/* BLIS An object-based framework for developing high-performance BLAS-like libraries. Copyright (C) 2014, The University of Texas at Austin Copyright (C) 2019, Forschunszentrum Juelich Copyright (C) 2020, The University of Tokyo 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" // Double-precision composite instructions. #include "armsve_asm_macros_double.h" // 2vx10 microkernels. #include "armsve_asm_2vx10.h" void bli_dgemm_armsve_asm_2vx10_unindexed ( dim_t m, \ dim_t n, \ dim_t k, \ const void* alpha, \ const void* a, \ const void* b, \ const void* beta, \ void* c, inc_t rs_c0, inc_t cs_c0, \ auxinfo_t* data, \ const cntx_t* cntx \ ) { const void* a_next = bli_auxinfo_next_a( data ); const void* b_next = bli_auxinfo_next_b( data ); // Typecast local copies of integers in case dim_t and inc_t are a // different size than is expected by load instructions. uint64_t k_mker = k / 4; uint64_t k_left = k % 4; uint64_t rs_c = rs_c0; uint64_t cs_c = cs_c0; GEMM_UKR_SETUP_CT_ANY( d, m, 10, false ); __asm__ volatile ( " mov x0, xzr \n\t" " ldr x1, %[m] \n\t" " whilelo p0.d, x0, x1 \n\t" " incd x0 \n\t" " whilelo p1.d, x0, x1 \n\t" " \n\t" " ldr x0, %[a] \n\t" " ldr x1, %[b] \n\t" " mov x2, xzr \n\t" " incd x2, ALL, MUL #2 \n\t" // Column-skip of A. " mov x3, #10 \n\t" // Row-skip of B. " \n\t" " ldr x5, %[c] \n\t" " ldr x6, %[rs_c] \n\t" // Row-skip of C. " ldr x7, %[cs_c] \n\t" // Column-skip of C. #ifdef _A64FX " mov x8, 0x3 \n\t" // Tag C address. " lsl x8, x8, #56 \n\t" " orr x5, x5, x8 \n\t" " mov x8, 0x2 \n\t" // Tag B address. " lsl x8, x8, #56 \n\t" " orr x1, x1, x8 \n\t" " mov x8, 0x1 \n\t" // Tag A address. " lsl x8, x8, #56 \n\t" " orr x0, x0, x8 \n\t" #endif " \n\t" " mov x8, #8 \n\t" // Multiply some address skips by sizeof(double). " madd x2, x8, x2, xzr \n\t" // cs_a " madd x3, x8, x3, xzr \n\t" // rs_b " madd x7, x8, x7, xzr \n\t" // cs_c " \n\t" " ldr x4, %[k_mker] \n\t" // Number of loops. " ldr x8, %[k_left] \n\t" " \n\t" LABEL(LOAD_ABC) " cmp x4, #0 \n\t" // Don't preload if no microkernel there. BEQ(END_CCOL_PRFM) " ld1rd z20.d, p0/z, [x1] \n\t" // Load 8/10 of first B row. " ld1rd z21.d, p0/z, [x1, 8] \n\t" " ld1rd z22.d, p0/z, [x1, 16] \n\t" " ld1rd z23.d, p0/z, [x1, 24] \n\t" " ld1rd z24.d, p0/z, [x1, 32] \n\t" " ld1rd z25.d, p0/z, [x1, 40] \n\t" " ld1rd z26.d, p0/z, [x1, 48] \n\t" " ld1rd z27.d, p0/z, [x1, 56] \n\t" " \n\t" GEMM_ACOL_CONTIGUOUS_LOAD(z28,z29,p0,p1,x0) " \n\t" LABEL(CCOL_PRFM) " cmp x6, #1 \n\t" BNE(END_CCOL_PRFM) // Do not prefetch for generic