Gitiles
Code Review Sign In
realtimeroboticsgroup.org / RealtimeRoboticsGroup / aos / 9a24b370cbc07492fa6c205f1027367fa55b6742 / . / kernel / avx2 / kernel_sgemm_8x4_lib8.S
blob: 44946f1dc3852ccb2938295189bfaf9ed7960d2d [file] [log] [blame]
/**************************************************************************************************
* *
* This file is part of BLASFEO. *
* *
* BLASFEO -- BLAS For Embedded Optimization. *
* Copyright (C) 2016-2017 by Gianluca Frison. *
* Developed at IMTEK (University of Freiburg) under the supervision of Moritz Diehl. *
* All rights reserved. *
* *
* HPMPC is free software; you can redistribute it and/or *
* modify it under the terms of the GNU Lesser General Public *
* License as published by the Free Software Foundation; either *
* version 2.1 of the License, or (at your option) any later version. *
* *
* HPMPC is distributed in the hope that it will be useful, *
* but WITHOUT ANY WARRANTY; without even the implied warranty of *
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. *
* See the GNU Lesser General Public License for more details. *
* *
* You should have received a copy of the GNU Lesser General Public *
* License along with HPMPC; if not, write to the Free Software *
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA *
* *
* Author: Gianluca Frison, giaf (at) dtu.dk *
* gianluca.frison (at) imtek.uni-freiburg.de *
* *
**************************************************************************************************/
#if defined(OS_LINUX) | defined(OS_MAC)
//#define STACKSIZE 96
#define STACKSIZE 64
#define ARG1 %rdi
#define ARG2 %rsi
#define ARG3 %rdx
#define ARG4 %rcx
#define ARG5 %r8
#define ARG6 %r9
#define ARG7 STACKSIZE + 8(%rsp)
#define ARG8 STACKSIZE + 16(%rsp)
#define ARG9 STACKSIZE + 24(%rsp)
#define ARG10 STACKSIZE + 32(%rsp)
#define ARG11 STACKSIZE + 40(%rsp)
#define ARG12 STACKSIZE + 48(%rsp)
#define ARG13 STACKSIZE + 56(%rsp)
#define ARG14 STACKSIZE + 64(%rsp)
#define ARG15 STACKSIZE + 72(%rsp)
#define ARG16 STACKSIZE + 80(%rsp)
#define ARG17 STACKSIZE + 88(%rsp)
#define ARG18 STACKSIZE + 96(%rsp)
#define PROLOGUE \
subq $STACKSIZE, %rsp; \
movq %rbx, (%rsp); \
movq %rbp, 8(%rsp); \
movq %r12, 16(%rsp); \
movq %r13, 24(%rsp); \
movq %r14, 32(%rsp); \
movq %r15, 40(%rsp); \
vzeroupper;
#define EPILOGUE \
vzeroupper; \
movq (%rsp), %rbx; \
movq 8(%rsp), %rbp; \
movq 16(%rsp), %r12; \
movq 24(%rsp), %r13; \
movq 32(%rsp), %r14; \
movq 40(%rsp), %r15; \
addq $STACKSIZE, %rsp;
#elif defined(OS_WINDOWS)
#define STACKSIZE 256
#define ARG1 %rcx
#define ARG2 %rdx
#define ARG3 %r8
#define ARG4 %r9
#define ARG5 STACKSIZE + 40(%rsp)
#define ARG6 STACKSIZE + 48(%rsp)
#define ARG7 STACKSIZE + 56(%rsp)
#define ARG8 STACKSIZE + 64(%rsp)
#define ARG9 STACKSIZE + 72(%rsp)
#define ARG10 STACKSIZE + 80(%rsp)
#define ARG11 STACKSIZE + 88(%rsp)
#define ARG12 STACKSIZE + 96(%rsp)
#define ARG13 STACKSIZE + 104(%rsp)
#define ARG14 STACKSIZE + 112(%rsp)
#define ARG15 STACKSIZE + 120(%rsp)
#define ARG16 STACKSIZE + 128(%rsp)
#define ARG17 STACKSIZE + 136(%rsp)
#define ARG18 STACKSIZE + 144(%rsp)
#define PROLOGUE \
subq $STACKSIZE, %rsp; \
movq %rbx, (%rsp); \
movq %rbp, 8(%rsp); \
movq %r12, 16(%rsp); \
movq %r13, 24(%rsp); \
movq %r14, 32(%rsp); \
movq %r15, 40(%rsp); \
movq %rdi, 48(%rsp); \
movq %rsi, 56(%rsp); \
vmovups %xmm6, 64(%rsp); \
vmovups %xmm7, 80(%rsp); \
vmovups %xmm8, 96(%rsp); \
vmovups %xmm9, 112(%rsp); \
vmovups %xmm10, 128(%rsp); \
vmovups %xmm11, 144(%rsp); \
vmovups %xmm12, 160(%rsp); \
vmovups %xmm13, 176(%rsp); \
vmovups %xmm14, 192(%rsp); \
vmovups %xmm15, 208(%rsp); \
vzeroupper;
#define EPILOGUE \
vzeroupper; \
movq (%rsp), %rbx; \
movq 8(%rsp), %rbp; \
movq 16(%rsp), %r12; \
movq 24(%rsp), %r13; \
movq 32(%rsp), %r14; \
movq 40(%rsp), %r15; \
movq 48(%rsp), %rdi; \
movq 56(%rsp), %rsi; \
vmovups 64(%rsp), %xmm6; \
vmovups 80(%rsp), %xmm7; \
vmovups 96(%rsp), %xmm8; \
vmovups 112(%rsp), %xmm9; \
vmovups 128(%rsp), %xmm10; \
vmovups 144(%rsp), %xmm11; \
vmovups 160(%rsp), %xmm12; \
vmovups 176(%rsp), %xmm13; \
vmovups 192(%rsp), %xmm14; \
vmovups 208(%rsp), %xmm15; \
addq $STACKSIZE, %rsp;
#else
#error wrong OS
#endif
#if defined(OS_LINUX) | defined(OS_WINDOWS)
.text
#elif defined(OS_MAC)
.section __TEXT,__text,regular,pure_instructions
#endif
// common inner routine with file scope
//
// input arguments:
// r10d <- k
// r11 <- A
// r12 <- B
// ymm0 <- [d00 d10 d20 d30 d40 d50 d60 d70]
// ymm1 <- [d01 d11 d21 d31 d41 d51 d61 d71]
// ymm2 <- [d02 d12 d22 d32 d42 d52 d62 d72]
// ymm3 <- [d03 d13 d23 d33 d43 d53 d63 d73]
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10d <- 0
// r11 <- A+4*k*sizeof(double)
// r12 <- B+4*k*sizeof(double)
// ymm0 <- [d00 d10 d20 d30 d40 d50 d60 d70]
// ymm1 <- [d01 d11 d21 d31 d41 d51 d61 d71]
// ymm2 <- [d02 d12 d22 d32 d42 d52 d62 d72]
// ymm3 <- [d03 d13 d23 d33 d43 d53 d63 d73]
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=2
.macro INNER_KERNEL_GEMM_ADD_NT_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_kernel_gemm_add_nt_8x4_lib8, @function
inner_kernel_gemm_add_nt_8x4_lib8:
#elif defined(OS_MAC)
_inner_kernel_gemm_add_nt_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_kernel_gemm_add_nt_8x4_lib8; .scl 2; .type 32; .endef
inner_kernel_gemm_add_nt_8x4_lib8:
#endif
#endif
// broadcast scheme
#if 1
cmpl $0, %r10d
jle 5f // return
// preload
vmovaps 0(%r11), %ymm13 // A
vxorps %ymm4, %ymm4, %ymm4
vmovaps %ymm4, %ymm5
vmovaps %ymm4, %ymm6
vmovaps %ymm4, %ymm7
cmpl $4, %r10d
jle 0f // consider clean-up loop
// main loop
.p2align 3
1: // main loop
// unroll 0
vbroadcastss 0(%r12), %ymm12 // B
vfmadd231ps %ymm13, %ymm12, %ymm0
vmovaps 32(%r11), %ymm14 // A
vbroadcastss 4(%r12), %ymm12 // B
vfmadd231ps %ymm13, %ymm12, %ymm1
vbroadcastss 8(%r12), %ymm12 // B
vfmadd231ps %ymm13, %ymm12, %ymm2
vbroadcastss 12(%r12), %ymm12 // B
vfmadd231ps %ymm13, %ymm12, %ymm3
subl $4, %r10d
// unroll 0
vbroadcastss 32(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm4
vmovaps 64(%r11), %ymm13 // A
vbroadcastss 36(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm5
vbroadcastss 40(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm6
vbroadcastss 44(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm7
addq $128, %r11
// unroll 0
vbroadcastss 64(%r12), %ymm12 // B
vfmadd231ps %ymm13, %ymm12, %ymm0
vmovaps -32(%r11), %ymm14 // A
vbroadcastss 68(%r12), %ymm12 // B
vfmadd231ps %ymm13, %ymm12, %ymm1
vbroadcastss 72(%r12), %ymm12 // B
vfmadd231ps %ymm13, %ymm12, %ymm2
vbroadcastss 76(%r12), %ymm12 // B
vfmadd231ps %ymm13, %ymm12, %ymm3
addq $128, %r12
// unroll 0
vbroadcastss -32(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm4
vmovaps 0(%r11), %ymm13 // A
vbroadcastss -28(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm5
vbroadcastss -24(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm6
vbroadcastss -20(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm7
cmpl $4, %r10d
jg 1b // main loop
0: // consider clean4-up
cmpl $3, %r10d
jle 4f // clean1
// unroll 0
vbroadcastss 0(%r12), %ymm12 // b
vfmadd231ps %ymm13, %ymm12, %ymm0
vmovaps 32(%r11), %ymm14 // a
vbroadcastss 4(%r12), %ymm12 // b
vfmadd231ps %ymm13, %ymm12, %ymm1
vbroadcastss 8(%r12), %ymm12 // b
vfmadd231ps %ymm13, %ymm12, %ymm2
vbroadcastss 12(%r12), %ymm12 // b
vfmadd231ps %ymm13, %ymm12, %ymm3
subl $4, %r10d
// unroll 0
vbroadcastss 32(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm4
vmovaps 64(%r11), %ymm13 // A
vbroadcastss 36(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm5
vbroadcastss 40(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm6
vbroadcastss 44(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm7
addq $128, %r11
// unroll 0
vbroadcastss 64(%r12), %ymm12 // B
vfmadd231ps %ymm13, %ymm12, %ymm0
vmovaps -32(%r11), %ymm14 // A
vbroadcastss 68(%r12), %ymm12 // B
vfmadd231ps %ymm13, %ymm12, %ymm1
vbroadcastss 72(%r12), %ymm12 // B
vfmadd231ps %ymm13, %ymm12, %ymm2
vbroadcastss 76(%r12), %ymm12 // B
vfmadd231ps %ymm13, %ymm12, %ymm3
addq $128, %r12
// unroll 0
vbroadcastss -32(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm4
// vmovaps 0(%r11), %ymm13 // A
vbroadcastss -28(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm5
vbroadcastss -24(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm6
vbroadcastss -20(%r12), %ymm12 // B
vfmadd231ps %ymm14, %ymm12, %ymm7
jmp 2f // return
4: // consider clean1-up loop
cmpl $0, %r10d
jle 2f // return
// clean-up loop
3: // clean up loop
// unroll 0
vmovaps 0(%r11), %ymm13 // a
vbroadcastss 0(%r12), %ymm12 // b
vfmadd231ps %ymm13, %ymm12, %ymm0
vbroadcastss 4(%r12), %ymm12 // b
vfmadd231ps %ymm13, %ymm12, %ymm1
subl $1, %r10d
vbroadcastss 8(%r12), %ymm12 // b
vfmadd231ps %ymm13, %ymm12, %ymm2
addq $32, %r11
vbroadcastss 12(%r12), %ymm12 // b
vfmadd231ps %ymm13, %ymm12, %ymm3
addq $32, %r12
cmpl $0, %r10d
jg 3b // clean up loop
2: // reduce
vaddps %ymm4, %ymm0, %ymm0
vaddps %ymm5, %ymm1, %ymm1
vaddps %ymm6, %ymm2, %ymm2
vaddps %ymm7, %ymm3, %ymm3
5: // return
// shuffle scheme
#else
cmpl $0, %r10d
jle 5f // return
// preload
vbroadcastf128 0(%r12), %ymm14 // B
vmovaps 0(%r11), %ymm12 // A
vbroadcastf128 32(%r12), %ymm15 // B
vmovaps 32(%r11), %ymm13 // A
vxorps %ymm4, %ymm4, %ymm4
vmovaps %ymm4, %ymm5
vmovaps %ymm4, %ymm6
vmovaps %ymm4, %ymm7
cmpl $4, %r10d
jle 0f // consider clean-up loop
// main loop
.p2align 3
1: // main loop
// unroll 0
vfmadd231ps %ymm12, %ymm14, %ymm0
vshufps $0xb1, %ymm14, %ymm14, %ymm14 // 10 11 00 01
vfmadd231ps %ymm12, %ymm14, %ymm1
vshufps $0x4e, %ymm14, %ymm14, %ymm14 // 01 00 11 10
vfmadd231ps %ymm12, %ymm14, %ymm2
vshufps $0xb1, %ymm14, %ymm14, %ymm14 // 10 11 00 01
vfmadd231ps %ymm12, %ymm14, %ymm3
vbroadcastf128 64(%r12), %ymm14 // B
vmovaps 64(%r11), %ymm12 // A
// unroll 1
vfmadd231ps %ymm13, %ymm15, %ymm4
vshufps $0xb1, %ymm15, %ymm15, %ymm15
vfmadd231ps %ymm13, %ymm15, %ymm5
vshufps $0x4e, %ymm15, %ymm15, %ymm15
vfmadd231ps %ymm13, %ymm15, %ymm6
vshufps $0xb1, %ymm15, %ymm15, %ymm15
vfmadd231ps %ymm13, %ymm15, %ymm7
vbroadcastf128 96(%r12), %ymm15 // B
vmovaps 96(%r11), %ymm13 // A
// unroll 2
vfmadd231ps %ymm12, %ymm14, %ymm0
vshufps $0xb1, %ymm14, %ymm14, %ymm14
vfmadd231ps %ymm12, %ymm14, %ymm1
vshufps $0x4e, %ymm14, %ymm14, %ymm14
vfmadd231ps %ymm12, %ymm14, %ymm2
vshufps $0xb1, %ymm14, %ymm14, %ymm14
vfmadd231ps %ymm12, %ymm14, %ymm3
vbroadcastf128 128(%r12), %ymm14 // B
vmovaps 128(%r11), %ymm12 // A
subl $4, %r10d
addq $128, %r11
addq $128, %r12
// unroll 3
vfmadd231ps %ymm13, %ymm15, %ymm4
vshufps $0xb1, %ymm15, %ymm15, %ymm15
vfmadd231ps %ymm13, %ymm15, %ymm5
vshufps $0x4e, %ymm15, %ymm15, %ymm15
vfmadd231ps %ymm13, %ymm15, %ymm6
vshufps $0xb1, %ymm15, %ymm15, %ymm15
vfmadd231ps %ymm13, %ymm15, %ymm7
vbroadcastf128 32(%r12), %ymm15 // B
vmovaps 32(%r11), %ymm13 // A
cmpl $4, %r10d
jg 1b // main loop
0: // consider clean4-up
cmpl $3, %r10d
jle 4f // clean1
// unroll 0
vfmadd231ps %ymm12, %ymm14, %ymm0
vshufps $0xb1, %ymm14, %ymm14, %ymm14
vfmadd231ps %ymm12, %ymm14, %ymm1
vshufps $0x4e, %ymm14, %ymm14, %ymm14
vfmadd231ps %ymm12, %ymm14, %ymm2
vshufps $0xb1, %ymm14, %ymm14, %ymm14
vfmadd231ps %ymm12, %ymm14, %ymm3
vbroadcastf128 64(%r12), %ymm14 // B
vmovaps 64(%r11), %ymm12 // A
// unroll 1
vfmadd231ps %ymm13, %ymm15, %ymm4
vshufps $0xb1, %ymm15, %ymm15, %ymm15
vfmadd231ps %ymm13, %ymm15, %ymm5
vshufps $0x4e, %ymm15, %ymm15, %ymm15
vfmadd231ps %ymm13, %ymm15, %ymm6
vshufps $0xb1, %ymm15, %ymm15, %ymm15
vfmadd231ps %ymm13, %ymm15, %ymm7
vbroadcastf128 96(%r12), %ymm15 // B
vmovaps 96(%r11), %ymm13 // A
// unroll 2
vfmadd231ps %ymm12, %ymm14, %ymm0
vshufps $0xb1, %ymm14, %ymm14, %ymm14
vfmadd231ps %ymm12, %ymm14, %ymm1
vshufps $0x4e, %ymm14, %ymm14, %ymm14
vfmadd231ps %ymm12, %ymm14, %ymm2
vshufps $0xb1, %ymm14, %ymm14, %ymm14
vfmadd231ps %ymm12, %ymm14, %ymm3
// vbroadcastf128 128(%r12), %ymm14 // B
// vmovaps 128(%r11), %ymm12 // A
subl $4, %r10d
addq $128, %r11
addq $128, %r12
// unroll 3
vfmadd231ps %ymm13, %ymm15, %ymm4
vshufps $0xb1, %ymm15, %ymm15, %ymm15
vfmadd231ps %ymm13, %ymm15, %ymm5
vshufps $0x4e, %ymm15, %ymm15, %ymm15
vfmadd231ps %ymm13, %ymm15, %ymm6
vshufps $0xb1, %ymm15, %ymm15, %ymm15
vfmadd231ps %ymm13, %ymm15, %ymm7
// vbroadcastf128 32(%r12), %ymm15 // B
// vmovaps 32(%r11), %ymm13 // A
// cmpl $4, %r10d
jmp 2f // return
4: // consider clean1-up loop
cmpl $0, %r10d
jle 2f // return
// clean-up loop
3: // clean up loop
// unroll 0
vbroadcastf128 0(%r12), %ymm14 // B
vmovaps 0(%r11), %ymm12 // A
vmulps %ymm12, %ymm14, %ymm11
vaddps %ymm11, %ymm0, %ymm0
vshufps $0xb1, %ymm14, %ymm14, %ymm14
vmulps %ymm12, %ymm14, %ymm11
vaddps %ymm11, %ymm1, %ymm1
vshufps $0x4e, %ymm14, %ymm14, %ymm14
vmulps %ymm12, %ymm14, %ymm11
vaddps %ymm11, %ymm2, %ymm2
subl $1, %r10d
addq $32, %r11
addq $32, %r12
vshufps $0xb1, %ymm14, %ymm14, %ymm14
vmulps %ymm12, %ymm14, %ymm11
vaddps %ymm11, %ymm3, %ymm3
cmpl $0, %r10d
jg 3b // clean up loop
2: // reduce
vaddps %ymm4, %ymm0, %ymm0
vaddps %ymm5, %ymm1, %ymm1
vaddps %ymm6, %ymm2, %ymm2
vaddps %ymm7, %ymm3, %ymm3
5: // return
#endif
#if MACRO_LEVEL>=2
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_kernel_gemm_add_nt_8x4_lib8, .-inner_kernel_gemm_add_nt_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// input arguments:
// r10d <- k
// r11 <- A
// r12 <- B
// ymm0 <- [d00 d10 d20 d30 d40 d50 d60 d70]
// ymm1 <- [d01 d11 d21 d31 d41 d51 d61 d71]
// ymm2 <- [d02 d12 d22 d32 d42 d52 d62 d72]
// ymm3 <- [d03 d13 d23 d33 d43 d53 d63 d73]
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10d <- 0
// r11 <- A+4*k*sizeof(double)
// r12 <- B+4*k*sizeof(double)
// ymm0 <- [d00 d10 d20 d30 d40 d50 d60 d70]
// ymm1 <- [d01 d11 d21 d31 d41 d51 d61 d71]
// ymm2 <- [d02 d12 d22 d32 d42 d52 d62 d72]
// ymm3 <- [d03 d13 d23 d33 d43 d53 d63 d73]
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=2
.macro INNER_KERNEL_GEMM_SUB_NT_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_kernel_gemm_sub_nt_8x4_lib8, @function
inner_kernel_gemm_sub_nt_8x4_lib8:
#elif defined(OS_MAC)
_inner_kernel_gemm_sub_nt_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_kernel_gemm_sub_nt_8x4_lib8; .scl 2; .type 32; .endef
inner_kernel_gemm_sub_nt_8x4_lib8:
#endif
#endif
cmpl $0, %r10d
jle 5f // return
// preload
vmovaps 0(%r11), %ymm13 // A
vxorps %ymm4, %ymm4, %ymm4
vmovaps %ymm4, %ymm5
vmovaps %ymm4, %ymm6
vmovaps %ymm4, %ymm7
cmpl $4, %r10d
jle 0f // consider clean-up loop
// main loop
.p2align 3
1: // main loop
// unroll 0
vbroadcastss 0(%r12), %ymm12 // B
vfnmadd231ps %ymm13, %ymm12, %ymm0
vmovaps 32(%r11), %ymm14 // A
vbroadcastss 4(%r12), %ymm12 // B
vfnmadd231ps %ymm13, %ymm12, %ymm1
vbroadcastss 8(%r12), %ymm12 // B
vfnmadd231ps %ymm13, %ymm12, %ymm2
vbroadcastss 12(%r12), %ymm12 // B
vfnmadd231ps %ymm13, %ymm12, %ymm3
subl $4, %r10d
// unroll 0
vbroadcastss 32(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm4
vmovaps 64(%r11), %ymm13 // A
