// ===== REFERENCES =====
// [1] 正文 lineage:本账号现役 mine **#114335** <https://duck.ac/submission/114335>(判题机 565.338753 ms)
// = 对手 pdoom **#112312** <https://duck.ac/submission/112312> 的逐字节正文
// + 三个 pragma(`web` <https://duck.ac/submission/114317> · `rename-registers`
// <https://duck.ac/submission/114328> · `#pragma GCC unroll 4` <https://duck.ac/submission/114327>)。
// duck.ac 提交正文按站点规则公开可见;其自带 credit 块(下方 `/* Credit: pdoom ... */`)逐字保留。
// [2] 本席(`s4af_`)依据:
// * `work/s4ae_tmp/ring{A..H}.cpp` / `work/s4ad_tmp/ring.cpp`(判题机叶片环,custom-test,不占发数)
// * 本席新装置 `work/s4af_ring.py`(**活锁深度阶梯 d0..d12**)· `work/s4af_ring2.py`(结构臂)
// * 本席新仪器 `work/s4af_tmp/ph_*.cpp` = `s4ad_probe.py` 的 rdtscp 逐相探针(叶/逐级 combine/
// pack_root/根融合),**用它做叶级 A/B**——本席证明它的重复性 = **±0.04%**(见 `work/s4af_pricing.md`)
// ======================
// ===== 思路(s4aj_ 席 · 新写一节)=====
// 【本件性质】**真机器码候选(s4aj_F1)**:只改 micro24 / micro16 **尾声**的掩码承载方式。
// 【靶标】尾声里有一个**每次调用重新物化的填充**:
// vpcmpeqd %%ymm13,%%ymm13,%%ymm13 <- 2 条,每次调用重建 0x0000FFFF
// vpsrld $16,%%ymm13,%%ymm13
// vpand %%ymm13, %%ymm0..11 <- 12 条(micro16 是 8 条),**全部串行依赖 ymm13**
// 调用密度:micro24 32 次/叶 + micro16 16 次/叶 ⇒ **5.65e6 次/run · 96 条/叶**。
// 【本件做法】把 `vpand %%ymm13, ymm, ymm` 换成 `vpand s4aj_m16(%%rip), ymm, ymm`
// (`s4aj_m16` = 32 B 常量,每个 dword = 0x0000FFFF)⇒ **填充归零(−2 条/调用)**,
// 且 **每条 vpand 自带常量、彼此解耦**(不再等同一个掩码寄存器)。
// ★ 注意:该常量**不能写 `const`**——命名空间 const 数组在 C++ 里是内部链接,而 GCC 看不见
// asm 模板里的引用,会链接失败(`undefined reference to 's4aj_m16'`)。
// 【判题机真值(本席探针,reps=3 取 min)】judge `times_ns`:
// base [651 958 161 · 651 932 227 · **651 813 871**] ↔ 本件 [651 177 276 · 651 249 357 · **651 154 196**]
// ⇒ **−0.660 ms(−0.101%)**,且 **本件最差值 < base 最好值(565 µs 分离)⇒ 非重叠、真臂**;
// 两批复现:ALL −2 350 544 拍 / −2 387 014 拍。
// 【反号对照(在档,证本形态独有)】`micro24` 同型改造 +0.26% · `micro16` 全内存操作数版 +0.90%;
// 同席的 `F2`(填充 2 条→1 条 `vmovdqa`,保留寄存器形态)= **+0.109 ms(零)**
// ⇒ ★ **收益的载体是"依赖解耦",不是"指令数"**(归因纠正,见 `work/s4aj_cand.md` §二)。
// 【定价与判词】尾声 1 条真指令 = **0.39~0.41 拍 = 0.107 ns**(E2/E4/E8 剂量梯),
// 本件整对象上界 = 96 条/叶 × 0.107 ns × 117 649 = **1.27 ms < 2× 缺口 5.841840 ms**
// ⇒ 按派单 (d) **已闸未发**(`.text` `6e36972f61c1` / 21 809 B、ck 逐位 `17627627990235469774`)。
// 【闸门】判题同款编译 rc=0 · `.text` ≠ 基座(真改机器码)· **ck 逐位 = 现役 mine(数值逐位不变)** ·
// asm 文本已过 (1172)③ 的严格性正则(无 `(`/`,` 后紧跟 `+`)。
// ======================
// ===== 思路 =====
// 【本件性质】**真机器码候选(`s4af_E3`)**:只改 16 列尾块内核 `micro16` 的内环展开形态。
// 【机理】判题机叶片环(`s4af_ring.py` 阶梯 + 依赖边解剖):
// * 现役 `micro24`(4 行×24 列,12 累加器,16/16 ymm 全满)每 k-pair 只有 **1 个**产品暂存器,
// 于是 `vpmaddwd`(lat 5) → `vpaddd` 的边**必然**被下一条指令消费 ⇒ 5 拍延迟裸露。
// 实测:R0 = 70.22 拍/1536 MAC,而**去掉该依赖边**(反事实)= 65.4 ⇒ 该边价 **4.7 拍/轮**。
// * **活锁深度阶梯**(192 ALU、零载入):d0 = 65.56(零依赖地板)· d2 = 69.78 · d4 = 69.04 ·
// d8 = 68.00 ⇒ **深度 8 也只买回 2.4 拍**,且 16 ymm 下 12 累加器最多只留 4 个暂存器
// ⇒ ★ **`micro24` 的依赖池不可提**(本席用叶级 A/B 复核:任何增载入的 `micro24` 变体都**变慢**)。
// * `micro16`(4 行×16 列,8 累加器 ⇒ **8 个空寄存器**)却有余量:本件把 b 放 ymm8/9,
// 把 **前三行的 madd 全部先发射**(产品存 ymm10..15,每行占 2 个),再发它们的 add
// —— 单行 madd→add 发射距离从 1 槽抬到 **9 槽**,且**指令数/载入数逐条不变**。
// 【判题机叶级真值(本席仪器,重复性 ±0.04%)】LEAF 拍数:基座 1 437.7~1 438.3 M →
// 本件 **1 433.67 M = −0.303%**(= 整机 ≈ **−1.2 ms**);同批对照:`micro24` 同型改造 **+0.26%(反号)**、
// `micro16` 全内存操作数版 **+0.90%(反号)** ⇒ 只有这一条形态是**正号**。
// 【闸门】判题机同款编译器(gcc 9.3,`-O2 -static -std=c++17`):`.text` md5 **≠ 基座**(真改机器码)·
// **ck 逐位 `17627627990235469774`(数值逐位不变 = 同 a/b 面板、同累加器映射、同循环控制)**。
// ======================
#define ROOT_PAD 32
#pragma GCC optimize("O3,unroll-loops,web,rename-registers")
#pragma GCC target("avx2")
#include <immintrin.h>
#include <stdint.h>
#include <string.h>
#ifndef TILE_PAD
#define TILE_PAD 64
#endif
#ifndef BASE
#define BASE (1 << 6)
#endif
using U = uint16_t;
static constexpr int MAXN = 4096;
alignas(4096) static U pa_[MAXN * MAXN + (MAXN/BASE)*(MAXN/BASE)*TILE_PAD + 8192], pb_[MAXN * MAXN + (MAXN/BASE)*(MAXN/BASE)*TILE_PAD + 8192];
alignas(4096) static U pc_[MAXN * MAXN + (MAXN/BASE)*(MAXN/BASE)*TILE_PAD + 8192], ws_[MAXN * MAXN + (MAXN/BASE)*(MAXN/BASE)*TILE_PAD + 8192];
static U *pa = pa_ + 1024, *pb = pb_ + 1024, *pc = pc_ + 0, *ws = ws_ + 1024;
static U *final_out;
static inline __m256i ld(const U *p) {
return _mm256_load_si256((const __m256i *)p);
}
static inline void st(U *p, __m256i x) {
_mm256_store_si256((__m256i *)p, x);
}
// The recursive quadrant layout makes every Strassen addition contiguous.
static void pack(U *d, const U *s, int n, int stride, bool right) {
if (n > BASE) {
int h = n / 2, q=h*h+(h/BASE)*(h/BASE)*TILE_PAD;
pack(d, s, h, stride, right);
pack(d + q, s + h, h, stride, right);
pack(d + 2*q, s + h*stride, h, stride, right);
pack(d + 3*q, s + h*stride + h, h, stride, right);
} else if (!right) {
for(int i=0;i<n;i+=4) for(int k=0;k<n;k+=16) {
__m256i a=ld(s+i*stride+k), b=ld(s+(i+1)*stride+k);
__m256i c=ld(s+(i+2)*stride+k), e=ld(s+(i+3)*stride+k);
__m256i t0=_mm256_unpacklo_epi32(a,b),t1=_mm256_unpacklo_epi32(c,e);
__m256i t2=_mm256_unpackhi_epi32(a,b),t3=_mm256_unpackhi_epi32(c,e);
__m256i u0=_mm256_unpacklo_epi64(t0,t1),u1=_mm256_unpackhi_epi64(t0,t1);
__m256i u2=_mm256_unpacklo_epi64(t2,t3),u3=_mm256_unpackhi_epi64(t2,t3);
st(d,_mm256_permute2x128_si256(u0,u1,0x20));
