gemmtc-1.3a/0000755000175000017500000000000013667424015013270 5ustar mukunokimukunokigemmtc-1.3a/include/0000755000175000017500000000000013647250516014714 5ustar mukunokimukunokigemmtc-1.3a/include/cuozblas.h0000644000175000017500000000414013647247617016716 0ustar mukunokimukunoki#ifndef CUOZBLAS_H #define CUOZBLAS_H #include typedef struct { cublasHandle_t ch; char* devWorkCommon; // for except split data, 128bytes char* devWork; // for split data size_t workSizeBytes; size_t memAddr; uint32_t numSplitArraysMax; uint32_t fastMode; uint32_t sumMode; uint32_t OFMode; uint32_t CRMode; // exec info float numSplitArraysA; float numSplitArraysB; float numSplitArraysC; float t_SplitA; float t_SplitB; float t_comp; float t_sum; float t_total; float n_comp; uint32_t mbk; uint32_t nbk; } cuozblasHandle_t; // helper routines extern void cuozblasCreate (cuozblasHandle_t*, size_t, uint32_t, uint32_t, uint32_t); extern void cuozblasSetNumSplitArraysMax (cuozblasHandle_t*, uint32_t); extern void cuozblasSetCRMode (cuozblasHandle_t*, uint32_t); extern void cuozblasSetOFMode (cuozblasHandle_t*, uint32_t); extern uint32_t cuozblasGetNumSplitArraysMax (cuozblasHandle_t*); extern uint32_t cuozblasGetCRMode (cuozblasHandle_t*); extern uint32_t cuozblasGetOFMode (cuozblasHandle_t*); extern void cuozblasDestroy (cuozblasHandle_t*); // BLAS routines (GEMM-TC) extern int32_t cuozblasSgemmTC (cuozblasHandle_t*, const cublasOperation_t, const cublasOperation_t, const size_t, const size_t, const size_t, const float*, const float*, const size_t, const float*, const size_t, const float*, float*, const size_t); extern int32_t cuozblasDgemmTC (cuozblasHandle_t*, const cublasOperation_t, const cublasOperation_t, const size_t, const size_t, const size_t, const double*, const double*, const size_t, const double*, const size_t, const double*, double*, const size_t); extern int32_t cuozblasDgemmSP (cuozblasHandle_t*, const cublasOperation_t, const cublasOperation_t, const size_t, const size_t, const size_t, const double*, const double*, const size_t, const double*, const size_t, const double*, double*, const size_t); extern int32_t cuozblasDgemmDP (cuozblasHandle_t*, const cublasOperation_t, const cublasOperation_t, const size_t, const size_t, const size_t, const double*, const double*, const size_t, const double*, const size_t, const double*, double*, const size_t); #endif gemmtc-1.3a/lib/0000755000175000017500000000000013667423717014046 5ustar mukunokimukunokigemmtc-1.3a/src/0000755000175000017500000000000013647243141014054 5ustar mukunokimukunokigemmtc-1.3a/src/cuda_tc/0000755000175000017500000000000013667423717015471 5ustar mukunokimukunokigemmtc-1.3a/src/cuda_tc/cuozblas_sum.cu0000644000175000017500000002231713634642404020523 0ustar mukunokimukunoki#include "cuozblas_common.h" /* // 2^(ceil(log2|p|)) __device__ __forceinline__ double NextPowTwo (double p) { return (__dsub_rn(__dmul_rn(53, p), __dmul_rn(52, p))); } */ // ========================================= // NearSum (and AccSum) // from "ACCURATE FLOATING-POINT SUMMATION" by S.M.RUMP, T.OGITA, S.OISHI (2005) // http://www.ti3.tu-harburg.de/paper/rump/RuOgOi06.pdf // ========================================= __device__ __forceinline__ void cuTransform (const size_t n, FP_TYPE3 *vec, const size_t ld, double rho, double &tau1, double &tau2) { size_t i, m; double tmp, mu, sigma, t, tau; mu = fabs(vec[0]); for (i = 1; i < n; i++) mu = max (mu, fabs(vec[i*ld])); if ((n == 0) || (mu == 0.)) { tau1 = rho; tau2 = 0.; return; } m = ceil (log2((double)(n+2))); sigma = scalbn (1., m+ceil(log2(mu))); t = rho; double t2 = scalbn (1., -1022); while (1) { // ExtractVector tau = 0.; for (i = 0; i < n; i++) { tmp = (sigma + vec[i*ld]) - sigma; vec[i*ld] -= tmp; // <- output tau += tmp; } // here, tau1 = t1 tau1 = t + tau; if ((sigma <= t2) || (fabs(tau1) >= scalbn (1., 2*m+1-53)*sigma)) { //FastTwoSum (t, tau, &tau1, &tau2); //tau1 = t + tau tau2 = tau - (tau1 - t); return; } sigma = scalbn (1., m-53) * sigma; t = tau1; } } __device__ __forceinline__ void cuTransformK (const size_t n, FP_TYPE3 *vec, const size_t ld, double rho, double &res, double &r) { size_t i; double tmp, tau1, tau2; cuTransform (n, vec, ld, rho, tau1, tau2); tmp = 0.; for (i = 0; i < n; i++) tmp += vec[i*ld]; res = tau1 + (tau2 + tmp); r = tau2 - (res - tau1); } __device__ __forceinline__ double cuSign (double v) { return (v < 0) ? -1.:1.; } __device__ double cuNearSum (const size_t n, FP_TYPE3 *vec, const size_t ld) { double tmp, res, res2, r, r2, mu, delta, delta2; double eps = scalbn (1., -53); cuTransformK (n, vec, ld, 0, res, r); cuTransformK (n, vec, ld, r, delta, r2); if (delta == 0) return res; res2 = res + cuSign (delta) * eps * fabs(res); if (res2 == res) { mu = cuSign (delta) * eps * fabs(res); res2 = res + 2. * cuSign (delta) * eps * fabs(res); } else { mu = (res2 - res) / 2.; } if (fabs(delta) < fabs(mu)) return res; if (fabs(delta) > fabs(mu)) return res2; cuTransformK (n, vec, ld, r2, delta2, tmp); if (delta2 == 0) return res + mu; if (cuSign (delta2) == cuSign (mu)) return res2; return res; } // ========================================= // NearSumMat // ========================================= __global__ void cuNearSumMatPhstKernel ( const size_t m, const size_t n, const short *devASpExp, const size_t ldas, const uint32_t numsplitA, const short *devBSpExp, const size_t ldbs, const uint32_t numsplitB, FP_TYPE3 *devCsplit, const size_t llsc, const size_t ldsc, FP_TYPE1 *devC, const size_t ldc, const FP_TYPE1 alpha, const FP_TYPE1 beta, const uint32_t maxlevel, uint32_t *check ) { const size_t iBx = blockIdx.x; const size_t iBy = blockIdx.y; const size_t nTx = blockDim.x; const size_t nTy = blockDim.y; const size_t iTx = threadIdx.x; const size_t iTy = threadIdx.y; const size_t addrx = iBx * nTx + iTx; const size_t addry = iBy * nTy + iTy; double t; if (addrx < m && addry < n) { uint32_t ia, ib, ic = 0; for (ia = 0; ia < numsplitA; ia++) { short seA = devASpExp[ldas*ia+addrx]; for (ib = 0; ib < numsplitB; ib++) { if (ia + ib <= maxlevel) { short seB = devBSpExp[ldbs*ib+addry]; FP_TYPE3 c = devCsplit[llsc * ic + addry * ldsc + addrx]; c = scalbn((FP_TYPE1)c, seA+seB); devCsplit[llsc * ic + addry * ldsc + addrx] = c; ic++; } } } t = cuNearSum (ic, &devCsplit[addry * ldsc + addrx], llsc); if (isnan (t) || isinf (t)) atomicAdd (check, 1); devC[addry * ldc + addrx] = alpha * t + ((beta == 0.) ? 0. : beta*devC[addry * ldc + addrx]); } } __global__ void cuNearSumMatPdevKernel ( const size_t m, const size_t n, const short *devASpExp, const size_t ldas, const uint32_t numsplitA, const short *devBSpExp, const size_t ldbs, const uint32_t numsplitB, FP_TYPE3 *devCsplit, const size_t llsc, const size_t ldsc, FP_TYPE1 *devC, const size_t ldc, const FP_TYPE1 *alpha, const FP_TYPE1 *beta, const uint32_t maxlevel, uint32_t *check ) { const size_t iBx = blockIdx.x; const size_t iBy = blockIdx.y; const size_t nTx = blockDim.x; const size_t nTy = blockDim.y; const size_t iTx = threadIdx.x; const size_t iTy = threadIdx.y; const size_t addrx = iBx * nTx + iTx; const size_t addry = iBy * nTy + iTy; double t; if (addrx < m && addry < n) { uint32_t ia, ib, ic = 0; for (ia = 0; ia < numsplitA; ia++) { short seA = devASpExp[ldas*ia+addrx]; for (ib = 0; ib < numsplitB; ib++) { if (ia + ib <= maxlevel) { short seB = devBSpExp[ldbs*ib+addry]; FP_TYPE3 c = devCsplit[llsc * ic + addry * ldsc + addrx]; c = scalbn((FP_TYPE1)c, seA+seB); devCsplit[llsc * ic + addry * ldsc + addrx] = c; ic++; } } } t = cuNearSum (ic, &devCsplit[addry * ldsc + addrx], llsc); if (isnan (t) || isinf (t)) atomicAdd (check, 1); devC[addry * ldc + addrx] = alpha[0] * t + ((beta[0] == 0.) ? 0. : beta[0]*devC[addry * ldc + addrx]); } } __global__ void cuSumMatPhstKernel ( const size_t m, const size_t n, const short *devASpExp, const size_t ldas, const uint32_t numsplitA, const short *devBSpExp, const size_t ldbs, const uint32_t numsplitB, const FP_TYPE3 * __restrict__ devCsplit, const size_t llsc, const size_t ldsc, FP_TYPE1 *devC, const size_t ldc, const FP_TYPE1 alpha, const FP_TYPE1 beta, const uint32_t maxlevel, uint32_t *check ) { const size_t iBx = blockIdx.x; const size_t iBy = blockIdx.y; const size_t nTx = blockDim.x; const size_t nTy = blockDim.y; const size_t iTx = threadIdx.x; const size_t iTy = threadIdx.y; const size_t addrx = iBx * nTx + iTx; const size_t addry = iBy * nTy + iTy; FP_TYPE1 t = 0.; if (addrx < m && addry < n) { uint32_t ia, ib, ic = 0; for (ia = 0; ia < numsplitA; ia++) { short seA = devASpExp[ldas*ia+addrx]; for (ib = 0; ib < numsplitB; ib++) { if (ia + ib <= maxlevel) { short seB = devBSpExp[ldbs*ib+addry]; FP_TYPE3 c = devCsplit[llsc * ic + addry * ldsc + addrx]; t += scalbn((FP_TYPE1)c, seA+seB); ic++; } } } if (isnan (t) || isinf (t)) atomicAdd (check, 1); devC[addry * ldc + addrx] = alpha * t + ((beta == 0.) ? 0. : beta*devC[addry * ldc + addrx]); } } __global__ void cuSumMatPdevKernel ( const size_t m, const size_t n, const short *devASpExp, const size_t ldas, const uint32_t numsplitA, const short *devBSpExp, const size_t ldbs, const uint32_t numsplitB, const FP_TYPE3 * __restrict__ devCsplit, const size_t llsc, const size_t ldsc, FP_TYPE1 *devC, const size_t ldc, const FP_TYPE1 *alpha, const FP_TYPE1 *beta, const uint32_t maxlevel, uint32_t *check ) { const size_t iBx = blockIdx.x; const size_t iBy = blockIdx.y; const size_t nTx = blockDim.x; const size_t nTy = blockDim.y; const size_t iTx = threadIdx.x; const size_t iTy = threadIdx.y; const size_t addrx = iBx * nTx + iTx; const size_t addry = iBy * nTy + iTy; FP_TYPE1 t = 0.; if (addrx < m && addry < n) { uint32_t ia, ib, ic = 0; for (ia = 0; ia < numsplitA; ia++) { short seA = devASpExp[ldas*ia+addrx]; for (ib = 0; ib < numsplitB; ib++) { if (ia + ib <= maxlevel) { short seB = devBSpExp[ldbs*ib+addry]; FP_TYPE3 c = devCsplit[llsc * ic + addry * ldsc + addrx]; t += scalbn((FP_TYPE1)c, seA+seB); ic++; } } } if (isnan (t) || isinf (t)) atomicAdd (check, 1); devC[addry * ldc + addrx] = alpha[0] * t + ((beta[0] == 0.) ? 0. : beta[0]*devC[addry * ldc + addrx]); } } __host__ int32_t cuSumMatDevice ( const size_t m, const size_t n, const short *devASpExp, const size_t ldas, const uint32_t numsplitA, const short *devBSpExp, const size_t ldbs, const uint32_t numsplitB, FP_TYPE3 *devCsplit, const size_t llsc, const size_t ldsc, FP_TYPE1 *devC, const size_t ldc, const FP_TYPE1 *alpha, const FP_TYPE1 *beta, cublasPointerMode_t pmode, const uint32_t maxlevel, const uint32_t sumMode, uint32_t *checkd ) { uint32_t izero = 0; int32_t check = 0; cudaMemcpy (checkd, &izero, sizeof(uint32_t), cudaMemcpyHostToDevice); dim3 threads, grid; int32_t ntx, nty, nbx, nby; ntx = SUM_NTX; nty = SUM_NTY; nbx = ceil (float(m) / ntx); nby = ceil (float(n) / nty); threads = dim3 (ntx, nty); grid = dim3 (nbx, nby); if (pmode == CUBLAS_POINTER_MODE_DEVICE) { if (sumMode) cuNearSumMatPdevKernel <<< grid, threads >>> (m, n, devASpExp, ldas, numsplitA, devBSpExp, ldbs, numsplitB, devCsplit, llsc, ldsc, devC, ldc, alpha, beta, maxlevel, checkd); else cuSumMatPdevKernel <<< grid, threads >>> (m, n, devASpExp, ldas, numsplitA, devBSpExp, ldbs, numsplitB, devCsplit, llsc, ldsc, devC, ldc, alpha, beta, maxlevel, checkd); } else { if (sumMode) cuNearSumMatPhstKernel <<< grid, threads >>> (m, n, devASpExp, ldas, numsplitA, devBSpExp, ldbs, numsplitB, devCsplit, llsc, ldsc, devC, ldc, *alpha, *beta, maxlevel, checkd); else cuSumMatPhstKernel <<< grid, threads >>> (m, n, devASpExp, ldas, numsplitA, devBSpExp, ldbs, numsplitB, devCsplit, llsc, ldsc, devC, ldc, *alpha, *beta, maxlevel, checkd); } cudaMemcpy (&check, checkd, sizeof(uint32_t), cudaMemcpyDeviceToHost); return check; } gemmtc-1.3a/src/cuda_tc/Makefile0000644000175000017500000000246713634412226017125 0ustar mukunokimukunoki# Makefile include ../../make.inc NVCC = $(CUDA_PATH)/bin/nvcc #-G -g AR = ar LIB := ../../lib/libcuozblastc.a FLAGS := -O3 -fPIC NVCCFLAGS := -Xcompiler -fPIC -O3 all: blas ifeq ($(CC60),yes) NVCCFLAGS += -gencode arch=compute_60,code=sm_60 endif ifeq ($(CC61),yes) NVCCFLAGS += -gencode arch=compute_61,code=sm_61 endif ifeq ($(CC70),yes) NVCCFLAGS += -gencode arch=compute_60,code=sm_70 endif ifeq ($(CC75),yes) NVCCFLAGS += -gencode arch=compute_60,code=sm_75 endif ALLSRC += \ cuozblas_split.cu \ cuozblas_sum.cu \ cuozblas_gemm.cu \ cuozblas_aux_tc.cu ALLHDR = \ cuozblas_common.h \ cuozblas_internal.h ALLOBJ = cuozblas.o ALLOBJ += $(ALLSRC:.cu=.st.o) ALLOBJ += $(ALLSRC:.cu=.dt.o) ALLOBJ += $(ALLSRC:.cu=.ds.o) ALLOBJ += $(ALLSRC:.cu=.dd.o) ifeq ($(QGEMMDP),yes) #ALLOBJ += $(ALLSRC:.cu=.qd.o) endif $(ALLOBJ) : $(ALLHDR) blas: $(ALLOBJ) $(AR) rv $(LIB) $(ALLOBJ) %.o: %.cu $(NVCC) $(NVCCFLAGS) -o $@ -c $< %.st.o: %.cu $(NVCC) $(NVCCFLAGS) -o $@ -c $< -DSGEMMTC %.dt.o: %.cu $(NVCC) $(NVCCFLAGS) -o $@ -c $< -DDGEMMTC %.ds.o: %.cu $(NVCC) $(NVCCFLAGS) -o $@ -c $< -DDGEMMSP %.dd.o: %.cu $(NVCC) $(NVCCFLAGS) -o $@ -c $< -DDGEMMDP %.qd.o: %.cu $(NVCC) $(NVCCFLAGS) -o $@ -c $< -DQGEMMDP cuozblas.o: cuozblas.cpp $(CXX) $(FLAGS) -o $@ -c $< -I$(CUDA_PATH)/include clean : $(RM) *.o $(RM) $(LIB) gemmtc-1.3a/src/cuda_tc/cuozblas_aux_tc.cu0000644000175000017500000000400313634610566021176 0ustar mukunokimukunoki#include "cuozblas_common.h" __host__ void cuozblasMatAddrAlloc ( cuozblasHandle_t *oh, const size_t m, const size_t n, const size_t size, void **dev, size_t &ld ) { ld = getPitchSize (m); dev[0] = oh->devWork + oh->memAddr; oh->memAddr += size * ld * n; } __host__ void cuozblasVecAddrAlloc ( cuozblasHandle_t *oh, const size_t n, const size_t size, void **dev ) { size_t ld = getPitchSize (n); dev[0] = oh->devWork + oh->memAddr; oh->memAddr += size * ld; } // for SplitMatA __global__ void cuCompVecAllKernel ( const size_t n, const FP_TYPE1 *devA, const FP_TYPE1 *devB, int32_t *val ) { const size_t iBx = blockIdx.x; const size_t nTx = blockDim.x; const size_t iTx = threadIdx.x; const size_t addrx = iBx * nTx + iTx; if (addrx < n) { if (devA[addrx] > devB[addrx]) atomicAdd (val, 1.); } } __host__ void cuCompVecAll ( const size_t n, const FP_TYPE1 *devA, const FP_TYPE1 *devB, int32_t *val ) { dim3 threads, grid; int32_t ntx, nbx; ntx = VEC_NTX; nbx = ceil (float(n) / ntx); threads = dim3 (ntx); grid = dim3 (nbx); cuCompVecAllKernel <<< grid, threads >>> (n, devA, devB, val); } __global__ void cuAbsMatKernel ( const size_t m, const size_t n, const FP_TYPE1 *devA, const size_t lda, FP_TYPE1 *devB, const size_t ldb ) { const size_t iBx = blockIdx.x; const size_t iBy = blockIdx.y; const size_t nTx = blockDim.x; const size_t nTy = blockDim.y; const size_t iTx = threadIdx.x; const size_t iTy = threadIdx.y; const size_t addrx = iBx * nTx + iTx; const size_t addry = iBy * nTy + iTy; if (addrx < m && addry < n) devB[addry*ldb+addrx] = fabs(devA[addry*lda+addrx]); } __host__ void cuAbsMat ( const size_t m, const size_t n, const FP_TYPE1 *devA, const size_t lda, FP_TYPE1 *devB, const size_t ldb ) { dim3 threads, grid; int32_t ntx, nty, nbx, nby; ntx = MAT_NTX; nty = MAT_NTY; nbx = ceil (float(m) / ntx); nby = ceil (float(n) / nty); threads = dim3 (ntx, nty); grid = dim3 (nbx, nby); cuAbsMatKernel <<< grid, threads >>> (m, n, devA, lda, devB, ldb); } gemmtc-1.3a/src/cuda_tc/cuozblas_common.h0000644000175000017500000000215613634611462021026 0ustar mukunokimukunoki// setting #define Profiling 1 #define DefaultWorkSize 8e9 // 8GB #define NumSplitDefaultMax 18 // num of threads #define SPLIT_N_NTX 32 #define SPLIT_N_NTY 16 #define SPLIT_T_NTX 512 #define SPLIT_T_NTY 1 #define SPLIT_VEC_NTX 512 #define SPLIT_VEC_NBX 512 #define SUM_NTX 64 #define SUM_NTY 1 #define VEC_NTX 128 #define MAT_NTX 256 #define MAT_NTY 2 #include #include #define __STDC_FORMAT_MACROS #include #include #include #include #include #include #include #include "../../include/cuozblas.h" #include "cuozblas_internal.h" #if defined (SGEMMTC) #define FP_TYPE1 float #define FP_TYPE2 half #define FP_TYPE3 float #elif defined (DGEMMTC) #define FP_TYPE1 double #define FP_TYPE2 half #define FP_TYPE3 float #elif defined (DGEMMSP) #define FP_TYPE1 double #define FP_TYPE2 float #define FP_TYPE3 float #elif defined (DGEMMDP) #define FP_TYPE1 double #define FP_TYPE2 double #define FP_TYPE3 double #elif defined (QGEMMDP) #define FP_TYPE1 __float128 #define FP_TYPE2 double #define FP_TYPE3 double #endif gemmtc-1.3a/src/cuda_tc/cuozblas_internal.h0000644000175000017500000001453713634412372021357 0ustar mukunokimukunoki#ifndef CUOZBLAS_INTERNAL_H #define CUOZBLAS_INTERNAL_H double timer (); size_t getPitchSize (size_t n); void counterInit (cuozblasHandle_t *oh); uint32_t memCheck (cuozblasHandle_t *oh); void PrintMat ( const size_t m, const size_t n, const double *devC, const size_t ldd); void PrintMatHalf ( const size_t m, const size_t n, const half *devC, const size_t ldd); void PrintMatFloat ( const size_t m, const size_t n, const float *devC, const size_t ldd); void PrintMat ( const size_t m, const size_t n, const double *devC, const size_t ldd); void cuozblasMatAddrAlloc ( cuozblasHandle_t *oh, const size_t m, const size_t n, const size_t size, void **dev, size_t &lds); void cuozblasVecAddrAlloc ( cuozblasHandle_t *oh, const size_t m, const size_t size, void **dev); cublasOperation_t ToCublasOp (char tran); char FromCublasOp (cublasOperation_t tran); #if defined (SGEMMTC) void cuInitVec ( const size_t n, float *devA, float val); void cuAbsMat ( const size_t m, const size_t n, const float *devA, const size_t lda, float *devB, const size_t ldb); void cuCompVecAll ( const size_t n, const float *devA, const float *devB, int32_t *val); void cuSplitMat ( cuozblasHandle_t *oh, const char major, const size_t m, const size_t n, const float *devInput, const size_t ldi, float *devTmp, const size_t ldt, half *devSplit, const size_t lds, short *devSpExp, const size_t ldse, float *devMax, uint32_t &numSplitArrays); void cuSplitMatA ( cuozblasHandle_t *oh, const char major, const size_t m, const size_t n, const size_t k, const float *devAInput, const size_t ldai, const float *devBInput, const size_t ldbi, float *devAOutput, const size_t ldao, half *devASplit, const size_t ldas, short *devASpExp, const