vbroadcastss 36(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm5
vbroadcastss 40(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm6
vbroadcastss 44(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm7
addq $128, %r11
// unroll 0
vbroadcastss 64(%r12), %ymm12 // B
vfnmadd231ps %ymm13, %ymm12, %ymm0
vmovaps -32(%r11), %ymm14 // A
vbroadcastss 68(%r12), %ymm12 // B
vfnmadd231ps %ymm13, %ymm12, %ymm1
vbroadcastss 72(%r12), %ymm12 // B
vfnmadd231ps %ymm13, %ymm12, %ymm2
vbroadcastss 76(%r12), %ymm12 // B
vfnmadd231ps %ymm13, %ymm12, %ymm3
addq $128, %r12
// unroll 0
vbroadcastss -32(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm4
vmovaps 0(%r11), %ymm13 // A
vbroadcastss -28(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm5
vbroadcastss -24(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm6
vbroadcastss -20(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm7
cmpl $4, %r10d
jg 1b // main loop
0: // consider clean4-up
cmpl $3, %r10d
jle 4f // clean1
// unroll 0
vbroadcastss 0(%r12), %ymm12 // b
vfnmadd231ps %ymm13, %ymm12, %ymm0
vmovaps 32(%r11), %ymm14 // a
vbroadcastss 4(%r12), %ymm12 // b
vfnmadd231ps %ymm13, %ymm12, %ymm1
vbroadcastss 8(%r12), %ymm12 // b
vfnmadd231ps %ymm13, %ymm12, %ymm2
vbroadcastss 12(%r12), %ymm12 // b
vfnmadd231ps %ymm13, %ymm12, %ymm3
subl $4, %r10d
// unroll 0
vbroadcastss 32(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm4
vmovaps 64(%r11), %ymm13 // A
vbroadcastss 36(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm5
vbroadcastss 40(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm6
vbroadcastss 44(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm7
addq $128, %r11
// unroll 0
vbroadcastss 64(%r12), %ymm12 // B
vfnmadd231ps %ymm13, %ymm12, %ymm0
vmovaps -32(%r11), %ymm14 // A
vbroadcastss 68(%r12), %ymm12 // B
vfnmadd231ps %ymm13, %ymm12, %ymm1
vbroadcastss 72(%r12), %ymm12 // B
vfnmadd231ps %ymm13, %ymm12, %ymm2
vbroadcastss 76(%r12), %ymm12 // B
vfnmadd231ps %ymm13, %ymm12, %ymm3
addq $128, %r12
// unroll 0
vbroadcastss -32(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm4
// vmovaps 0(%r11), %ymm13 // A
vbroadcastss -28(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm5
vbroadcastss -24(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm6
vbroadcastss -20(%r12), %ymm12 // B
vfnmadd231ps %ymm14, %ymm12, %ymm7
jmp 2f // return
4: // consider clean1-up loop
cmpl $0, %r10d
jle 2f // return
// clean-up loop
3: // clean up loop
// unroll 0
vmovaps 0(%r11), %ymm13 // a
vbroadcastss 0(%r12), %ymm12 // b
vfnmadd231ps %ymm13, %ymm12, %ymm0
vbroadcastss 4(%r12), %ymm12 // b
vfnmadd231ps %ymm13, %ymm12, %ymm1
subl $1, %r10d
vbroadcastss 8(%r12), %ymm12 // b
vfnmadd231ps %ymm13, %ymm12, %ymm2
addq $32, %r11
vbroadcastss 12(%r12), %ymm12 // b
vfnmadd231ps %ymm13, %ymm12, %ymm3
addq $32, %r12
cmpl $0, %r10d
jg 3b // clean up loop
2: // reduce
vaddps %ymm4, %ymm0, %ymm0
vaddps %ymm5, %ymm1, %ymm1
vaddps %ymm6, %ymm2, %ymm2
vaddps %ymm7, %ymm3, %ymm3
5: // return
#if MACRO_LEVEL>=2
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_kernel_gemm_sub_nt_8x4_lib8, .-inner_kernel_gemm_sub_nt_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// input arguments:
// r10d <- k
// r11 <- A
// r12 <- B
// r13 <- 4*sdb*sizeof(double)
// r14 <= dirty
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10d <- 0
// r11 <- A+4*k*sizeof(double)
// r12 <- B+(k/4)*sdb*sizeof(double)+(k%4)
// r13 <- 4*sdb*sizeof(double)
// r14 <= dirty
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=2
.macro INNER_KERNEL_GEMM_ADD_NN_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_kernel_gemm_add_nn_8x4_lib8, @function
inner_kernel_gemm_add_nn_8x4_lib8:
#elif defined(OS_MAC)
_inner_kernel_gemm_add_nn_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_kernel_gemm_add_nn_8x4_lib8; .scl 2; .type 32; .endef
inner_kernel_gemm_add_nn_8x4_lib8:
#endif
#endif
cmpl $0, %r10d
jle 2f // return
cmpl $8, %r10d
jl 0f // consider clean-up loop
vxorps %ymm4, %ymm4, %ymm4
vmovaps %ymm4, %ymm5
vmovaps %ymm4, %ymm6
vmovaps %ymm4, %ymm7
// main loop
.p2align 3
1: // main loop
// prefetcht0 0(%r12, %r13, 1) // software prefetch
// prefetcht0 64(%r12, %r13, 1) // software prefetch
// unroll 0
vmovaps 0(%r11), %ymm12 // A[0]
vbroadcastss 0(%r12), %ymm13 // B[0]
vfmadd231ps %ymm12, %ymm13, %ymm0
vbroadcastss 32(%r12), %ymm13 // B[1]
vfmadd231ps %ymm12, %ymm13, %ymm1
vbroadcastss 64(%r12), %ymm13 // B[2]
vfmadd231ps %ymm12, %ymm13, %ymm2
vbroadcastss 96(%r12), %ymm13 // B[3]
vfmadd231ps %ymm12, %ymm13, %ymm3
// unroll 1
vmovaps 32(%r11), %ymm12 // A[0]
vbroadcastss 4(%r12), %ymm13 // B[0]
vfmadd231ps %ymm12, %ymm13, %ymm4
vbroadcastss 36(%r12), %ymm13 // B[1]
vfmadd231ps %ymm12, %ymm13, %ymm5
vbroadcastss 68(%r12), %ymm13 // B[2]
vfmadd231ps %ymm12, %ymm13, %ymm6
vbroadcastss 100(%r12), %ymm13 // B[3]
vfmadd231ps %ymm12, %ymm13, %ymm7
// unroll 2
vmovaps 64(%r11), %ymm12 // A[0]
vbroadcastss 8(%r12), %ymm13 // B[0]
vfmadd231ps %ymm12, %ymm13, %ymm0
vbroadcastss 40(%r12), %ymm13 // B[1]
vfmadd231ps %ymm12, %ymm13, %ymm1
vbroadcastss 72(%r12), %ymm13 // B[2]
vfmadd231ps %ymm12, %ymm13, %ymm2
vbroadcastss 104(%r12), %ymm13 // B[3]
vfmadd231ps %ymm12, %ymm13, %ymm3
// unroll 3
vmovaps 96(%r11), %ymm12 // A[0]
vbroadcastss 12(%r12), %ymm13 // B[0]
vfmadd231ps %ymm12, %ymm13, %ymm4
vbroadcastss 44(%r12), %ymm13 // B[1]
vfmadd231ps %ymm12, %ymm13, %ymm5
vbroadcastss 76(%r12), %ymm13 // B[2]
vfmadd231ps %ymm12, %ymm13, %ymm6
vbroadcastss 108(%r12), %ymm13 // B[3]
vfmadd231ps %ymm12, %ymm13, %ymm7
// unroll 4
vmovaps 128(%r11), %ymm12 // A[0]
vbroadcastss 16(%r12), %ymm13 // B[0]
vfmadd231ps %ymm12, %ymm13, %ymm0
vbroadcastss 48(%r12), %ymm13 // B[1]
vfmadd231ps %ymm12, %ymm13, %ymm1
vbroadcastss 80(%r12), %ymm13 // B[2]
vfmadd231ps %ymm12, %ymm13, %ymm2
vbroadcastss 112(%r12), %ymm13 // B[3]
vfmadd231ps %ymm12, %ymm13, %ymm3
// unroll 5
vmovaps 160(%r11), %ymm12 // A[0]
vbroadcastss 20(%r12), %ymm13 // B[0]
vfmadd231ps %ymm12, %ymm13, %ymm4
vbroadcastss 52(%r12), %ymm13 // B[1]
vfmadd231ps %ymm12, %ymm13, %ymm5
vbroadcastss 84(%r12), %ymm13 // B[2]
vfmadd231ps %ymm12, %ymm13, %ymm6
vbroadcastss 116(%r12), %ymm13 // B[3]
vfmadd231ps %ymm12, %ymm13, %ymm7
subl $8, %r10d
// unroll 6
vmovaps 192(%r11), %ymm12 // A[0]
vbroadcastss 24(%r12), %ymm13 // B[0]
vfmadd231ps %ymm12, %ymm13, %ymm0
vbroadcastss 56(%r12), %ymm13 // B[1]
vfmadd231ps %ymm12, %ymm13, %ymm1
vbroadcastss 88(%r12), %ymm13 // B[2]
vfmadd231ps %ymm12, %ymm13, %ymm2
vbroadcastss 120(%r12), %ymm13 // B[3]
vfmadd231ps %ymm12, %ymm13, %ymm3
addq $256, %r11
// unroll 7
vmovaps -32(%r11), %ymm12 // A[0]
vbroadcastss 28(%r12), %ymm13 // B[0]
vfmadd231ps %ymm12, %ymm13, %ymm4
vbroadcastss 60(%r12), %ymm13 // B[1]
vfmadd231ps %ymm12, %ymm13, %ymm5
vbroadcastss 92(%r12), %ymm13 // B[2]
vfmadd231ps %ymm12, %ymm13, %ymm6
vbroadcastss 124(%r12), %ymm13 // B[3]
vfmadd231ps %ymm12, %ymm13, %ymm7
addq %r12, %r13
cmpl $7, %r10d
jg 1b // main loop
vaddps %ymm4, %ymm0, %ymm0
vaddps %ymm5, %ymm1, %ymm1
vaddps %ymm6, %ymm2, %ymm2
vaddps %ymm7, %ymm3, %ymm3
0: // consider clean1-up loop
cmpl $0, %r10d
jle 2f // return
3: // clean1-up loop
vmovaps 0(%r11), %ymm12 // A[0]
vbroadcastss 0(%r12), %ymm13 // B[0]
vfmadd231ps %ymm12, %ymm13, %ymm0
vbroadcastss 32(%r12), %ymm13 // B[1]
vfmadd231ps %ymm12, %ymm13, %ymm1
vbroadcastss 64(%r12), %ymm13 // B[2]
vfmadd231ps %ymm12, %ymm13, %ymm2
vbroadcastss 96(%r12), %ymm13 // B[3]
vfmadd231ps %ymm12, %ymm13, %ymm3
subl $1, %r10d
addq $32, %r11
addq $4, %r12
cmpl $0, %r10d
jg 3b // clean up loop
2: // return
#if MACRO_LEVEL>=2
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_kernel_gemm_add_nn_8x4_lib8, .-inner_kernel_gemm_add_nn_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// input arguments:
// r10d <- k
// r11 <- A
// r12 <- B
// r13 <- 4*sdb*sizeof(double)
// r14 <= dirty
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10d <- 0
// r11 <- A+4*k*sizeof(double)
// r12 <- B+(k/4)*sdb*sizeof(double)+(k%4)
// r13 <- 4*sdb*sizeof(double)
// r14 <= dirty
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=2
.macro INNER_KERNEL_GEMM_SUB_NN_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_kernel_gemm_sub_nn_8x4_lib8, @function
inner_kernel_gemm_sub_nn_8x4_lib8:
#elif defined(OS_MAC)
_inner_kernel_gemm_sub_nn_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_kernel_gemm_sub_nn_8x4_lib8; .scl 2; .type 32; .endef
inner_kernel_gemm_sub_nn_8x4_lib8:
#endif
#endif
cmpl $0, %r10d
jle 2f // return
cmpl $8, %r10d
jl 0f // consider clean-up loop
vxorps %ymm4, %ymm4, %ymm4
vmovaps %ymm4, %ymm5
vmovaps %ymm4, %ymm6
vmovaps %ymm4, %ymm7
// main loop
.p2align 3
1: // main loop
prefetcht0 0(%r12, %r13, 1) // software prefetch
prefetcht0 64(%r12, %r13, 1) // software prefetch
// unroll 0
vmovaps 0(%r11), %ymm12 // A[0]
vbroadcastss 0(%r12), %ymm13 // B[0]
vfnmadd231ps %ymm12, %ymm13, %ymm0
vbroadcastss 32(%r12), %ymm13 // B[1]
vfnmadd231ps %ymm12, %ymm13, %ymm1
vbroadcastss 64(%r12), %ymm13 // B[2]
vfnmadd231ps %ymm12, %ymm13, %ymm2
vbroadcastss 96(%r12), %ymm13 // B[3]
vfnmadd231ps %ymm12, %ymm13, %ymm3
// unroll 1
vmovaps 32(%r11), %ymm12 // A[0]
vbroadcastss 4(%r12), %ymm13 // B[0]
vfnmadd231ps %ymm12, %ymm13, %ymm4
vbroadcastss 36(%r12), %ymm13 // B[1]
vfnmadd231ps %ymm12, %ymm13, %ymm5
vbroadcastss 68(%r12), %ymm13 // B[2]
vfnmadd231ps %ymm12, %ymm13, %ymm6
vbroadcastss 100(%r12), %ymm13 // B[3]
vfnmadd231ps %ymm12, %ymm13, %ymm7
// unroll 2
vmovaps 64(%r11), %ymm12 // A[0]
vbroadcastss 8(%r12), %ymm13 // B[0]
vfnmadd231ps %ymm12, %ymm13, %ymm0
vbroadcastss 40(%r12), %ymm13 // B[1]
vfnmadd231ps %ymm12, %ymm13, %ymm1
vbroadcastss 72(%r12), %ymm13 // B[2]
vfnmadd231ps %ymm12, %ymm13, %ymm2
vbroadcastss 104(%r12), %ymm13 // B[3]
vfnmadd231ps %ymm12, %ymm13, %ymm3
// unroll 3
vmovaps 96(%r11), %ymm12 // A[0]
vbroadcastss 12(%r12), %ymm13 // B[0]
vfnmadd231ps %ymm12, %ymm13, %ymm4
vbroadcastss 44(%r12), %ymm13 // B[1]
vfnmadd231ps %ymm12, %ymm13, %ymm5
vbroadcastss 76(%r12), %ymm13 // B[2]
vfnmadd231ps %ymm12, %ymm13, %ymm6
vbroadcastss 108(%r12), %ymm13 // B[3]
vfnmadd231ps %ymm12, %ymm13, %ymm7
// unroll 4
vmovaps 128(%r11), %ymm12 // A[0]
vbroadcastss 16(%r12), %ymm13 // B[0]
vfnmadd231ps %ymm12, %ymm13, %ymm0
vbroadcastss 48(%r12), %ymm13 // B[1]
vfnmadd231ps %ymm12, %ymm13, %ymm1
vbroadcastss 80(%r12), %ymm13 // B[2]
vfnmadd231ps %ymm12, %ymm13, %ymm2
vbroadcastss 112(%r12), %ymm13 // B[3]
vfnmadd231ps %ymm12, %ymm13, %ymm3
// unroll 5
vmovaps 160(%r11), %ymm12 // A[0]
vbroadcastss 20(%r12), %ymm13 // B[0]
vfnmadd231ps %ymm12, %ymm13, %ymm4
vbroadcastss 52(%r12), %ymm13 // B[1]
vfnmadd231ps %ymm12, %ymm13, %ymm5
vbroadcastss 84(%r12), %ymm13 // B[2]
vfnmadd231ps %ymm12, %ymm13, %ymm6
vbroadcastss 116(%r12), %ymm13 // B[3]
vfnmadd231ps %ymm12, %ymm13, %ymm7
subl $8, %r10d
// unroll 6
vmovaps 192(%r11), %ymm12 // A[0]
vbroadcastss 24(%r12), %ymm13 // B[0]
vfnmadd231ps %ymm12, %ymm13, %ymm0
vbroadcastss 56(%r12), %ymm13 // B[1]
vfnmadd231ps %ymm12, %ymm13, %ymm1
vbroadcastss 88(%r12), %ymm13 // B[2]
vfnmadd231ps %ymm12, %ymm13, %ymm2
vbroadcastss 120(%r12), %ymm13 // B[3]
vfnmadd231ps %ymm12, %ymm13, %ymm3
addq $256, %r11
// unroll 7
vmovaps -32(%r11), %ymm12 // A[0]
vbroadcastss 28(%r12), %ymm13 // B[0]
vfnmadd231ps %ymm12, %ymm13, %ymm4
vbroadcastss 60(%r12), %ymm13 // B[1]
vfnmadd231ps %ymm12, %ymm13, %ymm5
vbroadcastss 92(%r12), %ymm13 // B[2]
vfnmadd231ps %ymm12, %ymm13, %ymm6
vbroadcastss 124(%r12), %ymm13 // B[3]
vfnmadd231ps %ymm12, %ymm13, %ymm7
addq %r12, %r13
cmpl $7, %r10d
jg 1b // main loop
vaddps %ymm4, %ymm0, %ymm0
vaddps %ymm5, %ymm1, %ymm1
vaddps %ymm6, %ymm2, %ymm2
vaddps %ymm7, %ymm3, %ymm3
0: // consider clean1-up loop
cmpl $0, %r10d
jle 2f // return
3: // clean1-up loop
vmovaps 0(%r11), %ymm12 // A[0]
vbroadcastss 0(%r12), %ymm13 // B[0]
vfnmadd231ps %ymm12, %ymm13, %ymm0
vbroadcastss 32(%r12), %ymm13 // B[1]
vfnmadd231ps %ymm12, %ymm13, %ymm1
vbroadcastss 64(%r12), %ymm13 // B[2]
vfnmadd231ps %ymm12, %ymm13, %ymm2
vbroadcastss 96(%r12), %ymm13 // B[3]
vfnmadd231ps %ymm12, %ymm13, %ymm3
subl $1, %r10d
addq $32, %r11
addq $4, %r12
cmpl $0, %r10d
jg 3b // clean up loop
2: // return
#if MACRO_LEVEL>=2
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_kernel_gemm_sub_nn_8x4_lib8, .-inner_kernel_gemm_sub_nn_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// edge for B unaligned
//
// input arguments:
// r10 <- k
// r11 <- A
// r12 <- B
// r13 <- bs*sdb*sizeof(double)
// r14 <- offB
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm12 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- k-(4-offB)
// r11 <- A+(4-offB)*bs*sizeof(double)
// r12 <- B-offB+bs*sdb*sizeof(double)
// r13 <- bs*sdb*sizeof(double)
// r14 <- offB
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm12 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_EDGE_GEMM_ADD_NN_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_edge_gemm_add_nn_8x4_lib8, @function
inner_edge_gemm_add_nn_8x4_lib8:
#elif defined(OS_MAC)
_inner_edge_gemm_add_nn_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_edge_gemm_add_nn_8x4_lib8; .scl 2; .type 32; .endef
inner_edge_gemm_add_nn_8x4_lib8:
#endif
#endif
cmpl $0, %r14d // offset==0
jle 2f // end
cmpl $0, %r10d // k==0
jle 2f // end
movl $8, %r15d
subl %r14d, %r15d // 8-offsetB
cmpl %r10d, %r15d
// jle 0f
// movl %r10d, %r15d // kend=min(k,8-offsetB)
//0:
cmovgl %r10d, %r15d // kend=min(k,8-offsetB)
movl %r14d, %eax
sall $2, %eax // offsetB*sizeof(float)
addq %rax, %r12 // B+offsetB*sizeof(float)
1:
vmovaps 0(%r11), %ymm8
vbroadcastss 0(%r12), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm0, %ymm0
vbroadcastss 32(%r12), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm1, %ymm1
vbroadcastss 64(%r12), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm2, %ymm2
vbroadcastss 96(%r12), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm3, %ymm3
subl $1, %r10d // k-1
subl $1, %r15d // kend-1
addq $32, %r11 // A+1*bs*sizeof(float)
addq $4, %r12 // B+1*sizeof(float)
cmpl $0, %r15d
jg 1b
cmpl $0, %r10d
jle 2f // end
addq %r13, %r12
subq $32, %r12 // B+bs*(sdb-1)*sizeof(float)
2:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_edge_gemm_add_nn_8x4_lib8, .-inner_edge_gemm_add_nn_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// edge for B lower triangular
//
// input arguments:
// r10 <- k
// r11 <- A
// r12 <- B
// r13 <- bs*sdb*sizeof(double)
// r14 <- offB
// ymm0 <- [d00 d10 d20 d30]
// ymm1 <- [d01 d11 d21 d31]
// ymm2 <- [d02 d12 d22 d32]
// ymm3 <- [d03 d13 d23 d33]
// ymm8 <- dirty
// ymm12 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- k-(4-offB)
// r11 <- A+(4-offB)*bs*sizeof(double)
// r12 <- B-offB+bs*sdb*sizeof(double)
// r13 <- bs*sdb*sizeof(double)
// r14 <- offB
// ymm0 <- [d00 d10 d20 d30]
// ymm1 <- [d01 d11 d21 d31]
// ymm2 <- [d02 d12 d22 d32]
// ymm3 <- [d03 d13 d23 d33]
// ymm8 <- dirty