st(d+16,_mm256_permute2x128_si256(u2,u3,0x20));
st(d+32,_mm256_permute2x128_si256(u0,u1,0x31));
st(d+48,_mm256_permute2x128_si256(u2,u3,0x31));d+=64;
}
} else {
for(int k=0;k<n;k+=2) for(int j=0;j<n;) {
int w=n-j>=24?24:n-j;
U *v=d+j*n+k*w;
for(int t=0;t<w;t+=16) {
if(t+16<=w) {
__m256i x=_mm256_loadu_si256((const __m256i*)(s+k*stride+j+t));
__m256i y=_mm256_loadu_si256((const __m256i*)(s+(k+1)*stride+j+t));
st(v,_mm256_unpacklo_epi16(x,y));
st(v+16,_mm256_unpackhi_epi16(x,y));v+=32;
} else {
__m128i x=_mm_load_si128((const __m128i*)(s+k*stride+j+t));
__m128i y=_mm_load_si128((const __m128i*)(s+(k+1)*stride+j+t));
_mm_store_si128((__m128i*)v,_mm_unpacklo_epi16(x,y));
_mm_store_si128((__m128i*)(v+8),_mm_unpackhi_epi16(x,y));v+=16;
}
}
j+=w;
}
}
}
static void unpack(U *d, const U *s, int n, int stride) {
if (n > BASE) {
int h = n/2, q=h*h+(h/BASE)*(h/BASE)*TILE_PAD;
unpack(d,s,h,stride);
unpack(d+h,s+q,h,stride);
unpack(d+h*stride,s+2*q,h,stride);
unpack(d+h*stride+h,s+3*q,h,stride);
} else {
for (int i=0;i<n;++i) {
U *dp = d + (size_t)i*stride; const U *sp = s + (size_t)i*n;
int k=0;
for (;k+32<=n;k+=32) {
_mm256_stream_si256((__m256i*)(dp+k), _mm256_load_si256((const __m256i*)(sp+k)));
_mm256_stream_si256((__m256i*)(dp+k+16), _mm256_load_si256((const __m256i*)(sp+k+16)));
}
for (;k<n;k+=16) _mm256_storeu_si256((__m256i*)(dp+k), _mm256_loadu_si256((const __m256i*)(sp+k)));
}
}
}
alignas(32) unsigned short s4aj_m16[16] = {0xFFFF,0,0xFFFF,0,0xFFFF,0,0xFFFF,0,0xFFFF,0,0xFFFF,0,0xFFFF,0,0xFFFF,0};
static inline void micro24(const U *a,const U *b,U *c,int n) {
const U *end=a+4*n; U *c3=c+3*n; intptr_t stride=2*n;
asm volatile(
"vmovdqa 0(%[b]), %%ymm12\n\t"
"vmovdqa 32(%[b]), %%ymm13\n\t"
"vpbroadcastd 0(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm1\n\t"
"vpmaddwd 64(%[b]), %%ymm14, %%ymm2\n\t"
"vpbroadcastd 4(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm4\n\t"
"vpmaddwd 64(%[b]), %%ymm14, %%ymm5\n\t"
"vpbroadcastd 8(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm7\n\t"
"vpmaddwd 64(%[b]), %%ymm14, %%ymm8\n\t"
"vpbroadcastd 12(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm10\n\t"
"vpmaddwd 64(%[b]), %%ymm14, %%ymm11\n\t"
"vmovdqa 96(%[b]), %%ymm12\n\t"
"vmovdqa 128(%[b]), %%ymm13\n\t"
"vpbroadcastd 16(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 160(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 20(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 160(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 24(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 160(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 28(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 160(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"vmovdqa 192(%[b]), %%ymm12\n\t"
"vmovdqa 224(%[b]), %%ymm13\n\t"
"vpbroadcastd 32(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 256(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 36(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 256(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 40(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 256(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 44(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 256(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"vmovdqa 288(%[b]), %%ymm12\n\t"
"vmovdqa 320(%[b]), %%ymm13\n\t"
"vpbroadcastd 48(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 352(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 52(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 352(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 56(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 352(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 60(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 352(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"vmovdqa 384(%[b]), %%ymm12\n\t"
"vmovdqa 416(%[b]), %%ymm13\n\t"
"vpbroadcastd 64(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 448(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 68(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 448(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 72(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 448(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 76(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 448(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"vmovdqa 480(%[b]), %%ymm12\n\t"
"vmovdqa 512(%[b]), %%ymm13\n\t"
"vpbroadcastd 80(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 544(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 84(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 544(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 88(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 544(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 92(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 544(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"vmovdqa 576(%[b]), %%ymm12\n\t"
"vmovdqa 608(%[b]), %%ymm13\n\t"