size_t ldase, float *devAMax, float *devAtmp, const size_t ldat, float *devBtmp, const size_t ldbt, float *devE, float *devBe, float *devB1, float *devB2, uint32_t &numSplitArrays); int32_t cuSumMatDevice ( const size_t m, const size_t n, const short *devASpExp, const size_t ldas, const uint32_t numsplitA, const short *devBSpExp, const size_t ldbs, const uint32_t numsplitB, float *devCsplit, const size_t llsc, const size_t ldsc, float *devC, const size_t ldc, const float*, const float*, const cublasPointerMode_t, const uint32_t maxlevel, const uint32_t, uint32_t*); #elif defined (DGEMMTC) void cuAbsMat ( const size_t m, const size_t n, const double *devA, const size_t lda, double *devB, const size_t ldb); void cuCompVecAll ( const size_t n, const double *devA, const double *devB, int32_t *val); void cuSplitMat ( cuozblasHandle_t *oh, const char major, const size_t m, const size_t n, const double *devInput, const size_t ldi, double *devTmp, const size_t ldt, half *devSplit, const size_t lds, short *devSpExp, const size_t ldse, double *devMax, uint32_t &numSplitArrays); void cuSplitMatA ( cuozblasHandle_t *oh, const char major, const size_t m, const size_t n, const size_t k, const double *devAInput, const size_t ldai, const double *devBInput, const size_t ldbi, double *devAOutput, const size_t ldao, half *devASplit, const size_t ldas, short *devASpExp, const size_t ldase, double *devAMax, double *devAtmp, const size_t ldat, double *devBtmp, const size_t ldbt, double *devE, double *devBe, double *devB1, double *devB2, uint32_t &numSplitArrays); int32_t cuSumMatDevice ( const size_t m, const size_t n, const short *devASpExp, const size_t ldas, const uint32_t numsplitA, const short *devBSpExp, const size_t ldbs, const uint32_t numsplitB, float *devCsplit, const size_t llsc, const size_t ldsc, double *devC, const size_t ldc, const double*, const double*, const cublasPointerMode_t, const uint32_t maxlevel, const uint32_t, uint32_t*); #elif defined (DGEMMSP) void cuAbsMat ( const size_t m, const size_t n, const double *devA, const size_t lda, double *devB, const size_t ldb); void cuCompVecAll ( const size_t n, const double *devA, const double *devB, int32_t *val); void cuSplitMat ( cuozblasHandle_t *oh, const char major, const size_t m, const size_t n, const double *devInput, const size_t ldi, double *devTmp, const size_t ldt, float *devSplit, const size_t lds, short *devSpExp, const size_t ldse, double *devMax, uint32_t &numSplitArrays); void cuSplitMatA ( cuozblasHandle_t *oh, const char major, const size_t m, const size_t n, const size_t k, const double *devAInput, const size_t ldai, const double *devBInput, const size_t ldbi, double *devAOutput, const size_t ldao, float *devASplit, const size_t ldas, short *devASpExp, const size_t ldase, double *devAMax, double *devAtmp, const size_t ldat, double *devBtmp, const size_t ldbt, double *devE, double *devBe, double *devB1, double *devB2, uint32_t &numSplitArrays); int32_t cuSumMatDevice ( const size_t m, const size_t n, const short *devASpExp, const size_t ldas, const uint32_t numsplitA, const short *devBSpExp, const size_t ldbs, const uint32_t numsplitB, float *devCsplit, const size_t llsc, const size_t ldsc, double *devC, const size_t ldc, const double*, const double*, const cublasPointerMode_t, const uint32_t maxlevel, const uint32_t, uint32_t*); #elif defined (DGEMMDP) void cuAbsMat ( const size_t m, const size_t n, const double *devA, const size_t lda, double *devB, const size_t ldb); void cuCompVecAll ( const size_t n, const double *devA, const double *devB, int32_t *val); void cuSplitMat ( cuozblasHandle_t *oh, const char major, const size_t m, const size_t n, const double *devInput, const size_t ldi, double *devTmp, const size_t ldt, double *devSplit, const size_t lds, short *devSpExp, const size_t ldse, double *devMax, uint32_t &numSplitArrays); void cuSplitMatA ( cuozblasHandle_t *oh, const char major, const size_t m, const size_t n, const size_t k, const double *devAInput, const size_t ldai, const double *devBInput, const size_t ldbi, double *devAOutput, const size_t ldao, double *devASplit, const size_t ldas, short *devASpExp, const size_t ldase, double *devAMax, double *devAtmp, const size_t ldat, double *devBtmp, const size_t ldbt, double *devE, double *devBe, double *devB1, double *devB2, uint32_t &numSplitArrays); int32_t cuSumMatDevice ( const size_t m, const size_t n, const short *devASpExp, const size_t ldas, const uint32_t numsplitA, const short *devBSpExp, const size_t ldbs, const uint32_t numsplitB, double *devCsplit, const size_t llsc, const size_t ldsc, double *devC, const size_t ldc, const double*, const double*, const cublasPointerMode_t, const uint32_t maxlevel, const uint32_t, uint32_t*); #elif defined (QGEMMDP) #endif #endif gemmtc-1.3a/src/cuda_tc/cuozblas_gemm.cu0000644000175000017500000001507013660232775020647 0ustar mukunokimukunoki#include "cuozblas_common.h" //#include "cuozblas_common.cu" #if defined (SGEMMTC) || defined (DGEMMTC) #define CU_TYPE2 CUDA_R_16F #define CU_TYPE3 CUDA_R_32F #define CU_ALGO CUBLAS_GEMM_DFALT_TENSOR_OP #elif defined (DGEMMSP) #define CU_TYPE2 CUDA_R_32F #define CU_TYPE3 CUDA_R_32F #define CU_ALGO CUBLAS_GEMM_DEFAULT #elif defined (DGEMMDP) || defined (QGEMMDP) #define CU_TYPE2 CUDA_R_64F #define CU_TYPE3 CUDA_R_64F #define CU_ALGO CUBLAS_GEMM_DEFAULT #endif #if defined (SGEMMTC) #define OZGEMM cuozblasSgemmTC #elif defined (DGEMMTC) #define OZGEMM cuozblasDgemmTC #elif defined (DGEMMSP) #define OZGEMM cuozblasDgemmSP #elif defined (DGEMMDP) #define OZGEMM cuozblasDgemmDP #elif defined (QGEMMDP) #define OZGEMM cuozblasQgemmDP #endif __host__ int32_t OZGEMM ( cuozblasHandle_t *oh, const cublasOperation_t tranA, const cublasOperation_t tranB, const size_t m, const size_t n, const size_t k, const FP_TYPE1 *alpha, const FP_TYPE1 *devA, const size_t lda, const FP_TYPE1 *devB, const size_t ldb, const FP_TYPE1 *beta, FP_TYPE1 *devC, const size_t ldc ) { counterInit (oh); if (oh->numSplitArraysMax == 1) { #if defined (SGEMMTC) cublasSgemm (oh->ch, tranA, tranB, m, n, k, alpha, devA, lda, devB, ldb, beta, devC, ldc); #elif defined (DGEMMTC) || defined (DGEMMSP) || defined (DGEMMDP) cublasDgemm (oh->ch, tranA, tranB, m, n, k, alpha, devA, lda, devB, ldb, beta, devC, ldc); #elif defined (QGEMMTC) fprintf (stderr, "* Not available.\n"); #endif return 0; } if (tranA != CUBLAS_OP_N || tranB != CUBLAS_OP_N) { fprintf (stderr, "* CUBLAS_OP_T is not supported.\n"); exit (1); } // keep original settings uint32_t fastModeOrig = oh->fastMode; // need to keep the original setting for disabling fast-mode during SplitA cublasPointerMode_t pmode; cublasGetPointerMode(oh->ch, &pmode); cublasSetPointerMode(oh->ch, CUBLAS_POINTER_MODE_HOST); cublasMath_t mmode; cublasGetMathMode(oh->ch, &mmode); cublasSetMathMode(oh->ch, CUBLAS_TENSOR_OP_MATH); FP_TYPE1 *devAmax, *devBmax, *devATmp, *devBTmp, *devATmp2, *devE, *devBe, *devB1, *devB2; FP_TYPE2 *devASplit, *devBSplit; FP_TYPE3 *devCSplit; FP_TYPE3 fone_ = 1., fzero_ = 0.; short *devASpExp, *devBSpExp; uint32_t numSplitArraysMax = (oh->numSplitArraysMax > 0) ? oh->numSplitArraysMax : NumSplitDefaultMax; double t1, t0 = timer(); size_t ldas, ldbs, ldcs, ldase, ldbse, ldat, ldbt; size_t mbk = m; size_t nbk = n; size_t memAddrTmp = oh->memAddr; uint32_t *devCheckOF; cuozblasVecAddrAlloc (oh, 1, sizeof(uint32_t), (void**)&devCheckOF); while (mbk > 0) { cuozblasMatAddrAlloc (oh, mbk, k * numSplitArraysMax, sizeof(FP_TYPE2), (void**)&devASplit, ldas); cuozblasMatAddrAlloc (oh, k, nbk * numSplitArraysMax, sizeof(FP_TYPE2), (void**)&devBSplit, ldbs); cuozblasVecAddrAlloc (oh, mbk, sizeof(FP_TYPE1), (void**)&devAmax); cuozblasVecAddrAlloc (oh, k, sizeof(FP_TYPE1), (void**)&devBmax); cuozblasMatAddrAlloc (oh, mbk, k, sizeof(FP_TYPE1), (void**)&devATmp, ldat); cuozblasMatAddrAlloc (oh, mbk, k, sizeof(FP_TYPE1), (void**)&devATmp2, ldat); cuozblasMatAddrAlloc (oh, k, n, sizeof(FP_TYPE1), (void**)&devBTmp, ldbt); // note: use 'n', not 'nbk' cuozblasVecAddrAlloc (oh, n, sizeof(FP_TYPE1), (void**)&devE); cuozblasVecAddrAlloc (oh, k, sizeof(FP_TYPE1), (void**)&devBe); cuozblasVecAddrAlloc (oh, mbk, sizeof(FP_TYPE1), (void**)&devB1); cuozblasVecAddrAlloc (oh, mbk, sizeof(FP_TYPE1), (void**)&devB2); cuozblasMatAddrAlloc (oh, mbk, nbk * numSplitArraysMax * numSplitArraysMax, sizeof(FP_TYPE3), (void**)&devCSplit, ldcs); cuozblasMatAddrAlloc (oh, mbk, numSplitArraysMax, sizeof(short), (void**)&devASpExp, ldase); cuozblasMatAddrAlloc (oh, nbk, numSplitArraysMax, sizeof(short), (void**)&devBSpExp, ldbse); if (!memCheck (oh)) break; oh->memAddr = memAddrTmp; mbk = ceil (mbk / 2.); nbk = ceil (nbk / 2.); } oh->mbk = mbk; oh->nbk = nbk; uint32_t block_count = 0; for (size_t im = 0; im < ceil((FP_TYPE3)m/mbk); im++) { size_t mbk_ = (m-mbk*im >= mbk) ? mbk : m-mbk*im; // Split of A ----------------------------------- t1 = timer(); uint32_t numSplitArraysA = numSplitArraysMax; if (oh->sumMode) cuSplitMat (oh, 'r', mbk_, k, devA+im*mbk, lda, devATmp, ldat, devASplit, ldas, devASpExp, ldase, devAmax, numSplitArraysA); else cuSplitMatA (oh, 'r', mbk_, n, k, devA+im*mbk, lda, devB, ldb, devATmp, ldat, devASplit, ldas, devASpExp, ldase, devAmax, devATmp2, ldat, devBTmp, ldbt, devE, devBe, devB1, devB2, numSplitArraysA); oh->t_SplitA += timer() - t1; for (size_t in = 0; in < ceil((FP_TYPE3)n/nbk); in++) { size_t nbk_ = (n-nbk*in >= nbk) ? nbk : n-nbk*in; // Split of B ----------------------------------- t1 = timer(); uint32_t numSplitArraysB = (oh->sumMode) ? numSplitArraysMax : numSplitArraysA; cuSplitMat (oh, 'c', k, nbk_, devB+in*nbk*ldb, ldb, devBTmp, ldbt, devBSplit, ldbs, devBSpExp, ldbse, devBmax, numSplitArraysB); oh->t_SplitB += timer() - t1; // Compute -------------------------------------- uint32_t maxlevel = (numSplitArraysA-1) + (numSplitArraysB-1); if (oh->fastMode) { if (oh->numSplitArraysMax == 0) maxlevel = min (numSplitArraysA-1, numSplitArraysB-1); else maxlevel = numSplitArraysMax-1; } uint32_t ic = 0; for (uint32_t ia = 0; ia < numSplitArraysA; ia++) { for (uint32_t ib = 0; ib < numSplitArraysB; ib++) { if (ia + ib <= maxlevel) { t1 = timer(); cublasGemmEx (oh->ch, tranA, tranB, mbk_, nbk_, k, &fone_, devASplit+ldas*k*ia, CU_TYPE2, ldas, devBSplit+ldbs*nbk*ib, CU_TYPE2, ldbs, &fzero_, devCSplit+ldcs*nbk*ic, CU_TYPE3, ldcs, CU_TYPE3, CU_ALGO); oh->t_comp += timer() - t1; oh->n_comp += 2. * mbk_ * nbk_ * k; ic++; } } } // Sum ----------------------------------------- t1 = timer(); if (cuSumMatDevice (mbk_, nbk_, devASpExp, ldase, numSplitArraysA, devBSpExp, ldbse, numSplitArraysB, devCSplit, ldcs*nbk, ldcs, &devC[ldc*(in*nbk)+im*mbk], ldc, alpha, beta, pmode, maxlevel, oh->sumMode, devCheckOF)) { fprintf (stderr, "overflow-mode is failed\n"); return -1; } oh->t_sum += timer() - t1; block_count++; oh->numSplitArraysA += numSplitArraysA; oh->numSplitArraysB += numSplitArraysB; oh->numSplitArraysC += ic; } } oh->t_total = timer() - t0; oh->numSplitArraysA /= (float)block_count; oh->numSplitArraysB /= (float)block_count; oh->numSplitArraysC /= (float)block_count; // recovering original setting --------------- cublasSetPointerMode(oh->ch, pmode); cublasSetMathMode(oh->ch, mmode); oh->fastMode = fastModeOrig; return 0; } gemmtc-1.3a/src/cuda_tc/cuozblas.cpp0000644000175000017500000000370013647250553020011 0ustar mukunokimukunoki#include "cuozblas_common.h" double timer () { struct timeval tv; cudaDeviceSynchronize (); gettimeofday (&tv, NULL); return tv.tv_sec + (double) tv.tv_usec * 1.0e-6; } size_t getPitchSize (size_t n) { return ceil((double)n / 8) * 8; } uint32_t memCheck (cuozblasHandle_t *oh) { if ((float)oh->memAddr > (float)oh->workSizeBytes) return 1; return 0; } void counterInit (cuozblasHandle_t *oh) { oh->t_SplitA = 0.; oh->t_SplitB = 0.; oh->t_comp = 0.; oh->t_sum = 0.; oh->t_total = 0.; oh->n_comp = 0.; oh->numSplitArraysA = 0; oh->numSplitArraysB = 0; oh->numSplitArraysC = 0; oh->memAddr = 0; } void cuozblasCreate (cuozblasHandle_t *oh, size_t WorkSizeBytes, uint32_t NumSplitArraysMax, uint32_t CRMode, uint32_t OFMode) { oh->workSizeBytes = (WorkSizeBytes) ? WorkSizeBytes : DefaultWorkSize; oh->numSplitArraysMax = NumSplitArraysMax; oh->CRMode = (CRMode == 0) ? 0 : 1; if (CRMode) { oh->fastMode = 0; oh->sumMode = 1; } else { oh->fastMode = 1; oh->sumMode = 0; } oh->OFMode = (OFMode == 0) ? 0 : 1; // work memory allocation cudaMalloc ((void **) &oh->devWork, oh->workSizeBytes); // cuBLAS cublasCreate (&oh->ch); cublasSetPointerMode(oh->ch, CUBLAS_POINTER_MODE_HOST); } void cuozblasSetNumSplitArraysMax (cuozblasHandle_t *oh, uint32_t NumSplitArraysMax) { oh->numSplitArraysMax = NumSplitArraysMax; } void cuozblasSetOFMode (cuozblasHandle_t *oh, uint32_t OFMode) { oh->OFMode = (OFMode == 0) ? 0 : 1; } void cuozblasSetCRMode (cuozblasHandle_t *oh, uint32_t CRMode) { if (CRMode) { oh->fastMode = 0; oh->sumMode = 1; } else { oh->fastMode = 1; oh->sumMode = 0; } } uint32_t cuozblasGetNumSplitArraysMax (cuozblasHandle_t *oh) { return oh->numSplitArraysMax; } uint32_t cuozblasGetOFMode (cuozblasHandle_t *oh) { return oh->OFMode; } uint32_t cuozblasGetCRMode (cuozblasHandle_t *oh) { return oh->CRMode; } void cuozblasDestroy (cuozblasHandle_t *oh) { cudaFree (oh->devWork); cublasDestroy (oh->ch); } gemmtc-1.3a/src/cuda_tc/cuozblas_split.cu0000644000175000017500000004030013657543425021052 0ustar mukunokimukunoki#include "cuozblas_common.h" #if defined (SGEMMTC) #define DATAEPS 24 #define IxAMAX cublasIsamax #define xASUM cublasSasum #define xGEMV cublasSgemv #elif defined (DGEMMTC) || defined (DGEMMSP) || defined (DGEMMDP) #define DATAEPS 53 #define IxAMAX cublasIdamax #define xASUM cublasDasum #define xGEMV cublasDgemv #elif defined (QGEMMDP) #define DATAEPS 113 #define IxAMAX cublasIqamax #define xASUM cublasQasum #define xGEMV cublasQgemv #endif #if defined (SGEMMTC) || defined (DGEMMTC) #define ConvT3tT2(x) __float2half(x) #elif defined (DGEMMSP) || defined (QGEMMDP) #define ConvT3tT2(x) (x) #endif // ========================================= // SplitMat // ========================================= __global__ void cuFindMaxMatNKernel ( const size_t m, const size_t n, const FP_TYPE1 * __restrict__ devInput, const size_t ldi, FP_TYPE1 *devMax ) { const size_t iBx = blockIdx.x; const size_t nTx = blockDim.x; const size_t nTy = blockDim.y; const size_t iTx = threadIdx.x; const size_t iTy = threadIdx.y; const size_t addrx = iBx * nTx + iTx; FP_TYPE1 max, input, tmp; extern __shared__ FP_TYPE1 shm[]; size_t j; if (addrx < m){ max = 0.; for (j = iTy; j < n; j+=nTy) { input = devInput[j * ldi + addrx]; if (max < fabs(input)) max = fabs(input); } shm[nTx * iTy + iTx] = max; __syncthreads (); if (iTy == 0) { max = shm[iTx]; #pragma unroll for (j = 1; j < SPLIT_N_NTY; j++) { tmp = shm[nTx * j + iTx]; if (max < fabs(tmp)) max = fabs(tmp); } devMax[addrx] = max; } } } __global__ void cuFindMaxMatTKernel ( const size_t m, const size_t n, const FP_TYPE1 * __restrict__ devInput, const size_t ldi, FP_TYPE1 *devMax ) { const size_t iBx = blockIdx.x; const size_t iBy = blockIdx.y; const size_t nTx = blockDim.x; const size_t nTy = blockDim.y; const size_t iTx = threadIdx.x; const size_t iTy = threadIdx.y; const size_t addrx = iBx * nTx + iTx; const size_t addry = iBy * nTy + iTy; FP_TYPE1 max, input, tmp;; extern __shared__ FP_TYPE1 shm[]; size_t i; if (addry < n){ max = 0.; for (i = addrx; i < m; i += nTx) { input = devInput[addry * ldi + i]; if (max < fabs(input)) max = fabs(input); } shm[iTx] = max; __syncthreads (); #pragma unroll for (i = nTx/2; i > 0; i>>=1) { tmp = shm[iTx+i]; if (iTx < i && shm[iTx] < tmp) shm[iTx] = tmp; __syncthreads (); } if (iTx == 0) devMax[addry] = shm[0]; } } __host__ void cuFindMaxMatDevice ( const char major, const size_t m, const size_t n, const FP_TYPE1 *devInput, const size_t ldi, FP_TYPE1 *devMax ) { size_t ntx, nty, nbx, nby; dim3 threads, grid; if (major == 'r') { ntx = SPLIT_N_NTX; nty = SPLIT_N_NTY; nbx = ceil (float(m) / ntx); nby = 1; threads = dim3 (ntx, nty); grid = dim3 (nbx, nby); cuFindMaxMatNKernel <<< grid, threads, sizeof(FP_TYPE1) * ntx * nty >>> (m, n, devInput, ldi, devMax); } else { ntx = SPLIT_T_NTX; nty = SPLIT_T_NTY; nbx = 1; nby = ceil (float(n) / nty); threads = dim3 (ntx, nty); grid = dim3 (nbx, nby); cuFindMaxMatTKernel <<< grid, threads, sizeof(FP_TYPE1) * ntx * 2 >>> (m, n, devInput, ldi, devMax); } } __global__ void cuSplitMatNKernel ( const size_t m, const size_t n, const int32_t rho, const FP_TYPE1 * __restrict__ devInput, const size_t ldi, FP_TYPE1 *devOutput, const size_t ldo, FP_TYPE2 *devSplit, const size_t lds, short *devSpExp, FP_TYPE1 *devMax ) { const size_t iBx = blockIdx.x; const size_t nTx = SPLIT_N_NTX; const size_t nTy = SPLIT_N_NTY; const size_t iTx = threadIdx.x; const size_t iTy = threadIdx.y; const size_t addrx = iBx * nTx + iTx; FP_TYPE1 max, tmp; volatile FP_TYPE1 sigma, input, split, tmp2; extern __shared__ FP_TYPE1 shm[]; size_t j; short tau; if (addrx < m){ // split ------------------------------------ // u=2^-53, log2(u^-1)=53 // z = scalbn (x, y): z = x * FLT_RADIX ** y (FLT_RADIX: 2) max = fabs(devMax[addrx]); // ============================================= tau = ceil(log2(max)); sigma = scalbn (1., rho+tau); // sigma = 2^(rho+log2(max)) #if !defined (DGEMMDP) devSpExp[addrx] = tau; #endif // ============================================= max = 0.; for (j = iTy; j < n; j+=nTy) { input = devInput[j * ldi + addrx]; tmp2 = input + sigma; split = tmp2 - sigma; input = input - split; // ============================================= #if defined (DGEMMDP) devSplit[j * lds + addrx] = split; #else FP_TYPE3 split_ = scalbn(split, -tau); devSplit[j * lds + addrx] = ConvT3tT2(split_); #endif // ============================================= devOutput[j * ldo + addrx] = input; if (max < fabs(input)) max = fabs(input); } shm[nTx * iTy + iTx] = max; __syncthreads (); if (iTy == 0) { max = shm[iTx]; #pragma unroll for (j = 1; j < nTy; j++) { tmp = shm[nTx * j + iTx]; if (max < fabs(tmp)) max = fabs(tmp); } devMax[addrx] = max; } } } __global__ void cuSplitMatNKernelA ( const size_t m, const