C storage. " mov x16, x5 \n\t" " prfm PLDL1KEEP, [x16] \n\t" " add x16, x16, x7 \n\t" " prfm PLDL1KEEP, [x16] \n\t" " add x16, x16, x7 \n\t" " prfm PLDL1KEEP, [x16] \n\t" " add x16, x16, x7 \n\t" " prfm PLDL1KEEP, [x16] \n\t" " add x16, x16, x7 \n\t" " prfm PLDL1KEEP, [x16] \n\t" " add x16, x16, x7 \n\t" " prfm PLDL1KEEP, [x16] \n\t" " add x16, x16, x7 \n\t" " prfm PLDL1KEEP, [x16] \n\t" " add x16, x16, x7 \n\t" " prfm PLDL1KEEP, [x16] \n\t" " add x16, x16, x7 \n\t" " prfm PLDL1KEEP, [x16] \n\t" " add x16, x16, x7 \n\t" " prfm PLDL1KEEP, [x16] \n\t" LABEL(END_CCOL_PRFM) " \n\t" CLEAR_COL20(z0,z1,z2,z3,z4,z5,z6,z7,z8,z9,z10,z11,z12,z13,z14,z15,z16,z17,z18,z19) " \n\t" " cmp x4, #0 \n\t" // If no 4-microkernel can be applied BEQ(K_LEFT_LOOP) " \n\t" LABEL(K_MKER_LOOP) " \n\t" " add x0, x0, x2 \n\t" // Forward A's address to the next column. GEMM_ACOL_CONTIGUOUS_LOAD(z30,z31,p0,p1,x0) GEMM_2VX10_MKER_LOOP_PLAIN_C_1(z0,z2,z4,z6,z8,z10,z12,z14,z16,z18,z1,z3,z5,z7,z9,z11,z13,z15,z17,z19,p0,z28,z29,z20,z21,z22,z23,z24,z25,z26,z27,x1,x3) " \n\t" " add x0, x0, x2 \n\t" // Forward A's address to the next column. GEMM_ACOL_CONTIGUOUS_LOAD(z28,z29,p0,p1,x0) GEMM_2VX10_MKER_LOOP_PLAIN_C_2(z0,z2,z4,z6,z8,z10,z12,z14,z16,z18,z1,z3,z5,z7,z9,z11,z13,z15,z17,z19,p0,z30,z31,z20,z21,z22,z23,z24,z25,z26,z27,x1,x3) " \n\t" " add x0, x0, x2 \n\t" // Forward A's address to the next column. GEMM_ACOL_CONTIGUOUS_LOAD(z30,z31,p0,p1,x0) GEMM_2VX10_MKER_LOOP_PLAIN_C_3(z0,z2,z4,z6,z8,z10,z12,z14,z16,z18,z1,z3,z5,z7,z9,z11,z13,z15,z17,z19,p0,z28,z29,z20,z21,z22,z23,z24,z25,z26,z27,x1,x3) " \n\t" " subs x4, x4, #1 \n\t" // Decrease counter before final replica. BEQ(FIN_MKER_LOOP) // Branch early to avoid reading excess mem. " \n\t" " add x0, x0, x2 \n\t" // Forward A's address to the next column. GEMM_ACOL_CONTIGUOUS_LOAD(z28,z29,p0,p1,x0) GEMM_2VX10_MKER_LOOP_PLAIN_C_4(z0,z2,z4,z6,z8,z10,z12,z14,z16,z18,z1,z3,z5,z7,z9,z11,z13,z15,z17,z19,p0,z30,z31,z20,z21,z22,z23,z24,z25,z26,z27,x1,x3) BRANCH(K_MKER_LOOP) " \n\t" LABEL(FIN_MKER_LOOP) GEMM_2VX10_MKER_LOOP_PLAIN_C_4_RESIDUAL(z0,z2,z4,z6,z8,z10,z12,z14,z16,z18,z1,z3,z5,z7,z9,z11,z13,z15,z17,z19,p0,z30,z31,z20,z21,z22,z23,z24,z25,z26,z27,x1,x3) " add x0, x0, x2 \n\t" // Forward A to fill the blank. " \n\t" LABEL(K_LEFT_LOOP) " cmp x8, #0 \n\t" // End of execution. BEQ(WRITE_MEM_PREP) " \n\t" GEMM_ACOL_CONTIGUOUS_LOAD(z30,z31,p0,p1,x0) " ld1rd z20.d, p0/z, [x1] \n\t" // Load 8/10 of first B row. " ld1rd z21.d, p0/z, [x1, 8] \n\t" " ld1rd z22.d, p0/z, [x1, 