//=================================================================================================================== //=================================================================================================================== //=================================================================================================================== extern "C" void magma_clarft_sm32x32_batched(magma_int_t n, magma_int_t k, magmaFloatComplex **v_array, magma_int_t ldv, magmaFloatComplex **tau_array, magmaFloatComplex **T_array, magma_int_t ldt, magma_int_t batchCount, magma_queue_t queue) { if ( k <= 0) return; //================================== // GEMV //================================== #define USE_GEMV2 #define use_gemm_larft_sm32 #if defined(use_gemm_larft_sm32) magma_cgemm_batched( MagmaConjTrans, MagmaNoTrans, k, k, n, MAGMA_C_ONE, v_array, ldv, v_array, ldv, MAGMA_C_ZERO, T_array, ldt, batchCount, queue ); magmablas_claset_batched( MagmaLower, k, k, MAGMA_C_ZERO, MAGMA_C_ZERO, T_array, ldt, batchCount, queue ); #else #if 1 for (magma_int_t i=0; i < k; i++) { //W(1:i-1) := - tau(i) * V(i:n,1:i-1)' * V(i:n,i) //T( i, i ) = tau( i ) //custom implementation. #ifdef USE_GEMV2 magmablas_clarft_gemvrowwise_batched( n-i, i, tau_array, v_array, ldv, T_array, ldt, batchCount, queue); #else magmablas_clarft_gemvcolwise_batched( n-i, i, v_array, ldv, T_array, ldt, tau_array, batchCount, queue); #endif } #else //seems to be very slow when k=32 while the one by one loop above is faster clarft_gemv_loop_inside_kernel_batched(n, k, tau_array, v_array, ldv, T_array, ldt, batchCount, queue); #endif #endif //================================== // TRMV //================================== //T(1:i-1,i) := T(1:i-1,1:i-1) * W(1:i-1) i=[1:k] magmablas_clarft_ctrmv_sm32x32_batched(k, k, tau_array, T_array, ldt, T_array, ldt, batchCount, queue); }
extern "C" magma_int_t magma_cgeqrf_panel_batched( magma_int_t m, magma_int_t n, magma_int_t nb, magmaFloatComplex** dA_array, magma_int_t ldda, magmaFloatComplex** tau_array, magmaFloatComplex** dT_array, magma_int_t ldt, magmaFloatComplex** dR_array, magma_int_t ldr, magmaFloatComplex** dW0_displ, magmaFloatComplex** dW1_displ, magmaFloatComplex *dwork, magmaFloatComplex** dW2_displ, magmaFloatComplex** dW3_displ, magma_int_t *info_array, magma_int_t batchCount, cublasHandle_t myhandle, magma_queue_t queue) { magma_int_t j, jb; magma_int_t ldw = nb; for( j=0; j<n; j+=nb) { jb = min(nb, n-j); magma_cdisplace_pointers(dW0_displ, dA_array, ldda, j, j, batchCount, queue); magma_cdisplace_pointers(dW2_displ, tau_array, 1, j, 0, batchCount, queue); magma_cdisplace_pointers(dW3_displ, dR_array, ldr, j, j, batchCount, queue); // //sub-panel factorization magma_cgeqr2_batched( m-j, jb, dW0_displ, ldda, dW2_displ, info_array, batchCount, queue); //copy upper part of dA to dR magma_cdisplace_pointers(dW0_displ, dA_array, ldda, j, j, batchCount, queue); magma_cdisplace_pointers(dW3_displ, dR_array, ldr, j, j, batchCount, queue); magmablas_clacpy_batched(MagmaUpper, jb, jb, dW0_displ, ldda, dW3_displ, ldr, batchCount, queue); magma_cdisplace_pointers(dW0_displ, dA_array, ldda, j, j, batchCount, queue); magma_cdisplace_pointers(dW3_displ, dR_array, ldr, j, j, batchCount, queue); magmablas_claset_batched(MagmaUpper, jb, jb, MAGMA_C_ZERO, MAGMA_C_ONE, dW0_displ, ldda, batchCount, queue); if( (n-j-jb) > 0) //update the trailing matrix inside the panel { magma_clarft_sm32x32_batched(m-j, jb, dW0_displ, ldda, dW2_displ, dT_array, ldt, batchCount, myhandle, queue); magma_cdisplace_pointers(dW1_displ, dA_array, ldda, j, j + jb, batchCount, queue); cset_pointer(dW2_displ, dwork, 1, 0, 0, ldw*n, batchCount, queue ); cset_pointer(dW3_displ, dwork + ldw*n*batchCount, 1, 0, 0, ldw*n, batchCount, queue ); magma_clarfb_gemm_batched( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise, m-j, n-j-jb, jb, (const magmaFloatComplex**)dW0_displ, ldda, (const magmaFloatComplex**)dT_array, ldt, dW1_displ, ldda, dW2_displ, ldw, dW3_displ, ldw, batchCount, myhandle, queue); } } return 0; }
extern "C" magma_int_t magma_cgeqrf_batched( magma_int_t m, magma_int_t n, magmaFloatComplex **dA_array, magma_int_t ldda, magmaFloatComplex **tau_array, magma_int_t *info_array, magma_int_t batchCount, magma_queue_t queue) { #define dA(i, j) (dA + (i) + (j)*ldda) // A(i, j) means at i row, j column magma_int_t min_mn = min(m, n); cudaMemset(info_array, 0, batchCount*sizeof(magma_int_t)); /* Check arguments */ magma_int_t arginfo = 0; if (m < 0) arginfo = -1; else if (n < 0) arginfo = -2; else if (ldda < max(1,m)) arginfo = -4; if (arginfo != 0) { magma_xerbla( __func__, -(arginfo) ); return arginfo; } /* Quick return if possible */ if (m == 0 || n == 0) if(min_mn == 0 ) return arginfo; if( m > 2048 || n > 2048 ){ printf("=========================================================================================\n"); printf(" WARNING batched routines are designed for small sizes it might be better to use the\n Native/Hybrid classical routines if you want performance\n"); printf("=========================================================================================\n"); } magma_int_t nb = 32; magma_int_t nnb = 8; magma_int_t i, k, ib=nb, jb=nnb; magma_int_t ldw, ldt, ldr, offset; cublasHandle_t myhandle; cublasCreate_v2(&myhandle); magmaFloatComplex **dW0_displ = NULL; magmaFloatComplex **dW1_displ = NULL; magmaFloatComplex **dW2_displ = NULL; magmaFloatComplex **dW3_displ = NULL; magmaFloatComplex **dW4_displ = NULL; magmaFloatComplex **dW5_displ = NULL; magmaFloatComplex *dwork = NULL; magmaFloatComplex *dT = NULL; magmaFloatComplex *dR = NULL; magmaFloatComplex **dR_array = NULL; magmaFloatComplex **dT_array = NULL; magmaFloatComplex **cpuAarray = NULL; magmaFloatComplex **cpuTarray = NULL; magma_malloc((void**)&dW0_displ, batchCount * sizeof(*dW0_displ)); magma_malloc((void**)&dW1_displ, batchCount * sizeof(*dW1_displ)); magma_malloc((void**)&dW2_displ, batchCount * sizeof(*dW2_displ)); magma_malloc((void**)&dW3_displ, batchCount * sizeof(*dW3_displ)); magma_malloc((void**)&dW4_displ, batchCount * sizeof(*dW4_displ)); // used in clarfb magma_malloc((void**)&dW5_displ, batchCount * sizeof(*dW5_displ)); magma_malloc((void**)&dR_array, batchCount * sizeof(*dR_array)); magma_malloc((void**)&dT_array, batchCount * sizeof(*dT_array)); ldt = ldr = min(nb, min_mn); magma_cmalloc(&dwork, (2 * nb * n) * batchCount); magma_cmalloc(&dR, ldr * n * batchCount); magma_cmalloc(&dT, ldt * ldt * batchCount); magma_malloc_cpu((void**) &cpuAarray, batchCount*sizeof(magmaFloatComplex*)); magma_malloc_cpu((void**) &cpuTarray, batchCount*sizeof(magmaFloatComplex*)); /* check allocation */ if ( dW0_displ == NULL || dW1_displ == NULL || dW2_displ == NULL || dW3_displ == NULL || dW4_displ == NULL || dW5_displ == NULL || dR_array == NULL || dT_array == NULL || dR == NULL || dT == NULL || dwork == NULL || cpuAarray == NULL || cpuTarray == NULL ) { magma_free(dW0_displ); magma_free(dW1_displ); magma_free(dW2_displ); magma_free(dW3_displ); magma_free(dW4_displ); magma_free(dW5_displ); magma_free(dR_array); magma_free(dT_array); magma_free(dR); magma_free(dT); magma_free(dwork); free(cpuAarray); free(cpuTarray); magma_int_t info = MAGMA_ERR_DEVICE_ALLOC; magma_xerbla( __func__, -(info) ); return info; } magmablas_claset_q(MagmaFull, ldr, n*batchCount , MAGMA_C_ZERO, MAGMA_C_ZERO, dR, ldr, queue); magmablas_claset_q(MagmaFull, ldt, ldt*batchCount, MAGMA_C_ZERO, MAGMA_C_ZERO, dT, ldt, queue); cset_pointer(dR_array, dR, 1, 0, 0, ldr*min(nb, min_mn), batchCount, queue); cset_pointer(dT_array, dT, 1, 0, 0, ldt*min(nb, min_mn), batchCount, queue); magma_queue_t cstream; magmablasGetKernelStream(&cstream); magma_int_t streamid; const magma_int_t nbstreams=32; magma_queue_t stream[nbstreams]; for(i=0; i<nbstreams; i++){ magma_queue_create( &stream[i] ); } magma_getvector( batchCount, sizeof(magmaFloatComplex*), dA_array, 1, cpuAarray, 1); magma_getvector( batchCount, sizeof(magmaFloatComplex*), dT_array, 1, cpuTarray, 1); magmablasSetKernelStream(NULL); for(i=0; i<min_mn;i+=nb) { ib = min(nb, min_mn-i); //=============================================== // panel factorization //=============================================== magma_cdisplace_pointers(dW0_displ, dA_array, ldda, i, i, batchCount, queue); magma_cdisplace_pointers(dW2_displ, tau_array, 1, i, 0, batchCount, queue); //dwork is used in panel factorization and trailing matrix update //dW4_displ, dW5_displ are used as workspace and configured inside magma_cgeqrf_panel_batched(m-i, ib, jb, dW0_displ, ldda, dW2_displ, dT_array, ldt, dR_array, ldr, dW1_displ, dW3_displ, dwork, dW4_displ, dW5_displ, info_array, batchCount, myhandle, queue); //=============================================== // end of panel //=============================================== //direct panel matrix V in dW0_displ, magma_cdisplace_pointers(dW0_displ, dA_array, ldda, i, i, batchCount, queue); // copy the upper part of V into dR cgeqrf_copy_upper_batched(ib, jb, dW0_displ, ldda, dR_array, ldr, batchCount, queue); //=============================================== // update trailing matrix //=============================================== //dwork is used in panel factorization and trailing matrix update //reset dW4_displ ldw = nb; cset_pointer(dW4_displ, dwork, 1, 0, 0, ldw*n, batchCount, queue ); offset = ldw*n*batchCount; cset_pointer(dW5_displ, dwork + offset, 1, 0, 0, ldw*n, batchCount, queue ); if( (n-ib-i) > 0) { // set the diagonal of v as one and the upper triangular part as zero magmablas_claset_batched(MagmaUpper, ib, ib, MAGMA_C_ZERO, MAGMA_C_ONE, dW0_displ, ldda, batchCount, queue); magma_cdisplace_pointers(dW2_displ, tau_array, 1, i, 0, batchCount, queue); // it is faster since it is using BLAS-3 GEMM routines, different from lapack implementation magma_clarft_batched(m-i, ib, 0, dW0_displ, ldda, dW2_displ, dT_array, ldt, dW4_displ, nb*ldt, batchCount, myhandle, queue); // perform C = (I-V T^H V^H) * C, C is the trailing matrix //------------------------------------------- // USE STREAM GEMM //------------------------------------------- if( (m-i) > 100 && (n-i-ib) > 100) { // But since the code use the NULL stream everywhere, // so I don't need it, because the NULL stream do the sync by itself //magma_device_sync(); for(k=0; k<batchCount; k++) { streamid = k%nbstreams; magmablasSetKernelStream(stream[streamid]); // the stream gemm must take cpu pointer magma_clarfb_gpu_gemm(MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise, m-i, n-i-ib, ib, cpuAarray[k] + i + i * ldda, ldda, cpuTarray[k], ldt, cpuAarray[k] + i + (i+ib) * ldda, ldda, dwork + nb * n * k, -1, dwork + nb * n * batchCount + nb * n * k, -1); } // need to synchronise to be sure that panel does not start before // finishing the update at least of the next panel // BUT no need for it as soon as the other portion of the code // use the NULL stream which do the sync by itself //magma_device_sync(); magmablasSetKernelStream(NULL); } //------------------------------------------- // USE BATCHED GEMM //------------------------------------------- else { //direct trailing matrix in dW1_displ magma_cdisplace_pointers(dW1_displ, dA_array, ldda, i, i+ib, batchCount, queue); magma_clarfb_gemm_batched( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise, m-i, n-i-ib, ib, (const magmaFloatComplex**)dW0_displ, ldda, (const magmaFloatComplex**)dT_array, ldt, dW1_displ, ldda, dW4_displ, ldw, dW5_displ, ldw, batchCount, myhandle, queue); } }// update the trailing matrix //=============================================== // copy dR back to V after the trailing matrix update magmablas_clacpy_batched(MagmaUpper, ib, ib, dR_array, ldr, dW0_displ, ldda, batchCount, queue); } for(k=0; k<nbstreams; k++){ magma_queue_destroy( stream[k] ); } magmablasSetKernelStream(cstream); cublasDestroy_v2(myhandle); magma_free(dW0_displ); magma_free(dW1_displ); magma_free(dW2_displ); magma_free(dW3_displ); magma_free(dW4_displ); magma_free(dW5_displ); magma_free(dR_array); magma_free(dT_array); magma_free(dR); magma_free(dT); magma_free(dwork); free(cpuAarray); free(cpuTarray); return arginfo; }
//=================================================================================================================== //=================================================================================================================== //=================================================================================================================== extern "C" magma_int_t magma_clarft_batched(magma_int_t n, magma_int_t k, magma_int_t stair_T, magmaFloatComplex **v_array, magma_int_t ldv, magmaFloatComplex **tau_array, magmaFloatComplex **T_array, magma_int_t ldt, magmaFloatComplex **work_array, magma_int_t lwork, magma_int_t batchCount, magma_queue_t queue) { magmaFloatComplex c_one = MAGMA_C_ONE; magmaFloatComplex c_zero = MAGMA_C_ZERO; if ( k <= 0) return 0; if ( stair_T > 0 && k <= stair_T) return 0; magma_int_t maxnb = max_shared_bsiz; if ( lwork < k*ldt) { magma_xerbla( __func__, -(10) ); return -10; } if ( stair_T > 0 && stair_T > maxnb) { magma_xerbla( __func__, -(3) ); return -3; } magma_int_t DEBUG=0; magma_int_t nb = stair_T == 0 ? min(k,maxnb) : stair_T; magma_int_t i, j, prev_n, mycol, rows; magmaFloatComplex **dW1_displ = NULL; magmaFloatComplex **dW2_displ = NULL; magmaFloatComplex **dW3_displ = NULL; magmaFloatComplex **dTstep_array = NULL; magma_malloc((void**)&dW1_displ, batchCount * sizeof(*dW1_displ)); magma_malloc((void**)&dW2_displ, batchCount * sizeof(*dW2_displ)); magma_malloc((void**)&dW3_displ, batchCount * sizeof(*dW3_displ)); magma_malloc((void**)&dTstep_array, batchCount * sizeof(*dTstep_array)); //magmaFloatComplex *Tstep = k > nb ? work : T; if (k > nb) { magma_cdisplace_pointers(dTstep_array, work_array, lwork, 0, 0, batchCount, queue); } else { magma_cdisplace_pointers(dTstep_array, T_array, ldt, 0, 0, batchCount, queue); } //magma_int_t ldtstep = k > nb ? k : ldt; magma_int_t ldtstep = ldt; //a enlever // stair_T = 0 meaning all T // stair_T > 0 meaning the triangular portion of T has been computed. // the value of stair_T is the nb of these triangulars //GEMV compute the whole triangular upper portion of T (phase 1) // TODO addcublas to check perf magma_cgemm_batched( MagmaConjTrans, MagmaNoTrans, k, k, n, c_one, v_array, ldv, v_array, ldv, c_zero, dTstep_array, ldtstep, batchCount, queue ); magmablas_claset_batched( MagmaLower, k, k, MAGMA_C_ZERO, MAGMA_C_ZERO, dTstep_array, ldtstep, batchCount, queue ); // no need for it as T is expected to be lower zero //if (k > nb) magmablas_claset_batched( MagmaLower, k, k, MAGMA_C_ZERO, MAGMA_C_ZERO, dTstep_array, ldtstep, batchCount, queue ); //TRMV //T(1:i-1,i) := T(1:i-1,1:i-1) * W(1:i-1) i=[1:k] // TRMV is split over block of column of size nb // the update should be done from top to bottom so: // 1- a gemm using the previous computed columns // of T to update rectangular upper protion above // the triangle of my columns // 2- the columns need to be updated by a serial // loop over of gemv over itself. since we limit the // shared memory to nb, this nb column // are split vertically by chunk of nb rows dim3 grid(1, 1, batchCount); for (j=0; j < k; j += nb) { prev_n = j; mycol = min(nb, k-j); // note that myrow = prev_n + mycol; if (prev_n > 0 && mycol > 0) { if (DEBUG == 3) { printf("doing gemm on the rectangular portion of size %d %d of T(%d,%d)\n", (int) prev_n, (int) mycol, 0, (int) j ); } magma_cdisplace_pointers(dW1_displ, dTstep_array, ldtstep, 0, j, batchCount, queue); magma_cdisplace_pointers(dW2_displ, T_array, ldt, 0, j, batchCount, queue); magma_cgemm_batched( MagmaNoTrans, MagmaNoTrans, prev_n, mycol, prev_n, c_one, T_array, ldt, dW1_displ, ldtstep, c_zero, dW2_displ, ldt, batchCount, queue ); // update my rectangular portion (prev_n,mycol) using sequence of gemv magma_cdisplace_pointers(dW1_displ, dTstep_array, ldtstep, j, j, batchCount, queue); magma_cdisplace_pointers(dW3_displ, tau_array, 1, j, 0, batchCount, queue); for (i=0; i < prev_n; i += nb) { rows = min(nb,prev_n-i); if (DEBUG == 3) { printf(" doing recctrmv on the rectangular portion of size %d %d of T(%d,%d)\n", (int) rows, (int) mycol, (int) i, (int) j ); } if (rows > 0 && mycol > 0) { magma_cdisplace_pointers(dW2_displ, T_array, ldt, i, j, batchCount, queue); magmablas_clarft_recctrmv_sm32x32_batched(rows, mycol, dW3_displ, dW2_displ, ldt, dW1_displ, ldtstep, batchCount, queue); } } } // the upper rectangular protion is updated, now if needed update the triangular portion if (stair_T == 0) { if (DEBUG == 3) { printf("doing ctrmv on the triangular portion of size %d %d of T(%d,%d)\n", (int) mycol, (int) mycol, (int) j, (int) j ); } if (mycol > 0) { magma_cdisplace_pointers(dW1_displ, dTstep_array, ldtstep, j, j, batchCount, queue); magma_cdisplace_pointers(dW3_displ, tau_array, 1, j, 0, batchCount, queue); magma_cdisplace_pointers(dW2_displ, T_array, ldt, j, j, batchCount, queue); magmablas_clarft_ctrmv_sm32x32_batched(mycol, mycol, dW3_displ, dW1_displ, ldtstep, dW2_displ, ldt, batchCount, queue); } } }// end of j magma_free(dW1_displ); magma_free(dW2_displ); magma_free(dW3_displ); magma_free(dTstep_array); return 0; }