size_t n, const int32_t rho, const FP_TYPE1 * __restrict__ devInput, const size_t ldi, FP_TYPE1 *devOutput, const size_t ldo, FP_TYPE2 *devSplit, const size_t lds, FP_TYPE1 *devSplitD, const size_t ldd, short *devSpExp, FP_TYPE1 *devMax ) { const size_t iBx = blockIdx.x; const size_t nTx = SPLIT_N_NTX; const size_t nTy = SPLIT_N_NTY; const size_t iTx = threadIdx.x; const size_t iTy = threadIdx.y; const size_t addrx = iBx * nTx + iTx; FP_TYPE1 max, tmp; volatile FP_TYPE1 sigma, input, split, tmp2; extern __shared__ FP_TYPE1 shm[]; size_t j; short tau; if (addrx < m){ // split ------------------------------------ // u=2^-53, log2(u^-1)=53 // z = scalbn (x, y): z = x * FLT_RADIX ** y (FLT_RADIX: 2) max = fabs(devMax[addrx]); // ============================================= tau = ceil(log2(max)); sigma = scalbn (1., rho + tau); // sigma = 2^(rho+log2(max)) #if !defined (DGEMMDP) devSpExp[addrx] = tau; #endif // ============================================= max = 0.; for (j = iTy; j < n; j+=nTy) { input = devInput[j * ldi + addrx]; tmp2 = input + sigma; split = tmp2 - sigma; input = input - split; // ============================================= devSplitD[j * ldd + addrx] = abs(split); #if defined (DGEMMDP) devSplit[j * lds + addrx] = split; #else FP_TYPE3 split_ = scalbn(split, -tau); devSplit[j * lds + addrx] = ConvT3tT2(split_); #endif // ============================================= devOutput[j * ldo + addrx] = input; if (max < fabs(input)) max = fabs(input); } shm[nTx * iTy + iTx] = max; __syncthreads (); if (iTy == 0) { max = shm[iTx]; #pragma unroll for (j = 1; j < nTy; j++) { tmp = shm[nTx * j + iTx]; if (max < fabs(tmp)) max = fabs(tmp); } devMax[addrx] = max; } } } __global__ void cuSplitMatTKernel ( const size_t m, const size_t n, const int32_t rho, const FP_TYPE1 * __restrict__ devInput, const size_t ldi, FP_TYPE1 *devOutput, const size_t ldo, FP_TYPE2 *devSplit, const size_t lds, short *devSpExp, FP_TYPE1 *devMax ) { const size_t iBx = blockIdx.x; const size_t iBy = blockIdx.y; const size_t nTx = blockDim.x; const size_t nTy = blockDim.y; const size_t iTx = threadIdx.x; const size_t iTy = threadIdx.y; const size_t addrx = iBx * nTx + iTx; const size_t addry = iBy * nTy + iTy; FP_TYPE1 max; volatile FP_TYPE1 sigma, input, split, tmp2; extern __shared__ FP_TYPE1 shm[]; short tau; if (addry < n){ max = fabs(devMax[addry]); // ============================================= tau = ceil(log2(max)); sigma = scalbn (1., rho + tau); #if !defined (DGEMMDP) devSpExp[addry] = tau; #endif // ============================================= max = 0.; for (size_t i = addrx; i < m; i += nTx) { input = devInput[addry * ldi + i]; tmp2 = input + sigma; split = tmp2 - sigma; input = input - split; // ============================================= #if defined (DGEMMDP) devSplit[addry * lds + i] = split; #else FP_TYPE3 split_ = scalbn(split, -tau); devSplit[addry * lds + i] = ConvT3tT2(split_); #endif // ============================================= devOutput[addry * ldo + i] = input; if (max < fabs(input)) max = fabs(input); } shm[iTx] = max; __syncthreads (); #pragma unroll for (uint32_t i = nTx/2; i > 0; i>>=1) { if (iTx < i && shm[iTx] < shm[iTx+i]) shm[iTx] = shm[iTx+i]; __syncthreads (); } if (iTx == 0) devMax[addry] = shm[0]; } } __host__ int32_t getRho (const size_t dim, const uint32_t OFMode) { int32_t rho; #if defined (SGEMMTC) if (OFMode) rho = ceil((24-(24-log2(2.*sqrt(dim)))/2)); // overflow-ver else rho = ceil((24-(24-log2(1.*dim))/2)); // standard-ver #elif defined (DGEMMTC) || defined (DGEMMSP) if (OFMode) rho = ceil((53-(24-log2(2.*sqrt(dim)))/2)); // overflow-ver else rho = ceil((53-(24-log2(1.*dim))/2)); // standard-ver #elif defined (DGEMMDP) if (OFMode) rho = ceil((53-(53-log2(2.*sqrt(dim)))/2)); // overflow-ver else rho = ceil((53-(53-log2(1.*dim))/2)); // standard-ver = ceil((53+log2(1.*dim))/2); #elif defined (QGEMMDP) if (OFMode) rho = ceil((113-(53-log2(2.*sqrt(dim)))/2)); // overflow-ver else rho = ceil((113-(53-log2(1.*dim))/2)); // standard-ver #endif return rho; } __host__ void cuSplitMatDevice ( const char major, const size_t m, const size_t n, const FP_TYPE1 *devInput, // input matrix (devAwrk) const size_t ldi, // leading dimension of input matrix FP_TYPE1 *devOutput, // output matrix (devAwrk) const size_t ldo, FP_TYPE2 *devSplit, // split matrices (output): this includes NumSplitMax matrices const size_t lds, // leading dimension of split matrix (# of cols) short *devSpExp, // exponent of split matrix FP_TYPE1 *devMax, uint32_t OFMode ) { size_t ntx, nty, nbx, nby; dim3 threads, grid; const size_t dim = ((major == 'r') && n > 1) ? n : m; int32_t rho = getRho (dim, OFMode); if (major == 'r') { ntx = SPLIT_N_NTX; nty = SPLIT_N_NTY; nbx = ceil (float(m) / ntx); nby = 1; threads = dim3 (ntx, nty); grid = dim3 (nbx, nby); cuSplitMatNKernel <<< grid, threads, sizeof(FP_TYPE1) * ntx * nty >>> (m, n, rho, devInput, ldi, devOutput, ldo, devSplit, lds, devSpExp, devMax); } else { ntx = SPLIT_T_NTX; nty = SPLIT_T_NTY; nbx = 1; nby = ceil (float(n) / nty); threads = dim3 (ntx, nty); grid = dim3 (nbx, nby); cuSplitMatTKernel <<< grid, threads, sizeof(FP_TYPE1) * ntx * 2 >>> (m, n, rho, devInput, ldi, devOutput, ldo, devSplit, lds, devSpExp, devMax); } } __host__ void cuSplitMatDeviceA ( const char major, const size_t m, const size_t n, const FP_TYPE1 *devInput, // input matrix (devAwrk) const size_t ldi, // leading dimension of input matrix FP_TYPE1 *devOutput, // output matrix (devAwrk) const size_t ldo, FP_TYPE2 *devSplit, // split matrices (output): this includes NumSplitMax matrices const size_t lds, // leading dimension of split matrix (# of cols) FP_TYPE1 *devSplitD, const size_t ldd, short *devSpExp, // exponent of split matrix FP_TYPE1 *devMax, uint32_t OFMode ) { size_t ntx, nty, nbx, nby; dim3 threads, grid; const size_t dim = ((major == 'r') && n > 1) ? n : m; int32_t rho = getRho (dim, OFMode); if (major == 'r') { ntx = SPLIT_N_NTX; nty = SPLIT_N_NTY; nbx = ceil (float(m) / ntx); nby = 1; threads = dim3 (ntx, nty); grid = dim3 (nbx, nby); cuSplitMatNKernelA <<< grid, threads, sizeof(FP_TYPE1) * ntx * nty >>> (m, n, rho, devInput, ldi, devOutput, ldo, devSplit, lds, devSplitD, ldd, devSpExp, devMax); } else { printf ("not implemented yet!!!!\n"); } } __host__ void cuSplitMat ( cuozblasHandle_t *oh, const char major, const size_t m, const size_t n, const FP_TYPE1 *devInput, const size_t ldi, FP_TYPE1 *devOutput, // devTmp const size_t ldo, FP_TYPE2 *devSplit, const size_t lds, short *devSpExp, // exponent of split matrix const size_t ldse, FP_TYPE1 *devMax, uint32_t &numSplitArrays ) { uint32_t numSplitArraysMax = numSplitArrays; uint32_t i = 1; FP_TYPE1 check; // FindMax^(0) if (n == 1 && major == 'r') { int32_t ptrMax = 0; IxAMAX (oh->ch, m, devInput, 1, &ptrMax); cudaMemcpy (devMax, &devInput[ptrMax-1], sizeof(FP_TYPE1), cudaMemcpyDeviceToDevice); } else { cuFindMaxMatDevice (major, m, n, devInput, ldi, devMax); } // printf ("ldo=%d, lds=%d\n", ldo, lds); // Split^(0) & FindMax^(1) cuSplitMatDevice (major, m, n, devInput, ldi, devOutput, ldo, &devSplit[0], lds, devSpExp, devMax, oh->OFMode); while (i < numSplitArraysMax) { // Check if (n == 1 && major == 'r') cudaMemcpy (&check, devMax, sizeof(FP_TYPE1), cudaMemcpyDeviceToHost); else xASUM (oh->ch, (major=='r')?m:n, devMax, 1, &check); if (fabs(check) == 0) break; // Split^(i) & FindMax^(i+1) cuSplitMatDevice (major, m, n, devOutput, ldo, devOutput, ldo, &devSplit[lds*n*i], lds, &devSpExp[ldse*i], devMax, oh->OFMode); i++; } numSplitArrays = i; } __global__ void cuGenEVecKernel ( const size_t n, FP_TYPE1 *devA ) { const size_t iBx = blockIdx.x; const size_t nTx = blockDim.x; const size_t iTx = threadIdx.x; const size_t addrx = iBx * nTx + iTx; FP_TYPE1 *addrA = &devA [addrx]; if (addrx < n) addrA[0] = 1.; } __host__ void cuGenEVec ( const size_t n, FP_TYPE1 *devA ) { dim3 threads, grid; int32_t ntx, nbx; ntx = VEC_NTX; nbx = ceil (float(n) / ntx); threads = dim3 (ntx); grid = dim3 (nbx); cuGenEVecKernel <<< grid, threads >>> (n, devA); } __host__ void cuSplitMatA ( cuozblasHandle_t *oh, const char major, const size_t m, const size_t n, const size_t k, const FP_TYPE1 *devAInput, const size_t ldai, const FP_TYPE1 *devBInput, const size_t ldbi, FP_TYPE1 *devAOutput, // devTmp const size_t ldao, FP_TYPE2 *devASplit, const size_t ldas, short *devASpExp, // exponent of split matrix const size_t ldase, FP_TYPE1 *devAMax, FP_TYPE1 *devAtmp, // work const size_t ldat, FP_TYPE1 *devBtmp, // work const size_t ldbt, FP_TYPE1 *devE, // work FP_TYPE1 *devBe, // work FP_TYPE1 *devB1, // work FP_TYPE1 *devB2, // work uint32_t &numSplitArrays ) { uint32_t numSplitArraysMax = numSplitArrays; int32_t icheck, izero = 0; FP_TYPE1 done = 1., dzero = 0., check; // For determining # of splits ------------------------------------------------------------- FP_TYPE1 c = 2. * scalbn (sqrt(k), -DATAEPS); cuGenEVec (n, devE); // devE = ones (n, 1) cuAbsMat (m, k, devAInput, ldai, devAtmp, ldat); // devAtmp = abs (devA) cuAbsMat (k, n, devBInput, ldbi, devBtmp, ldbt); // devBtmp = abs (devB) xGEMV (oh->ch, CUBLAS_OP_N, k, n, &done, devBtmp, ldbt, devE, 1, &dzero, devBe, 1); // devBe = devBtmp * e xGEMV (oh->ch, CUBLAS_OP_N, m, k, &c, devAtmp, ldat, devBe, 1, &dzero, devB1, 1); // devB1 = c * devAtmp * devBe // FindMax^(0) ----------------------------------------------------------------------------- if (k == 1 && major == 'r') { int32_t ptrMax = 0; IxAMAX (oh->ch, m, devAInput, 1, &ptrMax); cudaMemcpy (devAMax, &devAInput[ptrMax-1], sizeof(FP_TYPE1), cudaMemcpyDeviceToDevice); } else { cuFindMaxMatDevice (major, m, k, devAInput, ldai, devAMax); } // Split^(0) & FindMax^(1) ----------------------------------------------------------------- cuSplitMatDevice (major, m, k, devAInput, ldai, devAOutput, ldao, &devASplit[0], ldas, devASpExp, devAMax, oh->OFMode); // Split^(i) & FindMax^(i+1) --------------------------------------------------------------- uint32_t i = 1; while (i < numSplitArraysMax) { // Check if (k == 1 && major == 'r') cudaMemcpy (&check, devAMax, sizeof(FP_TYPE1), cudaMemcpyDeviceToHost); else xASUM (oh->ch, (major=='r')?m:k, devAMax, 1, &check); if (fabs(check) == 0) { oh->fastMode = 0; break; } // Split^(i) & FindMax^(i+1) cuSplitMatDeviceA (major, m, k, devAOutput, ldao, devAOutput, ldao, &devASplit[ldas*k*i], ldas, devAtmp, ldat, &devASpExp[ldase*i], devAMax, oh->OFMode); // For determining # of splits ------------------------------------------------------------- FP_TYPE1 di = (FP_TYPE1)(i+1); xGEMV (oh->ch, CUBLAS_OP_N, m, k, &di, devAtmp, ldat, devBe, 1, &dzero, devB2, 1); // devB2 = i * devASplit * devBe cudaMemcpy (devE, &izero, sizeof(int32_t), cudaMemcpyHostToDevice); cuCompVecAll (m, devB1, devB2, (int32_t*)devE); cudaMemcpy (&icheck, devE, sizeof(int32_t), cudaMemcpyDeviceToHost); i++; if (icheck == m) break; // ------------------------------------------------------------------------------------------ } numSplitArrays = i; } gemmtc-1.3a/Makefile0000644000175000017500000000042713647243115014731 0ustar mukunokimukunokiMAKE = make #--no-print-directory include make.inc MAKE_DIRS += src/cuda_tc MAKE_DIRS += testing default: @for subdir in $(MAKE_DIRS) ; do \ (cd $$subdir && $(MAKE)) ;\ done clean: @for subdir in $(MAKE_DIRS) ; do \ (cd $$subdir && $(MAKE) clean) ;\ done gemmtc-1.3a/testing/0000755000175000017500000000000013667423717014755 5ustar mukunokimukunokigemmtc-1.3a/testing/testing_common.cpp0000644000175000017500000003502313667406211020477 0ustar mukunokimukunoki// ----- testing handler --------------------------------------------------------------- struct testingHandle_t { // device parameter (single GPU only) cudaDeviceProp devprop; uint32_t devno; int32_t cc; double device_bytes_per_sec; double device_flops_per_sec_sp; double device_flops_per_sec_dp; int32_t Cores; int32_t DPCores; // testing parameter char mode; // c=check, p=perf double routine_flops; double routine_bytes; int32_t driverVersion; int32_t runtimeVersion; int32_t nodisp; int32_t profile; uint32_t numSplitArraysMax; uint32_t CRMode; uint32_t OFMode; // dim size_t dim_start; size_t dim_stop; size_t dim_step; size_t dim_m_const; size_t dim_n_const; size_t dim_k_const; size_t dim_m_dev; size_t dim_n_dev; size_t dim_k_dev; size_t dim_m_hst; size_t dim_n_hst; size_t dim_k_hst; size_t incx; size_t incy; size_t lda_dev; size_t ldb_dev; size_t ldc_dev; uint32_t dim_step_mode; cublasPointerMode_t pmode; // cublasAtomicsMode_t amode; // options char tranA; // N, T, C char tranB; // N, T, C float phi; // for input data uint32_t maxiter; // for iterative solvers }; // ----- CUBLAS const value function --------------------------------------------------------------- inline cublasOperation_t cublas_trans_t (char tran) { if (tran == 'N' || tran == 'n') return CUBLAS_OP_N; if (tran == 'T' || tran == 't') return CUBLAS_OP_T; if (tran == 'C' || tran == 'c') return CUBLAS_OP_C; return CUBLAS_OP_N; //default } void get_routine_theoretial_performance (testingHandle_t *th) { #if defined (PREC_S) size_t s = sizeof(float); #elif defined (PREC_D) size_t s = sizeof(double); #endif size_t n = th->dim_n_dev; size_t m = th->dim_m_dev; size_t k = th->dim_k_dev; th->routine_flops = (double) ROUTINE_FLOPS (m, n, k); th->routine_bytes = (double) ROUTINE_BYTES (m, n, k, s); } void dim_hst_setup (testingHandle_t *th) { th->dim_m_hst = th->dim_stop; th->dim_n_hst = th->dim_stop; th->dim_k_hst = th->dim_stop; if (th->dim_m_const > 0) th->dim_m_hst = th->dim_m_const; if (th->dim_n_const > 0) th->dim_n_hst = th->dim_n_const; if (th->dim_k_const > 0) th->dim_k_hst = th->dim_k_const; } void dim_dev_setup (testingHandle_t *th) { th->dim_m_dev = th->dim_start; th->dim_n_dev = th->dim_start; th->dim_k_dev = th->dim_start; if (th->dim_m_const > 0) th->dim_m_dev = th->dim_m_const; if (th->dim_n_const > 0) th->dim_n_dev = th->dim_n_const; if (th->dim_k_const > 0) th->dim_k_dev = th->dim_k_const; } void dim_dev_increment (testingHandle_t *th) { if (th->dim_m_const == 0) { switch (th->dim_step_mode) { case 1: th->dim_m_dev *= 10; break; case 2: th->dim_m_dev *= 2; break; default: th->dim_m_dev += th->dim_step; break; } } if (th->dim_n_const == 0) { switch (th->dim_step_mode) { case 1: th->dim_n_dev *= 10; break; case 2: th->dim_n_dev *= 2; break; default: th->dim_n_dev += th->dim_step; break; } } if (th->dim_k_const == 0) { switch (th->dim_step_mode) { case 1: th->dim_k_dev *= 10; break; case 2: th->dim_k_dev *= 2; break; default: th->dim_k_dev += th->dim_step; break; } } } __inline__ double gettime () { struct timeval tv; cudaDeviceSynchronize (); gettimeofday (&tv, NULL); return tv.tv_sec + (double) tv.tv_usec * 1.0e-6; } int32_t get_num_sp_cores (int32_t cc) { switch (cc) { case 500: return 128; case 520: return 128; case 530: return 128; case 600: return 64; case 610: return 128; case 620: return 128; case 700: return 64; case 750: return 64; default: printf ("error: unknown architecture\n"); return 0; } } int32_t get_num_dp_cores (int32_t cc) { switch (cc) { case 500: return 0; case 520: return 4; case 530: return 0; case 600: return 32; case 610: return 4; case 620: return 4; case 700: return 32; case 750: return 2; default: printf ("error: unknown architecture\n"); return 0; } } void dev_setting (testingHandle_t *th) { int32_t driverVersion = 0; int32_t runtimeVersion = 0; int32_t devcnt; cudaGetDeviceCount (&devcnt); cudaDeviceProp devprop; cudaDriverGetVersion(&driverVersion); cudaRuntimeGetVersion(&runtimeVersion); th->driverVersion = driverVersion; th->runtimeVersion = runtimeVersion; cudaSetDevice (th->devno); cudaGetDeviceProperties (&devprop, th->devno); th->devprop = devprop; th->cc = devprop.major * 100 + devprop.minor * 10; th->Cores = get_num_sp_cores (th->cc); th->DPCores = get_num_dp_cores (th->cc); th->device_bytes_per_sec = (double)th->devprop.memoryClockRate * 2. * 1000. * (double)th->devprop.memoryBusWidth / 8.; th->device_flops_per_sec_sp = (double)th->devprop.clockRate * 1000. * 2. * (double)th->devprop.multiProcessorCount * th->Cores; th->device_flops_per_sec_dp = (double)th->devprop.clockRate * 1000. * 2. * (double)th->devprop.multiProcessorCount * th->DPCores; } void testingCreate (int32_t argc, char **argv, testingHandle_t *th) { // default --------------------------- th->devno = 0; th->mode = 'c'; // options th->tranA = 'N'; th->tranB = 'N'; th->nodisp = 0; th->profile = 0; th->numSplitArraysMax = 0; th->OFMode = 0; th->CRMode = 0; th->phi = 0; th->dim_start = 128; th->dim_stop = 1024; th->dim_step = 128; th->dim_m_dev = 0; th->dim_n_dev = 0; th->dim_k_dev = 0; th->dim_m_hst = 0; th->dim_n_hst = 0; th->dim_k_hst = 0; th->dim_m_const = 0; th->dim_n_const = 0; th->dim_k_const = 0; th->incx = 1; th->incy = 1; th->lda_dev = 0; th->ldb_dev = 0; th->ldc_dev = 0; th->dim_step_mode = 0; th->pmode = CUBLAS_POINTER_MODE_VALUE; // args --------------------------- int32_t opt, opt_index; optind = 1; struct option long_options[] = { {"transa", required_argument, 0, 'a'}, {"transb", required_argument, 0, 'b'}, {"degree", required_argument, 0, 'd'}, {"ld", required_argument, NULL, 'e'}, {"devno", required_argument, NULL, 'g'}, {"CRmode", required_argument, NULL, 'h'}, {"OFmode", required_argument, NULL, 'i'}, {"mnk", required_argument, NULL, 'j'}, {"k", required_argument, NULL, 'k'}, {"m", required_argument, NULL, 'm'}, {"n", required_argument, NULL, 'n'}, {"mode", required_argument, NULL, 'o'}, {"phi", required_argument, NULL, 'p'}, {"range", required_argument, NULL, 'r'}, {"nodisp", required_argument, NULL, 'w'}, {"profile", required_argument, NULL, 'z'}, {0, 0, 0, 0} }; while ((opt = getopt_long (argc, argv, "abdefghijklmnoprwxz", long_options, &opt_index)) != -1){ switch (opt) { case 'a':{ // N, T, C if (optarg[0]) th->tranA= optarg[0]; break; } case 'b':{ // N, T, C if (optarg[0]) th->tranB= optarg[0]; break; } case 'd':{ th->numSplitArraysMax = atoi(optarg); break; } case 'e':{ char *str; char par[] = ":"; str = optarg; th->lda_dev = atoi (strtok (str, par)); th->ldb_dev = atoi (strtok (NULL, par)); th->ldc_dev = atoi (strtok (NULL, par)); break; } case 