"vpbroadcastd 96(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 640(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 100(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 640(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 104(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 640(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 108(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 640(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"vmovdqa 672(%[b]), %%ymm12\n\t"
"vmovdqa 704(%[b]), %%ymm13\n\t"
"vpbroadcastd 112(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 736(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 116(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 736(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 120(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 736(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 124(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 736(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"add $128, %[a]\n\t"
"add $768, %[b]\n\t"
"1:\n\t"
"vmovdqa 0(%[b]), %%ymm12\n\t"
"vmovdqa 32(%[b]), %%ymm13\n\t"
"vpbroadcastd 0(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 64(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 4(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 64(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 8(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 64(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 12(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 64(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"vmovdqa 96(%[b]), %%ymm12\n\t"
"vmovdqa 128(%[b]), %%ymm13\n\t"
"vpbroadcastd 16(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 160(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 20(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 160(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 24(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 160(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 28(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 160(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"vmovdqa 192(%[b]), %%ymm12\n\t"
"vmovdqa 224(%[b]), %%ymm13\n\t"
"vpbroadcastd 32(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 256(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 36(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 256(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 40(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 256(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 44(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 256(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"vmovdqa 288(%[b]), %%ymm12\n\t"
"vmovdqa 320(%[b]), %%ymm13\n\t"
"vpbroadcastd 48(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 352(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 52(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 352(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 56(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 352(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 60(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 352(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"vmovdqa 384(%[b]), %%ymm12\n\t"
"vmovdqa 416(%[b]), %%ymm13\n\t"
"vpbroadcastd 64(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 448(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 68(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 448(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 72(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 448(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 76(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 448(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"vmovdqa 480(%[b]), %%ymm12\n\t"
"vmovdqa 512(%[b]), %%ymm13\n\t"
"vpbroadcastd 80(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 544(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 84(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 544(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 88(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 544(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 92(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 544(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"vmovdqa 576(%[b]), %%ymm12\n\t"
"vmovdqa 608(%[b]), %%ymm13\n\t"
"vpbroadcastd 96(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 640(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 100(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 640(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 104(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 640(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 108(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 640(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"vmovdqa 672(%[b]), %%ymm12\n\t"
"vmovdqa 704(%[b]), %%ymm13\n\t"
"vpbroadcastd 112(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 736(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm14, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 116(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 736(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 120(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 736(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm7, %%ymm7\n\t"
