I need to build OpenBLAS static library which contains assembly functions - https://github.com/xianyi/OpenBLAS/releases . Personally, I'm trying to build 0.3.6 version.
Although, compilation process is successful, .o files are missing actual implementations when it comes to assembly functions.
I have problems with all my assembly functiosn, but I will give as an example one of them - _sdot_k.
How can I define that implementation is missing? I can define it by that fact that when I do nm [library_name].a, output for this function is following:
libopenblas_armv8p-r0.3.6.a(sdot_k.o):
0000000000000050 t .Ldot_kernel_F1
0000000000000058 t .Ldot_kernel_F10
0000000000000028 t .Ldot_kernel_F4
000000000000001c t .Ldot_kernel_F_BEGIN
00000000000000d8 t .Ldot_kernel_L999
00000000000000bc t .Ldot_kernel_S1
00000000000000c4 t .Ldot_kernel_S10
0000000000000084 t .Ldot_kernel_S4
0000000000000070 t .Ldot_kernel_S_BEGIN
0000000000000000 t ltmp0
There's no T identifier which tells that function implementation is here, in .o file. And of course, if I put this library into my iOS project I have such error:
Undefined symbols for architecture arm64:
"_sdot_k", referenced from:
_strmv_TLN in libopenblas.a(strmv_TLN.o)
_strmv_TLU in libopenblas.a(strmv_TLU.o)
_strmv_TUN in libopenblas.a(strmv_TUN.o)
_strmv_TUU in libopenblas.a(strmv_TUU.o)
_trmv_kernel in libopenblas.a(strmv_thread_TLN.o)
_trmv_kernel in libopenblas.a(strmv_thread_TLU.o)
_trmv_kernel in libopenblas.a(strmv_thread_TUN.o)
...
Here's the actual dot.S file:
#define ASSEMBLER
#include "common.h"
#define N x0 /* vector length */
#define X x1 /* X vector address */
#define INC_X x2 /* X stride */
#define Y x3 /* Y vector address */
#define INC_Y x4 /* Y stride */
#define I x5 /* loop variable */
/*******************************************************************************
* Macro definitions
*******************************************************************************/
#if !defined(DOUBLE)
#if !defined(DSDOT)
#define REG0 wzr
#define DOTF s0
#else // DSDOT
#define REG0 xzr
#define DOTF d0
#endif
#define DOTI s1
#define TMPX s2
#define LD1VX {v2.s}[0]
#define TMPY s3
#define LD1VY {v3.s}[0]
#define TMPVY v3.s[0]
#define SZ 4
#else
#define REG0 xzr
#define DOTF d0
#define DOTI d1
#define TMPX d2
#define LD1VX {v2.d}[0]
#define TMPY d3
#define LD1VY {v3.d}[0]
#define TMPVY v3.d[0]
#define SZ 8
#endif
/******************************************************************************/
.macro KERNEL_F1
ldr TMPX, [X], #SZ
ldr TMPY, [Y], #SZ
#if !defined(DSDOT)
fmadd DOTF, TMPX, TMPY, DOTF
#else // DSDOT
fcvt d3, TMPY
fcvt d2, TMPX
fmul d2, d2, d3
fadd DOTF, DOTF, d2
#endif
.endm
.macro KERNEL_F4
#if !defined(DOUBLE)
ld1 {v2.4s}, [X], #16
ld1 {v3.4s}, [Y], #16
#if !defined(DSDOT)
fmla v0.4s, v2.4s, v3.4s
#else
fcvtl2 v5.2d, v3.4s
fcvtl2 v4.2d, v2.4s
fcvtl v3.2d, v3.2s
fcvtl v2.2d, v2.2s
fmul v4.2d, v4.2d, v5.2d
fmul v2.2d, v2.2d, v3.2d
fadd v2.2d, v2.2d, v4.2d
fadd v0.2d, v0.2d, v2.2d
#endif
#else //DOUBLE
ld1 {v2.2d, v3.2d}, [X], #32
ld1 {v4.2d, v5.2d}, [Y], #32
fmul v2.2d, v2.2d, v4.2d
fmul v3.2d, v3.2d, v5.2d
fadd v0.2d, v0.2d, v2.2d
fadd v0.2d, v0.2d, v3.2d
#endif
PRFM PLDL1KEEP, [X, #1024]
PRFM PLDL1KEEP, [Y, #1024]
.endm
