github.com/jingcheng-WU/gonum@v0.9.1-0.20210323123734-f1a2a11a8f7b/internal/asm/f32/axpyunitary_amd64.s (about)

     1  // Copyright ©2016 The Gonum Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  // +build !noasm,!gccgo,!safe
     6  
     7  #include "textflag.h"
     8  
     9  // func AxpyUnitary(alpha float32, x, y []float32)
    10  TEXT ·AxpyUnitary(SB), NOSPLIT, $0
    11  	MOVQ    x_base+8(FP), SI  // SI = &x
    12  	MOVQ    y_base+32(FP), DI // DI = &y
    13  	MOVQ    x_len+16(FP), BX  // BX = min( len(x), len(y) )
    14  	CMPQ    y_len+40(FP), BX
    15  	CMOVQLE y_len+40(FP), BX
    16  	CMPQ    BX, $0            // if BX == 0 { return }
    17  	JE      axpy_end
    18  	MOVSS   alpha+0(FP), X0
    19  	SHUFPS  $0, X0, X0        // X0 = { a, a, a, a }
    20  	XORQ    AX, AX            // i = 0
    21  	PXOR    X2, X2            // 2 NOP instructions (PXOR) to align
    22  	PXOR    X3, X3            // loop to cache line
    23  	MOVQ    DI, CX
    24  	ANDQ    $0xF, CX          // Align on 16-byte boundary for ADDPS
    25  	JZ      axpy_no_trim      // if CX == 0 { goto axpy_no_trim }
    26  
    27  	XORQ $0xF, CX // CX = 4 - floor( BX % 16 / 4 )
    28  	INCQ CX
    29  	SHRQ $2, CX
    30  
    31  axpy_align: // Trim first value(s) in unaligned buffer  do {
    32  	MOVSS (SI)(AX*4), X2 // X2 = x[i]
    33  	MULSS X0, X2         // X2 *= a
    34  	ADDSS (DI)(AX*4), X2 // X2 += y[i]
    35  	MOVSS X2, (DI)(AX*4) // y[i] = X2
    36  	INCQ  AX             // i++
    37  	DECQ  BX
    38  	JZ    axpy_end       // if --BX == 0 { return }
    39  	LOOP  axpy_align     // } while --CX > 0
    40  
    41  axpy_no_trim:
    42  	MOVUPS X0, X1           // Copy X0 to X1 for pipelining
    43  	MOVQ   BX, CX
    44  	ANDQ   $0xF, BX         // BX = len % 16
    45  	SHRQ   $4, CX           // CX = int( len / 16 )
    46  	JZ     axpy_tail4_start // if CX == 0 { return }
    47  
    48  axpy_loop: // Loop unrolled 16x   do {
    49  	MOVUPS (SI)(AX*4), X2   // X2 = x[i:i+4]
    50  	MOVUPS 16(SI)(AX*4), X3
    51  	MOVUPS 32(SI)(AX*4), X4
    52  	MOVUPS 48(SI)(AX*4), X5
    53  	MULPS  X0, X2           // X2 *= a
    54  	MULPS  X1, X3
    55  	MULPS  X0, X4
    56  	MULPS  X1, X5
    57  	ADDPS  (DI)(AX*4), X2   // X2 += y[i:i+4]
    58  	ADDPS  16(DI)(AX*4), X3
    59  	ADDPS  32(DI)(AX*4), X4
    60  	ADDPS  48(DI)(AX*4), X5
    61  	MOVUPS X2, (DI)(AX*4)   // dst[i:i+4] = X2
    62  	MOVUPS X3, 16(DI)(AX*4)
    63  	MOVUPS X4, 32(DI)(AX*4)
    64  	MOVUPS X5, 48(DI)(AX*4)
    65  	ADDQ   $16, AX          // i += 16
    66  	LOOP   axpy_loop        // while (--CX) > 0
    67  	CMPQ   BX, $0           // if BX == 0 { return }
    68  	JE     axpy_end
    69  
    70  axpy_tail4_start: // Reset loop counter for 4-wide tail loop
    71  	MOVQ BX, CX          // CX = floor( BX / 4 )
    72  	SHRQ $2, CX
    73  	JZ   axpy_tail_start // if CX == 0 { goto axpy_tail_start }
    74  
    75  axpy_tail4: // Loop unrolled 4x   do {
    76  	MOVUPS (SI)(AX*4), X2 // X2 = x[i]
    77  	MULPS  X0, X2         // X2 *= a
    78  	ADDPS  (DI)(AX*4), X2 // X2 += y[i]
    79  	MOVUPS X2, (DI)(AX*4) // y[i] = X2
    80  	ADDQ   $4, AX         // i += 4
    81  	LOOP   axpy_tail4     // } while --CX > 0
    82  
    83  axpy_tail_start: // Reset loop counter for 1-wide tail loop
    84  	MOVQ BX, CX   // CX = BX % 4
    85  	ANDQ $3, CX
    86  	JZ   axpy_end // if CX == 0 { return }
    87  
    88  axpy_tail:
    89  	MOVSS (SI)(AX*4), X1 // X1 = x[i]
    90  	MULSS X0, X1         // X1 *= a
    91  	ADDSS (DI)(AX*4), X1 // X1 += y[i]
    92  	MOVSS X1, (DI)(AX*4) // y[i] = X1
    93  	INCQ  AX             // i++
    94  	LOOP  axpy_tail      // } while --CX > 0
    95  
    96  axpy_end:
    97  	RET