | |||
| $(XGEMMONCOPYOBJ) $(XGEMMOTCOPYOBJ) | |||
| ifeq ($(BUILD_HALF),1) | |||
| SHBLASOBJS += $(SHKERNELOBJS) | |||
| endif | |||
| SBLASOBJS += $(SKERNELOBJS) | |||
| DBLASOBJS += $(DKERNELOBJS) | |||
| QBLASOBJS += $(QKERNELOBJS) | |||
| @@ -118,7 +122,10 @@ CBLASOBJS += $(CKERNELOBJS) | |||
| ZBLASOBJS += $(ZKERNELOBJS) | |||
| XBLASOBJS += $(XKERNELOBJS) | |||
| ifeq ($(BUILD_HALF),1) | |||
| SHBLASOBJS += shgemm_beta$(TSUFFIX).$(SUFFIX) | |||
| endif | |||
| SBLASOBJS += \ | |||
| sgemm_beta$(TSUFFIX).$(SUFFIX) \ | |||
| strmm_kernel_LN$(TSUFFIX).$(SUFFIX) strmm_kernel_LT$(TSUFFIX).$(SUFFIX) \ | |||
| @@ -408,11 +415,13 @@ ZBLASOBJS += \ | |||
| zimatcopy_k_ctc$(TSUFFIX).$(SUFFIX) zimatcopy_k_rtc$(TSUFFIX).$(SUFFIX) \ | |||
| zgeadd_k$(TSUFFIX).$(SUFFIX) | |||
| ifeq ($(BUILD_HALF), 1) | |||
| SHGEMMINCOPYOBJ_P = $(SHGEMMINCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX)) | |||
| SHGEMMITCOPYOBJ_P = $(SHGEMMITCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX)) | |||
| SHGEMMONCOPYOBJ_P = $(SHGEMMONCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX)) | |||
| SHGEMMOTCOPYOBJ_P = $(SHGEMMOTCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX)) | |||
| endif | |||
| SGEMMINCOPYOBJ_P = $(SGEMMINCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX)) | |||
| SGEMMITCOPYOBJ_P = $(SGEMMITCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX)) | |||
| SGEMMONCOPYOBJ_P = $(SGEMMONCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX)) | |||
| @@ -438,8 +447,10 @@ XGEMMITCOPYOBJ_P = $(XGEMMITCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX)) | |||
| XGEMMONCOPYOBJ_P = $(XGEMMONCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX)) | |||
| XGEMMOTCOPYOBJ_P = $(XGEMMOTCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX)) | |||
| ifeq ($(BUILD_HALF),1) | |||
| $(KDIR)shgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SHGEMM_BETA) | |||
| $(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@ | |||
| endif | |||
| $(KDIR)sgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMM_BETA) | |||
| $(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@ | |||
| @@ -459,10 +470,14 @@ $(KDIR)zgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMM_BETA) | |||
| $(KDIR)xgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(XGEMM_BETA) | |||
| $(CC) $(CFLAGS) -c -DXDOUBLE -DCOMPLEX $< -o $@ | |||
| ifeq ($(BUILD_HALF), 1) | |||
| $(KDIR)$(SHGEMMONCOPYOBJ) : $(KERNELDIR)/$(SHGEMMONCOPY) | |||
| $(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@ | |||
| $(KDIR)$(SHGEMMOTCOPYOBJ) : $(KERNELDIR)/$(SHGEMMOTCOPY) | |||
| ifeq ($(OS), AIX) | |||
| $(CC) $(CFLAGS) -E -DHALF -UDOUBLE -UCOMPLEX $< -o shgemmotcopy.s | |||
| m4 shgemmotcopy.s > shgemmotcopy_nomacros.s | |||
| @@ -487,6 +502,7 @@ else | |||
| $(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@ | |||
| endif | |||
| endif | |||
| endif | |||
| $(KDIR)$(SGEMMONCOPYOBJ) : $(KERNELDIR)/$(SGEMMONCOPY) | |||
| @@ -646,6 +662,8 @@ else | |||
| $(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@ | |||
| endif | |||
| ifeq ($(BUILD_HALF), 1) | |||
| $(KDIR)shgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SHGEMMKERNEL) $(SHGEMMDEPEND) | |||
| ifeq ($(OS), AIX) | |||
| $(CC) $(CFLAGS) -E -DHALF -UDOUBLE -UCOMPLEX $< -o shgemm_kernel$(TSUFFIX).s | |||
| @@ -655,6 +673,7 @@ ifeq ($(OS), AIX) | |||
| else | |||
| $(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@ | |||
| endif | |||
| endif | |||
| $(KDIR)dgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMMKERNEL) $(DGEMMDEPEND) | |||
| ifeq ($(OS), AIX) | |||
| @@ -2272,8 +2291,10 @@ $(KDIR)xtrsm_oltncopy$(TSUFFIX).$(SUFFIX) : generic/ztrsm_ltcopy_$(XGEMM_UNROLL_ | |||
| $(KDIR)sgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SGEMM_BETA) | |||
| $(CC) $(PFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@ | |||
| ifeq ($(BUILD_HALF),1) | |||
| $(KDIR)shgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SHGEMM_BETA) | |||
| $(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@ | |||
| endif | |||
| $(KDIR)dgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(DGEMM_BETA) | |||
| $(CC) $(PFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@ | |||
| @@ -2290,6 +2311,8 @@ $(KDIR)zgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(ZGEMM_BETA) | |||
| $(KDIR)xgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(XGEMM_BETA) | |||
| $(CC) $(PFLAGS) -c -DXDOUBLE -DCOMPLEX $< -o $@ | |||
| ifeq ($(BUILD_HALF), 1) | |||
| $(SHGEMMONCOPYOBJ_P) : $(KERNELDIR)/$(SHGEMMONCOPY) | |||
| $(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@ | |||
| @@ -2304,6 +2327,8 @@ $(SHGEMMITCOPYOBJ_P) : $(KERNELDIR)/$(SHGEMMITCOPY) | |||
| $(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@ | |||
| endif | |||
| endif | |||
| $(SGEMMONCOPYOBJ_P) : $(KERNELDIR)/$(SGEMMONCOPY) | |||
| $(CC) $(PFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@ | |||
| @@ -2408,8 +2433,11 @@ endif | |||
| endif | |||
| ifeq ($(BUILD_HALF), 1) | |||
| $(KDIR)shgemm_kernel$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SHGEMMKERNEL) $(SHGEMMDEPEND) | |||
| $(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@ | |||
| endif | |||
| $(KDIR)sgemm_kernel$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL) $(SGEMMDEPEND) | |||
| $(CC) $(PFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@ | |||