// ymm12 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_EDGE_TRMM_NN_RL_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_edge_trmm_nn_rl_8x4_lib8, @function
inner_edge_trmm_nn_rl_8x4_lib8:
#elif defined(OS_MAC)
_inner_edge_trmm_nn_rl_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_edge_trmm_nn_rl_8x4_lib8; .scl 2; .type 32; .endef
inner_edge_trmm_nn_rl_8x4_lib8:
#endif
#endif
cmpl $0, %r10d
jle 0f // end
movl %r14d, %eax
sall $2, %eax // offsetB*sizeof(float)
movq %r12, %rbx // B
addq %rax, %rbx // B+offsetB*sizeof(float)
cmpl $4, %r14d
jg 1f
// offB==0, 1, 2, 3, 4
vmovaps 0(%r11), %ymm8
vbroadcastss 0(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm0, %ymm0
subl $1, %r10d // k-1
addq $32, %r11 // A+1*bs*sizeof(float)
addl $1, %r14d // offsetB+1
cmpl $0, %r10d
jle 0f // end
vmovaps 0(%r11), %ymm8
vbroadcastss 4(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm0, %ymm0
vbroadcastss 36(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm1, %ymm1
subl $1, %r10d // k-1
addq $32, %r11 // A+1*bs*sizeof(float)
addl $1, %r14d // offsetB+1
cmpl $0, %r10d
jle 0f // end
vmovaps 0(%r11), %ymm8
vbroadcastss 8(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm0, %ymm0
vbroadcastss 40(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm1, %ymm1
vbroadcastss 72(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm2, %ymm2
subl $1, %r10d // k-1
addq $32, %r11 // A+1*bs*sizeof(float)
addl $1, %r14d // offsetB+1
jmp 0f // end
1:
cmpl $5, %r14d
jg 1f
// offB==5
vmovaps 0(%r11), %ymm8
vbroadcastss 0(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm0, %ymm0
subl $1, %r10d // k-1
addq $32, %r11 // A+1*bs*sizeof(float)
addl $1, %r14d // offsetB+1
cmpl $0, %r10d
jle 0f // end
vmovaps 0(%r11), %ymm8
vbroadcastss 4(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm0, %ymm0
vbroadcastss 36(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm1, %ymm1
subl $1, %r10d // k-1
addq $32, %r11 // A+1*bs*sizeof(float)
addl $1, %r14d // offsetB+1
cmpl $0, %r10d
jle 0f // end
vmovaps 0(%r11), %ymm8
vbroadcastss 8(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm0, %ymm0
vbroadcastss 40(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm1, %ymm1
vbroadcastss 72(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm2, %ymm2
subl $1, %r10d // k-1
addq $32, %r11 // A+1*bs*sizeof(float)
addq %r13, %r12 // B+8*sdb*sizeof(float)
movl $0, %r14d // offsetB=0
jmp 0f // end
1:
cmpl $6, %r14d
jg 1f
// offB==6
vmovaps 0(%r11), %ymm8
vbroadcastss 0(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm0, %ymm0
subl $1, %r10d // k-1
addq $32, %r11 // A+1*bs*sizeof(float)
addl $1, %r14d // offsetB+1
cmpl $0, %r10d
jle 0f // end
vmovaps 0(%r11), %ymm8
vbroadcastss 4(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm0, %ymm0
vbroadcastss 36(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm1, %ymm1
subl $1, %r10d // k-1
addq $32, %r11 // A+1*bs*sizeof(float)
addq %r13, %r12 // B+8*sdb*sizeof(float)
movq %r12, %rbx // B
movl $0, %r14d // offsetB=0
cmpl $0, %r10d
jle 0f // end
vmovaps 0(%r11), %ymm8
vbroadcastss 0(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm0, %ymm0
vbroadcastss 32(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm1, %ymm1
vbroadcastss 64(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm2, %ymm2
subl $1, %r10d // k-1
addq $32, %r11 // A+1*bs*sizeof(float)
addl $1, %r14d // offsetB+1
jmp 0f // end
1:
// cmpl $7, %r14d
// jg 0f
// offB==6
vmovaps 0(%r11), %ymm8
vbroadcastss 0(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm0, %ymm0
subl $1, %r10d // k-1
addq $32, %r11 // A+1*bs*sizeof(float)
addq %r13, %r12 // B+8*sdb*sizeof(float)
movq %r12, %rbx // B
movl $0, %r14d // offsetB=0
cmpl $0, %r10d
jle 0f // end
vmovaps 0(%r11), %ymm8
vbroadcastss 0(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm0, %ymm0
vbroadcastss 32(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm1, %ymm1
subl $1, %r10d // k-1
addq $32, %r11 // A+1*bs*sizeof(float)
addl $1, %r14d // offsetB+1
cmpl $0, %r10d
jle 0f // end
vmovaps 0(%r11), %ymm8
vbroadcastss 4(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm0, %ymm0
vbroadcastss 36(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm1, %ymm1
vbroadcastss 68(%rbx), %ymm12
vmulps %ymm8, %ymm12, %ymm15
vaddps %ymm15, %ymm2, %ymm2
subl $1, %r10d // k-1
addq $32, %r11 // A+1*bs*sizeof(float)
addl $1, %r14d // offsetB+1
// jmp 0f // end
// end
0:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_edge_trmm_nn_rl_8x4_lib8, .-inner_edge_trmm_nn_rl_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// strsm
// right
// lower
// transposed
// not-unit
//
// input arguments:
// r10 <- inv_diag_E
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- inv_diag_E
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_EDGE_TRSM_RLT_INV_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_edge_trsm_rlt_inv_8x4_lib8, @function
inner_edge_trsm_rlt_inv_8x4_lib8:
#elif defined(OS_MAC)
_inner_edge_trsm_rlt_inv_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_edge_trsm_rlt_inv_8x4_lib8; .scl 2; .type 32; .endef
inner_edge_trsm_rlt_inv_8x4_lib8:
#endif
#endif
vbroadcastss 0(%r11), %ymm13
vmulps %ymm0, %ymm13, %ymm0
vbroadcastss 4(%r10), %ymm13
vmulps %ymm0, %ymm13, %ymm12
vsubps %ymm12, %ymm1, %ymm1
vbroadcastss 8(%r10), %ymm13
vmulps %ymm0, %ymm13, %ymm12
vsubps %ymm12, %ymm2, %ymm2
vbroadcastss 12(%r10), %ymm13
vmulps %ymm0, %ymm13, %ymm12
vsubps %ymm12, %ymm3, %ymm3
vbroadcastss 4(%r11), %ymm13
vmulps %ymm1, %ymm13, %ymm1
vbroadcastss 40(%r10), %ymm13
vmulps %ymm1, %ymm13, %ymm12
vsubps %ymm12, %ymm2, %ymm2
vbroadcastss 44(%r10), %ymm13
vmulps %ymm1, %ymm13, %ymm12
vsubps %ymm12, %ymm3, %ymm3
vbroadcastss 8(%r11), %ymm13
vmulps %ymm2, %ymm13, %ymm2
vbroadcastss 76(%r10), %ymm13
vmulps %ymm2, %ymm13, %ymm12
vsubps %ymm12, %ymm3, %ymm3
vbroadcastss 12(%r11), %ymm13
vmulps %ymm3, %ymm13, %ymm3
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_edge_trsm_rlt_inv_8x4_lib8, .-inner_edge_trsm_rlt_inv_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// strsm
// right
// lower
// transposed
// not-unit
//
// input arguments:
// r10 <- D
// r11 <- inv_diag_D
// r12d <- kn
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- D
// r11 <- inv_diag_D
// r12d <- kn
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_EDGE_TRSM_RLT_INV_8X4_VS_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_edge_trsm_rlt_inv_8x4_vs_lib8, @function
inner_edge_trsm_rlt_inv_8x4_vs_lib8:
#elif defined(OS_MAC)
_inner_edge_trsm_rlt_inv_8x4_vs_lib8:
#elif defined(OS_WINDOWS)
.def inner_edge_trsm_rlt_inv_8x4_vs_lib8; .scl 2; .type 32; .endef
inner_edge_trsm_rlt_inv_8x4_vs_lib8:
#endif
#endif
vbroadcastss 0(%r11), %ymm13
vmulps %ymm0, %ymm13, %ymm0
cmpl $2, %r12d
jl 0f // ret
vbroadcastss 4(%r10), %ymm13
vmulps %ymm0, %ymm13, %ymm12
vsubps %ymm12, %ymm1, %ymm1
vbroadcastss 8(%r10), %ymm13
vmulps %ymm0, %ymm13, %ymm12
vsubps %ymm12, %ymm2, %ymm2
vbroadcastss 12(%r10), %ymm13
vmulps %ymm0, %ymm13, %ymm12
vsubps %ymm12, %ymm3, %ymm3
vbroadcastss 4(%r11), %ymm13
vmulps %ymm1, %ymm13, %ymm1
cmpl $3, %r12d
jl 0f // ret
vbroadcastss 40(%r10), %ymm13
vmulps %ymm1, %ymm13, %ymm12
vsubps %ymm12, %ymm2, %ymm2
vbroadcastss 44(%r10), %ymm13
vmulps %ymm1, %ymm13, %ymm12
vsubps %ymm12, %ymm3, %ymm3
vbroadcastss 8(%r11), %ymm13
vmulps %ymm2, %ymm13, %ymm2
cmpl $4, %r12d
jl 0f // ret
vbroadcastss 76(%r10), %ymm13
vmulps %ymm2, %ymm13, %ymm12
vsubps %ymm12, %ymm3, %ymm3
vbroadcastss 12(%r11), %ymm13
vmulps %ymm3, %ymm13, %ymm3
0:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_edge_trsm_rlt_inv_8x4_vs_lib8, .-inner_edge_trsm_rlt_inv_8x4_vs_lib8
#endif
#endif
// common inner routine with file scope
//
// strsm
// right
// lower
// transposed
// not-unit
//
// input arguments:
// r10 <- D
// r11 <- inv_diag_D
// r12d <- kn
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm4 <- []
// ymm5 <- []
// ymm6 <- []
// ymm7 <- []
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- D
// r11 <- inv_diag_D
// r12d <- kn
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm4 <- []
// ymm5 <- []
// ymm6 <- []
// ymm7 <- []
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_EDGE_TRSM_RLT_INV_4X8_VS_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_edge_trsm_rlt_inv_4x8_vs_lib8, @function
inner_edge_trsm_rlt_inv_4x8_vs_lib8:
#elif defined(OS_MAC)
_inner_edge_trsm_rlt_inv_4x8_vs_lib8:
#elif defined(OS_WINDOWS)
.def inner_edge_trsm_rlt_inv_4x8_vs_lib8; .scl 2; .type 32; .endef
inner_edge_trsm_rlt_inv_4x8_vs_lib8:
#endif
#endif
vbroadcastss 0(%r11), %xmm13
vmulps %xmm0, %xmm13, %xmm0
vbroadcastss 4(%r10), %xmm13
vfnmadd231ps %xmm0, %xmm13, %xmm1
vbroadcastss 8(%r10), %xmm13
vfnmadd231ps %xmm0, %xmm13, %xmm2
vbroadcastss 12(%r10), %xmm13
vfnmadd231ps %xmm0, %xmm13, %xmm3
vbroadcastss 16(%r10), %xmm13
vfnmadd231ps %xmm0, %xmm13, %xmm4
vbroadcastss 20(%r10), %xmm13
vfnmadd231ps %xmm0, %xmm13, %xmm5
vbroadcastss 24(%r10), %xmm13
vfnmadd231ps %xmm0, %xmm13, %xmm6
vbroadcastss 28(%r10), %xmm13
vfnmadd231ps %xmm0, %xmm13, %xmm7
vbroadcastss 4(%r11), %xmm13
vmulps %xmm1, %xmm13, %xmm1
vbroadcastss 40(%r10), %xmm13
vfnmadd231ps %xmm1, %xmm13, %xmm2
vbroadcastss 44(%r10), %xmm13
vfnmadd231ps %xmm1, %xmm13, %xmm3
vbroadcastss 48(%r10), %xmm13
vfnmadd231ps %xmm1, %xmm13, %xmm4
vbroadcastss 52(%r10), %xmm13
vfnmadd231ps %xmm1, %xmm13, %xmm5
vbroadcastss 56(%r10), %xmm13
vfnmadd231ps %xmm1, %xmm13, %xmm6
vbroadcastss 60(%r10), %xmm13
vfnmadd231ps %xmm1, %xmm13, %xmm7
vbroadcastss 8(%r11), %xmm13
vmulps %xmm2, %xmm13, %xmm2
vbroadcastss 76(%r10), %xmm13
vfnmadd231ps %xmm2, %xmm13, %xmm3
vbroadcastss 80(%r10), %xmm13
vfnmadd231ps %xmm2, %xmm13, %xmm4
vbroadcastss 84(%r10), %xmm13
vfnmadd231ps %xmm2, %xmm13, %xmm5
vbroadcastss 88(%r10), %xmm13
vfnmadd231ps %xmm2, %xmm13, %xmm6
vbroadcastss 92(%r10), %xmm13
vfnmadd231ps %xmm2, %xmm13, %xmm7
vbroadcastss 12(%r11), %xmm13
vmulps %xmm3, %xmm13, %xmm3
vbroadcastss 112(%r10), %xmm13
vfnmadd231ps %xmm3, %xmm13, %xmm4
vbroadcastss 116(%r10), %xmm13
vfnmadd231ps %xmm3, %xmm13, %xmm5
vbroadcastss 120(%r10), %xmm13
vfnmadd231ps %xmm3, %xmm13, %xmm6
vbroadcastss 124(%r10), %xmm13
vfnmadd231ps %xmm3, %xmm13, %xmm7
vbroadcastss 16(%r11), %xmm13
vmulps %xmm4, %xmm13, %xmm4
cmpl $6, %r12d
jl 0f // ret
vbroadcastss 148(%r10), %xmm13
vfnmadd231ps %xmm4, %xmm13, %xmm5
vbroadcastss 152(%r10), %xmm13
vfnmadd231ps %xmm4, %xmm13, %xmm6
vbroadcastss 156(%r10), %xmm13
vfnmadd231ps %xmm4, %xmm13, %xmm7
vbroadcastss 20(%r11), %xmm13
vmulps %xmm5, %xmm13, %xmm5
cmpl $7, %r12d
jl 0f // ret
vbroadcastss 184(%r10), %xmm13
vfnmadd231ps %xmm5, %xmm13, %xmm6
vbroadcastss 188(%r10), %xmm13
vfnmadd231ps %xmm5, %xmm13, %xmm7
vbroadcastss 24(%r11), %xmm13
vmulps %xmm6, %xmm13, %xmm6
cmpl $8, %r12d
jl 0f // ret
vbroadcastss 220(%r10), %xmm13
vfnmadd231ps %xmm6, %xmm13, %xmm7
vbroadcastss 28(%r11), %xmm13
vmulps %xmm7, %xmm13, %xmm7
0:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_edge_trsm_rlt_inv_4x8_vs_lib8, .-inner_edge_trsm_rlt_inv_4x8_vs_lib8
#endif
#endif
// common inner routine with file scope
//
// cholesky factorization
//
// input arguments:
// r10 <- inv_diag_E
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm4 <- []
// ymm5 <- []
// ymm6 <- []
// ymm7 <- []
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- inv_diag_E
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm4 <- []
// ymm5 <- []
// ymm6 <- []
// ymm7 <- []
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_EDGE_POTRF_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_edge_potrf_8x4_lib8, @function
inner_edge_potrf_8x4_lib8:
#elif defined(OS_MAC)
_inner_edge_potrf_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_edge_potrf_8x4_lib8; .scl 2; .type 32; .endef
inner_edge_potrf_8x4_lib8:
#endif
#endif
vxorps %ymm15, %ymm15, %ymm15 // 0.0
#if defined(OS_LINUX) | defined(OS_WINDOWS)
vmovss .LC03(%rip), %xmm14 // 1.0
#elif defined(OS_MAC)
vmovss LC03(%rip), %xmm14 // 1.0
#endif
vmovss %xmm0, %xmm0, %xmm13
vucomiss %xmm15, %xmm13 // d_00 > 0.0 ?
jbe 1f
vsqrtss %xmm13, %xmm13, %xmm13
vdivss %xmm13, %xmm14, %xmm13
2:
vmovss %xmm13, 0(%r10)
vpermilps $0x00, %xmm13, %xmm13
vinsertf128 $0x1, %xmm13, %ymm13, %ymm13
vmulps %ymm0, %ymm13, %ymm0
vperm2f128 $0x00, %ymm0, %ymm0, %ymm11
vpermilps $0x55, %ymm11, %ymm13
vmulps %ymm0, %ymm13, %ymm12
vsubps %ymm12, %ymm1, %ymm1
vpermilps $0xaa, %ymm11, %ymm13
vmulps %ymm0, %ymm13, %ymm12
vsubps %ymm12, %ymm2, %ymm2
vpermilps $0xff, %ymm11, %ymm13
vmulps %ymm0, %ymm13, %ymm12
vsubps %ymm12, %ymm3, %ymm3
vpermilps $0x55, %xmm1, %xmm13
vucomiss %xmm15, %xmm13 // d_11 > 0.0 ?
jbe 3f
vsqrtss %xmm13, %xmm13, %xmm13
vdivss %xmm13, %xmm14, %xmm13
4:
vmovss %xmm13, 4(%r10)
vpermilps $0x00, %xmm13, %xmm13
vinsertf128 $0x1, %xmm13, %ymm13, %ymm13
vmulps %ymm1, %ymm13, %ymm1
vperm2f128 $0x00, %ymm1, %ymm1, %ymm11
vpermilps $0xaa, %ymm11, %ymm13
vmulps %ymm1, %ymm13, %ymm12
vsubps %ymm12, %ymm2, %ymm2
vpermilps $0xff, %ymm11, %ymm13
vmulps %ymm1, %ymm13, %ymm12
vsubps %ymm12, %ymm3, %ymm3
vpermilps $0xaa, %xmm2, %xmm13
vucomiss %xmm15, %xmm13 // d_22 > 0.0 ?
jbe 5f
vsqrtss %xmm13, %xmm13, %xmm13
vdivss %xmm13, %xmm14, %xmm13
6:
vmovss %xmm13, 8(%r10)
vpermilps $0x00, %xmm13, %xmm13
vinsertf128 $0x1, %xmm13, %ymm13, %ymm13
vmulps %ymm2, %ymm13, %ymm2
vperm2f128 $0x00, %ymm2, %ymm2, %ymm11
vpermilps $0xff, %ymm11, %ymm13
vmulps %ymm2, %ymm13, %ymm12
vsubps %ymm12, %ymm3, %ymm3
vpermilps $0xff, %xmm3, %xmm13
vucomiss %xmm15, %xmm13 // d_33 > 0.0 ?
jbe 7f
vsqrtss %xmm13, %xmm13, %xmm13
vdivss %xmm13, %xmm14, %xmm13
8:
vmovsd %xmm13, 12(%r10)
vpermilps $0x00, %xmm13, %xmm13
vinsertf128 $0x1, %xmm13, %ymm13, %ymm13
vmulps %ymm3, %ymm13, %ymm3
jmp 0f
1:
vxorps %ymm13, %ymm13, %ymm13
jmp 2b
3:
vxorpd %ymm13, %ymm13, %ymm13
jmp 4b
5:
vxorpd %ymm13, %ymm13, %ymm13
jmp 6b
7:
vxorpd %ymm13, %ymm13, %ymm13
jmp 8b
0:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_edge_potrf_8x4_lib8, .-inner_edge_potrf_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// cholesky factorization gen
//
// input arguments:
// r10 <- inv_diag_E
// r11d <- kn
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm4 <- []
// ymm5 <- []
// ymm6 <- []
// ymm7 <- []
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- inv_diag_E
// r11d <- kn
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm4 <- []
// ymm5 <- []
// ymm6 <- []
// ymm7 <- []
// ymm12 <- dirty
// ymm13 <- dirty
// ymm14 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_EDGE_POTRF_8X4_VS_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_edge_potrf_8x4_vs_lib8, @function
inner_edge_potrf_8x4_vs_lib8:
#elif defined(OS_MAC)
_inner_edge_potrf_8x4_vs_lib8:
#elif defined(OS_WINDOWS)
.def inner_edge_potrf_8x4_vs_lib8; .scl 2; .type 32; .endef
inner_edge_potrf_8x4_vs_lib8:
#endif
#endif
vxorps %ymm15, %ymm15, %ymm15 // 0.0
#if defined(OS_LINUX) | defined(OS_WINDOWS)
vmovss .LC03(%rip), %xmm14 // 1.0
#elif defined(OS_MAC)
vmovss LC03(%rip), %xmm14 // 1.0
#endif
vmovss %xmm0, %xmm0, %xmm13
vucomiss %xmm15, %xmm13 // d_00 > 0.0 ?