'g':{ th->devno = atoi(optarg); break; } case 'h':{ th->CRMode = atoi(optarg); break; } case 'i':{ th->OFMode = atoi(optarg); break; } case 'j':{ th->dim_start = atoi (optarg); th->dim_stop = atoi (optarg); th->dim_step = 1; break; } case 'k':{ th->dim_k_const = (size_t)atoi(optarg); break; } case 'm':{ th->dim_m_const = (size_t)atoi(optarg); break; } case 'n':{ th->dim_n_const = (size_t)atoi(optarg); break; } case 'o':{ th->mode = optarg[0]; break; } case 'p':{ th->phi = atof(optarg); break; } case 'r':{ char *str, *c_start, *c_stop, *c_step; char par[] = ":"; str = optarg; c_start = strtok (str, par); c_stop = strtok (NULL, par); c_step = strtok (NULL, par); th->dim_start = atoi (c_start); th->dim_stop = atoi (c_stop); if (strcmp (c_step, "pow") == 0) th->dim_step_mode = 2; else th->dim_step = atoi (c_step); break; } case 'w':{ th->nodisp = atoi(optarg); break; } case 'z':{ th->profile = atoi(optarg); break; } } } dev_setting (th); } void mublasInitMatPhi (size_t m, size_t n, size_t ld, FP_TYPE *mat, float phi) { size_t i, j; double val, r1, r2, x1; double mu = 0.0; double sigma = 1.0; double amax = 0., amin = 1.; if (ld == 0) ld = m; for (j = 0; j < n; j++) { for (i = 0; i < m; i++) { r1 = rand()/((double)RAND_MAX+1.); r2 = rand()/((double)RAND_MAX+1.); x1 = mu + (std::sqrt (-2. * log(r1)) * sin (2. * M_PI * r2)) * sigma; val = (r1 - 0.5) * exp ((double)phi * x1); if (phi < 0) val = rand()/((double)RAND_MAX+1.); mat[j*ld+i] = (FP_TYPE)val; if (amax < fabs(val)) amax = fabs(val); if (amin > fabs(val)) amin = fabs(val); } } printf ("#\tinput range: max = %1.3e, min = %1.3e\n", amax, amin); } void mublasInitMat (size_t rd, size_t cd, size_t ld, FP_TYPE *mat, FP_TYPE val) { size_t i, j; if (ld == 0) ld = rd; for (j = 0; j < cd; j++) { for (i = 0; i < rd; i++) mat[j*ld+i] = val; } } #ifdef USE_MPFR void mublasConvert2D (size_t rd, size_t cd, size_t ld, FP_TYPE *src, mpreal *dst) { size_t i, j; if (ld == 0) ld = rd; for (j = 0; j < cd; j++) { for (i = 0; i < rd; i++) mpfr_set_d (dst[j*ld+i], src[j*ld+i], MPFR_RNDN); } } #endif // check routine (|trg-ref|/|ref|) ------------------------------------------------- #ifdef USE_MPFR void mublasCheckMatrix (size_t m, size_t n, FP_TYPE *target, size_t ldt, mpreal *ref, size_t ldr) { #else void mublasCheckMatrix (size_t m, size_t n, FP_TYPE *target, size_t ldt, FP_TYPE *ref, size_t ldr) { #endif FP_TYPE target_d, ref_d; FP_TYPE max = 0.0, val; size_t errcnt = 0; for(size_t j = 0; j < n; j++) { for(size_t i = 0; i < m; i++) { target_d = target[j*ldt+i]; #ifdef USE_MPFR ref_d = mpfr_get_d (ref[j*ldr+i], MPFR_RNDN); #else ref_d = ref[j*ldr+i]; #endif //if (target_d != ref_d) printf(" -- [%zd,%zd] ref=%1.3e trg=%1.3e diff=%1.3e\n", j, i, ref_d, target_d, ref_d-target_d); val = fabs ((target_d - ref_d) / ref_d); //if (target_d != ref_d && val > 1e-1) printf(" -- [%zd,%zd] ref=%1.3e trg=%1.3e diff=%1.3e, diffr=%1.3e\n", j, i, ref_d, target_d, ref_d-target_d, val); //printf ("[%lu,%lu] ref=%+1.18e trg=%+1.18e dif=%+1.18e rel=%+1.18e\n", j, i, ref_d, target_d, ref_d-target_d, val); if (max < val) max = val; if (ref_d - target_d != 0) errcnt++; } } if (isinf(abs(max)) || isnan(abs(max))) printf("\terr: inf or nan\t"); else printf("\t%1.6e\t%zd", max, errcnt); } // print routine ------------------------------------------------- void print_routine_name () { #if defined (PREC_S) printf ("S"); #elif defined (PREC_D) printf ("D"); #else printf ("unknown-precision,"); #endif printf ("GEMM\n"); } #include void print_info1 (testingHandle_t *th) { if (!th->nodisp) { system("echo -n '# Testing info -----------------------------\n'"); system("echo -n '# \tDate:\t'; date"); system("echo -n '# \tCPU:'; cat /proc/cpuinfo | grep 'model name' | head -n 1"); system("echo -n '# \tOS:\t'; uname -a"); system("echo -n '# \tHost:\t'; hostname"); printf("# \tDevice[%d]\t%s\n", th->devno, th->devprop.name); printf("# \t- Theoretical Peak Memory Bandwidth:\t%3.1f [GB/s]\n", th->device_bytes_per_sec * 1.e-9); printf("# \t- Theoretical Peak Performance (SP):\t%4.1f [GFlops]\n", th->device_flops_per_sec_sp * 1.e-9); printf("# \t- Theoretical Peak Performance (DP):\t%4.1f [GFlops]\n", th->device_flops_per_sec_dp * 1.e-9); printf("# \t- Total Gmem:\t%4.1f [GB]\n", th->devprop.totalGlobalMem * 1e-9); printf("# \t- CC:\t%d.%d\n", th->devprop.major, th->devprop.minor); printf("# \t- ECC Enabled:\t%d\n", th->devprop.ECCEnabled); printf("# \t- MultiProcessorCount:\t%d\n", th->devprop.multiProcessorCount); printf("# \tDriver / Runtime:\t%d, %d\n", th->driverVersion, th->runtimeVersion); printf ("#\tRoutine:\t"); print_routine_name (); printf ("#\tLibrary: GEMMTC\n"); if (th->mode == 'c') { printf ("#\tmode: checking with "); #ifdef USE_MPFR printf ("MPFR %zd-bit\n", mpfr_get_default_prec()); #else printf ("cuBLAS\n"); #endif } else { printf ("#\tmode: performance evaluation\n"); } printf ("#\tNLOOP:\t%d\n", NLOOP); printf ("# GEMM-TC ----------------------------------\n"); printf ("#\tWork-mem size = %1.1e (bytes)\n", WORK_MEM_SIZE); printf ("#\tPhi = %f\n", th->phi); printf ("#\tDegree = %d\n", th->numSplitArraysMax); printf ("#\tCRMode = %d\n", th->CRMode); // printf ("#\tOFMode = %d\n", th->OFMode); printf ("# BLAS parameter ---------------------------\n"); printf ("#\ttranA:\t%c\n", th->tranA); printf ("#\ttranB:\t%c\n", th->tranB); printf ("#\tdim_start:\t%zd\n", (size_t)th->dim_start); printf ("#\tdim_stop:\t%zd\n", (size_t)th->dim_stop); printf ("#\tdim_step:\t%zd\n", (size_t)th->dim_step); if (th->dim_m_const == 0) printf ("#\tM:\trange\n"); else printf ("#\tM:\t%zd\n", (size_t)th->dim_m_const); if (th->dim_n_const == 0) printf ("#\tN:\trange\n"); else printf ("#\tN:\t%zd\n", (size_t)th->dim_n_const); if (th->dim_k_const == 0) printf ("#\tK:\trange\n"); else printf ("#\tK:\t%zd\n", (size_t)th->dim_k_const); if (th->pmode) printf ("#\tcublaspointer:\tDEVICE\n"); else printf ("#\tcublaspointer:\tHOST\n"); } } void print_info2 (testingHandle_t *th) { if (!th->nodisp) { printf ("# Evaluation result ------------------------\n"); if (th->mode == 'p') printf ("(* shows the value at the last execution, ** shows the value of the last block at the last execution)\n"); printf ("# M\tN\tK"); if (th->mode == 'p') { if (th->profile) printf ("\tTime[sec]\tPerf[GFlops]\tPerf[GB/s]\titr\t*t_SpltA\t*t_SpltB\t*t_Comp \t*t_Sum \t*t_Other\t*t_Total\t**#SpltA\t**#SpltB\t**#Comp\t**Mem[MB]\tTC[TFlops]"); else printf ("\tTime[sec]\tPerf[GFlops]\tPerf[GB/s]\titr\t**#SpltA\t**#SpltB\t**#Comp"); } #if PREC_D else printf ("\tERROR\t#ofERROR: (1)cuBLAS, (2)DGEMM-TC-DP, (3)DGEMM-TC-CR"); #else else printf ("\tERROR\t#ofERROR: (1)cuBLAS, (2)SGEMM-TC-SP"); #endif printf ("\n"); } } void print_cuozblas_info1 (testingHandle_t *th, cuozblasHandle_t *ha) { double t_other = ha->t_total - ha->t_SplitA - ha->t_SplitB - ha->t_comp - ha->t_sum; if (th->profile) { printf ("\t%1.3e\t%1.3e\t%1.3e\t%1.3e\t%1.3e\t%1.3e", ha->t_SplitA, ha->t_SplitB, ha->t_comp, ha->t_sum, t_other, ha->t_total); printf ("\t%1.2f\t%1.2f\t%1.2f\t%1.3e", ha->numSplitArraysA, ha->numSplitArraysB, ha->numSplitArraysC, ha->memAddr*1.e-6); printf ("\t%1.3e [TFlops] (comp only)", (1.e-12*ha->n_comp)/ha->t_comp); printf ("\t(%1.2f)", ha->n_comp); printf ("\t%d\t%d (mbk,nbk)", ha->mbk, ha->nbk); } else { printf ("\t%1.2f\t%1.2f\t%1.2f", ha->numSplitArraysA, ha->numSplitArraysB, ha->numSplitArraysC); printf ("\t%1.3e [TFlops] (comp only)", (1.e-12*ha->n_comp)/ha->t_comp); } } gemmtc-1.3a/testing/testing_level3.cpp0000644000175000017500000002044713667406035020411 0ustar mukunokimukunoki#define ROUTINE_FLOPS(m,n,k) (2*m*n*k+3*m*n) #define ROUTINE_BYTES(m,n,k,s) ((m*k+k*n+2*m*n)*s) #include "testing_common.h" #include "testing_common.cpp" #if defined (PREC_S) && defined (USE_TC) #define GEMMTC cuozblasSgemmTC #elif defined (PREC_S) && defined (USE_SP) #define GEMMTC cuozblasSgemmSP #elif defined (PREC_D) && defined (USE_TC) #define GEMMTC cuozblasDgemmTC #elif defined (PREC_D) && defined (USE_SP) #define GEMMTC cuozblasDgemmSP #elif defined (PREC_D) && defined (USE_DP) #define GEMMTC cuozblasDgemmDP #endif int32_t main (int32_t argc, char **argv) { size_t i; size_t lda_dev, lda_hst, rda_dev, rda_hst, cda_dev, cda_hst; size_t ldb_dev, ldb_hst, rdb_dev, rdb_hst, cdb_dev, cdb_hst; size_t ldc_dev, ldc_hst, rdc_dev, rdc_hst, cdc_dev, cdc_hst; double t0, t1; #ifdef USE_MPFR mpfr_set_default_prec(MPFR_PREC); mpreal::set_default_prec(MPFR_PREC); #endif // testing setup ------------------------------ testingHandle_t th; testingCreate (argc, argv, &th); print_info1 (&th); // library setup ------------------------------ cuozblasHandle_t ha; cuozblasCreate (&ha, WORK_MEM_SIZE, th.numSplitArraysMax, th.CRMode, th.OFMode); cublasSetPointerMode(ha.ch, CUBLAS_POINTER_MODE_VALUE); // memory setup ------------------------------- dim_hst_setup (&th); // dim setup // A rda_hst = th.dim_stop; cda_hst = th.dim_stop; if (th.tranA == 'N' || th.tranA == 'n') { if (th.dim_m_const > 0) rda_hst = th.dim_m_const; if (th.dim_k_const > 0) cda_hst = th.dim_k_const; } else { if (th.dim_k_const > 0) rda_hst = th.dim_k_const; if (th.dim_m_const > 0) cda_hst = th.dim_m_const; } lda_hst = rda_hst; // B rdb_hst = th.dim_stop; cdb_hst = th.dim_stop; if (th.tranB == 'N' || th.tranB == 'n') { if (th.dim_k_const > 0) rdb_hst = th.dim_k_const; if (th.dim_n_const > 0) cdb_hst = th.dim_n_const; } else { if (th.dim_n_const > 0) rdb_hst = th.dim_n_const; if (th.dim_k_const > 0) cdb_hst = th.dim_k_const; } ldb_hst = rdb_hst; // C rdc_hst = th.dim_stop; cdc_hst = th.dim_stop; if (th.dim_m_const > 0) rdc_hst = th.dim_m_const; if (th.dim_n_const > 0) cdc_hst = th.dim_n_const; ldc_hst = rdc_hst; // malloc host memory FP_TYPE alpha, beta; FP_TYPE *hst_A = new FP_TYPE[lda_hst * cda_hst]; FP_TYPE *hst_B = new FP_TYPE[ldb_hst * cdb_hst]; FP_TYPE *hst_C = new FP_TYPE[ldc_hst * cdc_hst]; FP_TYPE *hst_C_t = new FP_TYPE[ldc_hst * cdc_hst]; #ifdef USE_MPFR mpreal alpha_r, beta_r; mpreal *hst_A_r, *hst_B_r, *hst_C_r; if (th.mode == 'c') { alpha_r = 0; beta_r = 0; hst_A_r = new mpreal[lda_hst * cda_hst]; hst_B_r = new mpreal[ldb_hst * cdb_hst]; hst_C_r = new mpreal[ldc_hst * cdc_hst]; } #else FP_TYPE *hst_C_r; if (th.mode == 'c') hst_C_r = new FP_TYPE[ldc_hst * cdc_hst]; #endif // initialize mublasInitMat (1, 1, 0, &alpha, 1.); mublasInitMat (1, 1, 0, &beta, 0.); mublasInitMatPhi (lda_hst, cda_hst, lda_hst, hst_A, th.phi); mublasInitMatPhi (ldb_hst, cdb_hst, ldb_hst, hst_B, th.phi); mublasInitMat (ldc_hst, cdc_hst, ldc_hst, hst_C, 0.); // -------------------------------------------- print_info2 (&th); // evaluation --------------------------------- dim_dev_setup (&th); while (1) { if ((th.dim_m_const == 0 && th.dim_m_dev > th.dim_stop) || (th.dim_n_const == 0 && th.dim_n_dev > th.dim_stop) || (th.dim_k_const == 0 && th.dim_k_dev > th.dim_stop)) break; // dim setup // A if (th.tranA == 'N' || th.tranA == 'n') { rda_dev = th.dim_m_dev; cda_dev = th.dim_k_dev; } else { rda_dev = th.dim_k_dev; cda_dev = th.dim_m_dev; } // B if (th.tranB == 'N' || th.tranB == 'n') { rdb_dev = th.dim_k_dev; cdb_dev = th.dim_n_dev; } else { rdb_dev = th.dim_n_dev; cdb_dev = th.dim_k_dev; } // C rdc_dev = th.dim_m_dev; cdc_dev = th.dim_n_dev; // malloc device memory FP_TYPE *dev_A, *dev_B, *dev_C; size_t pitch; cudaMallocPitch ((void **) &dev_A, &pitch, sizeof(FP_TYPE) * rda_dev, cda_dev); lda_dev = pitch/sizeof(FP_TYPE); cudaMallocPitch ((void **) &dev_B, &pitch, sizeof(FP_TYPE) * rdb_dev, cdb_dev); ldb_dev = pitch/sizeof(FP_TYPE); cudaMallocPitch ((void **) &dev_C, &pitch, sizeof(FP_TYPE) * rdc_dev, cdc_dev); ldc_dev = pitch/sizeof(FP_TYPE); // memcpy from host to device cublasSetMatrix(rda_dev, cda_dev, sizeof(FP_TYPE), hst_A, lda_hst, dev_A, lda_dev); cublasSetMatrix(rdb_dev, cdb_dev, sizeof(FP_TYPE), hst_B, ldb_hst, dev_B, ldb_dev); cublasSetMatrix(rdc_dev, cdc_dev, sizeof(FP_TYPE), hst_C, ldc_hst, dev_C, ldc_dev); printf ("%zd\t%zd\t%zd", (size_t)th.dim_m_dev, (size_t)th.dim_n_dev, (size_t)th.dim_k_dev); get_routine_theoretial_performance (&th); // execution --------------------------------- if (th.mode == 'p') { // warm up for (i = 0; i < WLOOP; i++) { GEMMTC (&ha, cublas_trans_t(th.tranA), cublas_trans_t(th.tranB), th.dim_m_dev, th.dim_n_dev, th.dim_k_dev, &alpha, dev_A, lda_dev, dev_B, ldb_dev, &beta, dev_C, ldc_dev); } t0 = gettime (); for (i = 0; i < NLOOP; i++) { GEMMTC (&ha, cublas_trans_t(th.tranA), cublas_trans_t(th.tranB), th.dim_m_dev, th.dim_n_dev, th.dim_k_dev, &alpha, dev_A, lda_dev, dev_B, ldb_dev, &beta, dev_C, ldc_dev); } t1 = gettime (); float sec = (t1 - t0) / (float)NLOOP; printf ("\t%1.3e\t%1.3e\t%1.3e\t%zd", sec, (th.routine_flops / sec) * 1.0e-9, (th.routine_bytes / sec) * 1.0e-9, (size_t)NLOOP); } // ------------------------------------------- if (th.mode == 'c') { #ifdef USE_MPFR // MBLAS mublasConvert2D (1, 1, 0, &alpha, &alpha_r); mublasConvert2D (1, 1, 0, &beta, &beta_r); mublasConvert2D (rda_dev, cda_dev, lda_hst, hst_A, hst_A_r); mublasConvert2D (rdb_dev, cdb_dev, ldb_hst, hst_B, hst_B_r); mublasConvert2D (rdc_dev, cdc_dev, ldc_hst, hst_C, hst_C_r); Rgemm (&th.tranA,&th.tranB,th.dim_m_dev,th.dim_n_dev,th.dim_k_dev,alpha_r,hst_A_r,lda_hst,hst_B_r,ldb_hst,beta_r,hst_C_r,ldc_hst); #else cuozblasSetNumSplitArraysMax (&ha, 1); cublasSetMatrix(rdc_dev, cdc_dev, sizeof(FP_TYPE), hst_C, ldc_hst, dev_C, ldc_dev); GEMMTC (&ha, cublas_trans_t(th.tranA), cublas_trans_t(th.tranB), th.dim_m_dev, th.dim_n_dev, th.dim_k_dev, &alpha, dev_A, lda_dev, dev_B, ldb_dev, &beta, dev_C, ldc_dev); cublasGetMatrix (rdc_dev, cdc_dev, sizeof(FP_TYPE), dev_C, ldc_dev, hst_C_r, ldc_hst); #endif // cuBLAS (through GEMM-TC) cuozblasSetNumSplitArraysMax (&ha, 1); cublasSetMatrix(rdc_dev, cdc_dev, sizeof(FP_TYPE), hst_C, ldc_hst, dev_C, ldc_dev); GEMMTC (&ha, cublas_trans_t(th.tranA), cublas_trans_t(th.tranB), th.dim_m_dev, th.dim_n_dev, th.dim_k_dev, &alpha, dev_A, lda_dev, dev_B, ldb_dev, &beta, dev_C, ldc_dev); cublasGetMatrix (rdc_dev, cdc_dev, sizeof(FP_TYPE), dev_C, ldc_dev, hst_C_t, ldc_hst); mublasCheckMatrix (th.dim_m_dev, th.dim_n_dev, hst_C_t, ldc_hst, hst_C_r, ldc_hst); // GEMM-TC with DP-mode cuozblasSetNumSplitArraysMax (&ha, th.numSplitArraysMax); cuozblasSetCRMode (&ha, 0); cublasSetMatrix(rdc_dev, cdc_dev, sizeof(FP_TYPE), hst_C, ldc_hst, dev_C, ldc_dev); GEMMTC (&ha, cublas_trans_t(th.tranA), cublas_trans_t(th.tranB), th.dim_m_dev, th.dim_n_dev, th.dim_k_dev, &alpha, dev_A, lda_dev, dev_B, ldb_dev, &beta, dev_C, ldc_dev); cublasGetMatrix (rdc_dev, cdc_dev, sizeof(FP_TYPE), dev_C, ldc_dev, hst_C_t, ldc_hst); mublasCheckMatrix (th.dim_m_dev, th.dim_n_dev, hst_C_t, ldc_hst, hst_C_r, ldc_hst); // GEMM-TC with CR-mode cuozblasSetNumSplitArraysMax (&ha, th.numSplitArraysMax); cuozblasSetCRMode (&ha, 1); cublasSetMatrix(rdc_dev, cdc_dev, sizeof(FP_TYPE), hst_C, ldc_hst, dev_C, ldc_dev); GEMMTC (&ha, cublas_trans_t(th.tranA), cublas_trans_t(th.tranB), th.dim_m_dev, th.dim_n_dev, th.dim_k_dev, &alpha, dev_A, lda_dev, dev_B, ldb_dev, &beta, dev_C, ldc_dev); cublasGetMatrix (rdc_dev, cdc_dev, sizeof(FP_TYPE), dev_C, ldc_dev, hst_C_t, ldc_hst); mublasCheckMatrix (th.dim_m_dev, th.dim_n_dev, hst_C_t, ldc_hst, hst_C_r, ldc_hst); } cudaFree (dev_A); cudaFree (dev_B); cudaFree (dev_C); if (th.mode == 'p') { print_cuozblas_info1 (&th, &ha); // sleep (10); } printf ("\n"); dim_dev_increment (&th); if (th.dim_m_const && th.dim_n_const && th.dim_k_const) break; } // -------------------------------------------- delete[]hst_A; delete[]hst_B; delete[]hst_C; delete[]hst_C_t; if (th.mode == 'c') { #ifdef USE_MPFR delete[]hst_A_r; delete[]hst_B_r; #endif delete[]hst_C_r; } cuozblasDestroy (&ha); return 0; } gemmtc-1.3a/testing/testing_setting.h0000644000175000017500000000057113647251725020337 0ustar mukunokimukunoki#ifndef TESTING_SETTING #define TESTING_SETTING // ***** SETTING *********************************** #define SIZE_T size_t #define CUBLAS_POINTER_MODE_VALUE CUBLAS_POINTER_MODE_HOST #define WLOOP 3 // (times) #define NLOOP 10 // (times) #define MPFR_PREC 2048 // (bits) #define WORK_MEM_SIZE 20e9 // (bytes) // ***** SETTING *********************************** #endif gemmtc-1.3a/testing/isc2020_dat/0000775000175000017500000000000013667423712016664 5ustar mukunokimukunokigemmtc-1.3a/testing/isc2020_dat/_testing_dgemm_cublas_perf.dat0000664000175000017500000000575513660242262024714 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 10:00:28 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: DGEMM # Library: GEMMTC # mode: performance evaluation # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 0.100000 # Degree = 1 # CRMode = 0 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 512 # dim_stop: 10240 # dim_step: 512 # M: range # N: range # K: range # cublaspointer: HOST # input range: max = 8.539e-01, min = 9.750e-10 # input range: max = 8.877e-01, min = 1.991e-09 # Evaluation result ------------------------ (* shows the value at the last execution, ** shows the value of the last block at the last execution) # M N K Time[sec] Perf[GFlops] Perf[GB/s] itr **#SpltA **#SpltB **#Comp 512 512 512 7.553e-04 3.564e+02 1.111e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 1024 1024 1024 6.257e-03 3.437e+02 5.362e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 1536 1536 1536 1.351e-02 5.368e+02 5.586e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 2048 2048 2048 3.364e-02 5.110e+02 3.989e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 2560 2560 2560 6.447e-02 5.208e+02 3.253e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 3072 3072 3072 1.077e-01 5.384e+02 2.803e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 3584 3584 3584 1.726e-01 5.336e+02 2.381e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 4096 4096 4096 2.600e-01 5.289e+02 2.065e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 4608 4608 4608 3.633e-01 5.388e+02 1.870e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 5120 5120 5120 5.082e-01 5.283e+02 1.651e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 5632 5632 5632 6.710e-01 5.326e+02 1.513e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 6144 