16] \n\t" " ld1rd z23.d, p0/z, [x1, 24] \n\t" " ld1rd z24.d, p0/z, [x1, 32] \n\t" " ld1rd z25.d, p0/z, [x1, 40] \n\t" " ld1rd z26.d, p0/z, [x1, 48] \n\t" " ld1rd z27.d, p0/z, [x1, 56] \n\t" " ld1rd z28.d, p0/z, [x1, 64] \n\t" " ld1rd z29.d, p0/z, [x1, 72] \n\t" GEMM_FMLA2(z0,z1,p0,z30,z31,z20) GEMM_FMLA2(z2,z3,p0,z30,z31,z21) GEMM_FMLA2(z4,z5,p0,z30,z31,z22) GEMM_FMLA2(z6,z7,p0,z30,z31,z23) GEMM_FMLA2(z8,z9,p0,z30,z31,z24) GEMM_FMLA2(z10,z11,p0,z30,z31,z25) GEMM_FMLA2(z12,z13,p0,z30,z31,z26) GEMM_FMLA2(z14,z15,p0,z30,z31,z27) GEMM_FMLA2(z16,z17,p0,z30,z31,z28) GEMM_FMLA2(z18,z19,p0,z30,z31,z29) " add x0, x0, x2 \n\t" // Forward A. " add x1, x1, x3 \n\t" // Forward B. " sub x8, x8, #1 \n\t" BRANCH(K_LEFT_LOOP) " \n\t" LABEL(WRITE_MEM_PREP) " \n\t" " ldr x4, %[alpha] \n\t" // Load alpha & beta (address). " ldr x8, %[beta] \n\t" " ldr x4, [x4] \n\t" // Load alpha & beta (value). " ldr x8, [x8] \n\t" " dup z30.d, x4 \n\t" // Broadcast alpha & beta into vectors. " dup z31.d, x8 \n\t" " fmov d28, #1.0 \n\t" // Prepare FP 1.0. " fmov x16, d28 \n\t" " \n\t" LABEL(PREFETCH_ABNEXT) " ldr x0, %[a_next] \n\t" " ldr x1, %[b_next] \n\t" #ifdef _A64FX " mov x8, 0x2 \n\t" // Tag B address. " lsl x8, x8, #56 \n\t" " orr x1, x1, x8 \n\t" " mov x8, 0x1 \n\t" // Tag A address. " lsl x8, x8, #56 \n\t" " orr x0, x0, x8 \n\t" #endif " prfm PLDL1STRM, [x0] \n\t" " prfm PLDL1STRM, [x0, 256*1] \n\t" // " prfm PLDL2KEEP, [x0, 256*2] \n\t" // " prfm PLDL2KEEP, [x0, 256*3] \n\t" // " prfm PLDL2KEEP, [x0, 256*4] \n\t" // " prfm PLDL2KEEP, [x0, 256*5] \n\t" // " prfm PLDL2KEEP, [x0, 256*6] \n\t" // " prfm PLDL2KEEP, [x0, 256*7] \n\t" // " prfm PLDL2KEEP, [x0, 256*8] \n\t" // " prfm PLDL2KEEP, [x0, 256*9] \n\t" // " prfm PLDL2KEEP, [x0, 256*10] \n\t" // " prfm PLDL2KEEP, [x0, 256*11] \n\t" // " prfm PLDL2KEEP, [x0, 256*12] \n\t" // " prfm PLDL2KEEP, [x0, 256*13] \n\t" // " prfm PLDL2KEEP, [x0, 256*14] \n\t" // " prfm PLDL2KEEP, [x0, 256*15] \n\t" " prfm PLDL1STRM, [x1] \n\t" " prfm PLDL1STRM, [x1, 256*1] \n\t" // " prfm PLDL2KEEP, [x1, 256*2] \n\t" // " prfm PLDL2KEEP, [x1, 256*3] \n\t" // " prfm PLDL2KEEP, [x1, 256*4] \n\t" // " prfm PLDL2KEEP, [x1, 256*5] \n\t" // " prfm PLDL2KEEP, [x1, 256*6] \n\t" // " prfm PLDL2KEEP, [x1, 256*7] \n\t" // " prfm PLDL2KEEP, [x1, 256*8] \n\t" // " prfm PLDL2KEEP, [x1, 256*9] \n\t" " \n\t" " mov x9, x5 \n\t" // C address for loading. " \n\t" // C address for storing is x5 itself. " cmp x6, #1 \n\t" // Preload first half of C for contiguous