jbe 1f
vsqrtss %xmm13, %xmm13, %xmm13
vdivss %xmm13, %xmm14, %xmm13
2:
vmovss %xmm13, 0(%r10)
vpermilps $0x00, %xmm13, %xmm13
vinsertf128 $0x1, %xmm13, %ymm13, %ymm13
vmulps %ymm0, %ymm13, %ymm0
cmpl $2, %r11d
jl 0f // ret
vperm2f128 $0x00, %ymm0, %ymm0, %ymm11
vpermilps $0x55, %ymm11, %ymm13
vmulps %ymm0, %ymm13, %ymm12
vsubps %ymm12, %ymm1, %ymm1
vpermilps $0xaa, %ymm11, %ymm13
vmulps %ymm0, %ymm13, %ymm12
vsubps %ymm12, %ymm2, %ymm2
vpermilps $0xff, %ymm11, %ymm13
vmulps %ymm0, %ymm13, %ymm12
vsubps %ymm12, %ymm3, %ymm3
vpermilps $0x55, %xmm1, %xmm13
vucomiss %xmm15, %xmm13 // d_11 > 0.0 ?
jbe 3f
vsqrtss %xmm13, %xmm13, %xmm13
vdivss %xmm13, %xmm14, %xmm13
4:
vmovss %xmm13, 4(%r10)
vpermilps $0x00, %xmm13, %xmm13
vinsertf128 $0x1, %xmm13, %ymm13, %ymm13
vmulps %ymm1, %ymm13, %ymm1
cmpl $3, %r11d
jl 0f // ret
vperm2f128 $0x00, %ymm1, %ymm1, %ymm11
vpermilps $0xaa, %ymm11, %ymm13
vmulps %ymm1, %ymm13, %ymm12
vsubps %ymm12, %ymm2, %ymm2
vpermilps $0xff, %ymm11, %ymm13
vmulps %ymm1, %ymm13, %ymm12
vsubps %ymm12, %ymm3, %ymm3
vpermilps $0xaa, %xmm2, %xmm13
vucomiss %xmm15, %xmm13 // d_22 > 0.0 ?
jbe 5f
vsqrtss %xmm13, %xmm13, %xmm13
vdivss %xmm13, %xmm14, %xmm13
6:
vmovss %xmm13, 8(%r10)
vpermilps $0x00, %xmm13, %xmm13
vinsertf128 $0x1, %xmm13, %ymm13, %ymm13
vmulps %ymm2, %ymm13, %ymm2
cmpl $4, %r11d
jl 0f // ret
vperm2f128 $0x00, %ymm2, %ymm2, %ymm11
vpermilps $0xff, %ymm11, %ymm13
vmulps %ymm2, %ymm13, %ymm12
vsubps %ymm12, %ymm3, %ymm3
vpermilps $0xff, %xmm3, %xmm13
vucomiss %xmm15, %xmm13 // d_33 > 0.0 ?
jbe 7f
vsqrtss %xmm13, %xmm13, %xmm13
vdivss %xmm13, %xmm14, %xmm13
8:
vmovsd %xmm13, 12(%r10)
vpermilps $0x00, %xmm13, %xmm13
vinsertf128 $0x1, %xmm13, %ymm13, %ymm13
vmulps %ymm3, %ymm13, %ymm3
jmp 0f
1:
vxorps %ymm13, %ymm13, %ymm13
jmp 2b
3:
vxorpd %ymm13, %ymm13, %ymm13
jmp 4b
5:
vxorpd %ymm13, %ymm13, %ymm13
jmp 6b
7:
vxorpd %ymm13, %ymm13, %ymm13
jmp 8b
0:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_edge_potrf_8x4_vs_lib8, .-inner_edge_potrf_8x4_vs_lib8
#endif
#endif
// common inner routine with file scope
//
// scale for generic alpha and beta
//
// input arguments:
// r10 <- alpha
// r11 <- beta
// r12 <- C
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- alpha
// r11 <- beta
// r12 <- C
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_SCALE_AB_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_scale_ab_8x4_lib8, @function
inner_scale_ab_8x4_lib8:
#elif defined(OS_MAC)
_inner_scale_ab_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_scale_ab_8x4_lib8; .scl 2; .type 32; .endef
inner_scale_ab_8x4_lib8:
#endif
#endif
// alpha
vbroadcastss 0(%r10), %ymm15
vmulps %ymm0, %ymm15, %ymm0
vmulps %ymm1, %ymm15, %ymm1
vmulps %ymm2, %ymm15, %ymm2
vmulps %ymm3, %ymm15, %ymm3
// beta
vbroadcastss 0(%r11), %ymm14
vxorps %ymm15, %ymm15, %ymm15 // 0.0
vucomiss %xmm15, %xmm14 // beta==0.0 ?
je 0f // end
vmovaps 0(%r12), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm0
vmovaps 32(%r12), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm1
vmovaps 64(%r12), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm2
vmovaps 96(%r12), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm3
0:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_scale_ab_8x4_lib8, .-inner_scale_ab_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// scale for generic alpha and beta
//
// input arguments:
// r10 <- alpha
// r11 <- beta
// r12 <- C
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- alpha
// r11 <- beta
// r12 <- C
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_TRAN_SCALE_AB_4X8_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_tran_scale_ab_4x8_lib8, @function
inner_tran_scale_ab_4x8_lib8:
#elif defined(OS_MAC)
_inner_tran_scale_ab_4x8_lib8:
#elif defined(OS_WINDOWS)
.def inner_tran_scale_ab_4x8_lib8; .scl 2; .type 32; .endef
inner_tran_scale_ab_4x8_lib8:
#endif
#endif
// alpha
vbroadcastss 0(%r10), %ymm15
vmulps %ymm0, %ymm15, %ymm0
vmulps %ymm1, %ymm15, %ymm1
vmulps %ymm2, %ymm15, %ymm2
vmulps %ymm3, %ymm15, %ymm3
// transpose
vunpcklps %ymm1, %ymm0, %ymm5
vunpckhps %ymm1, %ymm0, %ymm4
vunpcklps %ymm3, %ymm2, %ymm7
vunpckhps %ymm3, %ymm2, %ymm6
vunpcklpd %ymm7, %ymm5, %ymm0
vunpckhpd %ymm7, %ymm5, %ymm1
vunpcklpd %ymm6, %ymm4, %ymm2
vunpckhpd %ymm6, %ymm4, %ymm3
vextractf128 $0x1, %ymm0, %xmm4
vextractf128 $0x1, %ymm1, %xmm5
vextractf128 $0x1, %ymm2, %xmm6
vextractf128 $0x1, %ymm3, %xmm7
// beta
vbroadcastss 0(%r11), %ymm14
vxorps %ymm15, %ymm15, %ymm15 // 0.0
vucomiss %xmm15, %xmm14 // beta==0.0 ?
je 0f // end
vmovaps 0(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm0
vmovaps 32(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm1
vmovaps 64(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm2
vmovaps 96(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm3
vmovaps 128(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm4
vmovaps 160(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm5
vmovaps 192(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm6
vmovaps 224(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm7
0:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_tran_scale_ab_4x8_lib8, .-inner_tran_scale_ab_4x8_lib8
#endif
#endif
// common inner routine with file scope
//
// blend scale for generic alpha and beta
//
// input arguments:
// r10 <- alpha
// r11 <- beta
// r12 <- offset
// r13 <- C
// r14 <- 4*sdc*sizeof(double)
// ymm0 <- [d00 d11 d22 d33]
// ymm1 <- [d01 d10 d23 d32]
// ymm2 <- [d03 d12 d21 d30]
// ymm3 <- [d02 d13 d20 d31]
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- alpha
// r11 <- beta
// r12 <- offset
// r13 <- C
// r14 <- 4*sdc*sizeof(double)
// r15 <- n0 // col index: start from (inc)
// ymm0 <- [d00 d10 d20 d30]
// ymm1 <- [d01 d11 d21 d31]
// ymm2 <- [d02 d12 d22 d32]
// ymm3 <- [d03 d13 d23 d33]
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_SCALE_AB_8X4_GEN_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_scale_ab_8x4_gen_lib8, @function
inner_scale_ab_8x4_gen_lib8:
#elif defined(OS_MAC)
_inner_scale_ab_8x4_gen_lib8:
#elif defined(OS_WINDOWS)
.def inner_scale_ab_8x4_gen_lib8; .scl 2; .type 32; .endef
inner_scale_ab_8x4_gen_lib8:
#endif
#endif
// alpha
vbroadcastss 0(%r10), %ymm15
vmulps %ymm0, %ymm15, %ymm0
vmulps %ymm1, %ymm15, %ymm1
vmulps %ymm2, %ymm15, %ymm2
vmulps %ymm3, %ymm15, %ymm3
// beta
vbroadcastss 0(%r11), %ymm15
vxorps %ymm14, %ymm14, %ymm14 // 0.0
vucomiss %xmm15, %xmm14 // beta==0.0 ?
je 3f // end
cmpl $0, %r12d
jg 0f
// offset==0
vmovaps 0(%r13), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm0
vmovaps 32(%r13), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm1
vmovaps 64(%r13), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm2
vmovaps 96(%r13), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm3
jmp 7f
0:
// offset > 0
// 1 2 3 4 5 6 7
movq %r13, %r15 // C0
addq %r14, %r15 // C1 <- C0 + 4*sdc*sizeof(double)
cmpl $4, %r10d
jl 1f
jg 2f
// offset==4
// TODO
jmp 7f
1:
// 1 2 3
cmpl $2, %r10d
jl 3f
jg 4f
// offset==2
// TODO
jmp 7f
3:
// offset==1
// TODO
jmp 7f
4:
// offset==3
// TODO
jmp 7f
2:
// 5 6 7
cmpl $6, %r10d
jl 5f
jg 6f
// offset==6
// TODO
jmp 7f
5:
// offset==5
// TODO
jmp 7f
6:
// offset==7
// TODO
jmp 7f
// end
7:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_scale_ab_8x4_gen_lib8, .-inner_scale_ab_8x4_gen_lib8
#endif
#endif
// common inner routine with file scope
//
// blend scale for generic alpha and beta
//
// input arguments:
// r10 <- alpha
// r11 <- beta
// r12 <- offset
// r13 <- C
// r14 <- 4*sdc*sizeof(double)
// ymm0 <- [d00 d11 d22 d33]
// ymm1 <- [d01 d10 d23 d32]
// ymm2 <- [d03 d12 d21 d30]
// ymm3 <- [d02 d13 d20 d31]
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- alpha
// r11 <- beta
// r12 <- offset
// r13 <- C
// r14 <- 4*sdc*sizeof(double)
// r15 <- n0 // col index: start from (inc)
// ymm0 <- [d00 d10 d20 d30]
// ymm1 <- [d01 d11 d21 d31]
// ymm2 <- [d02 d12 d22 d32]
// ymm3 <- [d03 d13 d23 d33]
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_TRAN_SCALE_AB_4X8_GEN_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_tran_scale_ab_4x8_gen_lib8, @function
inner_tran_scale_ab_4x8_gen_lib8:
#elif defined(OS_MAC)
_inner_tran_scale_ab_4x8_gen_lib8:
#elif defined(OS_WINDOWS)
.def inner_tran_scale_ab_4x8_gen_lib8; .scl 2; .type 32; .endef
inner_tran_scale_ab_4x8_gen_lib8:
#endif
#endif
// alpha
vbroadcastss 0(%r10), %ymm15
vmulps %ymm0, %ymm15, %ymm0
vmulps %ymm1, %ymm15, %ymm1
vmulps %ymm2, %ymm15, %ymm2
vmulps %ymm3, %ymm15, %ymm3
// transpose
vunpcklps %ymm1, %ymm0, %ymm5
vunpckhps %ymm1, %ymm0, %ymm4
vunpcklps %ymm3, %ymm2, %ymm7
vunpckhps %ymm3, %ymm2, %ymm6
vunpcklpd %ymm7, %ymm5, %ymm0
vunpckhpd %ymm7, %ymm5, %ymm1
vunpcklpd %ymm6, %ymm4, %ymm2
vunpckhpd %ymm6, %ymm4, %ymm3
vextractf128 $0x1, %ymm0, %xmm4
vextractf128 $0x1, %ymm1, %xmm5
vextractf128 $0x1, %ymm2, %xmm6
vextractf128 $0x1, %ymm3, %xmm7
// beta
vbroadcastss 0(%r11), %ymm15
vxorps %ymm14, %ymm14, %ymm14 // 0.0
vucomiss %xmm15, %xmm14 // beta==0.0 ?
je 3f // end
cmpl $0, %r12d
jg 0f
// offset==0
vmovaps 0(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm0
vmovaps 32(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm1
vmovaps 64(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm2
vmovaps 96(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm3
vmovaps 128(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm4
vmovaps 160(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm5
vmovaps 192(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm6
vmovaps 224(%r12), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm7
jmp 7f
0:
// offset > 0
// 1 2 3 4 5 6 7
movq %r13, %r15 // C0
addq %r14, %r15 // C1 <- C0 + 4*sdc*sizeof(double)
cmpl $4, %r10d
jl 1f
jg 2f
// offset==4
// TODO
jmp 7f
1:
// 1 2 3
cmpl $2, %r10d
jl 3f
jg 4f
// offset==2
// TODO
jmp 7f
3:
// offset==1
// TODO
jmp 7f
4:
// offset==3
// TODO
jmp 7f
2:
// 5 6 7
cmpl $6, %r10d
jl 5f
jg 6f
// offset==6
// TODO
jmp 7f
5:
// offset==5
// TODO
jmp 7f
6:
// offset==7
// TODO
jmp 7f
// end
7:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_tran_scale_ab_4x8_gen_lib8, .-inner_tran_scale_ab_4x8_gen_lib8
#endif
#endif
// common inner routine with file scope
//
// scale for generic alpha and beta=0
//
// input arguments:
// r10 <- alpha
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- alpha
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_SCALE_A0_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_scale_a0_8x4_lib8, @function
inner_scale_a0_8x4_lib8:
#elif defined(OS_MAC)
_inner_scale_a0_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_scale_a0_8x4_lib8; .scl 2; .type 32; .endef
inner_scale_a0_8x4_lib8:
#endif
#endif
// alpha
vbroadcastss 0(%r10), %ymm15
vmulps %ymm0, %ymm15, %ymm0
vmulps %ymm1, %ymm15, %ymm1
vmulps %ymm2, %ymm15, %ymm2
vmulps %ymm3, %ymm15, %ymm3
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_scale_a0_8x4_lib8, .-inner_scale_a0_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// scale for generic alpha and beta
//
// input arguments:
// r10 <- C
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- C
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_SCALE_11_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_scale_11_8x4_lib8, @function
inner_scale_11_8x4_lib8:
#elif defined(OS_MAC)
_inner_scale_11_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_scale_11_8x4_lib8; .scl 2; .type 32; .endef
inner_scale_11_8x4_lib8:
#endif
#endif
#if defined(OS_LINUX) | defined(OS_WINDOWS)
vmovaps .LC03(%rip), %ymm14
#elif defined(OS_MAC)
vmovaps LC03(%rip), %ymm14
#endif
vmovaps 0(%r10), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm0
vmovaps 32(%r10), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm1
vmovaps 64(%r10), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm2
vmovaps 96(%r10), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm3
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_scale_11_8x4_lib8, .-inner_scale_11_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// scale for generic alpha and beta
//
// input arguments:
// r10 <- C
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- C
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_TRAN_SCALE_11_4X8_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_tran_scale_11_4x8_lib8, @function
inner_tran_scale_11_4x8_lib8:
#elif defined(OS_MAC)
_inner_tran_scale_11_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_tran_scale_11_4x8_lib8; .scl 2; .type 32; .endef
inner_tran_scale_11_4x8_lib8:
#endif
#endif
// transpose
vunpcklps %ymm1, %ymm0, %ymm5
vunpckhps %ymm1, %ymm0, %ymm4
vunpcklps %ymm3, %ymm2, %ymm7
vunpckhps %ymm3, %ymm2, %ymm6
vunpcklpd %ymm7, %ymm5, %ymm0
vunpckhpd %ymm7, %ymm5, %ymm1
vunpcklpd %ymm6, %ymm4, %ymm2
vunpckhpd %ymm6, %ymm4, %ymm3
vextractf128 $0x1, %ymm0, %xmm4
vextractf128 $0x1, %ymm1, %xmm5
vextractf128 $0x1, %ymm2, %xmm6
vextractf128 $0x1, %ymm3, %xmm7
#if defined(OS_LINUX) | defined(OS_WINDOWS)
vmovaps .LC03(%rip), %ymm14
#elif defined(OS_MAC)
vmovaps LC03(%rip), %ymm14
#endif
vmovaps 0(%r10), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm0
vmovaps 32(%r10), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm1
vmovaps 64(%r10), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm2
vmovaps 96(%r10), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm3
vmovaps 128(%r10), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm4
vmovaps 160(%r10), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm5
vmovaps 192(%r10), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm6
vmovaps 224(%r10), %xmm15
vfmadd231ps %xmm15, %xmm14, %xmm7
0:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_tran_scale_11_4x8_lib8, .-inner_tran_scale_11_4x8_lib8
#endif
#endif
// common inner routine with file scope
//
// blend scale for generic alpha and beta
//
// input arguments:
// r10 <- offset
// r11 <- C
// r12 <- 4*sdc*sizeof(double)
// ymm0 <- [d00 d11 d22 d33]
// ymm1 <- [d01 d10 d23 d32]
// ymm2 <- [d03 d12 d21 d30]
// ymm3 <- [d02 d13 d20 d31]
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- offset
// r11 <- C
// r12 <- 4*sdc*sizeof(double)
// r15 <- n0 // col index: start from (inc)
// ymm0 <- [d00 d10 d20 d30]
// ymm1 <- [d01 d11 d21 d31]
// ymm2 <- [d02 d12 d22 d32]
// ymm3 <- [d03 d13 d23 d33]
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_SCALE_11_8X4_GEN_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_scale_11_8x4_gen_lib8, @function
inner_scale_11_8x4_gen_lib8:
#elif defined(OS_MAC)
_inner_scale_11_8x4_gen_lib8:
#elif defined(OS_WINDOWS)
.def inner_scale_11_8x4_gen_lib8; .scl 2; .type 32; .endef
inner_scale_11_8x4_gen_lib8:
#endif
#endif
// offset==0
#if defined(OS_LINUX) | defined(OS_WINDOWS)
vmovaps .LC03(%rip), %ymm14
#elif defined(OS_MAC)
vmovaps LC03(%rip), %ymm14
#endif
vmovaps 0(%r11), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm0
vmovaps 32(%r11), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm1
vmovaps 64(%r11), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm2
vmovaps 96(%r11), %ymm15
vfmadd231ps %ymm15, %ymm14, %ymm3
jmp 7f
0:
// offset > 0
// 1 2 3 4 5 6 7
movq %r13, %r15 // C0
addq %r14, %r15 // C1 <- C0 + 4*sdc*sizeof(double)
cmpl $4, %r10d
jl 1f
jg 2f
// offset==4
// TODO
jmp 7f
1:
// 1 2 3
cmpl $2, %r10d
jl 3f
jg 4f
// offset==2
// TODO
jmp 7f
3:
// offset==1
// TODO
jmp 7f
4:
// offset==3
// TODO
jmp 7f
2:
// 5 6 7
cmpl $6, %r10d
jl 5f
jg 6f
// offset==6
// TODO
jmp 7f
5:
// offset==5
// TODO
jmp 7f
6:
// offset==7
// TODO
jmp 7f
// end
7:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_scale_11_8x4_gen_lib8, .-inner_scale_11_8x4_gen_lib8
#endif
#endif
// common inner routine with file scope
//
// blend scale for generic alpha and beta
//
// input arguments:
// r10 <- alpha
// r11 <- beta
// r12 <- C
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- alpha
// r11 <- beta
// r12 <- C
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_BLEND_SCALE_AB_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_blend_scale_ab_8x4_lib8, @function
inner_blend_scale_ab_8x4_lib8:
#elif defined(OS_MAC)
_inner_blend_scale_ab_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_blend_scale_ab_8x4_lib8; .scl 2; .type 32; .endef
inner_blend_scale_ab_8x4_lib8:
#endif
#endif
vblendps $0xaa, %ymm1, %ymm0, %ymm8 // 1010 1010
vblendps $0x55, %ymm1, %ymm0, %ymm9 // 0101 0101
vblendps $0xaa, %ymm3, %ymm2, %ymm10
vblendps $0x55, %ymm3, %ymm2, %ymm11
vblendps $0xcc, %ymm11, %ymm8, %ymm0 // 1100 1100
vblendps $0x33, %ymm11, %ymm8, %ymm2 // 0011 0011
vblendps $0xcc, %ymm10, %ymm9, %ymm1
vblendps $0x33, %ymm10, %ymm9, %ymm3
// alpha
vbroadcastss 0(%r10), %ymm15
vmulps %ymm0, %ymm15, %ymm0
vmulps %ymm1, %ymm15, %ymm1
vmulps %ymm2, %ymm15, %ymm2
vmulps %ymm3, %ymm15, %ymm3
// beta
vbroadcastss 0(%r11), %ymm14
vxorps %ymm15, %ymm15, %ymm15 // 0.0
vucomiss %xmm15, %xmm14 // beta==0.0 ?