.macro KERNEL_F4_FINALIZE
#if !defined(DOUBLE)
#if !defined(DSDOT)
ext v1.16b, v0.16b, v0.16b, #8
fadd v0.2s, v0.2s, v1.2s
faddp DOTF, v0.2s
#else
faddp DOTF, v0.2d
#endif
#else //DOUBLE
faddp DOTF, v0.2d
#endif
.endm
.macro INIT_S
#if !defined(DOUBLE)
lsl INC_X, INC_X, #2
lsl INC_Y, INC_Y, #2
#else
lsl INC_X, INC_X, #3
lsl INC_Y, INC_Y, #3
#endif
.endm
.macro KERNEL_S1
ld1 LD1VX, [X], INC_X
ld1 LD1VY, [Y], INC_Y
#if !defined(DSDOT)
fmadd DOTF, TMPX, TMPY, DOTF
#else // DSDOT
fcvt d3, TMPY
fcvt d2, TMPX
fmul d2, d2, d3
fadd DOTF, DOTF, d2
#endif
.endm
/*******************************************************************************
* End of macro definitions
*******************************************************************************/
PROLOGUE
fmov DOTF, REG0
#if defined(DOUBLE)
fmov d6, DOTF
#endif
cmp N, xzr
ble .Ldot_kernel_L999
cmp INC_X, #1
bne .Ldot_kernel_S_BEGIN
cmp INC_Y, #1
bne .Ldot_kernel_S_BEGIN
.Ldot_kernel_F_BEGIN:
asr I, N, #2
cmp I, xzr
beq .Ldot_kernel_F1
.Ldot_kernel_F4:
KERNEL_F4
subs I, I, #1
bne .Ldot_kernel_F4
KERNEL_F4_FINALIZE
.Ldot_kernel_F1:
ands I, N, #3
ble .Ldot_kernel_L999
.Ldot_kernel_F10:
KERNEL_F1
subs I, I, #1
bne .Ldot_kernel_F10
ret
.Ldot_kernel_S_BEGIN:
INIT_S
asr I, N, #2
cmp I, xzr
ble .Ldot_kernel_S1
.Ldot_kernel_S4:
KERNEL_S1
KERNEL_S1
KERNEL_S1
KERNEL_S1
subs I, I, #1
bne .Ldot_kernel_S4
.Ldot_kernel_S1:
ands I, N, #3
ble .Ldot_kernel_L999
.Ldot_kernel_S10:
KERNEL_S1
subs I, I, #1
bne .Ldot_kernel_S10
.Ldot_kernel_L999:
ret
EPILOGUE
As you can see, there's PROLOGUE and EPILOGUE which are macroses, defined in C header:
#ifndef F_INTERFACE
#define REALNAME ASMNAME
#else
#define REALNAME ASMFNAME
#endif
#if defined(ASSEMBLER) && !defined(NEEDPARAM)
#define PROLOGUE \
.text ;\
.align 2 ;\
.globl REALNAME ;\
REALNAME:
#define EPILOGUE
#define PROFCODE
#endif
My guts are telling me that the problem lies somewhere in EPILOGUE and PROLOGUE, but I don't have decent Assembler knowledge to figure out a problem on my own.
Also I found this article - https://developer.apple.com/library/archive/documentation/DeveloperTools/Reference/Assembler/040-Assembler_Directives/asm_directives.html, but it didn't help me much. Maybe because of my lack of Assembly skill.
NOTE: If anyone is struggling with the same library trying to make it work on iOS, here's my thread on it's Github - https://github.com/xianyi/OpenBLAS/issues/2275#issuecomment-536982253is. It contains descriptions of all problems I overcomed.
NOTE 2: I really need OpenBLAS library and using Accelerate.framework is not an option in my case, unfortunately.
I'm happy to say that I overcome that issue. The problem was indeed in my PROLOGUE macros.
PROLOGUE macros I was using was converting to the assembly code in this manner:
.text; .align 2; .globl REALNAME; REALNAME:
And that was an issue. In order to make Assembly label work it should look like this:
.text;
.align 2;
.globl REALNAME;
REALNAME:
//assembly code
So in order to achieve that result I changed C macro to GAS macro as such:
.macro PROLOGUE
.text
.align 2
.globl REALNAME
REALNAME:
.endm
After that issue was gone!
NOTE: In order to remove any confusions, ; doesn't really play a role here. It's new line after each comma what matters.