case. BNE(WRITE_MEM) GEMM_C_LOAD_UKER_C(z20,z22,z24,z26,z28,z21,z23,z25,z27,z29,p0,p1,x9,x7) " \n\t" LABEL(WRITE_MEM) " \n\t" " cmp x16, x4 \n\t" BEQ(UNIT_ALPHA) " \n\t" SCALE_COL20(z0,z1,z2,z3,z4,z5,z6,z7,z8,z9,z10,z11,z12,z13,z14,z15,z16,z17,z18,z19,z30) " \n\t" LABEL(UNIT_ALPHA) " cmp x6, #1 \n\t" BNE(WRITE_MEM_G) " \n\t" LABEL(WRITE_MEM_C) " \n\t" // Available scratch: Z[20-30]. " \n\t" // Here used scratch: Z[20-29]. " fcmp d31, #0.0 \n\t" // Skip loading if *beta == 0 to override NaN. BEQ(BETA_ZERO_C) // First half of C is already loaded in this case. // GEMM_C_FMAD_LOAD_UKER_C(z20,z22,z24,z26,z28,z21,z23,z25,z27,z29,p0,p1,z0,z2,z4,z6,z8,z1,z3,z5,z7,z9,z31,x9,x7) GEMM_C_FMLA_UKER(z0,z2,z4,z6,z8,z1,z3,z5,z7,z9,p0,z20,z22,z24,z26,z28,z21,z23,z25,z27,z29,z31) GEMM_C_LOAD_UKER_C(z20,z22,z24,z26,z28,z21,z23,z25,z27,z29,p0,p1,x9,x7) GEMM_C_FMLA_UKER(z10,z12,z14,z16,z18,z11,z13,z15,z17,z19,p0,z20,z22,z24,z26,z28,z21,z23,z25,z27,z29,z31) LABEL(BETA_ZERO_C) GEMM_C_STORE_UKER_C(z0,z2,z4,z6,z8,z1,z3,z5,z7,z9,p0,p1,x5,x7) GEMM_C_STORE_UKER_C(z10,z12,z14,z16,z18,z11,z13,z15,z17,z19,p0,p1,x5,x7) BRANCH(END_EXEC) // Generic-storage case -- Mainly for transposed storage. LABEL(WRITE_MEM_G) " mov x8, xzr \n\t" " incb x8 \n\t" " madd x8, x8, x6, xzr \n\t" // C-column's logical 1-vector skip. " index z30.d, xzr, x6 \n\t" // Skips passed to index is not multiplied by 8. " \n\t" " fcmp d31, #0.0 \n\t" // Skip loading if *beta == 0 to override NaN. BEQ(BETA_ZERO_G) GEMM_C_LOAD_UKER_G(z20,z22,z24,z26,z28,z21,z23,z25,z27,z29,z30,p0,p1,x9,x7,x8,x16) GEMM_C_FMLA_UKER(z0,z2,z4,z6,z8,z1,z3,z5,z7,z9,p0,z20,z22,z24,z26,z28,z21,z23,z25,z27,z29,z31) GEMM_C_LOAD_UKER_G(z20,z22,z24,z26,z28,z21,z23,z25,z27,z29,z30,p0,p1,x9,x7,x8,x16) GEMM_C_FMLA_UKER(z10,z12,z14,z16,z18,z11,z13,z15,z17,z19,p0,z20,z22,z24,z26,z28,z21,z23,z25,z27,z29,z31) LABEL(BETA_ZERO_G) GEMM_C_STORE_UKER_G(z0,z2,z4,z6,z8,z1,z3,z5,z7,z9,z30,p0,p1,x5,x7,x8,x16) GEMM_C_STORE_UKER_G(z10,z12,z14,z16,z18,z11,z13,z15,z17,z19,z30,p0,p1,x5,x7,x8,x16) LABEL(END_EXEC) " mov x0, #0 \n\t" // Return normal. : : [m] "m" (m), [a] "m" (a), [b] "m" (b), [c] "m" (c), [rs_c] "m" (rs_c), [cs_c] "m" (cs_c), [k_mker] "m" (k_mker), [k_left] "m" (k_left), [alpha] "m" (alpha), [beta] "m" (beta), [a_next] "m" (a_next), [b_next] "m" (b_next) : "x0","x1","x2","x3","x4","x5","x6","x7","x8", "x9","x16", "z0","z1","z2","z3","z4","z5","z6","z7", "z8","z9","z10","z11","z12","z13","z14","z15", "z16","z17","z18","z19", "z20","z21","z22","z23", "z24","z25","z26","z27", "z28","z29","z30","z31" ); GEMM_UKR_FLUSH_CT( d ); }