je 0f // end
vmovaps 0(%r12), %ymm15
vmulps %ymm15, %ymm14, %ymm15
vaddps %ymm0, %ymm15, %ymm0
vmovaps 32(%r12), %ymm15
vmulps %ymm15, %ymm14, %ymm15
vaddps %ymm1, %ymm15, %ymm1
vmovaps 64(%r12), %ymm15
vmulps %ymm15, %ymm14, %ymm15
vaddps %ymm2, %ymm15, %ymm2
vmovaps 96(%r12), %ymm15
vmulps %ymm15, %ymm14, %ymm15
vaddps %ymm3, %ymm15, %ymm3
0:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_blend_scale_ab_8x4_lib8, .-inner_blend_scale_ab_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// blend scale for generic alpha and beta
//
// input arguments:
// r10 <- alpha
// r11 <- beta
// r12 <- offset
// r13 <- C
// r14 <- 4*sdc*sizeof(double)
// ymm0 <- [d00 d11 d22 d33]
// ymm1 <- [d01 d10 d23 d32]
// ymm2 <- [d03 d12 d21 d30]
// ymm3 <- [d02 d13 d20 d31]
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- alpha
// r11 <- beta
// r12 <- offset
// r13 <- C
// r14 <- 4*sdc*sizeof(double)
// r15 <- n0 // col index: start from (inc)
// ymm0 <- [d00 d10 d20 d30]
// ymm1 <- [d01 d11 d21 d31]
// ymm2 <- [d02 d12 d22 d32]
// ymm3 <- [d03 d13 d23 d33]
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_BLEND_SCALE_AB_8X4_GEN_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_blend_scale_ab_8x4_gen_lib8, @function
inner_blend_scale_ab_8x4_gen_lib8:
#elif defined(OS_MAC)
_inner_blend_scale_ab_8x4_gen_lib8:
#elif defined(OS_WINDOWS)
.def inner_blend_scale_ab_8x4_gen_lib8; .scl 2; .type 32; .endef
inner_blend_scale_ab_8x4_gen_lib8:
#endif
#endif
vblendps $0xaa, %ymm1, %ymm0, %ymm8 // 1010 1010
vblendps $0x55, %ymm1, %ymm0, %ymm9 // 0101 0101
vblendps $0xaa, %ymm3, %ymm2, %ymm10
vblendps $0x55, %ymm3, %ymm2, %ymm11
vblendps $0xcc, %ymm11, %ymm8, %ymm0 // 1100 1100
vblendps $0x33, %ymm11, %ymm8, %ymm2 // 0011 0011
vblendps $0xcc, %ymm10, %ymm9, %ymm1
vblendps $0x33, %ymm10, %ymm9, %ymm3
// alpha
vbroadcastss 0(%r10), %ymm15
vmulps %ymm0, %ymm15, %ymm0
vmulps %ymm1, %ymm15, %ymm1
vmulps %ymm2, %ymm15, %ymm2
vmulps %ymm3, %ymm15, %ymm3
// beta
vbroadcastss 0(%r11), %ymm15
vxorps %ymm14, %ymm14, %ymm14 // 0.0
vucomiss %xmm15, %xmm14 // beta==0.0 ?
je 3f // end
cmpl $0, %r12d
jg 0f
// offset==0
vmovaps 0(%r13), %ymm12
vmulps %ymm12, %ymm15, %ymm12
vaddps %ymm0, %ymm12, %ymm0
vmovaps 32(%r13), %ymm12
vmulps %ymm12, %ymm15, %ymm12
vaddps %ymm1, %ymm12, %ymm1
vmovaps 64(%r13), %ymm12
vmulps %ymm12, %ymm15, %ymm12
vaddps %ymm2, %ymm12, %ymm2
vmovaps 96(%r13), %ymm12
vmulps %ymm12, %ymm15, %ymm12
vaddps %ymm3, %ymm12, %ymm3
jmp 7f
0:
// offset > 0
// 1 2 3 4 5 6 7
movq %r13, %r15 // C0
addq %r14, %r15 // C1 <- C0 + 4*sdc*sizeof(double)
cmpl $4, %r10d
jl 1f
jg 2f
// offset==4
// TODO
jmp 7f
1:
// 1 2 3
cmpl $2, %r10d
jl 3f
jg 4f
// offset==2
// TODO
jmp 7f
3:
// offset==1
// TODO
jmp 7f
4:
// offset==3
// TODO
jmp 7f
2:
// 5 6 7
cmpl $6, %r10d
jl 5f
jg 6f
// offset==6
// TODO
jmp 7f
5:
// offset==5
// TODO
jmp 7f
6:
// offset==7
// TODO
jmp 7f
// end
7:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_blend_scale_ab_8x4_gen_lib8, .-inner_blend_scale_ab_8x4_gen_lib8
#endif
#endif
// common inner routine with file scope
//
// blend scale for generic alpha=1.0 and beta=1.0
//
// input arguments:
// r10 <- C
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- C
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_BLEND_SCALE_11_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_blend_scale_11_8x4_lib8, @function
inner_blend_scale_11_8x4_lib8:
#elif defined(OS_MAC)
_inner_blend_scale_11_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_blend_scale_11_8x4_lib8; .scl 2; .type 32; .endef
inner_blend_scale_11_8x4_lib8:
#endif
#endif
vblendps $0xaa, %ymm1, %ymm0, %ymm8 // 1010 1010
vblendps $0x55, %ymm1, %ymm0, %ymm9 // 0101 0101
vblendps $0xaa, %ymm3, %ymm2, %ymm10
vblendps $0x55, %ymm3, %ymm2, %ymm11
vblendps $0xcc, %ymm11, %ymm8, %ymm0 // 1100 1100
vblendps $0x33, %ymm11, %ymm8, %ymm2 // 0011 0011
vblendps $0xcc, %ymm10, %ymm9, %ymm1
vblendps $0x33, %ymm10, %ymm9, %ymm3
vmovaps 0(%r10), %ymm15
vaddps %ymm0, %ymm15, %ymm0
vmovaps 32(%r10), %ymm15
vaddps %ymm1, %ymm15, %ymm1
vmovaps 64(%r10), %ymm15
vaddps %ymm2, %ymm15, %ymm2
vmovaps 96(%r10), %ymm15
vaddps %ymm3, %ymm15, %ymm3
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_blend_scale_11_8x4_lib8, .-inner_blend_scale_11_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// blend scale for generic alpha=1.0 and beta=1.0
//
// input arguments:
// r10 <- offset
// r11 <- C
// r12 <- 4*sdc*sizeof(double)
// ymm0 <- [d00 d11 d22 d33]
// ymm1 <- [d01 d10 d23 d32]
// ymm2 <- [d03 d12 d21 d30]
// ymm3 <- [d02 d13 d20 d31]
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
//
// output arguments:
// r10 <- offset
// r11 <- C
// r12 <- 4*sdc*sizeof(double)
// ymm0 <- [d00 d10 d20 d30]
// ymm1 <- [d01 d11 d21 d31]
// ymm2 <- [d02 d12 d22 d32]
// ymm3 <- [d03 d13 d23 d33]
// ymm8 <- dirty
// ymm9 <- dirty
// ymm10 <- dirty
// ymm11 <- dirty
// ymm15 <- dirty
#if MACRO_LEVEL>=1
.macro INNER_BLEND_SCALE_11_8X4_GEN_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_blend_scale_11_8x4_gen_lib8, @function
inner_blend_scale_11_8x4_gen_lib8:
#elif defined(OS_MAC)
_inner_blend_scale_11_8x4_gen_lib8:
#elif defined(OS_WINDOWS)
.def inner_blend_scale_11_8x4_gen_lib8; .scl 2; .type 32; .endef
inner_blend_scale_11_8x4_gen_lib8:
#endif
#endif
vblendps $0xaa, %ymm1, %ymm0, %ymm8 // 1010 1010
vblendps $0x55, %ymm1, %ymm0, %ymm9 // 0101 0101
vblendps $0xaa, %ymm3, %ymm2, %ymm10
vblendps $0x55, %ymm3, %ymm2, %ymm11
vblendps $0xcc, %ymm11, %ymm8, %ymm0 // 1100 1100
vblendps $0x33, %ymm11, %ymm8, %ymm2 // 0011 0011
vblendps $0xcc, %ymm10, %ymm9, %ymm1
vblendps $0x33, %ymm10, %ymm9, %ymm3
// offset==0
vmovaps 0(%r11), %ymm12
vaddps %ymm0, %ymm12, %ymm0
vmovaps 32(%r11), %ymm12
vaddps %ymm1, %ymm12, %ymm1
vmovaps 64(%r11), %ymm12
vaddps %ymm2, %ymm12, %ymm2
vmovaps 96(%r11), %ymm12
vaddps %ymm3, %ymm12, %ymm3
jmp 7f
0:
// offset > 0
// 1 2 3 4 5 6 7
movq %r11, %r15 // C0
addq %r12, %r15 // C1 <- C0 + 4*sdc*sizeof(double)
cmpl $4, %r10d
jl 1f
jg 2f
// offset==4
// TODO
jmp 7f
1:
// 1 2 3
cmpl $2, %r10d
jl 3f
jg 4f
// offset==2
// TODO
jmp 7f
3:
// offset==1
// TODO
jmp 7f
4:
// offset==3
// TODO
jmp 7f
2:
// 5 6 7
cmpl $6, %r10d
jl 5f
jg 6f
// offset==6
// TODO
jmp 7f
5:
// offset==5
// TODO
jmp 7f
6:
// offset==7
// TODO
jmp 7f
// end
7:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_blend_scale_11_8x4_gen_lib8, .-inner_blend_scale_11_8x4_gen_lib8
#endif
#endif
// common inner routine with file scope
//
// store n
//
// input arguments:
// r10 <- D
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
//
// output arguments:
// r10 <- D
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
#if MACRO_LEVEL>=1
.macro INNER_STORE_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_store_8x4_lib8, @function
inner_store_8x4_lib8:
#elif defined(OS_MAC)
_inner_store_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_store_8x4_lib8; .scl 2; .type 32; .endef
inner_store_8x4_lib8:
#endif
#endif
vmovaps %ymm0, 0(%r10)
vmovaps %ymm1, 32(%r10)
vmovaps %ymm2, 64(%r10)
vmovaps %ymm3, 96(%r10)
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_store_8x4_lib8, .-inner_store_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// store n
//
// input arguments:
// r10 <- D
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
//
// output arguments:
// r10 <- D
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
#if MACRO_LEVEL>=1
.macro INNER_STORE_4X8_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_store_4x8_lib8, @function
inner_store_4x8_lib8:
#elif defined(OS_MAC)
_inner_store_4x8_lib8:
#elif defined(OS_WINDOWS)
.def inner_store_4x8_lib8; .scl 2; .type 32; .endef
inner_store_4x8_lib8:
#endif
#endif
vmovaps %xmm0, 0(%r10)
vmovaps %xmm1, 32(%r10)
vmovaps %xmm2, 64(%r10)
vmovaps %xmm3, 96(%r10)
vmovaps %xmm4, 128(%r10)
vmovaps %xmm5, 160(%r10)
vmovaps %xmm6, 192(%r10)
vmovaps %xmm7, 224(%r10)
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_store_4x8_lib8, .-inner_store_4x8_lib8
#endif
#endif
// common inner routine with file scope
//
// store n vs
//
// input arguments:
// r10 <- D
// r11 <- km
// r12 <- kn
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
//
// output arguments:
// r10 <- D
// r11 <- km
// r12 <- kn
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
#if MACRO_LEVEL>=1
.macro INNER_STORE_8X4_VS_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_store_8x4_vs_lib8, @function
inner_store_8x4_vs_lib8:
#elif defined(OS_MAC)
_inner_store_8x4_vs_lib8:
#elif defined(OS_WINDOWS)
.def inner_store_8x4_vs_lib8; .scl 2; .type 32; .endef
inner_store_8x4_vs_lib8:
#endif
#endif
// compute mask for rows
vcvtsi2ss %r11d, %xmm14, %xmm14
#if defined(OS_LINUX) | defined(OS_WINDOWS)
vmovups .LC00(%rip), %ymm12
#elif defined(OS_MAC)
vmovups LC00(%rip), %ymm12
#endif
vshufps $0x00, %xmm14, %xmm14, %xmm14
vinsertf128 $0x1, %xmm14, %ymm14, %ymm14
vsubps %ymm14, %ymm12, %ymm14
// offset==0
vmaskmovps %ymm0, %ymm14, 0(%r10)
cmpl $2, %r12d
jl 0f // end
vmaskmovps %ymm1, %ymm14, 32(%r10)
cmpl $3, %r12d
jl 0f // end
vmaskmovps %ymm2, %ymm14, 64(%r10)
je 0f // end
vmaskmovps %ymm3, %ymm14, 96(%r10)
0:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_store_8x4_vs_lib8, .-inner_store_8x4_vs_lib8
#endif
#endif
// common inner routine with file scope
//
// store n generalized
//
// input arguments:
// r10 <- D
// r11 <- km
// r12 <- kn
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
//
// output arguments:
// r10 <- D
// r11 <- km
// r12 <- kn
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
#if MACRO_LEVEL>=1
.macro INNER_STORE_4X8_VS_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_store_4x8_vs_lib8, @function
inner_store_4x8_vs_lib8:
#elif defined(OS_MAC)
_inner_store_4x8_vs_lib8:
#elif defined(OS_WINDOWS)
.def inner_store_4x8_vs_lib8; .scl 2; .type 32; .endef
inner_store_4x8_vs_lib8:
#endif
#endif
// compute mask for rows
vcvtsi2ss %r11d, %xmm14, %xmm14
#if defined(OS_LINUX) | defined(OS_WINDOWS)
vmovups .LC00(%rip), %xmm12
#elif defined(OS_MAC)
vmovups LC00(%rip), %xmm12
#endif
vshufps $0x00, %xmm14, %xmm14, %xmm14
vinsertf128 $0x1, %xmm14, %ymm14, %ymm14
vsubps %xmm14, %xmm12, %xmm14
// offset==0
vmaskmovps %xmm0, %xmm14, 0(%r10)
cmpl $2, %r12d
jl 0f // end
vmaskmovps %xmm1, %xmm14, 32(%r10)
cmpl $3, %r12d
jl 0f // end
vmaskmovps %xmm2, %xmm14, 64(%r10)
cmpl $4, %r12d
jl 0f // end
vmaskmovps %xmm3, %xmm14, 96(%r10)
cmpl $5, %r12d
jl 0f // end
vmaskmovps %xmm4, %xmm14, 128(%r10)
cmpl $6, %r12d
jl 0f // end
vmaskmovps %xmm5, %xmm14, 160(%r10)
cmpl $7, %r12d
jl 0f // end
vmaskmovps %xmm6, %xmm14, 192(%r10)
je 0f // end
vmaskmovps %xmm7, %xmm14, 224(%r10)
//
0:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_store_4x8_vs_lib8, .-inner_store_4x8_vs_lib8
#endif
#endif
// common inner routine with file scope
//
// store n generalized
//
// input arguments:
// r10 <- offset
// r11 <- D
// r12 <- 4*sdd*sizeof(double)
// r13 <- m0 // row index: start from (inc)
// r14 <- m1 // row index: up to (exc)
// r15 <- n0 // col index: start from (inc)
// rax <- n1 // col index: up to (exc)
// rbx <- dirty
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
//
// output arguments:
// r10 <- offset
// r11 <- D
// r12 <- 4*sdd*sizeof(double)
// r13 <- m0 // row index: start from (inc)
// r14 <- m1 // row index: up to (exc)
// r15 <- n1-n0
// rax <- n1-n0
// rbx <- dirty
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
#if MACRO_LEVEL>=1
.macro INNER_STORE_8X4_GEN_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_store_8x4_gen_lib8, @function
inner_store_8x4_gen_lib8:
#elif defined(OS_MAC)
_inner_store_8x4_gen_lib8:
#elif defined(OS_WINDOWS)
.def inner_store_8x4_gen_lib8; .scl 2; .type 32; .endef
inner_store_8x4_gen_lib8:
#endif
#endif
// compute mask for rows
vcvtsi2ss %r13d, %xmm14, %xmm14
vcvtsi2ss %r14d, %xmm15, %xmm15
#if defined(OS_LINUX) | defined(OS_WINDOWS)
vmovups .LC00(%rip), %ymm12
#elif defined(OS_MAC)
vmovups LC00(%rip), %ymm12
#endif
vshufps $0x00, %xmm14, %xmm14, %xmm14
vshufps $0x00, %xmm15, %xmm15, %xmm15
vinsertf128 $0x1, %xmm14, %ymm14, %ymm14
vinsertf128 $0x1, %xmm15, %ymm15, %ymm15
vsubps %ymm12, %ymm14, %ymm14
vsubps %ymm15, %ymm12, %ymm15
vandps %ymm14, %ymm15, %ymm15
// shift D and sol for cols
cmpl $0, %r15d
jle 0f
vmovaps %ymm1, %ymm0
vmovaps %ymm2, %ymm1
vmovaps %ymm3, %ymm2
addq $32, %r11
cmpl $1, %r15d
jle 0f
vmovaps %ymm1, %ymm0
vmovaps %ymm2, %ymm1
addq $32, %r11
cmpl $2, %r15d
jle 0f
vmovaps %ymm1, %ymm0
addq $32, %r11
0:
// compute number of cols
cmpl $4, %eax
jle 0f
movl $4, %eax
0:
subl %r15d, %eax
movl %eax, %r15d
cmpl $0, %r10d
jg 0f
// offset==0
vmaskmovps %ymm0, %ymm15, 0(%r11)
cmpl $2, %r15d
jl 7f // end
vmaskmovps %ymm1, %ymm15, 32(%r11)
cmpl $3, %r15d
jl 7f // end
vmaskmovps %ymm2, %ymm15, 64(%r11)
je 7f // end
vmaskmovps %ymm3, %ymm15, 96(%r11)
//
jmp 7f
0:
// offset > 0
// 1 2 3 4 5 6 7
movq %r11, %rbx // D0
addq %r12, %rbx // D1 <- D0 + 4*sdd*sizeof(double)
cmpl $4, %r10d
jl 1f
jg 2f
// offset==4
// TODO
jmp 7f
1:
// 1 2 3
cmpl $2, %r10d
jl 3f
jg 4f
// offset==2
// TODO
jmp 7f
3:
// offset==1
// TODO
jmp 7f
4:
// offset==3
// TODO
jmp 7f
2:
// 5 6 7
cmpl $6, %r10d
jl 5f
jg 6f
// offset==6
// TODO
jmp 7f
5:
// offset==5
// TODO
jmp 7f
6:
// offset==7
// TODO
jmp 7f
// end
7:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_store_8x4_gen_lib8, .-inner_store_8x4_gen_lib8
#endif
#endif
// common inner routine with file scope
//
// store n generalized
//
// input arguments:
// r10 <- offset
// r11 <- D
// r12 <- 4*sdd*sizeof(double)
// r13 <- m0 // row index: start from (inc)
// r14 <- m1 // row index: up to (exc)
// r15 <- n0 // col index: start from (inc)
// rax <- n1 // col index: up to (exc)
// rbx <- dirty
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
//
// output arguments:
// r10 <- offset
// r11 <- D
// r12 <- 4*sdd*sizeof(double)
// r13 <- m0 // row index: start from (inc)
// r14 <- m1 // row index: up to (exc)
// r15 <- n1-n0
// rax <- n1-n0
// rbx <- dirty
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
#if MACRO_LEVEL>=1
.macro INNER_STORE_4X8_GEN_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_store_4x8_gen_lib8, @function
inner_store_4x8_gen_lib8:
#elif defined(OS_MAC)
_inner_store_4x8_gen_lib8:
#elif defined(OS_WINDOWS)
.def inner_store_4x8_gen_lib8; .scl 2; .type 32; .endef
inner_store_4x8_gen_lib8:
#endif
#endif
// compute mask for rows
vcvtsi2ss %r13d, %xmm14, %xmm14
vcvtsi2ss %r14d, %xmm15, %xmm15
#if defined(OS_LINUX) | defined(OS_WINDOWS)
vmovups .LC00(%rip), %xmm12
#elif defined(OS_MAC)
vmovups LC00(%rip), %xmm12
#endif
vshufps $0x00, %xmm14, %xmm14, %xmm14
vshufps $0x00, %xmm15, %xmm15, %xmm15
vinsertf128 $0x1, %xmm14, %ymm14, %ymm14
vinsertf128 $0x1, %xmm15, %ymm15, %ymm15
vsubps %xmm12, %xmm14, %xmm14
vsubps %xmm15, %xmm12, %xmm15
vandps %xmm14, %xmm15, %xmm15
// shift D and sol for cols
cmpl $0, %r15d
jle 0f
vmovaps %xmm1, %xmm0
vmovaps %xmm2, %xmm1
vmovaps %xmm3, %xmm2
vmovaps %xmm4, %xmm3
vmovaps %xmm5, %xmm4
vmovaps %xmm6, %xmm5
vmovaps %xmm7, %xmm6
addq $32, %r11
cmpl $1, %r15d
jle 0f
vmovaps %xmm1, %xmm0
vmovaps %xmm2, %xmm1
vmovaps %xmm3, %xmm2
vmovaps %xmm4, %xmm3
vmovaps %xmm5, %xmm4
vmovaps %xmm6, %xmm5
addq $32, %r11
cmpl $2, %r15d
jle 0f
vmovaps %xmm1, %xmm0
vmovaps %xmm2, %xmm1
vmovaps %xmm3, %xmm2
vmovaps %xmm4, %xmm3
vmovaps %xmm5, %xmm4
addq $32, %r11
cmpl $3, %r15d
jle 0f
vmovaps %xmm1, %xmm0
vmovaps %xmm2, %xmm1
vmovaps %xmm3, %xmm2
vmovaps %xmm4, %xmm3
addq $32, %r11
cmpl $4, %r15d
jle 0f
vmovaps %xmm1, %xmm0
vmovaps %xmm2, %xmm1
vmovaps %xmm3, %xmm2
addq $32, %r11
cmpl $5, %r15d
jle 0f
vmovaps %xmm1, %xmm0
vmovaps %xmm2, %xmm1
addq $32, %r11
cmpl $6, %r15d
jle 0f
vmovaps %xmm1, %xmm0
addq $32, %r11
0:
// compute number of cols
cmpl $8, %eax
jle 0f
movl $8, %eax
0:
subl %r15d, %eax
movl %eax, %r15d
cmpl $0, %r10d
jg 0f
// offset==0
vmaskmovps %xmm0, %xmm15, 0(%r11)
cmpl $2, %r15d
jl 7f // end
vmaskmovps %xmm1, %xmm15, 32(%r11)
cmpl $3, %r15d
jl 7f // end
vmaskmovps %xmm2, %xmm15, 64(%r11)
cmpl $4, %r15d
jl 7f // end
vmaskmovps %xmm3, %xmm15, 96(%r11)
cmpl $5, %r15d
jl 7f // end
vmaskmovps %xmm4, %xmm15, 128(%r11)
cmpl $6, %r15d
jl 7f // end
vmaskmovps %xmm5, %xmm15, 160(%r11)
cmpl $7, %r15d
jl 7f // end
vmaskmovps %xmm6, %xmm15, 192(%r11)
je 7f // end
vmaskmovps %xmm7, %xmm15, 224(%r11)
//
jmp 7f
0:
// offset > 0
// 1 2 3 4 5 6 7
movq %r11, %rbx // D0
addq %r12, %rbx // D1 <- D0 + 4*sdd*sizeof(double)
cmpl $4, %r10d
jl 1f
jg 2f
// offset==4
// TODO
jmp 7f
1:
// 1 2 3
cmpl $2, %r10d
jl 3f
jg 4f
// offset==2
// TODO
jmp 7f
3:
// offset==1
// TODO
jmp 7f
4:
// offset==3
// TODO
jmp 7f
2:
// 5 6 7
cmpl $6, %r10d
jl 5f
jg 6f
// offset==6
// TODO
jmp 7f