6144 6144 8.678e-01 5.346e+02 1.392e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 6656 6656 6656 1.116e+00 5.287e+02 1.271e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 7168 7168 7168 1.391e+00 5.295e+02 1.182e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 7680 7680 7680 1.696e+00 5.344e+02 1.113e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 8192 8192 8192 2.066e+00 5.323e+02 1.039e+00 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 8704 8704 8704 2.480e+00 5.319e+02 9.777e-01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 9216 9216 9216 2.930e+00 5.344e+02 9.277e-01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 9728 9728 9728 3.454e+00 5.331e+02 8.766e-01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 10240 10240 10240 4.021e+00 5.341e+02 8.345e-01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) gemmtc-1.3a/testing/isc2020_dat/_testing_dgemm_perf.dat0000664000175000017500000000624313660253575023365 0ustar mukunokimukunoki## cublas # input range: max = 8.539e-01, min = 9.750e-10 # input range: max = 8.877e-01, min = 1.991e-09 10240 10240 10240 4.021e+00 5.342e+02 8.345e-01 10 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.00 0.00 0.00 0.000e+00 -nan [TFlops] (comp only) (0.00) 0 0 (mbk,nbk) ## dgemmtc-dp phi=0.1 # input range: max = 8.539e-01, min = 9.750e-10 # input range: max = 8.877e-01, min = 1.991e-09 10240 10240 10240 2.186e+00 9.823e+02 1.535e+00 10 1.602e-01 2.866e-01 1.567e+00 1.800e-01 9.534e-04 2.195e+00 11.00 11.00 66.00 1.164e+04 9.045e+01 [TFlops] (comp only) (141733920768000.00) 2560 2560 (mbk,nbk) ## dgemmtc-dp phi=1 # input range: max = 1.055e+02, min = 1.472e-09 # input range: max = 1.555e+02, min = 1.353e-09 10240 10240 10240 2.536e+00 8.470e+02 1.323e+00 10 1.735e-01 3.105e-01 1.849e+00 2.137e-01 1.057e-03 2.548e+00 12.00 12.00 78.00 1.164e+04 9.058e+01 [TFlops] (comp only) (167503724544000.00) 2560 2560 (mbk,nbk) ## dgemmtc-dp phi=2 # input range: max = 2.226e+04, min = 2.328e-09 # input range: max = 4.836e+04, min = 4.367e-10 10240 10240 10240 2.905e+00 7.393e+02 1.155e+00 10 1.850e-01 3.349e-01 2.145e+00 2.508e-01 1.158e-03 2.917e+00 13.00 13.00 91.00 1.164e+04 9.109e+01 [TFlops] (comp only) (195421011968000.00) 2560 2560 (mbk,nbk) ## dgemmtc-cr phi=0.1 # input range: max = 8.539e-01, min = 9.750e-10 # input range: max = 8.877e-01, min = 1.991e-09 10240 10240 10240 9.635e+00 2.229e+02 3.483e-01 10 1.483e-01 3.792e-01 6.041e+00 3.176e+00 3.259e-03 9.748e+00 16.00 15.75 252.00 1.164e+04 8.958e+01 [TFlops] (comp only) (541165879296000.00) 2560 2560 (mbk,nbk) ## dgemmtc-cr phi=1 # input range: max = 1.055e+02, min = 1.472e-09 # input range: max = 1.555e+02, min = 1.353e-09 10240 10240 10240 1.112e+01 1.932e+02 3.018e-01 10 1.612e-01 4.141e-01 6.999e+00 3.627e+00 4.760e-03 1.121e+01 17.00 17.00 289.00 1.164e+04 8.867e+01 [TFlops] (comp only) (620622774272000.00) 2560 2560 (mbk,nbk) ## dgemmtc-cr phi=2 # input range: max = 2.226e+04, min = 2.328e-09 # input range: max = 4.836e+04, min = 4.367e-10 10240 10240 10240 1.246e+01 1.724e+02 2.693e-01 10 1.711e-01 4.440e-01 7.854e+00 4.084e+00 4.605e-03 1.256e+01 18.00 18.00 324.00 1.164e+04 8.859e+01 [TFlops] (comp only) (695784701952000.00) 2560 2560 (mbk,nbk) ## dgemmdp-cr phi=0.1 # input range: max = 8.539e-01, min = 9.750e-10 # input range: max = 8.877e-01, min = 1.991e-09 10240 10240 10240 8.677e+01 2.475e+01 3.867e-02 10 2.622e-02 2.490e-01 8.623e+01 2.615e-01 1.547e-03 8.677e+01 4.50 4.75 21.38 9.071e+03 5.323e-01 [TFlops] (comp only) (45902462976000.00) 1280 1280 (mbk,nbk) ## dgemmdp-cr phi=1 # input range: max = 1.055e+02, min = 1.472e-09 # input range: max = 1.555e+02, min = 1.353e-09 10240 10240 10240 1.015e+02 2.117e+01 3.307e-02 10 2.883e-02 2.616e-01 1.009e+02 3.135e-01 1.899e-03 1.015e+02 5.00 5.00 25.00 9.071e+03 5.323e-01 [TFlops] (comp only) (53687091200000.00) 1280 1280 (mbk,nbk) ## dgemmdp-cr phi=2 # input range: max = 2.226e+04, min = 2.328e-09 # input range: max = 4.836e+04, min = 4.367e-10 10240 10240 10240 1.015e+02 2.117e+01 3.307e-02 10 2.880e-02 2.608e-01 1.009e+02 3.134e-01 2.001e-03 1.015e+02 5.00 5.00 25.00 9.071e+03 5.323e-01 [TFlops] (comp only) (53687091200000.00) 1280 1280 (mbk,nbk) gemmtc-1.3a/testing/isc2020_dat/_testing_dgemmtc-cr_phi01_perf.dat0000664000175000017500000001260213660235732025306 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 09:14:30 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: DGEMM # Library: GEMMTC # mode: performance evaluation # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 0.100000 # Degree = 0 # CRMode = 1 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 512 # dim_stop: 10240 # dim_step: 512 # M: range # N: range # K: range # cublaspointer: HOST # input range: max = 8.539e-01, min = 9.750e-10 # input range: max = 8.877e-01, min = 1.991e-09 # Evaluation result ------------------------ (* shows the value at the last execution, ** shows the value of the last block at the last execution) # M N K Time[sec] Perf[GFlops] Perf[GB/s] itr *t_SpltA *t_SpltB *t_Comp *t_Sum *t_Other *t_Total **#SpltA **#SpltB **#Comp **Mem[MB] TC[TFlops] 512 512 512 1.015e-02 2.651e+01 8.262e-01 10 1.121e-03 1.103e-03 2.569e-03 5.174e-03 9.513e-05 1.006e-02 11.00 11.00 121.00 3.650e+02 1.264e+01 [TFlops] (comp only) (32480690176.00) 512 512 (mbk,nbk) 1024 1024 1024 2.405e-02 8.944e+01 1.395e+00 10 1.441e-03 1.645e-03 4.651e-03 1.499e-02 1.011e-04 2.283e-02 11.00 11.00 121.00 1.460e+03 5.587e+01 [TFlops] (comp only) (259845521408.00) 1024 1024 (mbk,nbk) 1536 1536 1536 7.085e-02 1.024e+02 1.066e+00 10 2.482e-03 3.218e-03 1.647e-02 4.836e-02 1.330e-04 7.066e-02 13.00 13.00 169.00 3.284e+03 7.439e+01 [TFlops] (comp only) (1224870985728.00) 1536 1536 (mbk,nbk) 2048 2048 2048 1.322e-01 1.300e+02 1.015e+00 10 3.192e-03 4.813e-03 3.851e-02 8.552e-02 1.352e-04 1.322e-01 13.00 13.00 169.00 5.839e+03 7.539e+01 [TFlops] (comp only) (2903397892096.00) 2048 2048 (mbk,nbk) 2560 2560 2560 2.221e-01 1.512e+02 9.442e-01 10 7.373e-03 6.688e-03 7.206e-02 1.356e-01 1.316e-04 2.218e-01 13.00 13.00 169.00 9.123e+03 7.870e+01 [TFlops] (comp only) (5670699008000.00) 2560 2560 (mbk,nbk) 3072 3072 3072 3.316e-01 1.749e+02 9.106e-01 10 8.823e-03 8.868e-03 1.188e-01 1.966e-01 1.566e-04 3.332e-01 13.00 13.00 169.00 1.314e+04 8.248e+01 [TFlops] (comp only) (9798967885824.00) 3072 3072 (mbk,nbk) 3584 3584 3584 4.710e-01 1.956e+02 8.727e-01 10 1.109e-02 1.139e-02 1.848e-01 2.641e-01 1.380e-04 4.715e-01 13.00 13.00 169.00 1.788e+04 8.419e+01 [TFlops] (comp only) (15560398077952.00) 3584 3584 (mbk,nbk) 4096 4096 4096 6.540e-01 2.102e+02 8.209e-01 10 1.205e-02 2.967e-02 2.657e-01 3.485e-01 5.500e-04 6.564e-01 13.00 13.00 169.00 6.309e+03 8.743e+01 [TFlops] (comp only) (23227183136768.00) 2048 2048 (mbk,nbk) 4608 4608 4608 1.160e+00 1.687e+02 5.856e-01 10 1.585e-02 4.241e-02 5.299e-01 5.739e-01 7.355e-04 1.163e+00 15.00 15.00 225.00 7.984e+03 8.309e+01 [TFlops] (comp only) (44030125670400.00) 2304 2304 (mbk,nbk) 5120 5120 5120 1.571e+00 1.709e+02 5.339e-01 10 3.634e-02 5.090e-02 7.372e-01 7.468e-01 8.261e-04 1.572e+00 16.00 15.00 240.00 9.857e+03 8.740e+01 [TFlops] (comp only) (64424509440000.00) 2560 2560 (mbk,nbk) 5632 5632 5632 2.100e+00 1.702e+02 4.834e-01 10 4.097e-02 6.294e-02 1.058e+00 9.327e-01 8.519e-04 2.096e+00 16.00 15.50 248.00 1.193e+04 8.371e+01 [TFlops] (comp only) (88607322800128.00) 2816 2816 (mbk,nbk) 6144 6144 6144 2.579e+00 1.799e+02 4.684e-01 10 4.473e-02 7.304e-02 1.346e+00 1.117e+00 8.526e-04 2.581e+00 16.00 15.50 248.00 1.419e+04 8.549e+01 [TFlops] (comp only) (115036404056064.00) 3072 3072 (mbk,nbk) 6656 6656 6656 3.276e+00 1.800e+02 4.327e-01 10 4.893e-02 8.782e-02 1.751e+00 1.384e+00 8.698e-04 3.273e+00 16.00 16.00 256.00 1.666e+04 8.621e+01 [TFlops] (comp only) (150976690388992.00) 3328 3328 (mbk,nbk) 7168 7168 7168 3.853e+00 1.912e+02 4.267e-01 10 5.367e-02 1.016e-01 2.112e+00 1.602e+00 9.334e-04 3.870e+00 16.00 16.00 256.00 1.932e+04 8.930e+01 [TFlops] (comp only) (188566244163584.00) 3584 3584 (mbk,nbk) 7680 7680 7680 5.071e+00 1.787e+02 3.722e-01 10 5.675e-02 2.254e-01 3.001e+00 1.790e+00 3.488e-03 5.077e+00 16.00 15.75 252.00 6.547e+03 7.607e+01 [TFlops] (comp only) (228304355328000.00) 1920 1920 (mbk,nbk) 8192 8192 8192 5.530e+00 1.989e+02 3.884e-01 10 6.091e-02 2.569e-01 3.158e+00 2.049e+00 3.465e-03 5.527e+00 16.00 15.75 252.00 7.449e+03 8.775e+01 [TFlops] (comp only) (277076930199552.00) 2048 2048 (mbk,nbk) 8704 8704 8704 6.727e+00 1.961e+02 3.604e-01 10 6.323e-02 2.887e-01 4.090e+00 2.265e+00 3.354e-03 6.711e+00 15.75 15.75 248.06 8.410e+03 7.999e+01 [TFlops] (comp only) (327141350703104.00) 2176 2176 (mbk,nbk) 9216 9216 9216 7.923e+00 1.976e+02 3.430e-01 10 6.645e-02 3.115e-01 5.065e+00 2.462e+00 3.410e-03 7.909e+00 15.50 15.75 244.12 9.428e+03 7.545e+01 [TFlops] (comp only) (382181490819072.00) 2304 2304 (mbk,nbk) 9728 9728 9728 8.775e+00 2.099e+02 3.451e-01 10 1.399e-01 3.452e-01 5.478e+00 2.809e+00 3.432e-03 8.775e+00 15.75 15.75 248.06 1.050e+04 8.338e+01 [TFlops] (comp only) (456757893660672.00) 2432 2432 (mbk,nbk) 10240 10240 10240 9.747e+00 2.204e+02 3.443e-01 10 1.471e-01 3.792e-01 6.048e+00 3.138e+00 3.542e-03 9.717e+00 16.00 15.75 252.00 1.164e+04 8.947e+01 [TFlops] (comp only) (541165879296000.00) 2560 2560 (mbk,nbk) gemmtc-1.3a/testing/isc2020_dat/_testing_dgemmtc-cr_phi1_perf.dat0000664000175000017500000001260313660237564025234 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 09:27:55 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: DGEMM # Library: GEMMTC # mode: performance evaluation # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 1.000000 # Degree = 0 # CRMode = 1 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 512 # dim_stop: 10240 # dim_step: 512 # M: range # N: range # K: range # cublaspointer: HOST # input range: max = 1.055e+02, min = 1.472e-09 # input range: max = 1.555e+02, min = 1.353e-09 # Evaluation result ------------------------ (* shows the value at the last execution, ** shows the value of the last block at the last execution) # M N K Time[sec] Perf[GFlops] Perf[GB/s] itr *t_SpltA *t_SpltB *t_Comp *t_Sum *t_Other *t_Total **#SpltA **#SpltB **#Comp **Mem[MB] TC[TFlops] 512 512 512 1.022e-02 2.634e+01 8.206e-01 10 1.147e-03 1.156e-03 2.610e-03 5.189e-03 9.894e-05 1.020e-02 11.00 11.00 121.00 3.650e+02 1.244e+01 [TFlops] (comp only) (32480690176.00) 512 512 (mbk,nbk) 1024 1024 1024 2.737e-02 7.858e+01 1.226e+00 10 1.600e-03 1.782e-03 5.531e-03 1.783e-02 1.135e-04 2.685e-02 12.00 12.00 144.00 1.460e+03 5.591e+01 [TFlops] (comp only) (309237645312.00) 1024 1024 (mbk,nbk) 1536 1536 1536 8.261e-02 8.782e+01 9.140e-01 10 2.667e-03 3.459e-03 1.898e-02 5.721e-02 1.864e-04 8.250e-02 14.00 14.00 196.00 3.284e+03 7.486e+01 [TFlops] (comp only) (1420560433152.00) 1536 1536 (mbk,nbk) 2048 2048 2048 1.534e-01 1.120e+02 8.747e-01 10 3.485e-03 5.167e-03 4.516e-02 9.960e-02 1.543e-04 1.536e-01 14.00 14.00 196.00 5.839e+03 7.457e+01 [TFlops] (comp only) (3367254360064.00) 2048 2048 (mbk,nbk) 2560 2560 2560 2.583e-01 1.300e+02 8.119e-01 10 8.090e-03 7.186e-03 8.400e-02 1.583e-01 1.626e-04 2.578e-01 14.00 14.00 196.00 9.123e+03 7.830e+01 [TFlops] (comp only) (6576668672000.00) 2560 2560 (mbk,nbk) 3072 3072 3072 3.852e-01 1.506e+02 7.841e-01 10 9.610e-03 9.545e-03 1.380e-01 2.298e-01 1.559e-04 3.872e-01 14.00 14.00 196.00 1.314e+04 8.233e+01 [TFlops] (comp only) (11364483465216.00) 3072 3072 (mbk,nbk) 3584 3584 3584 5.463e-01 1.686e+02 7.525e-01 10 1.179e-02 1.222e-02 2.139e-01 3.083e-01 1.712e-04 5.464e-01 14.00 14.00 196.00 1.788e+04 8.435e+01 [TFlops] (comp only) (18046378835968.00) 3584 3584 (mbk,nbk) 4096 4096 4096 7.546e-01 1.822e+02 7.115e-01 10 1.334e-02 3.254e-02 3.106e-01 3.978e-01 6.251e-04 7.549e-01 14.00 14.00 196.00 6.309e+03 8.674e+01 [TFlops] (comp only) (26938034880512.00) 2048 2048 (mbk,nbk) 4608 4608 4608 1.437e+00 1.363e+02 4.730e-01 10 1.822e-02 4.850e-02 6.624e-01 7.099e-01 9.418e-04 1.440e+00 16.50 17.00 280.50 7.984e+03 8.287e+01 [TFlops] (comp only) (54890890002432.00) 2304 2304 (mbk,nbk) 5120 5120 5120 1.782e+00 1.507e+02 4.708e-01 10 3.846e-02 5.662e-02 8.430e-01 8.446e-01 9.053e-04 1.784e+00 16.50 16.50 272.25 9.857e+03 8.669e+01 [TFlops] (comp only) (73081552896000.00) 2560 2560 (mbk,nbk) 5632 5632 5632 2.316e+00 1.543e+02 4.383e-01 10 4.329e-02 6.867e-02 1.167e+00 1.034e+00 9.336e-04 2.314e+00 16.50 16.50 272.25 1.193e+04 8.336e+01 [TFlops] (comp only) (97271546904576.00) 2816 2816 (mbk,nbk) 6144 6144 6144 3.010e+00 1.541e+02 4.013e-01 10 4.930e-02 8.136e-02 1.568e+00 1.317e+00 9.763e-04 3.017e+00 17.00 17.00 289.00 1.419e+04 8.550e+01 [TFlops] (comp only) (134054519242752.00) 3072 3072 (mbk,nbk) 6656 6656 6656 3.687e+00 1.600e+02 3.845e-01 10 5.245e-02 9.600e-02 1.988e+00 1.553e+00 9.875e-04 3.691e+00 17.00 17.00 289.00 1.666e+04 8.573e+01 [TFlops] (comp only) (170438529384448.00) 3328 3328 (mbk,nbk) 7168 7168 7168 4.363e+00 1.689e+02 3.769e-01 10 5.832e-02 1.108e-01 2.389e+00 1.823e+00 9.522e-04 4.382e+00 17.00 17.00 289.00 1.932e+04 8.911e+01 [TFlops] (comp only) (212873611575296.00) 3584 3584 (mbk,nbk) 7680 7680 7680 5.738e+00 1.579e+02 3.289e-01 10 6.341e-02 2.434e-01 3.399e+00 2.022e+00 3.654e-03 5.732e+00 17.00 16.75 284.75 6.547e+03 7.589e+01 [TFlops] (comp only) (257974861824000.00) 1920 1920 (mbk,nbk) 8192 8192 8192 6.288e+00 1.749e+02 3.415e-01 10 6.570e-02 2.774e-01 3.594e+00 2.322e+00 3.882e-03 6.263e+00 17.00 17.00 289.00 7.449e+03 8.842e+01 [TFlops] (comp only) (317758860427264.00) 2048 2048 (mbk,nbk) 8704 8704 8704 7.773e+00 1.697e+02 3.119e-01 10 7.001e-02 3.119e-01 4.758e+00 2.630e+00 4.236e-03 7.775e+00 17.00 17.00 289.00 8.410e+03 8.009e+01 [TFlops] (comp only) (381119694372864.00) 2176 2176 (mbk,nbk) 9216 9216 9216 9.314e+00 1.681e+02 2.918e-01 10 7.282e-02 3.384e-01 5.982e+00 2.918e+00 3.716e-03 9.315e+00 17.00 17.00 289.00 9.428e+03 7.564e+01 [TFlops] (comp only) (452434002444288.00) 2304 2304 (mbk,nbk) 9728 9728 9728 1.015e+01 1.813e+02 2.982e-01 10 1.513e-01 3.725e-01 6.375e+00 3.256e+00 3.843e-03 1.016e+01 17.00 17.00 289.00 1.050e+04 8.348e+01 [TFlops] (comp only) (532142958313472.00) 2432 2432 (mbk,nbk) 10240 10240 10240 1.110e+01 1.935e+02 3.023e-01 10 1.568e-01 4.131e-01 6.921e+00 3.605e+00 3.811e-03 1.110e+01 17.00 17.00 289.00 1.164e+04 8.967e+01 [TFlops] (comp only) (620622774272000.00) 2560 2560 (mbk,nbk) gemmtc-1.3a/testing/isc2020_dat/_testing_dgemmtc-cr_phi2_perf.dat0000664000175000017500000001260413660241573025231 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 09:43:17 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: DGEMM # Library: GEMMTC # mode: performance evaluation # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 2.000000 # Degree = 0 # CRMode = 1 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 512 # dim_stop: 10240 # dim_step: 512 # M: range # N: range # K: range # cublaspointer: HOST # input range: max = 2.226e+04, min = 2.328e-09 # input range: max = 4.836e+04, min = 4.367e-10 # Evaluation result ------------------------ (* shows the value at the last execution, ** shows the value of the last block at the last execution) # M N K Time[sec] Perf[GFlops] Perf[GB/s] itr *t_SpltA *t_SpltB *t_Comp *t_Sum *t_Other *t_Total **#SpltA **#SpltB **#Comp **Mem[MB] TC[TFlops] 512 512 512 1.161e-02 2.318e+01 7.222e-01 10 1.164e-03 1.152e-03 2.952e-03 5.673e-03 1.271e-04 1.107e-02 12.00 12.00 144.00 3.650e+02 1.310e+01 [TFlops] (comp only) (38654705664.00) 512 512 (mbk,nbk) 1024 1024 1024 2.998e-02 7.174e+01 1.119e+00 10 1.620e-03 1.909e-03 6.023e-03 1.931e-02 1.402e-04 2.900e-02 12.00 13.00 156.00 1.460e+03 5.562e+01 [TFlops] (comp only) (335007449088.00) 1024 1024 (mbk,nbk) 1536 1536 1536 1.021e-01 7.106e+01 7.395e-01 10 2.863e-03 3.872e-03 2.329e-02 7.206e-02 2.122e-04 1.023e-01 15.00 16.00 240.00 3.284e+03 7.470e+01 [TFlops] (comp only) (1739461754880.00) 1536 1536 (mbk,nbk) 2048 2048 2048 1.887e-01 9.110e+01 7.112e-01 10 3.735e-03 5.779e-03 5.497e-02 1.246e-01 2.084e-04 1.893e-01 15.00 16.00 240.00 5.839e+03 7.501e+01 [TFlops] (comp only) (4123168604160.00) 2048 2048 (mbk,nbk) 2560 2560 2560 3.174e-01 1.058e+02 6.608e-01 10 8.658e-03 8.100e-03 1.032e-01 1.972e-01 2.177e-04 3.174e-01 15.00 16.00 240.00 9.123e+03 7.802e+01 [TFlops] (comp only) (8053063680000.00) 2560 2560 (mbk,nbk) 3072 3072 3072 4.729e-01 1.227e+02 6.386e-01 10 1.037e-02 1.078e-02 1.677e-01 2.829e-01 2.139e-04 4.721e-01 15.00 16.00 240.00 1.314e+04 8.296e+01 [TFlops] (comp only) (13915694039040.00) 3072 3072 (mbk,nbk) 3584 3584 3584 6.691e-01 1.377e+02 6.143e-01 10 1.275e-02 1.401e-02 2.617e-01 3.852e-01 2.170e-04 6.738e-01 15.00 16.00 240.00 1.788e+04 8.445e+01 [TFlops] (comp only) (22097606737920.00) 3584 3584 (mbk,nbk) 4096 4096 4096 8.879e-01 1.549e+02 6.047e-01 10 1.468e-02 3.638e-02 3.715e-01 4.681e-01 8.159e-04 8.915e-01 15.00 15.50 232.50 6.309e+03 8.601e+01 [TFlops] (comp only) (31954556682240.00) 2048 2048 (mbk,nbk) 4608 4608 4608 1.642e+00 1.192e+02 4.138e-01 10 1.970e-02 5.171e-02 7.653e-01 8.057e-01 1.166e-03 1.644e+00 18.00 18.00 324.00 7.984e+03 8.284e+01 [TFlops] (comp only) (63403380965376.00) 2304 2304 (mbk,nbk) 5120 5120 5120 2.114e+00 1.270e+02 3.968e-01 10 4.255e-02 6.232e-02 1.001e+00 1.013e+00 1.139e-03 2.121e+00 18.00 18.00 324.00 9.857e+03 8.685e+01 [TFlops] (comp only) (86973087744000.00) 2560 2560 (mbk,nbk) 5632 5632 5632 2.669e+00 1.339e+02 3.803e-01 10 4.589e-02 7.480e-02 1.353e+00 1.197e+00 1.166e-03 2.672e+00 17.50 18.00 315.00 1.193e+04 8.316e+01 [TFlops] (comp only) (112545591459840.00) 2816 2816 (mbk,nbk) 6144 6144 6144 3.280e+00 1.414e+02 3.683e-01 10 5.084e-02 8.771e-02 1.725e+00 1.427e+00 1.168e-03 3.291e+00 17.50 18.00 315.00 1.419e+04 8.470e+01 [TFlops] (comp only) (146114787409920.00) 3072 3072 (mbk,nbk) 6656 6656 6656 4.003e+00 1.473e+02 3.541e-01 10 5.578e-02 1.027e-01 2.165e+00 1.673e+00 1.147e-03 3.998e+00 17.50 18.00 315.00 1.666e+04 8.582e+01 [TFlops] (comp only) (185772099502080.00) 3328 3328 (mbk,nbk) 7168 7168 7168 4.881e+00 1.509e+02 3.368e-01 10 6.218e-02 1.158e-01 2.667e+00 2.001e+00 1.200e-03 4.848e+00 18.00 18.00 324.00 1.932e+04 