/* BLIS An object-based framework for developing high-performance BLAS-like libraries. Copyright (C) 2014, The University of Texas at Austin Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: - Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. - Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. - Neither the name(s) of the copyright holder(s) nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include "blis.h" void bli_sgemm_opt_d2x4( dim_t k, float* restrict alpha, float* restrict a, float* restrict b, float* restrict beta, float* restrict c, inc_t rs_c, inc_t cs_c ) { bli_check_error_code( BLIS_NOT_YET_IMPLEMENTED ); } void bli_dgemm_opt_d2x4( dim_t k, double* restrict alpha, double* restrict a, double* restrict b, double* restrict beta, double* restrict c, inc_t rs_c, inc_t cs_c ) { dim_t k_iter; dim_t k_left; k_iter = k / 8; k_left = k % 8; __asm__ volatile ( " \n\t" "movl %6, %%ecx \n\t" // load address of c " \n\t" "movl %8, %%edi \n\t" // load cs_c "sall $3, %%edi \n\t" // cs_c *= sizeof(double) " \n\t" "leal (%%ecx,%%edi,2), %%edx \n\t" // load address of c + 2*cs_c " \n\t" "prefetcht0 (%%ecx) \n\t" // give a T0 prefetch hint for c00. "prefetcht0 (%%ecx,%%edi) \n\t" // give a T0 prefetch hint for c01. "prefetcht0 (%%edx) \n\t" // give a T0 prefetch hint for c02. "prefetcht0 (%%edx,%%edi) \n\t" // give a T0 prefetch hint for c03. " \n\t" "movl %2, %%eax \n\t" // load address of a. "movl %3, %%ebx \n\t" // load address of b. " \n\t" "addl $8 * 16, %%eax \n\t" // increment pointers to allow byte "addl $8 * 16, %%ebx \n\t" // offsets in the unrolled iterations. " \n\t" "movapd -8 * 16(%%eax), %%xmm0 \n\t" // initialize loop by pre-loading elements "movapd -8 * 16(%%ebx), %%xmm1 \n\t" // of a and b. " \n\t" "pxor %%xmm2, %%xmm2 \n\t" "pxor %%xmm3, %%xmm3 \n\t" " \n\t" "pxor %%xmm4, %%xmm4 \n\t" "pxor %%xmm5, %%xmm5 \n\t" "pxor %%xmm6, %%xmm6 \n\t" "pxor %%xmm7, %%xmm7 \n\t" " \n\t" " \n\t" " \n\t" "movl %0, %%esi \n\t" // i = k_iter; "testl %%esi, %%esi \n\t" // check i via logical AND. "je .CONSIDERKLEFT \n\t" // if i == 0, jump to code that " \n\t" // contains the k_left loop. " \n\t" " \n\t" ".LOOPKITER: \n\t" // MAIN LOOP " \n\t" "prefetcht0 (8*21+4)*8(%%eax) \n\t" " \n\t" " \n\t" "addpd %%xmm3, %%xmm7 \n\t" // iteration 0 "movapd -7 * 16(%%ebx), %%xmm3 \n\t" "addpd %%xmm2, %%xmm6 \n\t" "pshufd $0x4e, %%xmm1, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm1 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" " \n\t" "addpd %%xmm1, %%xmm5 \n\t" "movapd -6 * 16(%%ebx), %%xmm1 \n\t" "addpd %%xmm2, %%xmm4 \n\t" "pshufd $0x4e, %%xmm3, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm3 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" "movapd -7 * 16(%%eax), %%xmm0 \n\t" " \n\t" " \n\t" "addpd %%xmm3, %%xmm7 \n\t" // iteration 1 "movapd -5 * 16(%%ebx), %%xmm3 \n\t" "addpd %%xmm2, %%xmm6 \n\t" "pshufd $0x4e, %%xmm1, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm1 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" " \n\t" "addpd %%xmm1, %%xmm5 \n\t" "movapd -4 * 16(%%ebx), %%xmm1 \n\t" "addpd %%xmm2, %%xmm4 \n\t" "pshufd $0x4e, %%xmm3, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm3 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" "movapd -6 * 16(%%eax), %%xmm0 \n\t" " \n\t" " \n\t" "addpd %%xmm3, %%xmm7 \n\t" // iteration 2 "movapd -3 * 16(%%ebx), %%xmm3 \n\t" "addpd %%xmm2, %%xmm6 \n\t" "pshufd $0x4e, %%xmm1, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm1 