"vpaddd %%ymm14, %%ymm8, %%ymm8\n\t"
"vpbroadcastd 124(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm9, %%ymm9\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm15\n\t"
"vpmaddwd 736(%[b]), %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm14, %%ymm11, %%ymm11\n\t"
"add $128, %[a]\n\t"
"add $768, %[b]\n\t"
"cmp %[end], %[a]\n\t"
"jb 1b\n\t"
"2:\n\t"
"vpand s4aj_m16(%%rip), %%ymm0, %%ymm0\n\t"
"vpand s4aj_m16(%%rip), %%ymm1, %%ymm1\n\t"
"vpand s4aj_m16(%%rip), %%ymm2, %%ymm2\n\t"
"vpackusdw %%ymm1, %%ymm0, %%ymm0\n\t"
"vmovdqu %%ymm0, 0(%[c])\n\t"
"vpackusdw %%ymm2, %%ymm2, %%ymm2\n\t"
"vpermq $216, %%ymm2, %%ymm2\n\t"
"vmovdqu %%xmm2, 32(%[c])\n\t"
"vpand s4aj_m16(%%rip), %%ymm3, %%ymm3\n\t"
"vpand s4aj_m16(%%rip), %%ymm4, %%ymm4\n\t"
"vpand s4aj_m16(%%rip), %%ymm5, %%ymm5\n\t"
"vpackusdw %%ymm4, %%ymm3, %%ymm3\n\t"
"vmovdqu %%ymm3, 0(%[c],%[s],1)\n\t"
"vpackusdw %%ymm5, %%ymm5, %%ymm5\n\t"
"vpermq $216, %%ymm5, %%ymm5\n\t"
"vmovdqu %%xmm5, 32(%[c],%[s],1)\n\t"
"vpand s4aj_m16(%%rip), %%ymm6, %%ymm6\n\t"
"vpand s4aj_m16(%%rip), %%ymm7, %%ymm7\n\t"
"vpand s4aj_m16(%%rip), %%ymm8, %%ymm8\n\t"
"vpackusdw %%ymm7, %%ymm6, %%ymm6\n\t"
"vmovdqu %%ymm6, 0(%[c],%[s],2)\n\t"
"vpackusdw %%ymm8, %%ymm8, %%ymm8\n\t"
"vpermq $216, %%ymm8, %%ymm8\n\t"
"vmovdqu %%xmm8, 32(%[c],%[s],2)\n\t"
"vpand s4aj_m16(%%rip), %%ymm9, %%ymm9\n\t"
"vpand s4aj_m16(%%rip), %%ymm10, %%ymm10\n\t"
"vpand s4aj_m16(%%rip), %%ymm11, %%ymm11\n\t"
"vpackusdw %%ymm10, %%ymm9, %%ymm9\n\t"
"vmovdqu %%ymm9, 0(%[c3])\n\t"
"vpackusdw %%ymm11, %%ymm11, %%ymm11\n\t"
"vpermq $216, %%ymm11, %%ymm11\n\t"
"vmovdqu %%xmm11, 32(%[c3])\n\t"
: [a] "+&r"(a), [b] "+&r"(b)
: [c] "r"(c), [c3] "r"(c3), [s] "r"(stride), [end] "r"(end)
: "cc", "memory", "ymm0", "ymm1", "ymm2", "ymm3", "ymm4", "ymm5", "ymm6", "ymm7", "ymm8", "ymm9", "ymm10", "ymm11", "ymm12", "ymm13", "ymm14", "ymm15");
}
static inline void micro16(const U *a,const U *b,U *c,int n) {
const U *end=a+4*n; U *c3=c+3*n; intptr_t stride=2*n;
asm volatile(
"vmovdqa 0(%[b]), %%ymm8\n\t"
"vmovdqa 32(%[b]), %%ymm9\n\t"
"vpbroadcastd 0(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm0\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm1\n\t"
"vpbroadcastd 4(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm2\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm3\n\t"
"vpbroadcastd 8(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm4\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm5\n\t"
"vpbroadcastd 12(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm6\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm7\n\t"
"vmovdqa 64(%[b]), %%ymm8\n\t"
"vmovdqa 96(%[b]), %%ymm9\n\t"
"vpbroadcastd 16(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 20(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 24(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 28(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"vmovdqa 128(%[b]), %%ymm8\n\t"
"vmovdqa 160(%[b]), %%ymm9\n\t"
"vpbroadcastd 32(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 36(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 40(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 44(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"vmovdqa 192(%[b]), %%ymm8\n\t"
"vmovdqa 224(%[b]), %%ymm9\n\t"
"vpbroadcastd 48(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 52(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 56(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 60(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"vmovdqa 256(%[b]), %%ymm8\n\t"
"vmovdqa 288(%[b]), %%ymm9\n\t"
"vpbroadcastd 64(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 68(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 72(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 76(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"vmovdqa 320(%[b]), %%ymm8\n\t"
"vmovdqa 352(%[b]), %%ymm9\n\t"
"vpbroadcastd 80(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 84(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 88(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 92(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"vmovdqa 384(%[b]), %%ymm8\n\t"
"vmovdqa 416(%[b]), %%ymm9\n\t"
"vpbroadcastd 96(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 100(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 104(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 108(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"vmovdqa 448(%[b]), %%ymm8\n\t"
"vmovdqa 480(%[b]), %%ymm9\n\t"
"vpbroadcastd 112(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 116(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 120(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 124(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"add $128, %[a]\n\t"
"add $512, %[b]\n\t"
"1:\n\t"
"vmovdqa 0(%[b]), %%ymm8\n\t"
"vmovdqa 32(%[b]), %%ymm9\n\t"