8.948e+01 [TFlops] (comp only) (238654152769536.00) 3584 3584 (mbk,nbk) 7680 7680 7680 6.363e+00 1.424e+02 2.966e-01 10 6.509e-02 2.647e-01 3.742e+00 2.267e+00 4.554e-03 6.343e+00 17.75 18.00 319.50 6.547e+03 7.735e+01 [TFlops] (comp only) (289459656458240.00) 1920 1920 (mbk,nbk) 8192 8192 8192 6.987e+00 1.574e+02 3.074e-01 10 6.804e-02 2.936e-01 4.018e+00 2.582e+00 4.739e-03 6.966e+00 18.00 18.00 324.00 7.449e+03 8.867e+01 [TFlops] (comp only) (356241767399424.00) 2048 2048 (mbk,nbk) 8704 8704 8704 8.660e+00 1.523e+02 2.799e-01 10 7.347e-02 3.310e-01 5.322e+00 2.924e+00 4.797e-03 8.655e+00 18.00 18.00 324.00 8.410e+03 8.029e+01 [TFlops] (comp only) (427269117968384.00) 2176 2176 (mbk,nbk) 9216 9216 9216 1.040e+01 1.505e+02 2.612e-01 10 7.748e-02 3.628e-01 6.701e+00 3.258e+00 4.642e-03 1.040e+01 18.00 18.00 324.00 9.428e+03 7.570e+01 [TFlops] (comp only) (507227047723008.00) 2304 2304 (mbk,nbk) 9728 9728 9728 1.136e+01 1.622e+02 2.667e-01 10 1.616e-01 3.996e-01 7.135e+00 3.636e+00 4.711e-03 1.134e+01 18.00 18.00 324.00 1.050e+04 8.362e+01 [TFlops] (comp only) (596594311299072.00) 2432 2432 (mbk,nbk) 10240 10240 10240 1.241e+01 1.731e+02 2.704e-01 10 1.680e-01 4.396e-01 7.757e+00 4.029e+00 4.704e-03 1.240e+01 18.00 18.00 324.00 1.164e+04 8.969e+01 [TFlops] (comp only) (695784701952000.00) 2560 2560 (mbk,nbk) gemmtc-1.3a/testing/isc2020_dat/_testing_dgemmtc-dp_phi01_perf.dat0000664000175000017500000001253713660233320025303 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 09:03:13 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: DGEMM # Library: GEMMTC # mode: performance evaluation # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 0.100000 # Degree = 0 # CRMode = 0 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 512 # dim_stop: 10240 # dim_step: 512 # M: range # N: range # K: range # cublaspointer: HOST # input range: max = 8.539e-01, min = 9.750e-10 # input range: max = 8.877e-01, min = 1.991e-09 # Evaluation result ------------------------ (* shows the value at the last execution, ** shows the value of the last block at the last execution) # M N K Time[sec] Perf[GFlops] Perf[GB/s] itr *t_SpltA *t_SpltB *t_Comp *t_Sum *t_Other *t_Total **#SpltA **#SpltB **#Comp **Mem[MB] TC[TFlops] 512 512 512 3.591e-03 7.497e+01 2.336e+00 10 1.230e-03 9.542e-04 9.818e-04 3.970e-04 3.099e-05 3.594e-03 9.00 9.00 45.00 3.650e+02 1.230e+01 [TFlops] (comp only) (12079595520.00) 512 512 (mbk,nbk) 1024 1024 1024 7.818e-03 2.751e+02 4.292e+00 10 2.160e-03 1.876e-03 2.191e-03 1.468e-03 4.077e-05 7.736e-03 9.00 9.00 45.00 1.460e+03 4.410e+01 [TFlops] (comp only) (96636764160.00) 1024 1024 (mbk,nbk) 1536 1536 1536 1.519e-02 4.775e+02 4.969e+00 10 2.772e-03 2.535e-03 5.261e-03 2.896e-03 4.983e-05 1.351e-02 10.00 10.00 55.00 3.284e+03 7.577e+01 [TFlops] (comp only) (398626652160.00) 1536 1536 (mbk,nbk) 2048 2048 2048 2.520e-02 6.823e+02 5.326e+00 10 3.915e-03 3.785e-03 1.216e-02 5.117e-03 4.816e-05 2.503e-02 10.00 10.00 55.00 5.839e+03 7.767e+01 [TFlops] (comp only) (944892805120.00) 2048 2048 (mbk,nbk) 2560 2560 2560 4.597e-02 7.304e+02 4.562e+00 10 9.396e-03 5.819e-03 2.245e-02 8.745e-03 4.792e-05 4.646e-02 10.00 10.00 55.00 9.123e+03 8.220e+01 [TFlops] (comp only) (1845493760000.00) 2560 2560 (mbk,nbk) 3072 3072 3072 7.043e-02 8.237e+02 4.288e+00 10 1.382e-02 7.585e-03 3.669e-02 1.262e-02 5.078e-05 7.077e-02 10.00 10.00 55.00 1.314e+04 8.692e+01 [TFlops] (comp only) (3189013217280.00) 3072 3072 (mbk,nbk) 3584 3584 3584 1.033e-01 8.917e+02 3.979e+00 10 1.841e-02 9.504e-03 5.584e-02 1.918e-02 4.625e-05 1.030e-01 10.00 10.00 55.00 1.788e+04 9.069e+01 [TFlops] (comp only) (5064034877440.00) 3584 3584 (mbk,nbk) 4096 4096 4096 1.451e-01 9.474e+02 3.699e+00 10 1.527e-02 2.446e-02 8.360e-02 2.170e-02 1.860e-04 1.452e-01 10.00 10.00 55.00 6.309e+03 9.042e+01 [TFlops] (comp only) (7559142440960.00) 2048 2048 (mbk,nbk) 4608 4608 4608 2.777e-01 7.048e+02 2.446e+00 10 2.221e-02 3.857e-02 1.739e-01 4.252e-02 2.756e-04 2.775e-01 12.00 12.00 78.00 7.984e+03 8.778e+01 [TFlops] (comp only) (15263776899072.00) 2304 2304 (mbk,nbk) 5120 5120 5120 3.128e-01 8.585e+02 2.682e+00 10 3.795e-02 3.949e-02 1.874e-01 4.119e-02 2.165e-04 3.062e-01 11.00 11.00 66.00 9.857e+03 9.455e+01 [TFlops] (comp only) (17716740096000.00) 2560 2560 (mbk,nbk) 5632 5632 5632 4.170e-01 8.570e+02 2.434e+00 10 4.387e-02 4.958e-02 2.681e-01 5.518e-02 2.327e-04 4.170e-01 11.00 11.00 66.00 1.193e+04 8.795e+01 [TFlops] (comp only) (23580981067776.00) 2816 2816 (mbk,nbk) 6144 6144 6144 5.160e-01 8.992e+02 2.341e+00 10 5.129e-02 5.570e-02 3.433e-01 6.602e-02 2.303e-04 5.166e-01 11.00 11.00 66.00 1.419e+04 8.917e+01 [TFlops] (comp only) (30614526885888.00) 3072 3072 (mbk,nbk) 6656 6656 6656 6.376e-01 9.251e+02 2.223e+00 10 5.690e-02 6.356e-02 4.405e-01 7.718e-02 2.260e-04 6.384e-01 11.00 11.00 66.00 1.666e+04 8.836e+01 [TFlops] (comp only) (38923677990912.00) 3328 3328 (mbk,nbk) 7168 7168 7168 7.517e-01 9.801e+02 2.187e+00 10 6.309e-02 7.168e-02 5.298e-01 8.991e-02 2.165e-04 7.547e-01 11.00 11.00 66.00 1.932e+04 9.176e+01 [TFlops] (comp only) (48614734823424.00) 3584 3584 (mbk,nbk) 7680 7680 7680 1.077e+00 8.410e+02 1.752e+00 10 6.671e-02 1.690e-01 7.457e-01 9.636e-02 9.382e-04 1.079e+00 11.00 11.00 66.00 6.547e+03 8.018e+01 [TFlops] (comp only) (59793997824000.00) 1920 1920 (mbk,nbk) 8192 8192 8192 1.175e+00 9.358e+02 1.827e+00 10 7.330e-02 1.951e-01 7.949e-01 1.129e-01 8.891e-04 1.177e+00 11.00 11.00 66.00 7.449e+03 9.130e+01 [TFlops] (comp only) (72567767433216.00) 2048 2048 (mbk,nbk) 8704 8704 8704 1.500e+00 8.793e+02 1.616e+00 10 8.085e-02 2.233e-01 1.065e+00 1.325e-01 9.093e-04 1.503e+00 11.00 11.00 66.00 8.410e+03 8.170e+01 [TFlops] (comp only) (87042344091648.00) 2176 2176 (mbk,nbk) 9216 9216 9216 1.811e+00 8.648e+02 1.501e+00 10 9.196e-02 2.311e-01 1.348e+00 1.397e-01 9.048e-04 1.811e+00 11.00 11.00 66.00 9.428e+03 7.668e+01 [TFlops] (comp only) (103324028239872.00) 2304 2304 (mbk,nbk) 9728 9728 9728 2.016e+00 9.133e+02 1.502e+00 10 1.483e-01 2.615e-01 1.445e+00 1.620e-01 9.162e-04 2.018e+00 11.00 11.00 66.00 1.050e+04 8.407e+01 [TFlops] (comp only) (121519120318464.00) 2432 2432 (mbk,nbk) 10240 10240 10240 2.209e+00 9.724e+02 1.519e+00 10 1.608e-01 2.899e-01 1.578e+00 1.819e-01 9.012e-04 2.211e+00 11.00 11.00 66.00 1.164e+04 8.983e+01 [TFlops] (comp only) (141733920768000.00) 2560 2560 (mbk,nbk) gemmtc-1.3a/testing/isc2020_dat/_testing_dgemmtc-dp_phi1_perf.dat0000664000175000017500000001254613660233664025236 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 09:06:26 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: DGEMM # Library: GEMMTC # mode: performance evaluation # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 1.000000 # Degree = 0 # CRMode = 0 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 512 # dim_stop: 10240 # dim_step: 512 # M: range # N: range # K: range # cublaspointer: HOST # input range: max = 1.055e+02, min = 1.472e-09 # input range: max = 1.555e+02, min = 1.353e-09 # Evaluation result ------------------------ (* shows the value at the last execution, ** shows the value of the last block at the last execution) # M N K Time[sec] Perf[GFlops] Perf[GB/s] itr *t_SpltA *t_SpltB *t_Comp *t_Sum *t_Other *t_Total **#SpltA **#SpltB **#Comp **Mem[MB] TC[TFlops] 512 512 512 4.175e-03 6.449e+01 2.009e+00 10 1.388e-03 1.054e-03 1.207e-03 4.830e-04 4.792e-05 4.180e-03 10.00 10.00 55.00 3.650e+02 1.223e+01 [TFlops] (comp only) (14763950080.00) 512 512 (mbk,nbk) 1024 1024 1024 9.088e-03 2.367e+02 3.692e+00 10 2.416e-03 2.095e-03 2.677e-03 1.798e-03 4.625e-05 9.032e-03 10.00 10.00 55.00 1.460e+03 4.413e+01 [TFlops] (comp only) (118111600640.00) 1024 1024 (mbk,nbk) 1536 1536 1536 1.712e-02 4.237e+02 4.409e+00 10 3.107e-03 2.857e-03 6.389e-03 3.510e-03 5.269e-05 1.592e-02 11.00 11.00 66.00 3.284e+03 7.487e+01 [TFlops] (comp only) (478351982592.00) 1536 1536 (mbk,nbk) 2048 2048 2048 3.024e-02 5.685e+02 4.438e+00 10 4.303e-03 4.376e-03 1.505e-02 6.317e-03 5.436e-05 3.010e-02 11.00 11.00 66.00 5.839e+03 7.535e+01 [TFlops] (comp only) (1133871366144.00) 2048 2048 (mbk,nbk) 2560 2560 2560 5.502e-02 6.102e+02 3.811e+00 10 1.073e-02 6.319e-03 2.722e-02 1.043e-02 6.342e-05 5.477e-02 11.00 11.00 66.00 9.123e+03 8.136e+01 [TFlops] (comp only) (2214592512000.00) 2560 2560 (mbk,nbk) 3072 3072 3072 8.452e-02 6.863e+02 3.573e+00 10 1.551e-02 8.429e-03 4.486e-02 1.569e-02 5.937e-05 8.454e-02 11.00 11.00 66.00 1.314e+04 8.531e+01 [TFlops] (comp only) (3826815860736.00) 3072 3072 (mbk,nbk) 3584 3584 3584 1.237e-01 7.449e+02 3.324e+00 10 2.080e-02 1.053e-02 6.933e-02 2.356e-02 7.010e-05 1.243e-01 11.00 11.00 66.00 1.788e+04 8.765e+01 [TFlops] (comp only) (6076841852928.00) 3584 3584 (mbk,nbk) 4096 4096 4096 1.743e-01 7.890e+02 3.081e+00 10 1.717e-02 2.763e-02 1.022e-01 2.642e-02 2.556e-04 1.736e-01 11.00 11.00 66.00 6.309e+03 8.878e+01 [TFlops] (comp only) (9070970929152.00) 2048 2048 (mbk,nbk) 4608 4608 4608 3.258e-01 6.008e+02 2.086e+00 10 2.461e-02 4.208e-02 2.080e-01 5.246e-02 3.331e-04 3.275e-01 13.00 13.00 91.00 7.984e+03 8.561e+01 [TFlops] (comp only) (17807739715584.00) 2304 2304 (mbk,nbk) 5120 5120 5120 4.283e-01 6.269e+02 1.958e+00 10 4.495e-02 4.829e-02 2.732e-01 6.225e-02 3.488e-04 4.290e-01 13.00 13.00 91.00 9.857e+03 8.941e+01 [TFlops] (comp only) (24427626496000.00) 2560 2560 (mbk,nbk) 5632 5632 5632 5.693e-01 6.278e+02 1.783e+00 10 5.170e-02 5.790e-02 3.819e-01 7.786e-02 3.805e-04 5.697e-01 13.00 13.00 91.00 1.193e+04 8.514e+01 [TFlops] (comp only) (32513170866176.00) 2816 2816 (mbk,nbk) 6144 6144 6144 7.045e-01 6.586e+02 1.715e+00 10 5.970e-02 6.536e-02 4.878e-01 9.212e-02 3.936e-04 7.054e-01 13.00 13.00 91.00 1.419e+04 8.653e+01 [TFlops] (comp only) (42210938585088.00) 3072 3072 (mbk,nbk) 6656 6656 6656 7.541e-01 7.822e+02 1.880e+00 10 6.198e-02 7.008e-02 5.283e-01 9.332e-02 3.273e-04 7.540e-01 12.00 12.00 78.00 1.666e+04 8.708e+01 [TFlops] (comp only) (46000710352896.00) 3328 3328 (mbk,nbk) 7168 7168 7168 8.921e-01 8.259e+02 1.843e+00 10 6.850e-02 7.818e-02 6.362e-01 1.089e-01 3.283e-04 8.920e-01 12.00 12.00 78.00 1.932e+04 9.031e+01 [TFlops] (comp only) (57453777518592.00) 3584 3584 (mbk,nbk) 7680 7680 7680 1.277e+00 7.094e+02 1.478e+00 10 7.333e-02 1.891e-01 8.966e-01 1.184e-01 1.162e-03 1.279e+00 12.00 12.00 78.00 6.547e+03 7.881e+01 [TFlops] (comp only) (70665633792000.00) 1920 1920 (mbk,nbk) 8192 8192 8192 1.389e+00 7.919e+02 1.546e+00 10 8.023e-02 2.176e-01 9.510e-01 1.387e-01 1.185e-03 1.389e+00 12.00 12.00 78.00 7.449e+03 9.018e+01 [TFlops] (comp only) (85761906966528.00) 2048 2048 (mbk,nbk) 8704 8704 8704 1.767e+00 7.463e+02 1.372e+00 10 8.780e-02 2.461e-01 1.273e+00 1.602e-01 1.249e-03 1.769e+00 12.00 12.00 78.00 8.410e+03 8.078e+01 [TFlops] (comp only) (102868224835584.00) 2176 2176 (mbk,nbk) 9216 9216 9216 2.157e+00 7.259e+02 1.260e+00 10 9.732e-02 2.579e-01 1.633e+00 1.699e-01 1.247e-03 2.159e+00 12.00 12.00 78.00 9.428e+03 7.478e+01 [TFlops] (comp only) (122110215192576.00) 2304 2304 (mbk,nbk) 9728 9728 9728 2.384e+00 7.725e+02 1.270e+00 10 1.631e-01 2.903e-01 1.736e+00 1.963e-01 1.276e-03 2.387e+00 12.00 12.00 78.00 1.050e+04 8.273e+01 [TFlops] (comp only) (143613505830912.00) 2432 2432 (mbk,nbk) 10240 10240 10240 2.610e+00 8.230e+02 1.286e+00 10 1.758e-01 3.214e-01 1.899e+00 2.209e-01 1.285e-03 2.618e+00 12.00 12.00 78.00 1.164e+04 8.822e+01 [TFlops] (comp only) (167503724544000.00) 2560 2560 (mbk,nbk) gemmtc-1.3a/testing/isc2020_dat/_testing_dgemmtc-dp_phi2_perf.dat0000664000175000017500000001254713660234264025235 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 09:10:14 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: DGEMM # Library: GEMMTC # mode: performance evaluation # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 2.000000 # Degree = 0 # CRMode = 0 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 512 # dim_stop: 10240 # dim_step: 512 # M: range # N: range # K: range # cublaspointer: HOST # input range: max = 2.226e+04, min = 2.328e-09 # input range: max = 4.836e+04, min = 4.367e-10 # Evaluation result ------------------------ (* shows the value at the last execution, ** shows the value of the last block at the last execution) # M N K Time[sec] Perf[GFlops] Perf[GB/s] itr *t_SpltA *t_SpltB *t_Comp *t_Sum *t_Other *t_Total **#SpltA **#SpltB **#Comp **Mem[MB] TC[TFlops] 512 512 512 4.180e-03 6.441e+01 2.007e+00 10 1.383e-03 1.068e-03 1.203e-03 4.811e-04 4.601e-05 4.181e-03 10.00 10.00 55.00 3.650e+02 1.227e+01 [TFlops] (comp only) (14763950080.00) 512 512 (mbk,nbk) 1024 1024 1024 9.069e-03 2.371e+02 3.700e+00 10 2.421e-03 2.096e-03 2.679e-03 1.792e-03 4.768e-05 9.036e-03 10.00 10.00 55.00 1.460e+03 4.408e+01 [TFlops] (comp only) (118111600640.00) 1024 1024 (mbk,nbk) 1536 1536 1536 1.614e-02 4.495e+02 4.678e+00 10 3.111e-03 2.890e-03 6.430e-03 3.621e-03 5.484e-05 1.611e-02 11.00 11.00 66.00 3.284e+03 7.439e+01 [TFlops] (comp only) (478351982592.00) 1536 1536 (mbk,nbk) 2048 2048 2048 3.048e-02 5.640e+02 4.403e+00 10 4.364e-03 4.461e-03 1.522e-02 6.521e-03 5.054e-05 3.062e-02 11.00 11.00 66.00 5.839e+03 7.448e+01 [TFlops] (comp only) (1133871366144.00) 2048 2048 (mbk,nbk) 2560 2560 2560 5.538e-02 6.062e+02 3.787e+00 10 1.048e-02 6.476e-03 2.725e-02 1.105e-02 5.150e-05 5.532e-02 11.00 11.00 66.00 9.123e+03 8.127e+01 [TFlops] (comp only) (2214592512000.00) 2560 2560 (mbk,nbk) 3072 3072 3072 8.468e-02 6.851e+02 3.566e+00 10 1.551e-02 8.274e-03 4.489e-02 1.639e-02 5.603e-05 8.511e-02 11.00 11.00 66.00 1.314e+04 8.525e+01 [TFlops] (comp only) (3826815860736.00) 3072 3072 (mbk,nbk) 3584 3584 3584 1.243e-01 7.412e+02 3.308e+00 10 2.017e-02 1.070e-02 6.918e-02 2.386e-02 5.651e-05 1.240e-01 11.00 11.00 66.00 1.788e+04 8.785e+01 [TFlops] (comp only) (6076841852928.00) 3584 3584 (mbk,nbk) 4096 4096 4096 1.752e-01 7.845e+02 3.063e+00 10 1.712e-02 2.817e-02 1.025e-01 2.746e-02 2.120e-04 1.754e-01 11.00 11.00 66.00 6.309e+03 8.851e+01 [TFlops] (comp only) (9070970929152.00) 2048 2048 (mbk,nbk) 4608 4608 4608 3.278e-01 5.971e+02 2.073e+00 10 2.482e-02 4.239e-02 2.080e-01 5.330e-02 2.980e-04 3.289e-01 13.00 13.00 91.00 7.984e+03 8.560e+01 [TFlops] (comp only) (17807739715584.00) 2304 2304 (mbk,nbk) 5120 5120 5120 4.296e-01 6.250e+02 1.953e+00 10 4.498e-02 4.871e-02 2.722e-01 6.318e-02 3.152e-04 4.294e-01 13.00 13.00 91.00 9.857e+03 8.974e+01 [TFlops] (comp only) (24427626496000.00) 2560 2560 (mbk,nbk) 5632 5632 5632 5.703e-01 6.267e+02 1.780e+00 10 5.156e-02 5.834e-02 3.813e-01 7.915e-02 3.045e-04 5.706e-01 13.00 13.00 91.00 1.193e+04 8.528e+01 [TFlops] (comp only) (32513170866176.00) 2816 2816 (mbk,nbk) 6144 6144 6144 7.057e-01 6.574e+02 1.712e+00 10 5.925e-02 6.603e-02 4.874e-01 9.336e-02 3.037e-04 7.064e-01 13.00 13.00 91.00 1.419e+04 8.660e+01 [TFlops] (comp only) (42210938585088.00) 3072 3072 (mbk,nbk) 6656 6656 6656 8.712e-01 6.771e+02 1.627e+00 10 6.674e-02 7.545e-02 6.188e-01 1.096e-01 2.964e-04 8.709e-01 13.00 13.00 91.00 1.666e+04 8.672e+01 [TFlops] (comp only) (53667495411712.00) 3328 3328 (mbk,nbk) 7168 7168 7168 1.030e+00 7.152e+02 1.596e+00 10 7.327e-02 8.450e-02 7.448e-01 1.275e-01 3.040e-04 1.030e+00 13.00 13.00 91.00 1.932e+04 8.999e+01 [TFlops] (comp only) (67029407105024.00) 3584 3584 (mbk,nbk) 7680 7680 7680 1.483e+00 6.111e+02 1.273e+00 10 7.918e-02 2.058e-01 1.061e+00 1.405e-01 1.168e-03 1.488e+00 13.00 13.00 91.00 6.547e+03 7.768e+01 [TFlops] (comp only) (82443239424000.00) 1920 1920 (mbk,nbk) 8192 8192 8192 1.618e+00 6.795e+02 1.327e+00 10 8.793e-02 2.392e-01 1.143e+00 1.642e-01 1.173e-03 1.635e+00 13.00 13.00 91.00 7.449e+03 8.754e+01 [TFlops] (comp only) (100055558127616.00) 2048 2048 (mbk,nbk) 8704 8704 8704 2.057e+00 6.412e+02 1.178e+00 10 9.479e-02 2.672e-01 1.510e+00 1.899e-01 1.261e-03 2.063e+00 13.00 13.00 91.00 8.410e+03 7.949e+01 [TFlops] (comp only) (120012928974848.00) 2176 2176 (mbk,nbk) 9216 9216 9216 2.520e+00 6.214e+02 1.079e+00 10 1.045e-01 2.822e-01 1.929e+00 2.037e-01 1.230e-03 2.521e+00 13.00 13.00 91.00 9.428e+03 7.385e+01 [TFlops] (comp only) (142461917724672.00) 2304 2304 (mbk,nbk) 9728 9728 9728 2.764e+00 6.663e+02 1.096e+00 10 1.770e-01 3.150e-01 2.044e+00 2.317e-01 1.206e-03 2.769e+00 13.00 13.00 91.00 1.050e+04 8.197e+01 [TFlops] (comp only) (167549090136064.00) 2432 2432 (mbk,nbk) 10240 10240 10240 3.012e+00 7.131e+02 1.114e+00 10 1.876e-01 3.466e-01 2.214e+00 2.583e-01 1.204e-03 3.008e+00 13.00 13.00 91.00 1.164e+04 8.825e+01 [TFlops] (comp only) (195421011968000.00) 2560 2560 (mbk,nbk) gemmtc-1.3a/testing/isc2020_dat/_testing_dgemmtc_phi01_check.dat0000664000175000017500000000305513660242265025026 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 10:05:39 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: DGEMM # Library: GEMMTC # mode: checking with MPFR 2048-bit # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 0.100000 # Degree = 0 # CRMode = 0 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 128 # dim_stop: 10240 # dim_step: 128 # M: 32 # N: 32 # K: range # cublaspointer: HOST # input range: max = 7.900e-01, min = 4.144e-07 # input range: max = 7.961e-01, min = 4.556e-07 # Evaluation result ------------------------ # M N K ERROR #ofERROR: (1)cuBLAS, (2)DGEMM-TC-DP, (3)DGEMM-TC-CR 32 32 128 2.645854e-12 884 1.906209e-14 701 0.000000e+00 0 32 32 256 4.554028e-13 941 9.710082e-16 703 0.000000e+00 0 32 32 512 1.029221e-13 956 8.787555e-16 715 0.000000e+00 0 32 32 1024 1.105751e-12 962 9.362914e-16 719 0.000000e+00 0 32 32 2048 4.787853e-12 958 4.695328e-14 779 0.000000e+00 0 32 32 4096 6.766987e-13 957 5.936555e-15 761 0.000000e+00 0 32 32 8192 1.848163e-11 977 5.110989e-12 871 0.000000e+00 0 gemmtc-1.3a/testing/isc2020_dat/_testing_dgemmtc_phi1_check.dat0000664000175000017500000000305513660242267024750 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 10:05:42 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: DGEMM # Library: GEMMTC # mode: checking with MPFR 2048-bit # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 1.000000 # Degree = 0 # CRMode = 0 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 128 # dim_stop: 10240 # dim_step: 128 # M: 32 # N: 32 # K: range # cublaspointer: HOST # input range: max = 4.852e+01, min = 2.558e-07 # input range: max = 5.233e+01, min = 2.034e-07 # Evaluation result ------------------------ # M N K ERROR #ofERROR: (1)cuBLAS, (2)DGEMM-TC-DP, (3)DGEMM-TC-CR 32 32 128 1.356540e-13 863 4.635103e-15 704 0.000000e+00 0 32 32 256 4.188055e-12 908 2.078753e-14 695 0.000000e+00 0 32 32 512 2.538804e-13 930 1.391876e-14 742 0.000000e+00 0 32 32 1024 5.452898e-13 942 4.679311e-14 726 0.000000e+00 0 32 32 2048 4.115042e-14 913 3.748963e-15 719 0.000000e+00 0 32 32 4096 1.148332e-14 878 1.625516e-15 761 0.000000e+00 0 32 32 8192 1.469468e-14 897 6.223357e-15 833 0.000000e+00 0 gemmtc-1.3a/testing/isc2020_dat/_testing_dgemmtc_phi2_check.dat0000664000175000017500000000305513660242272024745 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 10:05:45 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: DGEMM # Library: GEMMTC # mode: checking with MPFR 2048-bit # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 2.000000 # Degree = 0 # CRMode = 0 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 128 # dim_stop: 10240 # dim_step: 128 # M: 32 # N: 32 # K: range # cublaspointer: HOST # input range: max = 4.709e+03, min = 1.497e-07 # input range: max = 5.478e+03, min = 8.303e-08 # Evaluation result ------------------------ # M N K ERROR #ofERROR: (1)cuBLAS, (2)DGEMM-TC-DP, (3)DGEMM-TC-CR 32 32 128 3.964229e-14 843 3.761647e-15 712 0.000000e+00 0 32 32 256 2.680740e-14 885 1.481108e-13 690 0.000000e+00 0 32 32 512 1.573645e-12 912 4.794855e-14 745 0.000000e+00 0 32 32 1024 9.901528e-13 910 1.529504e-14 741 0.000000e+00 0 32 32 2048 4.090601e-15 882 9.254196e-16 768 0.000000e+00 0 32 32 4096 2.136552e-15 869 1.418668e-15 768 0.000000e+00 0 32 32 8192 2.678222e-15 872 9.823080e-16 788 0.000000e+00 0 gemmtc-1.3a/testing/isc2020_dat/_testing_sgemm_cublas_perf.dat0000664000175000017500000000575513660242332024731 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 10:05:47 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: SGEMM # Library: GEMMTC # mode: performance evaluation # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 0.100000 # Degree = 1 # CRMode = 0 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 512 # dim_stop: 10240 # dim_step: 512 # M: range # N: range # K: range # cublaspointer: HOST # input range: max = 8.539e-01, min = 9.750e-10 # input range: max = 8.877e-01, min = 1.991e-09 # Evaluation result ------------------------ (* shows the value at the last execution, ** shows the value of the last block at the last execution) # M N K Time[sec] Perf[GFlops] Perf[GB/s] itr **#SpltA **#SpltB **#Comp 512 512 512 3.700e-05 7.276e+03 1.134e+02 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 1024 1024 1024 2.125e-04 1.012e+04 7.895e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 1536 1536 1536 6.267e-04 1.158e+04 6.023e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 2048 2048 2048 1.537e-03 1.119e+04 4.366e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 2560 2560 2560 2.913e-03 1.153e+04 3.600e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 3072 3072 3072 4.135e-03 1.403e+04 3.651e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 3584 3584 3584 6.764e-03 1.362e+04 3.039e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 4096 4096 4096 1.028e-02 1.338e+04 2.612e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 4608 4608 4608 1.480e-02 1.323e+04 2.296e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 5120 5120 5120 2.025e-02 1.326e+04 2.071e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 5632 5632 5632 2.678e-02 1.334e+04 1.895e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 6144 6144 6144 3.477e-02 1.335e+04 1.737e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 6656 6656 6656 4.428e-02 1.332e+04 1.601e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 7168 7168 7168 5.579e-02 1.321e+04 1.474e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 7680 7680 7680 6.871e-02 1.319e+04 1.373e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 8192 8192 8192 8.091e-02 1.359e+04 1.327e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 8704 8704 8704 9.701e-02 1.360e+04 1.250e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 9216 9216 9216 1.150e-01 1.362e+04 1.182e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 9728 9728 9728 1.350e-01 1.364e+04 1.121e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) 10240 10240 10240 1.589e-01 1.351e+04 1.056e+01 10 0.00 0.00 0.00 -nan [TFlops] (comp only) gemmtc-1.3a/testing/isc2020_dat/_testing_sgemmtc-sp_phi01_perf.dat0000664000175000017500000001246013660242423025340 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 10:06:20 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: SGEMM # Library: GEMMTC # mode: performance evaluation # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 0.100000 # Degree = 0 # CRMode = 0 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 512 # dim_stop: 10240 # dim_step: 512 # M: range # N: range # K: range # cublaspointer: HOST # input range: max = 8.539e-01, min = 9.750e-10 # input range: max = 8.877e-01, min = 1.991e-09 # Evaluation result ------------------------ (* shows the value at the last execution, ** shows the value of the last block at the last execution) # M N K Time[sec] Perf[GFlops] Perf[GB/s] itr *t_SpltA *t_SpltB *t_Comp *t_Sum *t_Other *t_Total **#SpltA **#SpltB **#Comp **Mem[MB] TC[TFlops] 512 512 512 8.439e-04 3.190e+02 4.970e+00 10 2.770e-04 1.640e-04 3.273e-04 6.199e-05 1.454e-05 8.450e-04 5.00 5.00 15.00 3.618e+02 1.230e+01 [TFlops] (comp only) (4026531840.00) 512 512 (mbk,nbk) 1024 1024 1024 1.300e-03 1.654e+03 1.290e+01 10 4.451e-04 2.129e-04 5.002e-04 1.259e-04 9.775e-06 1.294e-03 4.00 4.00 10.00 1.447e+03 4.293e+01 [TFlops] (comp only) (21474836480.00) 1024 1024 (mbk,nbk) 1536 1536 1536 3.527e-03 2.057e+03 1.070e+01 10 8.349e-04 4.182e-04 1.885e-03 3.748e-04 1.645e-05 3.529e-03 5.00 5.00 15.00 3.256e+03 5.768e+01 [TFlops] (comp only) (108716359680.00) 1536 1536 (mbk,nbk) 2048 2048 2048 6.944e-03 2.476e+03 9.665e+00 10 1.377e-03 6.120e-04 4.238e-03 6.430e-04 1.383e-05 6.884e-03 5.00 5.00 15.00 5.788e+03 6.080e+01 [TFlops] (comp only) (257698037760.00) 2048 2048 (mbk,nbk) 2560 2560 2560 1.143e-02 2.936e+03 9.171e+00 10 2.625e-03 7.901e-04 6.231e-03 8.910e-04 1.478e-05 1.055e-02 5.00 5.00 15.00 9.044e+03 8.078e+01 [TFlops] (comp only) (503316480000.00) 2560 2560 (mbk,nbk) 3072 3072 3072 1.670e-02 3.474e+03 9.043e+00 10 4.039e-03 1.105e-03 1.024e-02 1.273e-03 1.550e-05 1.667e-02 5.00 5.00 15.00 1.302e+04 8.495e+01 [TFlops] (comp only) (869730877440.00) 3072 3072 (mbk,nbk) 3584 3584 3584 2.469e-02 3.731e+03 8.325e+00 10 5.766e-03 1.484e-03 1.540e-02 1.746e-03 1.478e-05 2.442e-02 5.00 5.00 15.00 1.773e+04 8.965e+01 [TFlops] (comp only) (1381100421120.00) 3584 3584 (mbk,nbk) 4096 4096 4096 3.482e-02 3.949e+03 7.710e+00 10 5.064e-03 3.872e-03 2.363e-02 2.311e-03 5.865e-05 3.494e-02 5.00 5.00 15.00 6.174e+03 8.723e+01 [TFlops] (comp only) (2061584302080.00) 2048 2048 (mbk,nbk) 4608 4608 4608 4.887e-02 4.006e+03 6.952e+00 10 6.573e-03 4.845e-03 3.443e-02 2.938e-03 5.651e-05 4.884e-02 5.00 5.00 15.00 7.814e+03 8.526e+01 [TFlops] (comp only) (2935341711360.00) 2304 2304 (mbk,nbk) 5120 5120 5120 6.305e-02 4.259e+03 6.652e+00 10 8.548e-03 5.982e-03 4.495e-02 3.536e-03 5.627e-05 6.308e-02 5.00 5.00 15.00 9.647e+03 8.957e+01 [TFlops] (comp only) (4026531840000.00) 2560 2560 (mbk,nbk) 5632 5632 5632 8.448e-02 4.230e+03 6.008e+00 10 1.055e-02 7.244e-03 6.225e-02 4.326e-03 6.652e-05 8.444e-02 5.00 5.00 15.00 1.167e+04 8.609e+01 [TFlops] (comp only) (5359313879040.00) 2816 2816 (mbk,nbk) 6144 6144 6144 1.054e-01 4.402e+03 5.730e+00 10 1.255e-02 8.646e-03 7.907e-02 5.156e-03 5.722e-05 1.055e-01 5.00 5.00 15.00 1.389e+04 8.800e+01 [TFlops] (comp only) (6957847019520.00) 3072 3072 (mbk,nbk) 6656 6656 6656 1.315e-01 4.487e+03 5.392e+00 10 1.480e-02 1.008e-02 1.006e-01 6.036e-03 6.032e-05 1.316e-01 5.00 5.00 15.00 1.630e+04 8.794e+01 [TFlops] (comp only) (8846290452480.00) 3328 3328 (mbk,nbk) 7168 7168 7168 1.564e-01 4.711e+03 5.256e+00 10 1.739e-02 1.166e-02 1.204e-01 7.007e-03 5.198e-05 1.566e-01 5.00 5.00 15.00 1.891e+04 9.174e+01 [TFlops] (comp only) (11048803368960.00) 3584 3584 (mbk,nbk) 7680 7680 7680 2.283e-01 3.969e+03 4.134e+00 10 2.006e-02 2.816e-02 1.714e-01 8.142e-03 2.255e-04 2.280e-01 5.00 5.00 15.00 6.193e+03 7.928e+01 [TFlops] (comp only) (13589544960000.00) 1920 1920 (mbk,nbk) 8192 8192 8192 2.473e-01 4.447e+03 4.342e+00 10 2.391e-02 3.133e-02 1.832e-01 9.247e-03 2.222e-04 2.479e-01 5.00 5.00 15.00 7.047e+03 9.003e+01 [TFlops] (comp only) (16492674416640.00) 2048 2048 (mbk,nbk) 8704 8704 8704 3.167e-01 4.165e+03 3.827e+00 10 2.657e-02 3.640e-02 2.432e-01 1.059e-02 2.213e-04 3.169e-01 5.00 5.00 15.00 7.955e+03 8.136e+01 [TFlops] (comp only) (19782350929920.00) 2176 2176 (mbk,nbk) 9216 9216 9216 3.878e-01 4.037e+03 3.504e+00 10 3.046e-02 3.964e-02 3.060e-01 1.162e-02 2.267e-04 3.879e-01 5.00 5.00 15.00 8.918e+03 7.675e+01 [TFlops] (comp only) (23482733690880.00) 2304 2304 (mbk,nbk) 9728 9728 9728 4.215e-01 4.369e+03 3.592e+00 10 3.579e-02 4.544e-02 3.276e-01 1.296e-02 2.236e-04 4.220e-01 5.00 5.00 15.00 9.937e+03 8.430e+01 [TFlops] (comp only) (27617981890560.00) 2432 2432 (mbk,nbk) 10240 10240 10240 4.598e-01 4.671e+03 3.649e+00 10 3.965e-02 4.921e-02 3.565e-01 1.440e-02 2.253e-04 4.600e-01 5.00 5.00 15.00 1.101e+04 9.035e+01 [TFlops] (comp only) (32212254720000.00) 2560 2560 (mbk,nbk) gemmtc-1.3a/testing/isc2020_dat/_testing_sgemmtc-sp_phi1_perf.dat0000664000175000017500000001246413660242533025266 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 10:07:16 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: SGEMM # Library: GEMMTC # mode: performance evaluation # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 1.000000 # Degree = 0 # CRMode = 0 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 512 # dim_stop: 10240 # dim_step: 512 # M: range # N: range # K: range # cublaspointer: HOST # input range: max = 1.055e+02, min = 1.472e-09 # input range: max = 1.555e+02, min = 1.353e-09 # Evaluation result ------------------------ (* shows the value at the last execution, ** shows the value of the last block at the last execution) # M N K Time[sec] Perf[GFlops] Perf[GB/s] itr *t_SpltA *t_SpltB *t_Comp *t_Sum *t_Other *t_Total **#SpltA **#SpltB **#Comp **Mem[MB] TC[TFlops] 512 512 512 8.418e-04 3.198e+02 4.983e+00 10 2.811e-04 1.621e-04 3.250e-04 6.199e-05 1.383e-05 8.440e-04 5.00 5.00 15.00 3.618e+02 1.239e+01 [TFlops] (comp only) (4026531840.00) 512 512 (mbk,nbk) 1024 1024 1024 1.775e-03 1.212e+03 9.451e+00 10 5.589e-04 2.651e-04 7.489e-04 1.829e-04 1.812e-05 1.774e-03 5.00 5.00 15.00 1.447e+03 4.301e+01 [TFlops] (comp only) (32212254720.00) 1024 1024 (mbk,nbk) 1536 1536 1536 4.642e-03 1.563e+03 8.131e+00 10 9.949e-04 5.031e-04 2.639e-03 5.040e-04 1.884e-05 4.660e-03 6.00 6.00 21.00 3.256e+03 5.767e+01 [TFlops] (comp only) (152202903552.00) 1536 1536 (mbk,nbk) 2048 2048 2048 8.259e-03 2.082e+03 8.125e+00 10 1.706e-03 6.618e-04 5.083e-03 7.901e-04 1.860e-05 8.260e-03 6.00 6.00 21.00 5.788e+03 7.097e+01 [TFlops] (comp only) (360777252864.00) 2048 2048 (mbk,nbk) 2560 2560 2560 1.416e-02 2.371e+03 7.406e+00 10 3.124e-03 9.289e-04 8.705e-03 1.215e-03 2.098e-05 1.399e-02 6.00 6.00 21.00 9.044e+03 8.095e+01 [TFlops] (comp only) (704643072000.00) 2560 2560 (mbk,nbk) 3072 3072 3072 2.224e-02 2.608e+03 6.789e+00 10 4.834e-03 1.303e-03 1.425e-02 1.742e-03 1.621e-05 2.215e-02 6.00 6.00 21.00 1.302e+04 8.544e+01 [TFlops] (comp only) (1217623228416.00) 3072 3072 (mbk,nbk) 3584 3584 3584 3.308e-02 2.785e+03 6.214e+00 10 6.950e-03 1.789e-03 2.239e-02 2.417e-03 2.217e-05 3.357e-02 6.00 6.00 21.00 1.773e+04 8.634e+01 [TFlops] (comp only) (1933540589568.00) 3584 3584 (mbk,nbk) 4096 4096 4096 4.656e-02 2.953e+03 5.765e+00 10 6.000e-03 4.628e-03 3.277e-02 3.123e-03 7.939e-05 4.660e-02 6.00 6.00 21.00 6.174e+03 8.808e+01 [TFlops] (comp only) (2886218022912.00) 2048 2048 (mbk,nbk) 4608 4608 4608 8.524e-02 2.296e+03 3.985e+00 10 8.729e-03 6.710e-03 6.437e-02 5.212e-03 1.016e-04 8.512e-02 7.00 7.00 28.00 7.814e+03 8.513e+01 [TFlops] (comp only) (5479304527872.00) 2304 2304 (mbk,nbk) 5120 5120 5120 1.099e-01 2.442e+03 3.815e+00 10 1.150e-02 8.180e-03 8.396e-02 6.318e-03 1.037e-04 1.101e-01 7.00 7.00 28.00 9.647e+03 8.952e+01 [TFlops] (comp only) (7516192768000.00) 2560 2560 (mbk,nbk) 5632 5632 5632 1.485e-01 2.406e+03 3.417e+00 10 1.403e-02 9.980e-03 1.163e-01 7.769e-03 1.001e-04 1.482e-01 7.00 7.00 28.00 1.167e+04 8.600e+01 [TFlops] (comp only) (10004052574208.00) 2816 2816 (mbk,nbk) 6144 6144 6144 1.859e-01 2.496e+03 3.249e+00 10 1.692e-02 1.190e-02 1.478e-01 9.343e-03 9.871e-05 1.861e-01 7.00 7.00 28.00 1.389e+04 8.785e+01 [TFlops] (comp only) (12987981103104.00) 3072 3072 (mbk,nbk) 6656 6656 6656 1.777e-01 3.320e+03 3.989e+00 10 1.726e-02 1.199e-02 1.401e-01 8.325e-03 7.987e-05 1.778e-01 6.00 6.00 21.00 1.630e+04 8.839e+01 [TFlops] (comp only) (12384806633472.00) 3328 3328 (mbk,nbk) 7168 7168 7168 2.116e-01 3.482e+03 3.886e+00 10 2.037e-02 1.389e-02 1.677e-01 9.672e-03 7.772e-05 2.117e-01 6.00 6.00 21.00 1.891e+04 9.224e+01 [TFlops] (comp only) (15468324716544.00) 3584 3584 (mbk,nbk) 7680 7680 7680 3.070e-01 2.951e+03 3.074e+00 10 2.298e-02 3.345e-02 2.391e-01 1.109e-02 2.961e-04 3.069e-01 6.00 6.00 21.00 6.193e+03 7.956e+01 [TFlops] (comp only) (19025362944000.00) 1920 1920 (mbk,nbk) 8192 8192 8192 3.315e-01 3.318e+03 3.239e+00 10 2.720e-02 3.722e-02 2.544e-01 1.258e-02 3.011e-04 3.317e-01 6.00 6.00 21.00 7.047e+03 9.075e+01 [TFlops] (comp only) (23089744183296.00) 2048 2048 (mbk,nbk) 8704 8704 8704 4.264e-01 3.094e+03 2.843e+00 10 3.045e-02 4.312e-02 3.391e-01 1.443e-02 3.002e-04 4.274e-01 6.00 6.00 21.00 7.955e+03 8.168e+01 [TFlops] (comp only) (27695291301888.00) 2176 2176 (mbk,nbk) 9216 9216 9216 5.252e-01 2.981e+03 2.588e+00 10 3.516e-02 4.724e-02 4.271e-01 1.575e-02 3.026e-04 5.256e-01 6.00 6.00 21.00 8.918e+03 7.697e+01 [TFlops] (comp only) (32875827167232.00) 2304 2304 (mbk,nbk) 9728 9728 9728 5.694e-01 3.234e+03 2.659e+00 10 4.146e-02 5.398e-02 4.570e-01 1.768e-02 3.085e-04 5.704e-01 6.00 6.00 21.00 9.937e+03 8.461e+01 [TFlops] (comp only) (38665174646784.00) 2432 2432 (mbk,nbk) 10240 10240 10240 6.204e-01 3.462e+03 2.704e+00 10 4.555e-02 5.862e-02 4.962e-01 1.972e-02 3.092e-04 6.204e-01 6.00 6.00 21.00 1.101e+04 9.089e+01 [TFlops] (comp only) (45097156608000.00) 2560 2560 (mbk,nbk) gemmtc-1.3a/testing/isc2020_dat/_testing_sgemmtc-sp_phi2_perf.dat0000664000175000017500000001246413660242657025276 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 10:08:28 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: SGEMM # Library: GEMMTC # mode: performance evaluation # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 2.000000 # Degree = 0 # CRMode = 0 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 512 # dim_stop: 10240 # dim_step: 512 # M: range # N: range # K: range # cublaspointer: HOST # input range: max = 2.226e+04, min = 2.328e-09 # input range: max = 4.836e+04, min = 4.367e-10 # Evaluation result ------------------------ (* shows the value at the last execution, ** shows the value of the last block at the last execution) # M N K Time[sec] Perf[GFlops] Perf[GB/s] itr *t_SpltA *t_SpltB *t_Comp *t_Sum *t_Other *t_Total **#SpltA **#SpltB **#Comp **Mem[MB] TC[TFlops] 512 512 512 1.088e-03 2.474e+02 3.854e+00 10 3.390e-04 1.919e-04 4.542e-04 7.987e-05 2.098e-05 1.086e-03 6.00 6.00 21.00 3.618e+02 1.241e+01 [TFlops] (comp only) (5637144576.00) 512 512 (mbk,nbk) 1024 1024 1024 2.296e-03 9.368e+02 7.308e+00 10 6.738e-04 3.178e-04 1.041e-03 2.391e-04 2.146e-05 2.293e-03 6.00 6.00 21.00 1.447e+03 4.332e+01 [TFlops] (comp only) (45097156608.00) 1024 1024 (mbk,nbk) 1536 1536 1536 4.644e-03 1.562e+03 8.129e+00 10 9.918e-04 4.919e-04 2.626e-03 5.031e-04 1.860e-05 4.631e-03 6.00 6.00 21.00 3.256e+03 5.797e+01 [TFlops] (comp only) (152202903552.00) 1536 1536 (mbk,nbk) 2048 2048 2048 8.875e-03 1.937e+03 7.562e+00 10 1.701e-03 6.361e-04 4.867e-03 7.799e-04 1.979e-05 8.004e-03 6.00 6.00 21.00 5.788e+03 7.413e+01 [TFlops] (comp only) (360777252864.00) 2048 2048 (mbk,nbk) 2560 2560 2560 1.411e-02 2.379e+03 7.429e+00 10 3.174e-03 9.360e-04 8.790e-03 1.233e-03 2.408e-05 1.416e-02 6.00 6.00 21.00 9.044e+03 8.017e+01 [TFlops] (comp only) (704643072000.00) 2560 2560 (mbk,nbk) 3072 3072 3072 2.215e-02 2.619e+03 6.816e+00 10 4.801e-03 1.299e-03 1.419e-02 1.751e-03 1.764e-05 2.205e-02 6.00 6.00 21.00 1.302e+04 8.583e+01 [TFlops] (comp only) (1217623228416.00) 3072 3072 (mbk,nbk) 3584 3584 3584 3.307e-02 2.785e+03 6.215e+00 10 6.881e-03 1.770e-03 2.160e-02 2.406e-03 2.193e-05 3.268e-02 6.00 6.00 21.00 1.773e+04 8.953e+01 [TFlops] (comp only) (1933540589568.00) 3584 3584 (mbk,nbk) 4096 4096 4096 4.631e-02 2.969e+03 5.797e+00 10 5.862e-03 4.592e-03 3.252e-02 3.121e-03 7.391e-05 4.617e-02 6.00 6.00 21.00 6.174e+03 8.876e+01 [TFlops] (comp only) (2886218022912.00) 2048 2048 (mbk,nbk) 4608 4608 4608 8.478e-02 2.309e+03 4.007e+00 10 8.774e-03 6.713e-03 6.389e-02 5.214e-03 8.917e-05 8.468e-02 7.00 7.00 28.00 7.814e+03 8.577e+01 [TFlops] (comp only) (5479304527872.00) 2304 2304 (mbk,nbk) 5120 5120 5120 1.093e-01 2.458e+03 3.839e+00 10 1.145e-02 8.196e-03 8.298e-02 6.330e-03 9.918e-05 1.091e-01 7.00 7.00 28.00 9.647e+03 9.058e+01 [TFlops] (comp only) (7516192768000.00) 2560 2560 (mbk,nbk) 5632 5632 5632 1.476e-01 2.422e+03 3.439e+00 10 1.404e-02 9.993e-03 1.153e-01 7.745e-03 9.751e-05 1.472e-01 7.00 7.00 28.00 1.167e+04 8.677e+01 [TFlops] (comp only) (10004052574208.00) 2816 2816 (mbk,nbk) 6144 6144 6144 1.850e-01 2.507e+03 3.264e+00 10 1.685e-02 1.187e-02 1.465e-01 9.344e-03 1.037e-04 1.847e-01 7.00 7.00 28.00 1.389e+04 8.864e+01 [TFlops] (comp only) (12987981103104.00) 3072 3072 (mbk,nbk) 6656 6656 6656 2.312e-01 2.552e+03 3.066e+00 10 1.970e-02 1.391e-02 1.861e-01 1.093e-02 1.032e-04 2.308e-01 7.00 7.00 28.00 1.630e+04 8.872e+01 [TFlops] (comp only) (16513075511296.00) 3328 3328 (mbk,nbk) 7168 7168 7168 2.755e-01 2.674e+03 2.984e+00 10 2.347e-02 