5:
// offset==5
// TODO
jmp 7f
6:
// offset==7
// TODO
jmp 7f
// end
7:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_store_8x4_gen_lib8, .-inner_store_8x4_gen_lib8
#endif
#endif
// common inner routine with file scope
//
// store lower
//
// input arguments:
// r10 <- D
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
//
// output arguments:
// r10 <- D
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
#if MACRO_LEVEL>=1
.macro INNER_STORE_L_8X4_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_store_l_8x4_lib8, @function
inner_store_l_8x4_lib8:
#elif defined(OS_MAC)
_inner_store_l_8x4_lib8:
#elif defined(OS_WINDOWS)
.def inner_store_l_8x4_lib8; .scl 2; .type 32; .endef
inner_store_l_8x4_lib8:
#endif
#endif
vmovaps 32(%r10), %ymm12
vmovaps 64(%r10), %ymm13
vmovaps 96(%r10), %ymm14
vblendps $0x1, %ymm12, %ymm1, %ymm1
vblendps $0x3, %ymm13, %ymm2, %ymm2
vblendps $0x7, %ymm14, %ymm3, %ymm3
vmovaps %ymm0, 0(%r10)
vmovaps %ymm1, 32(%r10)
vmovaps %ymm2, 64(%r10)
vmovaps %ymm3, 96(%r10)
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_store_l_8x4_lib8, .-inner_store_l_8x4_lib8
#endif
#endif
// common inner routine with file scope
//
// store n vs
//
// input arguments:
// r10 <- D
// r11 <- km
// r12 <- kn
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
//
// output arguments:
// r10 <- D
// r11 <- km
// r12 <- kn
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
#if MACRO_LEVEL>=1
.macro INNER_STORE_L_8X4_VS_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_store_l_8x4_vs_lib8, @function
inner_store_l_8x4_vs_lib8:
#elif defined(OS_MAC)
_inner_store_l_8x4_vs_lib8:
#elif defined(OS_WINDOWS)
.def inner_store_l_8x4_vs_lib8; .scl 2; .type 32; .endef
inner_store_l_8x4_vs_lib8:
#endif
#endif
// compute mask for rows
vcvtsi2ss %r11d, %xmm15, %xmm15
#if defined(OS_LINUX) | defined(OS_WINDOWS)
vmovups .LC00(%rip), %ymm12
#elif defined(OS_MAC)
vmovups LC00(%rip), %ymm12
#endif
vshufps $0x00, %xmm15, %xmm15, %xmm15
vinsertf128 $0x1, %xmm15, %ymm15, %ymm15
vsubps %ymm15, %ymm12, %ymm15
vmaskmovps %ymm0, %ymm15, 0(%r10)
cmpl $2, %r12d
jl 0f // end
vmovaps 32(%r10), %ymm12
vblendps $0x1, %ymm12, %ymm1, %ymm1
vmaskmovps %ymm1, %ymm15, 32(%r10)
cmpl $3, %r12d
jl 0f // end
vmovaps 64(%r10), %ymm12
vblendps $0x3, %ymm12, %ymm2, %ymm2
vmaskmovps %ymm2, %ymm15, 64(%r10)
je 0f // end
vmovaps 96(%r10), %ymm12
vblendps $0x7, %ymm12, %ymm3, %ymm3
vmaskmovps %ymm3, %ymm15, 96(%r10)
//
0:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_store_l_8x4_vs_lib8, .-inner_store_l_8x4_vs_lib8
#endif
#endif
// common inner routine with file scope
//
// store n generalized
//
// input arguments:
// r10 <- offset
// r11 <- D
// r12 <- 4*sdd*sizeof(double)
// r13 <- m0 // row index: start from (inc)
// r14 <- m1 // row index: up to (exc)
// r15 <- n0 // col index: start from (inc)
// rax <- n1 // col index: up to (exc)
// rbx <- dirty
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
//
// output arguments:
// r10 <- offset
// r11 <- D
// r12 <- 4*sdd*sizeof(double)
// r13 <- m0 // row index: start from (inc)
// r14 <- m1 // row index: up to (exc)
// r15 <- n1-n0
// rax <- n1-n0
// rbx <- dirty
// ymm0 <- []
// ymm1 <- []
// ymm2 <- []
// ymm3 <- []
#if MACRO_LEVEL>=1
.macro INNER_STORE_L_8X4_GEN_LIB8
#else
.p2align 4,,15
#if defined(OS_LINUX)
.type inner_store_l_8x4_gen_lib8, @function
inner_store_l_8x4_gen_lib8:
#elif defined(OS_MAC)
_inner_store_l_8x4_gen_lib8:
#elif defined(OS_WINDOWS)
.def inner_store_l_8x4_gen_lib8; .scl 2; .type 32; .endef
inner_store_l_8x4_gen_lib8:
#endif
#endif
// compute mask for rows
vcvtsi2ss %r13d, %xmm14, %xmm14
vcvtsi2ss %r14d, %xmm15, %xmm15
#if defined(OS_LINUX) | defined(OS_WINDOWS)
vmovups .LC00(%rip), %ymm12
#elif defined(OS_MAC)
vmovups LC00(%rip), %ymm12
#endif
vshufps $0x00, %xmm14, %xmm14, %xmm14
vshufps $0x00, %xmm15, %xmm15, %xmm15
vinsertf128 $0x1, %xmm14, %ymm14, %ymm14
vinsertf128 $0x1, %xmm15, %ymm15, %ymm15
vsubps %ymm12, %ymm14, %ymm14
vsubps %ymm15, %ymm12, %ymm15
vandps %ymm14, %ymm15, %ymm15
// shift D and sol for cols
cmpl $0, %r15d
jle 0f
vmovaps %ymm1, %ymm0
vmovaps %ymm2, %ymm1
vmovaps %ymm3, %ymm2
addq $32, %r11
cmpl $1, %r15d
jle 0f
vmovaps %ymm1, %ymm0
vmovaps %ymm2, %ymm1
addq $32, %r11
cmpl $2, %r15d
jle 0f
vmovaps %ymm1, %ymm0
addq $32, %r11
0:
// compute number of cols
cmpl $4, %eax
jle 0f
movl $4, %eax
0:
subl %r15d, %eax
movl %eax, %r15d
cmpl $0, %r10d
jg 0f
// offset==0
vmaskmovps %ymm0, %ymm15, 0(%r11)
cmpl $2, %r15d
jl 7f // end
vmovaps 32(%r11), %ymm12
vblendps $0x1, %ymm12, %ymm1, %ymm1
vmaskmovps %ymm1, %ymm15, 32(%r11)
cmpl $3, %r15d
jl 7f // end
vmovaps 64(%r11), %ymm12
vblendps $0x3, %ymm12, %ymm2, %ymm2
vmaskmovps %ymm2, %ymm15, 64(%r11)
je 7f // end
vmovaps 96(%r11), %ymm12
vblendps $0x7, %ymm12, %ymm3, %ymm3
vmaskmovps %ymm3, %ymm15, 96(%r11)
//
jmp 7f
0:
// offset > 0
// 1 2 3 4 5 6 7
movq %r11, %rbx // D0
addq %r12, %rbx // D1 <- D0 + 4*sdd*sizeof(double)
cmpl $4, %r10d
jl 1f
jg 2f
// offset==4
// TODO
jmp 7f
1:
// 1 2 3
cmpl $2, %r10d
jl 3f
jg 4f
// offset==2
// TODO
jmp 7f
3:
// offset==1
// TODO
jmp 7f
4:
// offset==3
// TODO
jmp 7f
2:
// 5 6 7
cmpl $6, %r10d
jl 5f
jg 6f
// offset==6
// TODO
jmp 7f
5:
// offset==5
// TODO
jmp 7f
6:
// offset==7
// TODO
jmp 7f
// end
7:
#if MACRO_LEVEL>=1
.endm
#else
ret
#if defined(OS_LINUX)
.size inner_store_l_8x4_gen_lib8, .-inner_store_l_8x4_gen_lib8
#endif
#endif
// rdi rsi rdx rcx r8 r9 rsp+8
// void kernel_sgemm_nt_8x4_lib8(int k, float *alpha, float *A, float *B, float *beta, float *C, float *D);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_sgemm_nt_8x4_lib8
.type kernel_sgemm_nt_8x4_lib8, @function
kernel_sgemm_nt_8x4_lib8:
#elif defined(OS_MAC)
.globl _kernel_sgemm_nt_8x4_lib8
_kernel_sgemm_nt_8x4_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_sgemm_nt_8x4_lib8
.def kernel_sgemm_nt_8x4_lib8; .scl 2; .type 32; .endef
kernel_sgemm_nt_8x4_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorps %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt
movq ARG1, %r10 // k
movq ARG3, %r11 // A
movq ARG4, %r12 // B
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nt_8x4_lib8
#endif
#endif
// call inner scale
movq ARG2, %r10 // alpha
movq ARG5, %r11 // beta
movq ARG6, %r12 // C
#if MACRO_LEVEL>=1
INNER_SCALE_AB_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_ab_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_ab_8x4_lib8
#endif
#endif
// store n
movq ARG7, %r10 // D
#if MACRO_LEVEL>=1
INNER_STORE_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_8x4_lib8
#elif defined(OS_MAC)
callq _inner_store_8x4_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_sgemm_nt_8x4_lib8, .-kernel_sgemm_nt_8x4_lib8
#endif
// rdi rsi rdx rcx r8 r9 rsp+8
// void kernel_sgemm_nt_4x8_lib8(int k, float *alpha, float *A, float *B, float *beta, float *C, float *D);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_sgemm_nt_4x8_lib8
.type kernel_sgemm_nt_4x8_lib8, @function
kernel_sgemm_nt_4x8_lib8:
#elif defined(OS_MAC)
.globl _kernel_sgemm_nt_4x8_lib8
_kernel_sgemm_nt_4x8_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_sgemm_nt_4x8_lib8
.def kernel_sgemm_nt_4x8_lib8; .scl 2; .type 32; .endef
kernel_sgemm_nt_4x8_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorps %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt
movq ARG1, %r10 // k
movq ARG4, %r11 // B
movq ARG3, %r12 // A
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nt_8x4_lib8
#endif
#endif
// call inner scale
movq ARG2, %r10 // alpha
movq ARG5, %r11 // beta
movq ARG6, %r12 // C
#if MACRO_LEVEL>=1
INNER_TRAN_SCALE_AB_4X8_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_tran_scale_ab_4x8_lib8
#elif defined(OS_MAC)
callq _inner_tran_scale_ab_4x8_lib8
#endif
#endif
// store n
movq ARG7, %r10 // D
#if MACRO_LEVEL>=1
INNER_STORE_4X8_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_4x8_lib8
#elif defined(OS_MAC)
callq _inner_store_4x8_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_sgemm_nt_4x8_lib8, .-kernel_sgemm_nt_4x8_lib8
#endif
// rdi rsi rdx rcx r8 r9 rsp+8
// void kernel_sgemm_nt_8x4_vs_lib8(int k, float *alpha, float *A, float *B, float *beta, float *C, float *D, int km, int kn);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_sgemm_nt_8x4_vs_lib8
.type kernel_sgemm_nt_8x4_vs_lib8, @function
kernel_sgemm_nt_8x4_vs_lib8:
#elif defined(OS_MAC)
.globl _kernel_sgemm_nt_8x4_vs_lib8
_kernel_sgemm_nt_8x4_vs_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_sgemm_nt_8x4_vs_lib8
.def kernel_sgemm_nt_8x4_vs_lib8; .scl 2; .type 32; .endef
kernel_sgemm_nt_8x4_vs_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorps %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt
movq ARG1, %r10 // k
movq ARG3, %r11 // A
movq ARG4, %r12 // B
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nt_8x4_lib8
#endif
#endif
// call inner scale
movq ARG2, %r10 // alpha
movq ARG5, %r11 // beta
movq ARG6, %r12 // C
#if MACRO_LEVEL>=1
INNER_SCALE_AB_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_ab_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_ab_8x4_lib8
#endif
#endif
// store n
movq ARG7, %r10 // D
movq ARG8, %r11 // km
movq ARG9, %r12 // kn
#if MACRO_LEVEL>=1
INNER_STORE_8X4_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_8x4_vs_lib8
#elif defined(OS_MAC)
callq _inner_store_8x4_vs_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_sgemm_nt_8x4_vs_lib8, .-kernel_sgemm_nt_8x4_vs_lib8
#endif
// rdi rsi rdx rcx r8 r9 rsp+8
// void kernel_sgemm_nt_4x8_vs_lib8(int k, float *alpha, float *A, float *B, float *beta, float *C, float *D, int km, int kn);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_sgemm_nt_4x8_vs_lib8
.type kernel_sgemm_nt_4x8_vs_lib8, @function
kernel_sgemm_nt_4x8_vs_lib8:
#elif defined(OS_MAC)
.globl _kernel_sgemm_nt_4x8_vs_lib8
_kernel_sgemm_nt_4x8_vs_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_sgemm_nt_4x8_vs_lib8
.def kernel_sgemm_nt_4x8_vs_lib8; .scl 2; .type 32; .endef
kernel_sgemm_nt_4x8_vs_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorps %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt
movq ARG1, %r10 // k
movq ARG4, %r11 // B
movq ARG3, %r12 // A
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nt_8x4_lib8
#endif
#endif
// call inner scale
movq ARG2, %r10 // alpha
movq ARG5, %r11 // beta
movq ARG6, %r12 // C
#if MACRO_LEVEL>=1
INNER_TRAN_SCALE_AB_4X8_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_tran_scale_ab_4x8_lib8
#elif defined(OS_MAC)
callq _inner_tran_scale_ab_4x8_lib8
#endif
#endif
// store n
movq ARG7, %r10 // D
movq ARG8, %r11 // km
movq ARG9, %r12 // kn
#if MACRO_LEVEL>=1
INNER_STORE_4X8_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_4x8_vs_lib8
#elif defined(OS_MAC)
callq _inner_store_4x8_vs_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_sgemm_nt_4x8_vs_lib8, .-kernel_sgemm_nt_4x8_vs_lib8
#endif
// rdi rsi rdx rcx r8 r9 rsp+8 rsp+16 rsp+24 rsp+32 rsp+40 rsp+48 rsp+56 rsp+64 rsp+72
// void kernel_sgemm_nt_8x4_gen_lib8(int k, float *alpha, float *A, float *B, float *beta, int offsetC, float *C, int sdc, int offsetD, float *D, int sdd, int m0, int m1, int n0, int n1);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_sgemm_nt_8x4_gen_lib8
.type kernel_sgemm_nt_8x4_gen_lib8, @function
kernel_sgemm_nt_8x4_gen_lib8:
#elif defined(OS_MAC)
.globl _kernel_sgemm_nt_8x4_gen_lib8
_kernel_sgemm_nt_8x4_gen_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_sgemm_nt_8x4_gen_lib8
.def kernel_sgemm_nt_8x4_gen_lib8; .scl 2; .type 32; .endef
kernel_sgemm_nt_8x4_gen_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorps %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt
movq ARG1, %r10 // k
movq ARG3, %r11 // A
movq ARG4, %r12 // B
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nt_8x4_lib8
#endif
#endif
// call inner blend scale
movq ARG2, %r10 // alpha
movq ARG5, %r11 // beta
movq ARG6, %r12 // offsetC
movq ARG7, %r13 // C
movq ARG8, %r14 // sdc
sall $5, %r14d // 8*sdc*sizeof(float)
#if MACRO_LEVEL>=1
INNER_SCALE_AB_8X4_GEN_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_ab_8x4_gen_lib8
#elif defined(OS_MAC)
callq _inner_scale_ab_8x4_gen_lib8
#endif
#endif
// store n gen
movq ARG9, %r10 // offsetD
movq ARG10, %r11 // D
movq ARG11, %r12 // sdd
sall $5, %r12d // 8*sdb*sizeof(float)
movq ARG12, %r13 // m0
movq ARG13, %r14 // m1
movq ARG14, %r15 // n0
movq ARG15, %rax // n1
#if MACRO_LEVEL>=1
INNER_STORE_8X4_GEN_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_8x4_gen_lib8
#elif defined(OS_MAC)
callq _inner_store_8x4_gen_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_sgemm_nt_8x4_gen_lib8, .-kernel_sgemm_nt_8x4_gen_lib8
#endif
// rdi rsi rdx rcx r8 r9 rsp+8 rsp+16 rsp+24 rsp+32 rsp+40 rsp+48 rsp+56 rsp+64 rsp+72
// void kernel_sgemm_nt_4x8_gen_lib8(int k, float *alpha, float *A, float *B, float *beta, int offsetC, float *C, int sdc, int offsetD, float *D, int sdd, int m0, int m1, int n0, int n1);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_sgemm_nt_4x8_gen_lib8
.type kernel_sgemm_nt_4x8_gen_lib8, @function
kernel_sgemm_nt_4x8_gen_lib8:
#elif defined(OS_MAC)
.globl _kernel_sgemm_nt_4x8_gen_lib8
_kernel_sgemm_nt_4x8_gen_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_sgemm_nt_4x8_gen_lib8
.def kernel_sgemm_nt_4x8_gen_lib8; .scl 2; .type 32; .endef
kernel_sgemm_nt_4x8_gen_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorps %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt
movq ARG1, %r10 // k
movq ARG4, %r11 // A
movq ARG3, %r12 // B
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nt_8x4_lib8
#endif
#endif
// call inner blend scale
movq ARG2, %r10 // alpha
movq ARG5, %r11 // beta
movq ARG6, %r12 // offsetC
movq ARG7, %r13 // C
movq ARG8, %r14 // sdc
sall $5, %r14d // 8*sdc*sizeof(float)
#if MACRO_LEVEL>=1
INNER_TRAN_SCALE_AB_4X8_GEN_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_tran_scale_ab_4x8_gen_lib8
#elif defined(OS_MAC)
callq _inner_tran_scale_ab_4x8_gen_lib8
#endif
#endif
// store n gen
movq ARG9, %r10 // offsetD
movq ARG10, %r11 // D
movq ARG11, %r12 // sdd
sall $5, %r12d // 8*sdb*sizeof(float)
movq ARG12, %r13 // m0
movq ARG13, %r14 // m1
movq ARG14, %r15 // n0
movq ARG15, %rax // n1
#if MACRO_LEVEL>=1
INNER_STORE_4X8_GEN_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_4x8_gen_lib8
#elif defined(OS_MAC)
callq _inner_store_4x8_gen_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_sgemm_nt_4x8_gen_lib8, .-kernel_sgemm_nt_4x8_gen_lib8
#endif
// 0 1 2 3 4 5 6 7 8
// void kernel_sgemm_nn_8x4_lib8(int k, float *alpha, float *A, int offsetB, float *B, int sdb, float *beta, float *C, float *D);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_sgemm_nn_8x4_lib8
.type kernel_sgemm_nn_8x4_lib8, @function
kernel_sgemm_nn_8x4_lib8:
#elif defined(OS_MAC)
.globl _kernel_sgemm_nn_8x4_lib8
_kernel_sgemm_nn_8x4_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_sgemm_nn_8x4_lib8
.def kernel_sgemm_nn_8x4_lib8; .scl 2; .type 32; .endef
kernel_sgemm_nn_8x4_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorps %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nn
movq ARG1, %r10 // k
movq ARG3, %r11 // A
movq ARG5, %r12 // B
movq ARG6, %r13 // sdb
sall $5, %r13d // 4*sdb*sizeof(double)
movq ARG4, %r14 // offsetB
#if MACRO_LEVEL>=1
INNER_EDGE_GEMM_ADD_NN_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_gemm_add_nn_8x4_lib8
#elif defined(OS_MAC)
callq _inner_edge_gemm_add_nn_8x4_lib8
#endif
#endif
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NN_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nn_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nn_8x4_lib8
#endif
#endif
// call inner blend
movq ARG2, %r10 // alpha
movq ARG7, %r11 // beta
movq ARG8, %r12 // C
#if MACRO_LEVEL>=1
INNER_SCALE_AB_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_ab_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_ab_8x4_lib8
#endif
#endif
// store n
movq ARG9, %r10 // D
#if MACRO_LEVEL>=1
INNER_STORE_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_8x4_lib8
#elif defined(OS_MAC)
callq _inner_store_8x4_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_sgemm_nn_8x4_lib8, .-kernel_sgemm_nn_8x4_lib8
#endif
// 1 2 3 4 5 6 7 8 9 10 11
// void kernel_sgemm_nn_8x4_lib8(int k, float *alpha, float *A, int offsetB, float *B, int sdb, float *beta, float *C, float *D, int km, int kn);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_sgemm_nn_8x4_vs_lib8
.type kernel_sgemm_nn_8x4_vs_lib8, @function
kernel_sgemm_nn_8x4_vs_lib8:
#elif defined(OS_MAC)
.globl _kernel_sgemm_nn_8x4_vs_lib8
_kernel_sgemm_nn_8x4_vs_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_sgemm_nn_8x4_vs_lib8
.def kernel_sgemm_nn_8x4_vs_lib8; .scl 2; .type 32; .endef