"vpbroadcastd 0(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 4(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 8(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 12(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"vmovdqa 64(%[b]), %%ymm8\n\t"
"vmovdqa 96(%[b]), %%ymm9\n\t"
"vpbroadcastd 16(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 20(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 24(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 28(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"vmovdqa 128(%[b]), %%ymm8\n\t"
"vmovdqa 160(%[b]), %%ymm9\n\t"
"vpbroadcastd 32(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 36(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 40(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 44(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"vmovdqa 192(%[b]), %%ymm8\n\t"
"vmovdqa 224(%[b]), %%ymm9\n\t"
"vpbroadcastd 48(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 52(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 56(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 60(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"vmovdqa 256(%[b]), %%ymm8\n\t"
"vmovdqa 288(%[b]), %%ymm9\n\t"
"vpbroadcastd 64(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 68(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 72(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 76(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"vmovdqa 320(%[b]), %%ymm8\n\t"
"vmovdqa 352(%[b]), %%ymm9\n\t"
"vpbroadcastd 80(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 84(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 88(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 92(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"vmovdqa 384(%[b]), %%ymm8\n\t"
"vmovdqa 416(%[b]), %%ymm9\n\t"
"vpbroadcastd 96(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 100(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 104(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 108(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"vmovdqa 448(%[b]), %%ymm8\n\t"
"vmovdqa 480(%[b]), %%ymm9\n\t"
"vpbroadcastd 112(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpbroadcastd 116(%[a]), %%ymm12\n\t"
"vpmaddwd %%ymm8, %%ymm12, %%ymm13\n\t"
"vpmaddwd %%ymm9, %%ymm12, %%ymm12\n\t"
"vpbroadcastd 120(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm8, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm9, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm11, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm10, %%ymm1, %%ymm1\n\t"
"vpaddd %%ymm13, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm12, %%ymm3, %%ymm3\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 124(%[a]), %%ymm10\n\t"
"vpmaddwd %%ymm8, %%ymm10, %%ymm11\n\t"
"vpmaddwd %%ymm9, %%ymm10, %%ymm10\n\t"
"vpaddd %%ymm11, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm10, %%ymm7, %%ymm7\n\t"
"add $128, %[a]\n\t"
"add $512, %[b]\n\t"
"cmp %[end], %[a]\n\t"
"jb 1b\n\t"
"2:\n\t"
"vpand s4aj_m16(%%rip), %%ymm0, %%ymm0\n\t"
"vpand s4aj_m16(%%rip), %%ymm1, %%ymm1\n\t"
"vpackusdw %%ymm1, %%ymm0, %%ymm0\n\t"
"vmovdqu %%ymm0, 0(%[c])\n\t"
"vpand s4aj_m16(%%rip), %%ymm2, %%ymm2\n\t"
"vpand s4aj_m16(%%rip), %%ymm3, %%ymm3\n\t"
"vpackusdw %%ymm3, %%ymm2, %%ymm2\n\t"
"vmovdqu %%ymm2, 0(%[c],%[s],1)\n\t"
"vpand s4aj_m16(%%rip), %%ymm4, %%ymm4\n\t"
"vpand s4aj_m16(%%rip), %%ymm5, %%ymm5\n\t"
"vpackusdw %%ymm5, %%ymm4, %%ymm4\n\t"
"vmovdqu %%ymm4, 0(%[c],%[s],2)\n\t"
"vpand s4aj_m16(%%rip), %%ymm6, %%ymm6\n\t"
"vpand s4aj_m16(%%rip), %%ymm7, %%ymm7\n\t"
"vpackusdw %%ymm7, %%ymm6, %%ymm6\n\t"
"vmovdqu %%ymm6, 0(%[c3])\n\t"
: [a] "+&r"(a), [b] "+&r"(b)
: [c] "r"(c), [c3] "r"(c3), [s] "r"(stride), [end] "r"(end)
: "cc", "memory", "ymm0", "ymm1", "ymm2", "ymm3", "ymm4", "ymm5", "ymm6", "ymm7", "ymm8", "ymm9", "ymm10", "ymm11", "ymm12", "ymm13", "ymm14", "ymm15");
}
static inline void micro8(const U *a,const U *b,U *c,int n) {
const U *end=a+4*n; U *c3=c+3*n; intptr_t stride=2*n;
asm volatile(
"vmovdqa 0(%[b]), %%ymm12\n\t"
"vpbroadcastd 0(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm0\n\t"
"vpbroadcastd 4(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm1\n\t"
"vpbroadcastd 8(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm2\n\t"
"vpbroadcastd 12(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm3\n\t"
"vmovdqa 32(%[b]), %%ymm12\n\t"
"vpbroadcastd 16(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 20(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 24(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 28(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vmovdqa 64(%[b]), %%ymm12\n\t"
"vpbroadcastd 32(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 36(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 40(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 44(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vmovdqa 96(%[b]), %%ymm12\n\t"
"vpbroadcastd 48(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 52(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 56(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 60(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vmovdqa 128(%[b]), %%ymm12\n\t"