1.615e-02 2.221e-01 1.275e-02 1.130e-04 2.746e-01 7.00 7.00 28.00 1.891e+04 9.286e+01 [TFlops] (comp only) (20624432955392.00) 3584 3584 (mbk,nbk) 7680 7680 7680 3.981e-01 2.276e+03 2.371e+00 10 2.599e-02 3.874e-02 3.194e-01 1.442e-02 3.967e-04 3.990e-01 7.00 7.00 28.00 6.193e+03 7.941e+01 [TFlops] (comp only) (25367150592000.00) 1920 1920 (mbk,nbk) 8192 8192 8192 4.289e-01 2.564e+03 2.503e+00 10 3.071e-02 4.317e-02 3.391e-01 1.627e-02 3.798e-04 4.296e-01 7.00 7.00 28.00 7.047e+03 9.079e+01 [TFlops] (comp only) (30786325577728.00) 2048 2048 (mbk,nbk) 8704 8704 8704 5.543e-01 2.380e+03 2.187e+00 10 3.453e-02 5.019e-02 4.500e-01 1.892e-02 3.872e-04 5.541e-01 7.00 7.00 28.00 7.955e+03 8.205e+01 [TFlops] (comp only) (36927055069184.00) 2176 2176 (mbk,nbk) 9216 9216 9216 6.831e-01 2.292e+03 1.989e+00 10 3.963e-02 5.505e-02 5.731e-01 2.066e-02 3.841e-04 6.888e-01 7.00 7.00 28.00 8.918e+03 7.649e+01 [TFlops] (comp only) (43834436222976.00) 2304 2304 (mbk,nbk) 9728 9728 9728 7.396e-01 2.490e+03 2.047e+00 10 4.669e-02 6.250e-02 6.073e-01 2.312e-02 4.067e-04 7.400e-01 7.00 7.00 28.00 9.937e+03 8.489e+01 [TFlops] (comp only) (51553566195712.00) 2432 2432 (mbk,nbk) 10240 10240 10240 8.085e-01 2.657e+03 2.075e+00 10 5.158e-02 6.804e-02 6.605e-01 2.581e-02 4.072e-04 8.064e-01 7.00 7.00 28.00 1.101e+04 9.103e+01 [TFlops] (comp only) (60129542144000.00) 2560 2560 (mbk,nbk) gemmtc-1.3a/testing/isc2020_dat/_testing_sgemmtc_phi01_check.dat0000664000175000017500000000303513660242661025043 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 10:09:52 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: SGEMM # Library: GEMMTC # mode: checking with MPFR 2048-bit # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 0.100000 # Degree = 0 # CRMode = 0 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 128 # dim_stop: 10240 # dim_step: 128 # M: 32 # N: 32 # K: range # cublaspointer: HOST # input range: max = 7.900e-01, min = 4.144e-07 # input range: max = 7.961e-01, min = 4.556e-07 # Evaluation result ------------------------ # M N K ERROR #ofERROR: (1)cuBLAS, (2)SGEMM-TC-SP 32 32 128 7.961155e-05 768 9.653715e-06 538 0.000000e+00 0 32 32 256 1.316659e-04 729 8.238113e-06 554 0.000000e+00 0 32 32 512 1.062672e-04 792 3.258109e-07 540 0.000000e+00 0 32 32 1024 4.121947e-05 784 3.433267e-05 600 0.000000e+00 0 32 32 2048 3.175660e-03 831 5.041024e-05 651 0.000000e+00 0 32 32 4096 4.552654e-05 838 5.136597e-06 655 0.000000e+00 0 32 32 8192 6.873439e-04 867 1.577720e-03 887 0.000000e+00 0 gemmtc-1.3a/testing/isc2020_dat/_testing_sgemmtc_phi1_check.dat0000664000175000017500000000303513660242664024766 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 10:09:54 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: SGEMM # Library: GEMMTC # mode: checking with MPFR 2048-bit # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 1.000000 # Degree = 0 # CRMode = 0 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 128 # dim_stop: 10240 # dim_step: 128 # M: 32 # N: 32 # K: range # cublaspointer: HOST # input range: max = 4.852e+01, min = 2.558e-07 # input range: max = 5.233e+01, min = 2.034e-07 # Evaluation result ------------------------ # M N K ERROR #ofERROR: (1)cuBLAS, (2)SGEMM-TC-SP 32 32 128 6.907649e-05 745 9.706557e-06 569 0.000000e+00 0 32 32 256 1.140978e-04 719 4.755223e-06 579 0.000000e+00 0 32 32 512 1.226900e-04 768 4.459473e-06 632 0.000000e+00 0 32 32 1024 1.414796e-04 741 1.611114e-05 641 0.000000e+00 0 32 32 2048 4.794992e-06 688 1.445536e-06 654 0.000000e+00 0 32 32 4096 3.913083e-06 757 4.637820e-07 691 0.000000e+00 0 32 32 8192 5.634301e-07 762 3.796750e-06 804 0.000000e+00 0 gemmtc-1.3a/testing/isc2020_dat/_testing_sgemmtc_phi2_check.dat0000664000175000017500000000303513660242666024771 0ustar mukunokimukunoki# Testing info ----------------------------- # Date: Sun May 17 10:09:57 EDT 2020 # CPU:model name : Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz # OS: Linux turing 4.18.0-147.3.1.el8_1.x86_64 #1 SMP Fri Jan 3 23:55:26 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux # Host: turing # Device[0] TITAN RTX # - Theoretical Peak Memory Bandwidth: 672.1 [GB/s] # - Theoretical Peak Performance (SP): 16312.3 [GFlops] # - Theoretical Peak Performance (DP): 509.8 [GFlops] # - Total Gmem: 25.4 [GB] # - CC: 7.5 # - ECC Enabled: 0 # - MultiProcessorCount: 72 # Driver / Runtime: 10020, 10020 # Routine: SGEMM # Library: GEMMTC # mode: checking with MPFR 2048-bit # NLOOP: 10 # GEMM-TC ---------------------------------- # Work-mem size = 2.0e+10 (bytes) # Phi = 2.000000 # Degree = 0 # CRMode = 0 # BLAS parameter --------------------------- # tranA: N # tranB: N # dim_start: 128 # dim_stop: 10240 # dim_step: 128 # M: 32 # N: 32 # K: range # cublaspointer: HOST # input range: max = 4.709e+03, min = 1.497e-07 # input range: max = 5.478e+03, min = 8.303e-08 # Evaluation result ------------------------ # M N K ERROR #ofERROR: (1)cuBLAS, (2)SGEMM-TC-SP 32 32 128 5.317852e-05 704 3.077238e-06 612 0.000000e+00 0 32 32 256 5.482683e-05 662 1.154743e-05 617 0.000000e+00 0 32 32 512 4.573085e-04 678 3.217965e-05 663 0.000000e+00 0 32 32 1024 6.547578e-05 642 4.635167e-05 624 0.000000e+00 0 32 32 2048 1.242077e-06 651 5.264701e-07 647 0.000000e+00 0 32 32 4096 4.150480e-07 708 4.173166e-07 683 0.000000e+00 0 32 32 8192 5.081450e-07 770 4.656838e-07 721 0.000000e+00 0 gemmtc-1.3a/testing/isc2020_dat/_testing_dgemm_check_turing.sh0000700000175000017500000000064213660141352024711 0ustar mukunokimukunoki#!/bin/bash step=128 max=10240 file01=_testing_dgemmtc_phi01_check.dat file02=_testing_dgemmtc_phi1_check.dat file03=_testing_dgemmtc_phi2_check.dat ./testing_dgemmtc --range=${step}:${max}:pow --m=32 --n=32 --mode=c --phi=0.1 > ${file01} ./testing_dgemmtc --range=${step}:${max}:pow --m=32 --n=32 --mode=c --phi=1 > ${file02} ./testing_dgemmtc --range=${step}:${max}:pow --m=32 --n=32 --mode=c --phi=2 > ${file03} gemmtc-1.3a/testing/isc2020_dat/_testing_dgemm_perf2_cygnus.sh0000755000175000017500000000316213660141661024667 0ustar mukunokimukunoki#!/bin/bash #------- qsub option ----------- #PBS -A LSPANC #PBS -q fpga #PBS -l elapstim_req=01:00:00 #PBS -v OMP_NUM_THREADS=24 #------- Program execution ----------- cd $PBS_O_WORKDIR module load cuda/10.2 file01=_testing_dgemm_perf.dat echo "## cublas" > ${file01} ./testing_dgemmtc --nodisp=1 --mnk=10240 --mode=p --profile=1 --degree=1 --phi=0.1 >> ${file01} echo "" >> ${file01} echo "## dgemmtc-dp phi=0.1" >> ${file01} ./testing_dgemmtc --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=0 --phi=0.1 >> ${file01} echo "## dgemmtc-dp phi=1" >> ${file01} ./testing_dgemmtc --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=0 --phi=1 >> ${file01} echo "## dgemmtc-dp phi=2" >> ${file01} ./testing_dgemmtc --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=0 --phi=2 >> ${file01} echo "" >> ${file01} echo "## dgemmtc-cr phi=0.1" >> ${file01} ./testing_dgemmtc --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=1 --phi=0.1 >> ${file01} echo "## dgemmtc-cr phi=1" >> ${file01} ./testing_dgemmtc --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=1 --phi=1 >> ${file01} echo "## dgemmtc-cr phi=2" >> ${file01} ./testing_dgemmtc --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=1 --phi=2 >> ${file01} echo "" >> ${file01} echo "## dgemmdp-cr phi=0.1" >> ${file01} ./testing_dgemmdp --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=1 --phi=0.1 >> ${file01} echo "## dgemmdp-cr phi=1" >> ${file01} ./testing_dgemmdp --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=1 --phi=1 >> ${file01} echo "## dgemmdp-cr phi=2" >> ${file01} ./testing_dgemmdp --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=1 --phi=2 >> ${file01} gemmtc-1.3a/testing/isc2020_dat/_testing_dgemm_perf2_turing.sh0000755000175000017500000000265413660141516024673 0ustar mukunokimukunoki#!/bin/bash file01=_testing_dgemm_perf.dat echo "## cublas" > ${file01} ./testing_dgemmtc --nodisp=1 --mnk=10240 --mode=p --profile=1 --degree=1 --phi=0.1 >> ${file01} echo "" >> ${file01} echo "## dgemmtc-dp phi=0.1" >> ${file01} ./testing_dgemmtc --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=0 --phi=0.1 >> ${file01} echo "## dgemmtc-dp phi=1" >> ${file01} ./testing_dgemmtc --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=0 --phi=1 >> ${file01} echo "## dgemmtc-dp phi=2" >> ${file01} ./testing_dgemmtc --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=0 --phi=2 >> ${file01} echo "" >> ${file01} echo "## dgemmtc-cr phi=0.1" >> ${file01} ./testing_dgemmtc --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=1 --phi=0.1 >> ${file01} echo "## dgemmtc-cr phi=1" >> ${file01} ./testing_dgemmtc --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=1 --phi=1 >> ${file01} echo "## dgemmtc-cr phi=2" >> ${file01} ./testing_dgemmtc --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=1 --phi=2 >> ${file01} echo "" >> ${file01} echo "## dgemmdp-cr phi=0.1" >> ${file01} ./testing_dgemmdp --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=1 --phi=0.1 >> ${file01} echo "## dgemmdp-cr phi=1" >> ${file01} ./testing_dgemmdp --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=1 --phi=1 >> ${file01} echo "## dgemmdp-cr phi=2" >> ${file01} ./testing_dgemmdp --nodisp=1 --mnk=10240 --mode=p --profile=1 --CRmode=1 --phi=2 >> ${file01} gemmtc-1.3a/testing/isc2020_dat/_testing_dgemm_perf_turing.sh0000700000175000017500000000201213660232641024563 0ustar mukunokimukunoki#!/bin/bash step=512 max=10240 file01=_testing_dgemmtc-dp_phi01_perf.dat file02=_testing_dgemmtc-dp_phi1_perf.dat file03=_testing_dgemmtc-dp_phi2_perf.dat ./testing_dgemmtc --range=${step}:${max}:${step} --mode=p --profile=1 --CRmode=0 --phi=0.1 > ${file01} ./testing_dgemmtc --range=${step}:${max}:${step} --mode=p --profile=1 --CRmode=0 --phi=1 > ${file02} ./testing_dgemmtc --range=${step}:${max}:${step} --mode=p --profile=1 --CRmode=0 --phi=2 > ${file03} file01=_testing_dgemmtc-cr_phi01_perf.dat file02=_testing_dgemmtc-cr_phi1_perf.dat file03=_testing_dgemmtc-cr_phi2_perf.dat ./testing_dgemmtc --range=${step}:${max}:${step} --mode=p --profile=1 --CRmode=1 --phi=0.1 > ${file01} ./testing_dgemmtc --range=${step}:${max}:${step} --mode=p --profile=1 --CRmode=1 --phi=1 > ${file02} ./testing_dgemmtc --range=${step}:${max}:${step} --mode=p --profile=1 --CRmode=1 --phi=2 > ${file03} file01=_testing_dgemm_cublas_perf.dat ./testing_dgemmtc --range=${step}:${max}:${step} --mode=p --profile=0 --degree=1 --phi=0.1 > ${file01} gemmtc-1.3a/testing/isc2020_dat/_testing_sgemm_check_turing.sh0000700000175000017500000000064213660141417024732 0ustar mukunokimukunoki#!/bin/bash step=128 max=10240 file01=_testing_sgemmtc_phi01_check.dat file02=_testing_sgemmtc_phi1_check.dat file03=_testing_sgemmtc_phi2_check.dat ./testing_sgemmtc --range=${step}:${max}:pow --m=32 --n=32 --mode=c --phi=0.1 > ${file01} ./testing_sgemmtc --range=${step}:${max}:pow --m=32 --n=32 --mode=c --phi=1 > ${file02} ./testing_sgemmtc --range=${step}:${max}:pow --m=32 --n=32 --mode=c --phi=2 > ${file03} gemmtc-1.3a/testing/isc2020_dat/_testing_sgemm_perf_turing.sh0000700000175000017500000000113513660141436024610 0ustar mukunokimukunoki#!/bin/bash step=512 max=10240 file01=_testing_sgemm_cublas_perf.dat ./testing_sgemmtc --range=${step}:${max}:${step} --mode=p --profile=0 --degree=1 --phi=0.1 > ${file01} file01=_testing_sgemmtc-sp_phi01_perf.dat file02=_testing_sgemmtc-sp_phi1_perf.dat file03=_testing_sgemmtc-sp_phi2_perf.dat ./testing_sgemmtc --range=${step}:${max}:${step} --mode=p --profile=1 --CRmode=0 --phi=0.1 > ${file01} ./testing_sgemmtc --range=${step}:${max}:${step} --mode=p --profile=1 --CRmode=0 --phi=1 > ${file02} ./testing_sgemmtc --range=${step}:${max}:${step} --mode=p --profile=1 --CRmode=0 --phi=2 > ${file03} gemmtc-1.3a/testing/isc2020_dat/all.sh0000700000175000017500000000024313660140537017750 0ustar mukunokimukunoki./_testing_dgemm_perf_turing.sh ./_testing_dgemm_check_turing.sh ./_testing_sgemm_perf_turing.sh ./_testing_sgemm_check_turing.sh ./_testing_dgemm_perf2_turing.sh gemmtc-1.3a/testing/testing_common.h0000644000175000017500000000120513657536625020153 0ustar mukunokimukunoki#ifndef TESTING_COMMON #define TESTING_COMMON // header ------------------------------------------ #include "testing_setting.h" #include #include #include #include #include #define __STDC_FORMAT_MACROS #include #include #include #include #include #include #include #include #include #include #ifdef USE_MPFR //#include #include #endif #if defined (PREC_S) #define FP_TYPE float #elif defined (PREC_D) #define FP_TYPE double #endif #endif gemmtc-1.3a/testing/Makefile0000644000175000017500000000324613657607137016421 0ustar mukunokimukunoki# Makefile include ../make.inc # common ------------------------------------- FLAGS = -O0 -Wall -MD -fno-strict-aliasing # -------------------------------------------- GEMMTC_OBJS = \ testing_dgemmtc \ testing_dgemmsp \ testing_dgemmdp \ testing_sgemmtc # -------------------------------------------- OBJS += $(GEMMTC_OBJS) INCS += -I$(CUDA_PATH)/include -I../include LDFLAGS += -L$(CUDA_PATH)/lib64 -DCUDA -L../lib LIBS += -lcuda -lcudart -lcublas -lcusparse -lcuozblastc -fopenmp -lm -lstdc++ ifeq ($(USE_MPFR),yes) INCS += $(MPACK_INC) LDFLAGS += $(MPACK_LDFLAGS) -L$(MPACK_DIR)/external/i/MPC/lib/ -L$(MPACK_DIR)/external/i/MPFR/lib/ -L$(MPACK_DIR)/external/i/GMP/lib/ -DUSE_MPFR LIBS += -lmblas_mpfr -fopenmp -lm -lstdc++ -lcuozblastc MPFR_DIR = $(HOME)/local MPFR_LDFLAGS = -L$(MPFR_DIR)/lib MPFR_INC = -I$(MPFR_DIR)/include INCS += $(MPFR_INC) $(MPACK_INC) LDFLAGS += $(MPFR_LDFLAGS) $(MPACK_LDFLAGS) -L$(MPACK_DIR)/external/i/MPC/lib/ -L$(MPACK_DIR)/external/i/MPFR/lib/ -L$(MPACK_DIR)/external/i/GMP/lib/ -DUSE_MPFR LIBS += -lmpc -lgmp -lmblas_mpfr -lmpfr -lmpfrcxx endif default: $(OBJS) SRC_L3 = testing_level3.cpp testing_dgemmtc: $(SRC_L3) ../lib/libcuozblastc.a $(CXX) $< -o $@ $(FLAGS) $(INCS) $(LDFLAGS) $(LIBS) -DPREC_D -DUSE_TC testing_dgemmsp: $(SRC_L3) ../lib/libcuozblastc.a $(CXX) $< -o $@ $(FLAGS) $(INCS) $(LDFLAGS) $(LIBS) -DPREC_D -DUSE_SP testing_dgemmdp: $(SRC_L3) ../lib/libcuozblastc.a $(CXX) $< -o $@ $(FLAGS) $(INCS) $(LDFLAGS) $(LIBS) -DPREC_D -DUSE_DP testing_sgemmtc: $(SRC_L3) ../lib/libcuozblastc.a $(CXX) $< -o $@ $(FLAGS) $(INCS) $(LDFLAGS) $(LIBS) -DPREC_S -DUSE_TC clean: $(RM) $(OBJS) $(RM) *.d -include *.d .PHONY: all clean gemmtc-1.3a/LICENSE0000664000175000017500000000220213651734310014266 0ustar mukunokimukunokiThe MIT License (MIT) Copyright (c) 2018-2019 Daichi Mukunoki (Tokyo Woman's Christian University) Copyright (c) 2019-2020 RIKEN Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. gemmtc-1.3a/make.inc0000644000175000017500000000117513657536552014715 0ustar mukunokimukunoki############################################## CUDA_PATH = /usr/local/cuda-10.2 CXX = g++ # do not touch below CUDA = yes PREC_D = yes PREC_S = yes GEMM = yes USE_TC = yes USE_SP = yes USE_DP = yes # CUDA Compute Capability CC60 = no CC61 = no CC62 = no CC70 = yes CC75 = yes #--- for testing ----------------------------- USE_MPFR = yes # If 'yes', testing programs compare the # result with MPFR (using MPLAPACK). # https://github.com/nakatamaho/mplapack # Set the path. MPACK_DIR = $(HOME)/local MPACK_LDFLAGS = -L$(MPACK_DIR)/lib MPACK_INC = -I$(MPACK_DIR)/include/mpack ############################################## gemmtc-1.3a/README0000644000175000017500000000517313667424015014156 0ustar mukunokimukunokiGEMMTC 1.3 alpha May 20, 2020 Developers: - Daichi Mukunoki (RIKEN Center for Computational Science, daichi.mukunoki@riken.jp) - Takeshi Ogita (Tokyo Woman's Christian University) - Katsuhisa Ozaki (Shibaura Institute of Technology) - Toshiyuki Imamura (RIKEN Center for Computational Science) GEMMTC is developed based on OzBLAS [3] (1.2 alpha). License: MIT License. See attached file. Copyright: Copyright (c) 2018-2019 Daichi Mukunoki (Tokyo Woman's Christian University) Copyright (c) 2019-2020 RIKEN (1) Overview SGEMM and DGEMM implementations using Tensor Cores based on Ozaki scheme, presented in [4]. For more details, Ozaki scheme [1] and NearSum [2] (for the correctly-rounded operation). This implementation is a branch from OzBLAS [3]. (2) Requirements: libcuozblas.a: - cuBLAS and CUDA testing: - (option) Multiple precision arithmetic BLAS (MBLAS) and LAPACK (MLAPACK) (http://mplapack.sourceforge.net) (3) Build: After modifying Makefiles in src and testing directories, execute make in the current directory. (4) Usage: Compile your code with the following files: include/cuozblas.h lib/libcuozblas.a You must call cuozblasCreate and cuozblasDestroy before and after the use of the BLAS routines, respectively. For example, see testing/testing_level3.cpp. For cuozblasCreate: cuozblasCreate (&ha, workMemSize, numSplitArraysMax, CRMode, OFMode); - workMemSize (size_t): work memory size in bytes. - numSplitArraysMax (uint): computation degree (i.e., the maximum number of split matrices in the Ozaki scheme). 0 is default (auto) - CRMode (0/1): 0: DP-mode (DGEMM-equivalent accuracy) or SP-mode (SGEMM-equivalent accuracy), 1: CR-mode (correctly-rounded) - OFMode (0/1): overflow-mode (0: off, 1: on) Also, please check and modify "testing/testing_setting.h". (5) References: [1] Ozaki K., Ogita T., Oishi S., Rump S. M.: Error-free transformations of matrix multiplication by using fast routines of matrix multiplication and its applications, Numer. Algorithms, vol. 59, no. 1, pp. 95-118, 2012. [2] Rump S., Ogita T., Oishi S.: Accurate Floating-Point Summation Part II: Sign, K-Fold Faithful and Rounding to Nearest, SIAM Journal on Scientific Computing, Vol. 31, No. 2, pp. 1269-1302, 2009. [3] Mukunoki D., Ogita T., Ozaki K.: Reproducible BLAS Routines with Tunable Accuracy Using Ozaki Scheme for Many-core Architectures, 13th International Conference on Parallel Processing and Applied Mathematics (PPAM2019), Lecture Notes in Computer Science, Vol. 12043, pp. 516-527, 2020. [4] Mukunoki D., Ozaki K., Ogita T., Imamura T.: DGEMM using Tensor Cores, and Its Accurate and Reproducible Versions, ISC High Performance 2020, 2020 (to appear).