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" " \n\t" "addpd %%xmm1, %%xmm5 \n\t" "movapd -2 * 16(%%ebx), %%xmm1 \n\t" "addpd %%xmm2, %%xmm4 \n\t" "pshufd $0x4e, %%xmm3, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm3 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" "movapd -5 * 16(%%eax), %%xmm0 \n\t" " \n\t" " \n\t" "addpd %%xmm3, %%xmm7 \n\t" // iteration 3 "movapd -1 * 16(%%ebx), %%xmm3 \n\t" "addpd %%xmm2, %%xmm6 \n\t" "pshufd $0x4e, %%xmm1, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm1 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" " \n\t" "addpd %%xmm1, %%xmm5 \n\t" "movapd 0 * 16(%%ebx), %%xmm1 \n\t" "addpd %%xmm2, %%xmm4 \n\t" "pshufd $0x4e, %%xmm3, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm3 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" "movapd -4 * 16(%%eax), %%xmm0 \n\t" " \n\t" " \n\t" "prefetcht0 (8*21+12)*8(%%eax) \n\t" " \n\t" " \n\t" "addpd %%xmm3, %%xmm7 \n\t" // iteration 4 "movapd 1 * 16(%%ebx), %%xmm3 \n\t" "addpd %%xmm2, %%xmm6 \n\t" "pshufd $0x4e, %%xmm1, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm1 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" " \n\t" "addpd %%xmm1, %%xmm5 \n\t" "movapd 2 * 16(%%ebx), %%xmm1 \n\t" "addpd %%xmm2, %%xmm4 \n\t" "pshufd $0x4e, %%xmm3, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm3 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" "movapd -3 * 16(%%eax), %%xmm0 \n\t" " \n\t" " \n\t" "addpd %%xmm3, %%xmm7 \n\t" // iteration 5 "movapd 3 * 16(%%ebx), %%xmm3 \n\t" "addpd %%xmm2, %%xmm6 \n\t" "pshufd $0x4e, %%xmm1, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm1 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" " \n\t" "addpd %%xmm1, %%xmm5 \n\t" "movapd 4 * 16(%%ebx), %%xmm1 \n\t" "addpd %%xmm2, %%xmm4 \n\t" "pshufd $0x4e, %%xmm3, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm3 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" "movapd -2 * 16(%%eax), %%xmm0 \n\t" " \n\t" " \n\t" "addpd %%xmm3, %%xmm7 \n\t" // iteration 6 "movapd 5 * 16(%%ebx), %%xmm3 \n\t" "addpd %%xmm2, %%xmm6 \n\t" "pshufd $0x4e, %%xmm1, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm1 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" " \n\t" "addpd %%xmm1, %%xmm5 \n\t" "movapd 6 * 16(%%ebx), %%xmm1 \n\t" "addpd %%xmm2, %%xmm4 \n\t" "pshufd $0x4e, %%xmm3, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm3 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" "movapd -1 * 16(%%eax), %%xmm0 \n\t" " \n\t" " \n\t" "addpd %%xmm3, %%xmm7 \n\t" // iteration 7 "movapd 7 * 16(%%ebx), %%xmm3 \n\t" "addpd %%xmm2, %%xmm6 \n\t" "pshufd $0x4e, %%xmm1, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm1 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" " \n\t" "addpd %%xmm1, %%xmm5 \n\t" "movapd 8 * 16(%%ebx), %%xmm1 \n\t" "addpd %%xmm2, %%xmm4 \n\t" "pshufd $0x4e, %%xmm3, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm3 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" "movapd 0 * 16(%%eax), %%xmm0 \n\t" " \n\t" "addl $256, %%ebx \n\t" // b += 8*2*2 (unroll x nr x ndup) "addl $128, %%eax \n\t" // a += 8*4 (unroll x mr) " \n\t" " \n\t" "decl %%esi \n\t" // i -= 1; "jne .LOOPKITER \n\t" // iterate again if i != 0. " \n\t" " \n\t" " \n\t" ".CONSIDERKLEFT: \n\t" " \n\t" "movl %1, %%esi \n\t" // i = k_left; "testl %%esi, %%esi \n\t" // check i via logical AND. "je .POSTACCUM \n\t" // if i == 0, we're done; jump to end. " \n\t" // else, we prepare to enter k_left loop. " \n\t" " \n\t" ".LOOPKLEFT: \n\t" // EDGE LOOP " \n\t" "addpd %%xmm3, %%xmm7 \n\t" // iteration i "movapd -7 * 16(%%ebx), %%xmm3 \n\t" "addpd %%xmm2, %%xmm6 \n\t" "pshufd $0x4e, %%xmm1, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm1 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" " \n\t" "addpd %%xmm1, %%xmm5 \n\t" "movapd -6 * 16(%%ebx), %%xmm1 \n\t" "addpd %%xmm2, %%xmm4 \n\t" "pshufd $0x4e, %%xmm3, %%xmm2 \n\t" "mulpd %%xmm0, %%xmm3 \n\t" "mulpd %%xmm0, %%xmm2 \n\t" "movapd -7 * 16(%%eax), %%xmm0 \n\t" " \n\t" " \n\t" " \n\t" " \n\t" "addl $32, %%ebx \n\t" // b += 4 (1 x mr) "addl $16, %%eax \n\t" // a += 2*2 (1 x nr x ndup) " \n\t" "decl %%esi \n\t" // i -= 1; "jne .LOOPKLEFT \n\t" // iterate again if i != 0. " \n\t" " \n\t" " \n\t" ".POSTACCUM: \n\t" " \n\t" "addpd %%xmm2, %%xmm6 \n\t" "addpd %%xmm3, %%xmm7 \n\t" " \n\t" " \n\t" "movl %4, %%eax \n\t" // load address of alpha "movl %5, %%ebx \n\t" // load address of beta "movddup (%%eax), %%xmm2 \n\t" // load alpha and duplicate "movddup (%%ebx), %%xmm3 \n\t" // load beta and duplicate " \n\t" " \n\t" " \n\t" "movl %7, %%esi \n\t" // load rs_c "sall $3, %%esi \n\t" // rs_c *= sizeof(double) " \n\t" " \n\t" "movapd %%xmm4, %%xmm0 \n\t" "movsd %%xmm5, %%xmm4 \n\t" "movsd %%xmm0, %%xmm5 \n\t" " \n\t" "movapd %%xmm6, %%xmm0 \n\t" "movsd %%xmm7, %%xmm6 \n\t" "movsd %%xmm0, %%xmm7 \n\t" " \n\t" " \n\t" " \n\t" "movlpd (%%ecx), %%xmm0 \n\t" // load c00 and c10, "movhpd (%%ecx,%%esi), %%xmm0 \n\t" "mulpd %%xmm2, %%xmm4 \n\t" // scale by alpha, "mulpd %%xmm3, %%xmm0 \n\t" // scale by beta, "addpd %%xmm4, %%xmm0 \n\t" // add the gemm result, "movlpd %%xmm0, (%%ecx) \n\t" // and store back to memory. "movhpd %%xmm0, (%%ecx,%%esi) \n\t" "addl %%edi, %%ecx \n\t" " \n\t" "movlpd (%%edx), %%xmm1 \n\t" // load c02 and c12, "movhpd (%%edx,%%esi), %%xmm1 \n\t" "mulpd %%xmm2, %%xmm6 \n\t" // scale by alpha, "mulpd %%xmm3, %%xmm1 \n\t" // scale by beta, "addpd %%xmm6, %%xmm1 \n\t" // add the gemm result, "movlpd %%xmm1, (%%edx) \n\t" // and store back to memory. "movhpd %%xmm1, (%%edx,%%esi) \n\t" "addl %%edi, %%edx \n\t" " \n\t" "movlpd (%%ecx), %%xmm0 \n\t" // load c01 and c11, "movhpd (%%ecx,%%esi), %%xmm0 \n\t" "mulpd %%xmm2, %%xmm5 \n\t" // scale by alpha, "mulpd %%xmm3, %%xmm0 \n\t" // scale by beta, "addpd %%xmm5, %%xmm0 \n\t" // add the gemm result, "movlpd %%xmm0, (%%ecx) \n\t" // and store back to memory. "movhpd %%xmm0, (%%ecx,%%esi) \n\t" " \n\t" "movlpd (%%edx), %%xmm1 \n\t" // load c03 and c13, "movhpd (%%edx,%%esi), %%xmm1 \n\t" "mulpd %%xmm2, %%xmm7 \n\t" // scale by alpha, "mulpd %%xmm3, %%xmm1 \n\t" // scale by beta, "addpd %%xmm7, %%xmm1 \n\t" // add the gemm result, "movlpd %%xmm1, (%%edx) \n\t" // and store back to memory. "movhpd %%xmm1, (%%edx,%%esi) \n\t" " \n\t" " \n\t" " \n\t" : // output operands (none) : // input operands "m" (k_iter), "m" (k_left), "m" (a), "m" (b), "m" (alpha), "m" (beta), "m" (c), "m" (rs_c), "m" (cs_c) : // register clobber list "eax", "ebx", "ecx", "edx", "esi", "edi", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7", "memory" ); } void bli_cgemm_opt_d2x4( dim_t k, scomplex* restrict alpha, scomplex* restrict a, scomplex* restrict b, scomplex* restrict beta, scomplex* restrict c, inc_t rs_c, inc_t cs_c ) { bli_check_error_code( BLIS_NOT_YET_IMPLEMENTED ); } void bli_zgemm_opt_d2x4( dim_t k, dcomplex* restrict alpha, dcomplex* restrict a, dcomplex* restrict b, dcomplex* restrict beta, dcomplex* restrict c, inc_t rs_c, inc_t cs_c ) { bli_check_error_code( BLIS_NOT_YET_IMPLEMENTED ); }