"vpbroadcastd 64(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 68(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 72(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 76(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vmovdqa 160(%[b]), %%ymm12\n\t"
"vpbroadcastd 80(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 84(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 88(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 92(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vmovdqa 192(%[b]), %%ymm12\n\t"
"vpbroadcastd 96(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 100(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 104(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 108(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vmovdqa 224(%[b]), %%ymm12\n\t"
"vpbroadcastd 112(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 116(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 120(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 124(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"add $128, %[a]\n\t"
"add $256, %[b]\n\t"
"1:\n\t"
"vmovdqa 0(%[b]), %%ymm12\n\t"
"vpbroadcastd 0(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 4(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 8(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 12(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vmovdqa 32(%[b]), %%ymm12\n\t"
"vpbroadcastd 16(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 20(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 24(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 28(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vmovdqa 64(%[b]), %%ymm12\n\t"
"vpbroadcastd 32(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 36(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 40(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 44(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vmovdqa 96(%[b]), %%ymm12\n\t"
"vpbroadcastd 48(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 52(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 56(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 60(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vmovdqa 128(%[b]), %%ymm12\n\t"
"vpbroadcastd 64(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 68(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 72(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 76(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vmovdqa 160(%[b]), %%ymm12\n\t"
"vpbroadcastd 80(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 84(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 88(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 92(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vmovdqa 192(%[b]), %%ymm12\n\t"
"vpbroadcastd 96(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 100(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 104(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 108(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"vmovdqa 224(%[b]), %%ymm12\n\t"
"vpbroadcastd 112(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpbroadcastd 116(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 120(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpbroadcastd 124(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpaddd %%ymm15, %%ymm3, %%ymm3\n\t"
"add $128, %[a]\n\t"
"add $256, %[b]\n\t"
"cmp %[end], %[a]\n\t"
"jb 1b\n\t"
"2:\n\t"
"vpcmpeqd %%ymm13, %%ymm13, %%ymm13\n\t"
"vpsrld $16, %%ymm13, %%ymm13\n\t"
"vpand %%ymm13, %%ymm0, %%ymm0\n\t"
"vpackusdw %%ymm0, %%ymm0, %%ymm0\n\t"
"vpermq $216, %%ymm0, %%ymm0\n\t"
"vmovdqu %%xmm0, 0(%[c])\n\t"
"vpand %%ymm13, %%ymm1, %%ymm1\n\t"
"vpackusdw %%ymm1, %%ymm1, %%ymm1\n\t"
"vpermq $216, %%ymm1, %%ymm1\n\t"
"vmovdqu %%xmm1, 0(%[c],%[s],1)\n\t"
"vpand %%ymm13, %%ymm2, %%ymm2\n\t"
"vpackusdw %%ymm2, %%ymm2, %%ymm2\n\t"
"vpermq $216, %%ymm2, %%ymm2\n\t"
"vmovdqu %%xmm2, 0(%[c],%[s],2)\n\t"
"vpand %%ymm13, %%ymm3, %%ymm3\n\t"
"vpackusdw %%ymm3, %%ymm3, %%ymm3\n\t"
"vpermq $216, %%ymm3, %%ymm3\n\t"
"vmovdqu %%xmm3, 0(%[c3])\n\t"
: [a] "+&r"(a), [b] "+&r"(b)
: [c] "r"(c), [c3] "r"(c3), [s] "r"(stride), [end] "r"(end)
: "cc", "memory", "ymm0", "ymm1", "ymm2", "ymm3", "ymm4", "ymm5", "ymm6", "ymm7", "ymm8", "ymm9", "ymm10", "ymm11", "ymm12", "ymm13", "ymm14", "ymm15");
}
static void base(const U *a,const U *b,U *c,int n) {
int j=0;
for(;j+24<=n;j+=24)
for(int i=0;i<n;i+=4) micro24(a+i*n,b+j*n,c+i*n+j,n);
if(n-j==16)
for(int i=0;i<n;i+=4) micro16(a+i*n,b+j*n,c+i*n+j,n);
if(n-j==8)
for(int i=0;i<n;i+=4) micro8(a+i*n,b+j*n,c+i*n+j,n);
}
template<bool Sub>
static inline void combine(U *d,const U *a,const U *b,int len) {
#pragma GCC unroll 4
for(int i=0;i<len;i+=16) {
__m256i x=ld(a+i),y=ld(b+i);
st(d+i,Sub?_mm256_sub_epi16(x,y):_mm256_add_epi16(x,y));
}
}
// At the root, the three completed packed quadrants can be sent straight
// to their final row-major destinations rather than materialized in pc.
