.text .balign 4 .global sgemm_nn # RV64IDV system # # void # sgemm_nn(size_t n, # size_t m, # size_t k, # const float*a, // m * k matrix # size_t lda, # const float*b, // k * n matrix # size_t ldb, # float*c, // m * n matrix # size_t ldc) # # c += a*b (alpha=1, no transpose on input matrices) # matrices stored in C row-major order #define n a0 #define m a1 #define k a2 #define ap a3 #define astride a4 #define bp a5 #define bstride a6 #define cp a7 #define cstride t0 #define kt t1 #define nt t2 #define bnp t3 #define cnp t4 #define akp t5 #define bkp s0 #define nvl s1 #define ccp s2 #define amp s3 # Use args as additional temporaries #define ft12 fa0 #define ft13 fa1 #define ft14 fa2 #define ft15 fa3 # This version holds a 16*VLMAX block of C matrix in vector registers # in inner loop, but otherwise does not cache or TLB tiling. sgemm_nn: addi sp, sp, -FRAMESIZE sd s0, OFFSET(sp) sd s1, OFFSET(sp) sd s2, OFFSET(sp) # Check for zero size matrices beqz n, exit beqz m, exit beqz k, exit # Convert elements strides to byte strides. ld cstride, OFFSET(sp) # Get arg from stack frame slli astride, astride, 2 slli bstride, bstride, 2 slli cstride, cstride, 2 slti t6, m, 16 bnez t6, end_rows c_row_loop: # Loop across rows of C blocks mv nt, n # Initialize n counter for next row of C blocks mv bnp, bp # Initialize B n-loop pointer to start mv cnp, cp # Initialize C n-loop pointer c_col_loop: # Loop across one row of C blocks vsetvli nvl, nt, e32, ta, ma # 32-bit vectors, LMUL=1 mv akp, ap # reset pointer into A to beginning mv bkp, bnp # step to next column in B matrix # Initalize current C submatrix block from memory. vle32.v v0, (cnp); add ccp, cnp, cstride; vle32.v v1, (ccp); add ccp, ccp, cstride; vle32.v v2, (ccp); add ccp, ccp, cstride; vle32.v v3, (ccp); add ccp, ccp, cstride; vle32.v v4, (ccp); add ccp, ccp, cstride; vle32.v v5, (ccp); add ccp, ccp, cstride; vle32.v v6, (ccp); add ccp, ccp, cstride; vle32.v v7, (ccp); add ccp, ccp, cstride; vle32.v v8, (ccp); add ccp, ccp, cstride; vle32.v v9, (ccp); add ccp, ccp, cstride; vle32.v v10, (ccp); add ccp, ccp, cstride; vle32.v v11, (ccp); add ccp, ccp, cstride; vle32.v v12, (ccp); add ccp, ccp, cstride; vle32.v v13, (ccp); add ccp, ccp, cstride; vle32.v v14, (ccp); add ccp, ccp, cstride; vle32.v v15, (ccp) mv kt, k # Initialize inner loop counter # Inner loop scheduled assuming 4-clock occupancy of vfmacc instruction and single-issue pipeline # Software pipeline loads flw ft0, (akp); add amp, akp, astride; flw ft1, (amp); add amp, amp, astride; flw ft2, (amp); add amp, amp, astride; flw ft3, (amp); add amp, amp, astride; # Get vector from B matrix vle32.v v16, (bkp) # Loop on inner dimension for current C block k_loop: vfmacc.vf v0, ft0, v16 add bkp, bkp, bstride flw ft4, (amp) add amp, amp, astride vfmacc.vf v1, ft1, v16 addi kt, kt, -1 # Decrement k counter flw ft5, (amp) add amp, amp, astride vfmacc.vf v2, ft2, v16 flw ft6, (amp) add amp, amp, astride flw ft7, (amp) vfmacc.vf v3, ft3, v16 add amp, amp, astride flw ft8, (amp) add amp, amp, astride vfmacc.vf v4, ft4, v16 flw ft9, (amp) add amp, amp, astride vfmacc.vf v5, ft5, v16 flw ft10, (amp) add amp, amp, astride vfmacc.vf v6, ft6, v16 flw ft11, (amp) add amp, amp, astride vfmacc.vf v7, ft7, v16 flw ft12, (amp) add amp, amp, astride vfmacc.vf v8, ft8, v16 flw ft13, (amp) add amp, amp, astride vfmacc.vf v9, ft9, v16 flw ft14, (amp) add amp, amp, astride vfmacc.vf v10, ft10, v16 flw ft15, (amp) add amp, amp, astride addi akp, akp, 4 # Move to next column of a vfmacc.vf v11, ft11, v16 beqz kt, 1f # Don't load past end of matrix flw ft0, (akp) add amp, akp, astride 1: vfmacc.vf v12, ft12, v16 beqz kt, 1f flw ft1, (amp) add amp, amp, astride 1: vfmacc.vf v13, ft13, v16 beqz kt, 1f flw ft2, (amp) add amp, amp, astride 1: vfmacc.vf v14, ft14, v16 beqz kt, 1f # Exit out of loop flw ft3, (amp) add amp, amp, astride vfmacc.vf v15, ft15, v16 vle32.v v16, (bkp) # Get next vector from B matrix, overlap loads with jump stalls j k_loop 1: vfmacc.vf v15, ft15, v16 # Save C matrix block back to memory vse32.v v0, (cnp); add ccp, cnp, cstride; vse32.v v1, (ccp); add ccp, ccp, cstride; vse32.v v2, (ccp); add ccp, ccp, cstride; vse32.v v3, (ccp); add ccp, ccp, cstride; vse32.v v4, (ccp); add ccp, ccp, cstride; vse32.v v5, (ccp); add ccp, ccp, cstride; vse32.v v6, (ccp); add ccp, ccp, cstride; vse32.v v7, (ccp); add ccp, ccp, cstride; vse32.v v8, (ccp); add ccp, ccp, cstride; vse32.v v9, (ccp); add ccp, ccp, cstride; vse32.v v10, (ccp); add ccp, ccp, cstride; vse32.v v11, (ccp); add ccp, ccp, cstride; vse32.v v12, (ccp); add ccp, ccp, cstride; vse32.v v13, (ccp); add ccp, ccp, cstride; vse32.v v14, (ccp); add ccp, ccp, cstride; vse32.v v15, (ccp) # Following tail instructions should be scheduled earlier in free slots during C block save. # Leaving here for clarity. # Bump pointers for loop across blocks in one row slli t6, nvl, 2 add cnp, cnp, t6 # Move C block pointer over add bnp, bnp, t6 # Move B block pointer over sub nt, nt, nvl # Decrement element count in n dimension bnez nt, c_col_loop # Any more to do? # Move to next set of rows addi m, m, -16 # Did 16 rows above slli t6, astride, 4 # Multiply astride by 16 add ap, ap, t6 # Move A matrix pointer down 16 rows slli t6, cstride, 4 # Multiply cstride by 16 add cp, cp, t6 # Move C matrix pointer down 16 rows slti t6, m, 16 beqz t6, c_row_loop # Handle end of matrix with fewer than 16 rows. # Can use smaller versions of above decreasing in powers-of-2 depending on code-size concerns. end_rows: # Not done. exit: ld s0, OFFSET(sp) ld s1, OFFSET(sp) ld s2, OFFSET(sp) addi sp, sp, FRAMESIZE ret