kernel_sgemm_nn_8x4_vs_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorps %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nn
movq ARG1, %r10 // k
movq ARG3, %r11 // A
movq ARG5, %r12 // B
movq ARG6, %r13 // sdb
sall $5, %r13d // 4*sdb*sizeof(double)
movq ARG4, %r14 // offsetB
#if MACRO_LEVEL>=1
INNER_EDGE_GEMM_ADD_NN_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_gemm_add_nn_8x4_lib8
#elif defined(OS_MAC)
callq _inner_edge_gemm_add_nn_8x4_lib8
#endif
#endif
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NN_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nn_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nn_8x4_lib8
#endif
#endif
// call inner blend
movq ARG2, %r10 // alpha
movq ARG7, %r11 // beta
movq ARG8, %r12 // C
#if MACRO_LEVEL>=1
INNER_SCALE_AB_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_ab_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_ab_8x4_lib8
#endif
#endif
// store n
movq ARG9, %r10 // D
movq ARG10, %r11 // D
movq ARG11, %r12 // D
#if MACRO_LEVEL>=1
INNER_STORE_8X4_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_8x4_vs_lib8
#elif defined(OS_MAC)
callq _inner_store_8x4_vs_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_sgemm_nn_8x4_vs_lib8, .-kernel_sgemm_nn_8x4_vs_lib8
#endif
// rdi rsi rdx rcx r8 r9 rsp+8 rsp+16 rsp+24 rsp+32 rsp+40 rsp+48 rsp+56 rsp+64 rsp+72 rsp+80 rsp+88
// void kernel_sgemm_nn_8x4_gen_lib8(int k, float *alpha, float *A, int offB, float *B, int sdb, float *beta, int offC, float *C, int sdc, int offD, float *D, int sdd, int m0, int m1, int n0, int n1);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_sgemm_nn_8x4_gen_lib8
.type kernel_sgemm_nn_8x4_gen_lib8, @function
kernel_sgemm_nn_8x4_gen_lib8:
#elif defined(OS_MAC)
.globl _kernel_sgemm_nn_8x4_gen_lib8
_kernel_sgemm_nn_8x4_gen_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_sgemm_nn_8x4_gen_lib8
.def kernel_sgemm_nn_8x4_gen_lib8; .scl 2; .type 32; .endef
kernel_sgemm_nn_8x4_gen_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorps %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nn
movq ARG1, %r10 // k
movq ARG3, %r11 // A
movq ARG5, %r12 // B
movq ARG6, %r13 // sdb
sall $5, %r13d // 4*sdb*sizeof(double)
movq ARG4, %r14 // offsetB
#if MACRO_LEVEL>=1
INNER_EDGE_GEMM_ADD_NN_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_gemm_add_nn_8x4_lib8
#elif defined(OS_MAC)
callq _inner_edge_gemm_add_nn_8x4_lib8
#endif
#endif
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NN_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nn_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nn_8x4_lib8
#endif
#endif
// call inner blend scale
movq ARG2, %r10 // alpha
movq ARG7, %r11 // beta
movq ARG8, %r12 // offsetC
movq ARG9, %r13 // C
movq ARG10, %r14 // sdc
sall $5, %r14d // 4*sdc*sizeof(double)
#if MACRO_LEVEL>=1
INNER_SCALE_AB_8X4_GEN_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_ab_8x4_gen_lib8
#elif defined(OS_MAC)
callq _inner_scale_ab_8x4_gen_lib8
#endif
#endif
// store n gen
movq ARG11, %r10 // offsetD
movq ARG12, %r11 // D
movq ARG13, %r12 // sdd
sall $5, %r12d // 4*sdb*sizeof(double)
movq ARG14, %r13 // m0
movq ARG15, %r14 // m1
movq ARG16, %r15 // n0
movq ARG17, %rax // n1
#if MACRO_LEVEL>=1
INNER_STORE_8X4_GEN_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_8x4_gen_lib8
#elif defined(OS_MAC)
callq _inner_store_8x4_gen_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_sgemm_nn_8x4_gen_lib8, .-kernel_sgemm_nn_8x4_gen_lib8
#endif
// 1 2 3 4 5 6 7
// void kernel_ssyrk_nt_l_8x4_lib8(int k, float *alpha, float *A, float *B, float *beta, float *C, float *D);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_ssyrk_nt_l_8x4_lib8
.type kernel_ssyrk_nt_l_8x4_lib8, @function
kernel_ssyrk_nt_l_8x4_lib8:
#elif defined(OS_MAC)
.globl _kernel_ssyrk_nt_l_8x4_lib8
_kernel_ssyrk_nt_l_8x4_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_ssyrk_nt_l_8x4_lib8
.def kernel_ssyrk_nt_l_8x4_lib8; .scl 2; .type 32; .endef
kernel_ssyrk_nt_l_8x4_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorps %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt
movq ARG1, %r10 // k
movq ARG3, %r11 // A
movq ARG4, %r12 // B
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nt_8x4_lib8
#endif
#endif
// call inner scale
movq ARG2, %r10 // alpha
movq ARG5, %r11 // beta
movq ARG6, %r12 // C
#if MACRO_LEVEL>=1
INNER_SCALE_AB_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_ab_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_ab_8x4_lib8
#endif
#endif
// store n
movq ARG7, %r10 // D
#if MACRO_LEVEL>=1
INNER_STORE_L_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_l_8x4_lib8
#elif defined(OS_MAC)
callq _inner_store_l_8x4_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_ssyrk_nt_l_8x4_lib8, .-kernel_ssyrk_nt_l_8x4_lib8
#endif
// 1 2 3 4 5 6 7 8 9
// void kernel_ssyrk_nt_l_8x4_vs_lib8(int k, float *alpha, float *A, float *B, float *beta, float *C, float *D, int km, int kn);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_ssyrk_nt_l_8x4_vs_lib8
.type kernel_ssyrk_nt_l_8x4_vs_lib8, @function
kernel_ssyrk_nt_l_8x4_vs_lib8:
#elif defined(OS_MAC)
.globl _kernel_ssyrk_nt_l_8x4_vs_lib8
_kernel_ssyrk_nt_l_8x4_vs_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_ssyrk_nt_l_8x4_vs_lib8
.def kernel_ssyrk_nt_l_8x4_vs_lib8; .scl 2; .type 32; .endef
kernel_ssyrk_nt_l_8x4_vs_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorps %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt
movq ARG1, %r10 // k
movq ARG3, %r11 // A
movq ARG4, %r12 // B
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nt_8x4_lib8
#endif
#endif
// call inner scale
movq ARG2, %r10 // alpha
movq ARG5, %r11 // beta
movq ARG6, %r12 // C
#if MACRO_LEVEL>=1
INNER_SCALE_AB_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_ab_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_ab_8x4_lib8
#endif
#endif
// store n
movq ARG7, %r10 // D
movq ARG8, %r11 // km
movq ARG9, %r12 // kn
#if MACRO_LEVEL>=1
INNER_STORE_L_8X4_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_l_8x4_vs_lib8
#elif defined(OS_MAC)
callq _inner_store_l_8x4_vs_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_ssyrk_nt_l_8x4_vs_lib8, .-kernel_ssyrk_nt_l_8x4_vs_lib8
#endif
// edi rsi rdx ecx r8 r9 rsp+8
// void kernel_strsm_nt_rl_inv_8x4_lib8(int k, float *A, float *B, float *C, float *D, float *E, float *inv_diag_E);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_strsm_nt_rl_inv_8x4_lib8
.type kernel_strsm_nt_rl_inv_8x4_lib8, @function
kernel_strsm_nt_rl_inv_8x4_lib8:
#elif defined(OS_MAC)
.globl _kernel_strsm_nt_rl_inv_8x4_lib8
_kernel_strsm_nt_rl_inv_8x4_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_strsm_nt_rl_inv_8x4_lib8
.def kernel_strsm_nt_rl_inv_8x4_lib8; .scl 2; .type 32; .endef
kernel_strsm_nt_rl_inv_8x4_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorpd %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt
movq ARG1, %r10
movq ARG2, %r11
movq ARG3, %r12
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_SUB_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_sub_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_sub_nt_8x4_lib8
#endif
#endif
// call inner blender_loader nn
movq ARG4, %r10
#if MACRO_LEVEL>=1
INNER_SCALE_11_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_11_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_11_8x4_lib8
#endif
#endif
// solve
movq ARG6, %r10 // E
movq ARG7, %r11 // inv_diag_E
#if MACRO_LEVEL>=1
INNER_EDGE_TRSM_RLT_INV_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_trsm_rlt_inv_8x4_lib8
#elif defined(OS_MAC)
callq _inner_edge_trsm_rlt_inv_8x4_lib8
#endif
#endif
// store
movq ARG5, %r10 // D
#if MACRO_LEVEL>=1
INNER_STORE_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_8x4_lib8
#elif defined(OS_MAC)
callq _inner_store_8x4_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_strsm_nt_rl_inv_8x4_lib8, .-kernel_strsm_nt_rl_inv_8x4_lib8
#endif
// edi rsi rdx ecx r8 r9 rsp+8
// void kernel_strsm_nt_rl_inv_4x8_lib8(int k, float *A, float *B, float *C, float *D, float *E, float *inv_diag_E);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_strsm_nt_rl_inv_4x8_lib8
.type kernel_strsm_nt_rl_inv_4x8_lib8, @function
kernel_strsm_nt_rl_inv_4x8_lib8:
#elif defined(OS_MAC)
.globl _kernel_strsm_nt_rl_inv_4x8_lib8
_kernel_strsm_nt_rl_inv_4x8_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_strsm_nt_rl_inv_4x8_lib8
.def kernel_strsm_nt_rl_inv_4x8_lib8; .scl 2; .type 32; .endef
kernel_strsm_nt_rl_inv_4x8_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorpd %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt
movq ARG1, %r10
movq ARG3, %r11
movq ARG2, %r12
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_SUB_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_sub_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_sub_nt_8x4_lib8
#endif
#endif
// call inner blender_loader nn
movq ARG4, %r10
#if MACRO_LEVEL>=1
INNER_TRAN_SCALE_11_4X8_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_tran_scale_11_4x8_lib8
#elif defined(OS_MAC)
callq _inner_tran_scale_11_4x8_lib8
#endif
#endif
// solve
movq ARG6, %r10 // E
movq ARG7, %r11 // inv_diag_E
movq $8, %r12 // n1
#if MACRO_LEVEL>=1
INNER_EDGE_TRSM_RLT_INV_4X8_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_trsm_rlt_inv_4x8_vs_lib8
#elif defined(OS_MAC)
callq _inner_edge_trsm_rlt_inv_4x8_vs_lib8
#endif
#endif
// store
movq ARG5, %r10 // D
#if MACRO_LEVEL>=1
INNER_STORE_4X8_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_4x8_lib8
#elif defined(OS_MAC)
callq _inner_store_4x8_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_strsm_nt_rl_inv_4x8_lib8, .-kernel_strsm_nt_rl_inv_4x8_lib8
#endif
// edi rsi rdx ecx r8 r9 rsp+8 rsp+16 rsp+24
// void kernel_strsm_nt_rl_inv_8x4_vs_lib8(int k, float *A, float *B, float *C, float *D, float *E, float *inv_diag_E, int km, int kn);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_strsm_nt_rl_inv_8x4_vs_lib8
.type kernel_strsm_nt_rl_inv_8x4_vs_lib8, @function
kernel_strsm_nt_rl_inv_8x4_vs_lib8:
#elif defined(OS_MAC)
.globl _kernel_strsm_nt_rl_inv_8x4_vs_lib8
_kernel_strsm_nt_rl_inv_8x4_vs_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_strsm_nt_rl_inv_8x4_vs_lib8
.def kernel_strsm_nt_rl_inv_8x4_vs_lib8; .scl 2; .type 32; .endef
kernel_strsm_nt_rl_inv_8x4_vs_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorpd %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt
movq ARG1, %r10
movq ARG2, %r11
movq ARG3, %r12
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_SUB_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_sub_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_sub_nt_8x4_lib8
#endif
#endif
// call inner blender_loader nn // TODO scale gen
movq ARG4, %r10 // C
#if MACRO_LEVEL>=1
INNER_SCALE_11_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_11_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_11_8x4_lib8
#endif
#endif
// solve
movq ARG6, %r10 // E
movq ARG7, %r11 // inv_diag_E
movq ARG9, %r12 // kn
#if MACRO_LEVEL>=1
INNER_EDGE_TRSM_RLT_INV_8X4_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_trsm_rlt_inv_8x4_vs_lib8
#elif defined(OS_MAC)
callq _inner_edge_trsm_rlt_inv_8x4_vs_lib8
#endif
#endif
// store
movq ARG5, %r10 // D
movq ARG8, %r11 // km
movq ARG9, %r12 // kn
#if MACRO_LEVEL>=1
INNER_STORE_8X4_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_8x4_vs_lib8
#elif defined(OS_MAC)
callq _inner_store_8x4_vs_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_strsm_nt_rl_inv_8x4_vs_lib8, .-kernel_strsm_nt_rl_inv_8x4_vs_lib8
#endif
// edi rsi rdx ecx r8 r9 rsp+8 rsp+16 rsp+24
// void kernel_strsm_nt_rl_inv_4x8_vs_lib8(int k, float *A, float *B, float *C, float *D, float *E, float *inv_diag_E, int km, int kn);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_strsm_nt_rl_inv_4x8_vs_lib8
.type kernel_strsm_nt_rl_inv_4x8_vs_lib8, @function
kernel_strsm_nt_rl_inv_4x8_vs_lib8:
#elif defined(OS_MAC)
.globl _kernel_strsm_nt_rl_inv_4x8_vs_lib8
_kernel_strsm_nt_rl_inv_4x8_vs_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_strsm_nt_rl_inv_4x8_vs_lib8
.def kernel_strsm_nt_rl_inv_4x8_vs_lib8; .scl 2; .type 32; .endef
kernel_strsm_nt_rl_inv_4x8_vs_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorpd %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt
movq ARG1, %r10
movq ARG3, %r11
movq ARG2, %r12
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_SUB_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_sub_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_sub_nt_8x4_lib8
#endif
#endif
// call inner blender_loader nn // TODO scale gen
movq ARG4, %r10 // C
#if MACRO_LEVEL>=1
INNER_TRAN_SCALE_11_4X8_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_tran_scale_11_4x8_lib8
#elif defined(OS_MAC)
callq _inner_tran_scale_11_4x8_lib8
#endif
#endif
// solve
movq ARG6, %r10 // E
movq ARG7, %r11 // inv_diag_E
movq ARG9, %r12 // kn
#if MACRO_LEVEL>=1
INNER_EDGE_TRSM_RLT_INV_4X8_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_trsm_rlt_inv_4x8_vs_lib8
#elif defined(OS_MAC)
callq _inner_edge_trsm_rlt_inv_4x8_vs_lib8
#endif
#endif
// store
movq ARG5, %r10 // D
movq ARG8, %r11 // km
movq ARG9, %r12 // kn
#if MACRO_LEVEL>=1
INNER_STORE_4X8_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_4x8_vs_lib8
#elif defined(OS_MAC)
callq _inner_store_4x8_vs_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_strsm_nt_rl_inv_4x8_vs_lib8, .-kernel_strsm_nt_rl_inv_4x8_vs_lib8
#endif
// 1 2 3 4 5 6 7 8 9 10
// void kernel_sgemm_strsm_nt_rl_inv_8x4_lib8(int kp, float *Ap, float *Bp, int km, float *Am, float *Bm, float *C, float *D, float *E, float *inv_diag_E);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_sgemm_strsm_nt_rl_inv_8x4_lib8
.type kernel_sgemm_strsm_nt_rl_inv_8x4_lib8, @function
kernel_sgemm_strsm_nt_rl_inv_8x4_lib8:
#elif defined(OS_MAC)
.globl _kernel_sgemm_strsm_nt_rl_inv_8x4_lib8
_kernel_sgemm_strsm_nt_rl_inv_8x4_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_sgemm_strsm_nt_rl_inv_8x4_lib8
.def kernel_sgemm_strsm_nt_rl_inv_8x4_lib8; .scl 2; .type 32; .endef
kernel_sgemm_strsm_nt_rl_inv_8x4_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorpd %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt add
movq ARG1, %r10 // kp
movq ARG2, %r11 // Ap
movq ARG3, %r12 // Bp
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nt_8x4_lib8
#endif
#endif
// call inner dgemm kernel nt sub
movq ARG4, %r10 // km
movq ARG5, %r11 // Am
movq ARG6, %r12 // Bm
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_SUB_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_sub_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_sub_nt_8x4_lib8
#endif
#endif
// call inner blender_loader nn
movq ARG7, %r10 // C
#if MACRO_LEVEL>=1
INNER_SCALE_11_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_11_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_11_8x4_lib8
#endif
#endif
// solve
movq ARG9, %r10 // E
movq ARG10, %r11 // inv_diag_E
#if MACRO_LEVEL>=1
INNER_EDGE_TRSM_RLT_INV_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_trsm_rlt_inv_8x4_lib8
#elif defined(OS_MAC)
callq _inner_edge_trsm_rlt_inv_8x4_lib8
#endif
#endif
// store
movq ARG8, %r10 // D
#if MACRO_LEVEL>=1
INNER_STORE_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_8x4_lib8
#elif defined(OS_MAC)
callq _inner_store_8x4_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_sgemm_strsm_nt_rl_inv_8x4_lib8, .-kernel_sgemm_strsm_nt_rl_inv_8x4_lib8
#endif
// 1 2 3 4 5 6 7 8 9 10 11 12
// void kernel_sgemm_strsm_nt_rl_inv_8x4_vs_lib8(int kp, float *Ap, float *Bp, int km, float *Am, float *Bm, float *C, float *D, float *E, float *inv_diag_E, int km, int kn);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_sgemm_strsm_nt_rl_inv_8x4_vs_lib8
.type kernel_sgemm_strsm_nt_rl_inv_8x4_vs_lib8, @function
kernel_sgemm_strsm_nt_rl_inv_8x4_vs_lib8:
#elif defined(OS_MAC)
.globl _kernel_sgemm_strsm_nt_rl_inv_8x4_vs_lib8
_kernel_sgemm_strsm_nt_rl_inv_8x4_vs_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_sgemm_strsm_nt_rl_inv_8x4_vs_lib8
.def kernel_sgemm_strsm_nt_rl_inv_8x4_vs_lib8; .scl 2; .type 32; .endef
kernel_sgemm_strsm_nt_rl_inv_8x4_vs_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorpd %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt add
movq ARG1, %r10 // kp
movq ARG2, %r11 // Ap
movq ARG3, %r12 // Bp
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nt_8x4_lib8
#endif
#endif
// call inner dgemm kernel nt sub
movq ARG4, %r10 // km
movq ARG5, %r11 // Am
movq ARG6, %r12 // Bm
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_SUB_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_sub_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_sub_nt_8x4_lib8