template<int N>
static __attribute__((noinline)) void combine3_unpack(
U *d12,U *d21,U *d22,
const U *p6,const U *p7,const U *p5,
const U *p1,const U *p4,const U *p3,int stride) {
if constexpr (N > BASE) {
constexpr int h=N/2,q=h*h+(h/BASE)*(h/BASE)*TILE_PAD;
combine3_unpack<h>(d12,d21,d22,p6,p7,p5,p1,p4,p3,stride);
combine3_unpack<h>(d12+h,d21+h,d22+h,p6+q,p7+q,p5+q,p1+q,p4+q,p3+q,stride);
combine3_unpack<h>(d12+h*stride,d21+h*stride,d22+h*stride,
p6+2*q,p7+2*q,p5+2*q,p1+2*q,p4+2*q,p3+2*q,stride);
combine3_unpack<h>(d12+h*stride+h,d21+h*stride+h,d22+h*stride+h,
p6+3*q,p7+3*q,p5+3*q,p1+3*q,p4+3*q,p3+3*q,stride);
} else {
for(int row=0;row<N;++row) {
const int off=row*N;
U *o12=d12+(size_t)row*stride;
U *o21=d21+(size_t)row*stride;
U *o22=d22+(size_t)row*stride;
for(int col=0;col<N;col+=16) {
int idx=off+col;
__m256i u2=_mm256_add_epi16(ld(p1+idx),ld(p6+idx));
__m256i u3=_mm256_add_epi16(u2,ld(p7+idx));
__m256i v5=ld(p5+idx);
__m256i c12=_mm256_add_epi16(_mm256_add_epi16(u2,v5),ld(p3+idx));
__m256i c21=_mm256_sub_epi16(u3,ld(p4+idx));
__m256i c22=_mm256_add_epi16(u3,v5);
_mm256_stream_si256((__m256i*)(o12+col),c12);
_mm256_stream_si256((__m256i*)(o21+col),c21);
_mm256_stream_si256((__m256i*)(o22+col),c22);
}
}
}
}
// Add the surviving P1 into the root C11 product while mapping its packed
// quadrant layout to the final row-major C11 cells.
template<int N>
static __attribute__((noinline)) void unpack_add(U *d,const U *p2,const U *p1,int stride) {
if constexpr(N>BASE) {
constexpr int h=N/2,q=h*h+(h/BASE)*(h/BASE)*TILE_PAD;
unpack_add<h>(d,p2,p1,stride);
unpack_add<h>(d+h,p2+q,p1+q,stride);
unpack_add<h>(d+h*stride,p2+2*q,p1+2*q,stride);
unpack_add<h>(d+h*stride+h,p2+3*q,p1+3*q,stride);
} else {
for(int row=0;row<N;++row) {
U *o=d+(size_t)row*stride;
int off=row*N;
for(int col=0;col<N;col+=16)
_mm256_stream_si256((__m256i*)(o+col),
_mm256_add_epi16(ld(p2+off+col),ld(p1+off+col)));
}
}
}
// Strassen-Winograd: seven products, fifteen additions, two temporaries.
// All operations are in Z/(2^16); truncation never changes the answer.
template<int N>
static __attribute__((noinline)) void mul(const U *a,const U *b,U *c,U *work) {
if constexpr (N<=BASE) {
base(a,b,c,N);
} else {
constexpr int h=N/2,q=h*h+(h/BASE)*(h/BASE)*TILE_PAD;
const U *a11=a,*a12=a+q,*a21=a+2*q,*a22=a+3*q;
const U *b11=b,*b12=b+q,*b21=b+2*q,*b22=b+3*q;
U *c11=c,*c12=c+q,*c21=c+2*q,*c22=c+3*q;
U *x=work,*y=work+q+1024,*next=work+2*q+2048;
combine<true>(x,a11,a21,q);
combine<true>(y,b22,b12,q);
mul<h>(x,y,c21,next); // P7
combine<false>(x,a21,a22,q);
combine<true>(y,b12,b11,q);
mul<h>(x,y,c22,next); // P5
combine<true>(x,x,a11,q);
combine<true>(y,b22,y,q);
mul<h>(x,y,c12,next); // P6
combine<true>(x,a12,x,q);
mul<h>(x,b22,c11,next); // P3
combine<true>(y,y,b21,q);
mul<h>(a22,y,x,next); // P4
mul<h>(a11,b11,y,next); // P1
if constexpr (N == 4096) {
combine3_unpack<2048>(final_out+2048,
final_out+(size_t)2048*4096,
final_out+(size_t)2048*4096+2048,
c12,c21,c22,y,x,c11,4096);
} else {
for(int i=0;i<q;i+=16) {