#endif
#endif
// call inner blender_loader nn
movq ARG7, %r10 // C
#if MACRO_LEVEL>=1
INNER_SCALE_11_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_11_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_11_8x4_lib8
#endif
#endif
// solve
movq ARG9, %r10 // E
movq ARG10, %r11 // inv_diag_E
movq ARG12, %r12 // kn
#if MACRO_LEVEL>=1
INNER_EDGE_TRSM_RLT_INV_8X4_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_trsm_rlt_inv_8x4_vs_lib8
#elif defined(OS_MAC)
callq _inner_edge_trsm_rlt_inv_8x4_vs_lib8
#endif
#endif
// store
movq ARG8, %r10 // D
movq ARG11, %r11 // km
movq ARG12, %r12 // kn
#if MACRO_LEVEL>=1
INNER_STORE_8X4_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_8x4_vs_lib8
#elif defined(OS_MAC)
callq _inner_store_8x4_vs_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_sgemm_strsm_nt_rl_inv_8x4_vs_lib8, .-kernel_sgemm_strsm_nt_rl_inv_8x4_vs_lib8
#endif
// 1 2 3 4 5 6
// void kernel_spotrf_nt_l_8x4_lib8(int k, float *A, float *B, float *C, float *D, float *inv_diag_D);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_spotrf_nt_l_8x4_lib8
.type kernel_spotrf_nt_l_8x4_lib8, @function
kernel_spotrf_nt_l_8x4_lib8:
#elif defined(OS_MAC)
.globl _kernel_spotrf_nt_l_8x4_lib8
_kernel_spotrf_nt_l_8x4_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_spotrf_nt_l_8x4_lib8
.def kernel_spotrf_nt_l_8x4_lib8; .scl 2; .type 32; .endef
kernel_spotrf_nt_l_8x4_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorpd %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt
movq ARG1, %r10
movq ARG2, %r11
movq ARG3, %r12
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_SUB_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_sub_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_sub_nt_8x4_lib8
#endif
#endif
// call inner blender_loader nn
movq ARG4, %r10 // C
#if MACRO_LEVEL>=1
INNER_SCALE_11_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_11_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_11_8x4_lib8
#endif
#endif
// factorization
movq ARG6, %r10 // inv_diag_D
#if MACRO_LEVEL>=1
INNER_EDGE_POTRF_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_potrf_8x4_lib8
#elif defined(OS_MAC)
callq _inner_edge_potrf_8x4_lib8
#endif
#endif
// store
movq ARG5, %r10 // D
#if MACRO_LEVEL>=1
INNER_STORE_L_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_l_8x4_lib8
#elif defined(OS_MAC)
callq _inner_store_l_8x4_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_spotrf_nt_l_8x4_lib8, .-kernel_spotrf_nt_l_8x4_lib8
#endif
// edi rsi rdx rcx r8 r9 rsp+8 rsp+16
// void kernel_spotrf_nt_l_8x4_vs_lib8(int k, float *A, float *B, float *C, float *D, float *inv_diag_D, int km, int kn);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_spotrf_nt_l_8x4_vs_lib8
.type kernel_spotrf_nt_l_8x4_vs_lib8, @function
kernel_spotrf_nt_l_8x4_vs_lib8:
#elif defined(OS_MAC)
.globl _kernel_spotrf_nt_l_8x4_vs_lib8
_kernel_spotrf_nt_l_8x4_vs_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_spotrf_nt_l_8x4_vs_lib8
.def kernel_spotrf_nt_l_8x4_vs_lib8; .scl 2; .type 32; .endef
kernel_spotrf_nt_l_8x4_vs_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorpd %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt
movq ARG1, %r10
movq ARG2, %r11
movq ARG3, %r12
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_SUB_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_sub_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_sub_nt_8x4_lib8
#endif
#endif
// call inner blender_loader nn
movq ARG4, %r10 // C
#if MACRO_LEVEL>=1
INNER_SCALE_11_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_11_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_11_8x4_lib8
#endif
#endif
// factorization
movq ARG6, %r10 // inv_diag_D
movq ARG8, %r11 // kn
#if MACRO_LEVEL>=1
INNER_EDGE_POTRF_8X4_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_potrf_8x4_vs_lib8
#elif defined(OS_MAC)
callq _inner_edge_potrf_8x4_vs_lib8
#endif
#endif
// store
movq ARG5, %r10 // D
movq ARG7, %r11 // m1
movq ARG8, %r12 // n1
#if MACRO_LEVEL>=1
INNER_STORE_L_8X4_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_l_8x4_vs_lib8
#elif defined(OS_MAC)
callq _inner_store_l_8x4_vs_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_spotrf_nt_l_8x4_vs_lib8, .-kernel_spotrf_nt_l_8x4_vs_lib8
#endif
// 1 2 3 4 5 6 7 8 9
// void kernel_ssyrk_spotrf_nt_l_8x4_lib8(int kp, float *Ap, float *Bp, int km, float *Am, float *Bm, float *C, float *D, float *inv_diag_D);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_ssyrk_spotrf_nt_l_8x4_lib8
.type kernel_ssyrk_spotrf_nt_l_8x4_lib8, @function
kernel_ssyrk_spotrf_nt_l_8x4_lib8:
#elif defined(OS_MAC)
.globl _kernel_ssyrk_spotrf_nt_l_8x4_lib8
_kernel_ssyrk_spotrf_nt_l_8x4_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_ssyrk_spotrf_nt_l_8x4_lib8
.def kernel_ssyrk_spotrf_nt_l_8x4_lib8; .scl 2; .type 32; .endef
kernel_ssyrk_spotrf_nt_l_8x4_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorps %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt add
movq ARG1, %r10 // kp
movq ARG2, %r11 // Ap
movq ARG3, %r12 // Bp
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nt_8x4_lib8
#endif
#endif
// call inner dgemm kernel nt sub
movq ARG4, %r10 // km
movq ARG5, %r11 // Am
movq ARG6, %r12 // Bm
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_SUB_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_sub_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_sub_nt_8x4_lib8
#endif
#endif
// call inner blender_loader nn
movq ARG7, %r10 // C
#if MACRO_LEVEL>=1
INNER_SCALE_11_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_11_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_11_8x4_lib8
#endif
#endif
// factorization
movq ARG9, %r10 // inv_diag_D
movl $4, %r11d
#if MACRO_LEVEL>=1
INNER_EDGE_POTRF_8X4_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_potrf_8x4_vs_lib8
#elif defined(OS_MAC)
callq _inner_edge_potrf_8x4_vs_lib8
#endif
#endif
// store
movq ARG8, %r10 // D
#if MACRO_LEVEL>=1
INNER_STORE_L_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_l_8x4_lib8
#elif defined(OS_MAC)
callq _inner_store_l_8x4_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_ssyrk_spotrf_nt_l_8x4_lib8, .-kernel_ssyrk_spotrf_nt_l_8x4_lib8
#endif
// 1 2 3 4 5 6 7 8 9 10 11
// void kernel_ssyrk_spotrf_nt_l_8x4_vs_lib8(int kp, float *Ap, float *Bp, int km, float *Am, float *Bm, float *C, float *D, float *inv_diag_D, int km, int kn);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_ssyrk_spotrf_nt_l_8x4_vs_lib8
.type kernel_ssyrk_spotrf_nt_l_8x4_vs_lib8, @function
kernel_ssyrk_spotrf_nt_l_8x4_vs_lib8:
#elif defined(OS_MAC)
.globl _kernel_ssyrk_spotrf_nt_l_8x4_vs_lib8
_kernel_ssyrk_spotrf_nt_l_8x4_vs_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_ssyrk_spotrf_nt_l_8x4_vs_lib8
.def kernel_ssyrk_spotrf_nt_l_8x4_vs_lib8; .scl 2; .type 32; .endef
kernel_ssyrk_spotrf_nt_l_8x4_vs_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorpd %ymm0, %ymm0, %ymm0
vmovaps %ymm0, %ymm1
vmovaps %ymm0, %ymm2
vmovaps %ymm0, %ymm3
// call inner dgemm kernel nt add
movq ARG1, %r10 // kp
movq ARG2, %r11 // Ap
movq ARG3, %r12 // Bp
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nt_8x4_lib8
#endif
#endif
// call inner dgemm kernel nt sub
movq ARG4, %r10 // km
movq ARG5, %r11 // Am
movq ARG6, %r12 // Bm
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_SUB_NT_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_sub_nt_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_sub_nt_8x4_lib8
#endif
#endif
// call inner blender_loader nn
movq ARG7, %r10 // C
#if MACRO_LEVEL>=1
INNER_SCALE_11_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_11_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_11_8x4_lib8
#endif
#endif
// factorization
movq ARG9, %r10 // inv_diag_D
movq ARG11, %r11 // kn
#if MACRO_LEVEL>=1
INNER_EDGE_POTRF_8X4_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_potrf_8x4_vs_lib8
#elif defined(OS_MAC)
callq _inner_edge_potrf_8x4_vs_lib8
#endif
#endif
// store
movq ARG8, %r10 // D
movq ARG10, %r11 // km
movq ARG11, %r12 // kn
#if MACRO_LEVEL>=1
INNER_STORE_L_8X4_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_l_8x4_vs_lib8
#elif defined(OS_MAC)
callq _inner_store_l_8x4_vs_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_ssyrk_spotrf_nt_l_8x4_vs_lib8, .-kernel_ssyrk_spotrf_nt_l_8x4_vs_lib8
#endif
// 1 2 3 4 5 6 7
// void kernel_strmm_nn_rl_8x4_lib8(int k, float *alpha, float *A, int offsetB, float *B, int sdb, float *D);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_strmm_nn_rl_8x4_lib8
.type kernel_strmm_nn_rl_8x4_lib8, @function
kernel_strmm_nn_rl_8x4_lib8:
#elif defined(OS_MAC)
.globl _kernel_strmm_nn_rl_8x4_lib8
_kernel_strmm_nn_rl_8x4_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_strmm_nn_rl_8x4_lib8
.def kernel_strmm_nn_rl_8x4_lib8; .scl 2; .type 32; .endef
kernel_strmm_nn_rl_8x4_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorpd %ymm0, %ymm0, %ymm0
vmovapd %ymm0, %ymm1
vmovapd %ymm0, %ymm2
vmovapd %ymm0, %ymm3
// initial triangle
movq ARG1, %r10 // k
movq ARG3, %r11 // A
movq ARG5, %r12 // B
movq ARG6, %r13 // sdb
sall $5, %r13d // 4*sdb*sizeof(double)
movq ARG4, %r14 // offsetB
#if MACRO_LEVEL>=1
INNER_EDGE_TRMM_NN_RL_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_trmm_nn_rl_8x4_lib8
#elif defined(OS_MAC)
callq _inner_edge_trmm_nn_rl_8x4_lib8
#endif
#endif
#if MACRO_LEVEL>=1
INNER_EDGE_GEMM_ADD_NN_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_gemm_add_nn_8x4_lib8
#elif defined(OS_MAC)
callq _inner_edge_gemm_add_nn_8x4_lib8
#endif
#endif
// call inner dgemm kernel nt after initial triangle
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NN_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nn_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nn_8x4_lib8
#endif
#endif
// call inner scale
movq ARG2, %r10 // alpha
#if MACRO_LEVEL>=1
INNER_SCALE_A0_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_a0_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_a0_8x4_lib8
#endif
#endif
// store n
movq ARG7, %r10 // D
#if MACRO_LEVEL>=1
INNER_STORE_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_8x4_lib8
#elif defined(OS_MAC)
callq _inner_store_8x4_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_strmm_nn_rl_8x4_lib8, .-kernel_strmm_nn_rl_8x4_lib8
#endif
// 1 2 3 4 5 6 7 8 9
// void kernel_strmm_nn_rl_8x4_vs_lib8(int k, float *alpha, float *A, int offsetB, float *B, int sdb, float *D, int km, int kn);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_strmm_nn_rl_8x4_vs_lib8
.type kernel_strmm_nn_rl_8x4_vs_lib8, @function
kernel_strmm_nn_rl_8x4_vs_lib8:
#elif defined(OS_MAC)
.globl _kernel_strmm_nn_rl_8x4_vs_lib8
_kernel_strmm_nn_rl_8x4_vs_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_strmm_nn_rl_8x4_vs_lib8
.def kernel_strmm_nn_rl_8x4_vs_lib8; .scl 2; .type 32; .endef
kernel_strmm_nn_rl_8x4_vs_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorpd %ymm0, %ymm0, %ymm0
vmovapd %ymm0, %ymm1
vmovapd %ymm0, %ymm2
vmovapd %ymm0, %ymm3
// initial triangle
movq ARG1, %r10 // k
movq ARG3, %r11 // A
movq ARG5, %r12 // B
movq ARG6, %r13 // sdb
sall $5, %r13d // 4*sdb*sizeof(double)
movq ARG4, %r14 // offsetB
#if MACRO_LEVEL>=1
INNER_EDGE_TRMM_NN_RL_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_trmm_nn_rl_8x4_lib8
#elif defined(OS_MAC)
callq _inner_edge_trmm_nn_rl_8x4_lib8
#endif
#endif
#if MACRO_LEVEL>=1
INNER_EDGE_GEMM_ADD_NN_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_gemm_add_nn_8x4_lib8
#elif defined(OS_MAC)
callq _inner_edge_gemm_add_nn_8x4_lib8
#endif
#endif
// call inner dgemm kernel nt after initial triangle
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NN_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nn_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nn_8x4_lib8
#endif
#endif
// call inner scale
movq ARG2, %r10 // alpha
#if MACRO_LEVEL>=1
INNER_SCALE_A0_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_a0_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_a0_8x4_lib8
#endif
#endif
// store n
movq ARG7, %r10 // D
movq ARG8, %r11 // km
movq ARG9, %r12 // kn
#if MACRO_LEVEL>=1
INNER_STORE_8X4_VS_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_8x4_vs_lib8
#elif defined(OS_MAC)
callq _inner_store_8x4_vs_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_strmm_nn_rl_8x4_vs_lib8, .-kernel_strmm_nn_rl_8x4_vs_lib8
#endif
// 1 2 3 4 5 6 7 8 9 10 11 12 13
// void kernel_strmm_nn_rl_8x4_gen_lib8(int k, float *alpha, float *A, int offsetB, float *B, int sdb, int offsetD, float *D, int sdd, int m0, int m1, int n0, int n1);
.p2align 4,,15
#if defined(OS_LINUX)
.globl kernel_strmm_nn_rl_8x4_gen_lib8
.type kernel_strmm_nn_rl_8x4_gen_lib8, @function
kernel_strmm_nn_rl_8x4_gen_lib8:
#elif defined(OS_MAC)
.globl _kernel_strmm_nn_rl_8x4_gen_lib8
_kernel_strmm_nn_rl_8x4_gen_lib8:
#elif defined(OS_WINDOWS)
.globl kernel_strmm_nn_rl_8x4_gen_lib8
.def kernel_strmm_nn_rl_8x4_gen_lib8; .scl 2; .type 32; .endef
kernel_strmm_nn_rl_8x4_gen_lib8:
#endif
PROLOGUE
// zero accumulation registers
vxorpd %ymm0, %ymm0, %ymm0
vmovapd %ymm0, %ymm1
vmovapd %ymm0, %ymm2
vmovapd %ymm0, %ymm3
// initial triangle
movq ARG1, %r10 // k
movq ARG3, %r11 // A
movq ARG5, %r12 // B
movq ARG6, %r13 // sdb
sall $5, %r13d // 4*sdb*sizeof(double)
movq ARG4, %r14 // offsetB
#if MACRO_LEVEL>=1
INNER_EDGE_TRMM_NN_RL_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_trmm_nn_rl_8x4_lib8
#elif defined(OS_MAC)
callq _inner_edge_trmm_nn_rl_8x4_lib8
#endif
#endif
#if MACRO_LEVEL>=1
INNER_EDGE_GEMM_ADD_NN_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_edge_gemm_add_nn_8x4_lib8
#elif defined(OS_MAC)
callq _inner_edge_gemm_add_nn_8x4_lib8
#endif
#endif
// call inner dgemm kernel nt after initial triangle
#if MACRO_LEVEL>=2
INNER_KERNEL_GEMM_ADD_NN_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_kernel_gemm_add_nn_8x4_lib8
#elif defined(OS_MAC)
callq _inner_kernel_gemm_add_nn_8x4_lib8
#endif
#endif
// call inner scale
movq ARG2, %r10 // alpha
#if MACRO_LEVEL>=1
INNER_SCALE_A0_8X4_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_scale_a0_8x4_lib8
#elif defined(OS_MAC)
callq _inner_scale_a0_8x4_lib8
#endif
#endif
// store n
movq ARG7, %r10 // offsetD
movq ARG8, %r11 // D
movq ARG9, %r12 // sdd
sall $5, %r12d // 4*sdb*sizeof(double)
movq ARG10, %r13 // m0
movq ARG11, %r14 // m1
movq ARG12, %r15 // n0
movq ARG13, %rax // n1
#if MACRO_LEVEL>=1
INNER_STORE_8X4_GEN_LIB8
#else
#if defined(OS_LINUX) | defined(OS_WINDOWS)
call inner_store_8x4_gen_lib8
#elif defined(OS_MAC)
callq _inner_store_8x4_gen_lib8
#endif
#endif
EPILOGUE
ret
#if defined(OS_LINUX)
.size kernel_strmm_nn_rl_8x4_gen_lib8, .-kernel_strmm_nn_rl_8x4_gen_lib8
#endif
// read-only data
#if defined(OS_LINUX)
.section .rodata.cst32,"aM",@progbits,32
#elif defined(OS_MAC)
.section __TEXT,__const
#elif defined(OS_WINDOWS)
.section .rdata,"dr"
#endif
#if defined(OS_LINUX) | defined(OS_WINDOWS)
.align 32
.LC00: // { 7.5 6.5 5.5 4.5 3.5 2.5 1.5 0.5 }
#elif defined(OS_MAC)
.align 5
LC00: // { 7.5 6.5 5.5 4.5 3.5 2.5 1.5 0.5 }
#endif
.long 1056964608
.long 1069547520
.long 1075838976
.long 1080033280
.long 1083179008
.long 1085276160
.long 1087373312
.long 1089470464
#if defined(OS_LINUX) | defined(OS_WINDOWS)
.align 32
.LC01: // { 15.5 14.5 13.5 12.5 11.5 10.5 9.5 8.5 }
#elif defined(OS_MAC)
.align 5
LC01: // { 15.5 14.5 13.5 12.5 11.5 10.5 9.5 8.5 }
#endif
.long 1091043328
.long 1092091904
.long 1093140480
.long 1094189056
.long 1095237632
.long 1096286208
.long 1097334784
.long 1098383360
#if defined(OS_LINUX) | defined(OS_WINDOWS)
.align 32
.LC02: // { 23.5 22.5 21.5 20.5 19.5 18.5 17.5 16.5 }
#elif defined(OS_MAC)
.align 5
LC02: // { 23.5 22.5 21.5 20.5 19.5 18.5 17.5 16.5 }
#endif
.long 1099169792
.long 1099694080
.long 1100218368
.long 1100742656
.long 1101266944
.long 1101791232
.long 1102315520
.long 1102839808
#if defined(OS_LINUX) | defined(OS_WINDOWS)
.align 32
.LC03: // { 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 }
#elif defined(OS_MAC)
.align 5
LC03: // { 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 }
#endif
.long 1065353216
.long 1065353216
.long 1065353216
.long 1065353216
.long 1065353216
.long 1065353216
.long 1065353216
.long 1065353216
#if defined(OS_LINUX) | defined(OS_WINDOWS)
.align 32
.LC04: // { -1.0 -1.0 1.0 1.0 1.0 1.0 1.0 1.0 }
#elif defined(OS_MAC)
.align 5
LC04: // { -1.0 -1.0 1.0 1.0 1.0 1.0 1.0 1.0 }
#endif
.long 1065353216
.long 1065353216
.long 1065353216
.long 1065353216
.long 1065353216
.long 1065353216
.long 3212836864
.long 3212836864
#if defined(OS_LINUX)
.section .note.GNU-stack,"",@progbits
#elif defined(OS_MAC)
.subsections_via_symbols
#endif