__m256i u2=_mm256_add_epi16(ld(y+i),ld(c12+i));
__m256i u3=_mm256_add_epi16(u2,ld(c21+i));
__m256i p5=ld(c22+i);
st(c12+i,_mm256_add_epi16(_mm256_add_epi16(u2,p5),ld(c11+i)));
st(c21+i,_mm256_sub_epi16(u3,ld(x+i)));
st(c22+i,_mm256_add_epi16(u3,p5));
}
}
mul<h>(a12,b21,c11,next); // P2
if constexpr (N == 4096) unpack_add<2048>(final_out,c11,y,4096);
else combine<false>(c11,c11,y,q);
}
}
static constexpr int RQ=2048*2048+32*32*TILE_PAD,RS=RQ+ROOT_PAD;
alignas(4096) static U root_mem[15*RS+4096];
alignas(4096) static U root_tile[4*4096];
template<bool Right>
static void pack_root(const U*s,int n,int offset,U *dest) {
if(n>64){int h=n/2,q=h*h+(h/BASE)*(h/BASE)*TILE_PAD;
pack_root<Right>(s,h,offset,dest);
pack_root<Right>(s+h,h,offset+q,dest);
pack_root<Right>(s+h*4096,h,offset+2*q,dest);
pack_root<Right>(s+h*4096+h,h,offset+3*q,dest);
}else{
pack(root_tile,s,64,4096,Right);
pack(root_tile+4096,s+2048,64,4096,Right);
pack(root_tile+8192,s+2048*4096,64,4096,Right);
pack(root_tile+12288,s+2048*4096+2048,64,4096,Right);
U *r0=dest+offset,*r1=r0+RS,*r2=r1+RS,*v7=r2+RS,*v5=v7+RS,*v6=v5+RS,*v34=v6+RS;
for(int i=0;i<4096;i+=32){
__m256i a=ld(root_tile+i),b=ld(root_tile+4096+i),c=ld(root_tile+8192+i),d=ld(root_tile+12288+i);
__m256i aa=ld(root_tile+i+16),bb=ld(root_tile+4096+i+16),cc=ld(root_tile+8192+i+16),dd=ld(root_tile+12288+i+16);
_mm256_stream_si256((__m256i*)(r0+i),a);_mm256_stream_si256((__m256i*)(r0+i+16),aa);
_mm256_stream_si256((__m256i*)(r1+i),Right?c:b);_mm256_stream_si256((__m256i*)(r1+i+16),Right?cc:bb);
_mm256_stream_si256((__m256i*)(r2+i),d);_mm256_stream_si256((__m256i*)(r2+i+16),dd);
__m256i x7=Right?_mm256_sub_epi16(d,b):_mm256_sub_epi16(a,c);
__m256i xx7=Right?_mm256_sub_epi16(dd,bb):_mm256_sub_epi16(aa,cc);
_mm256_stream_si256((__m256i*)(v7+i),x7);_mm256_stream_si256((__m256i*)(v7+i+16),xx7);
__m256i x5=Right?_mm256_sub_epi16(b,a):_mm256_add_epi16(c,d);
__m256i xx5=Right?_mm256_sub_epi16(bb,aa):_mm256_add_epi16(cc,dd);
_mm256_stream_si256((__m256i*)(v5+i),x5);_mm256_stream_si256((__m256i*)(v5+i+16),xx5);
__m256i x6=Right?_mm256_sub_epi16(d,x5):_mm256_sub_epi16(x5,a);
__m256i xx6=Right?_mm256_sub_epi16(dd,xx5):_mm256_sub_epi16(xx5,aa);
_mm256_stream_si256((__m256i*)(v6+i),x6);_mm256_stream_si256((__m256i*)(v6+i+16),xx6);
__m256i x34=Right?_mm256_sub_epi16(x6,c):_mm256_sub_epi16(b,x6);
__m256i xx34=Right?_mm256_sub_epi16(xx6,cc):_mm256_sub_epi16(bb,xx6);
_mm256_stream_si256((__m256i*)(v34+i),x34);_mm256_stream_si256((__m256i*)(v34+i+16),xx34);
}
}
}
void matrix_multiply(int n,const short*A,const short*B,short*C){
U *ar=root_mem,*br=ar+7*RS,*pr=br+7*RS;
pack_root<false>((const U*)A,2048,0,ar);
pack_root<true>((const U*)B,2048,0,br);
_mm_mfence();
const U *a11=ar,*a12=ar+RS,*a22=ar+2*RS,*s7=ar+3*RS,*s5=ar+4*RS,*s6=ar+5*RS,*s3=ar+6*RS;
const U *b11=br,*b21=br+RS,*b22=br+2*RS,*t7=br+3*RS,*t5=br+4*RS,*t6=br+5*RS,*t4=br+6*RS;
U *p1=ar+5*RS,*p2=br+5*RS,*p3=ar+4*RS,*p4=br+4*RS,*p5=ar+3*RS,*p6=br+3*RS,*p7=pr;
mul<2048>(s7,t7,p7,ws);mul<2048>(s5,t5,p5,ws);mul<2048>(s6,t6,p6,ws);
mul<2048>(s3,b22,p3,ws);mul<2048>(a22,t4,p4,ws);mul<2048>(a11,b11,p1,ws);mul<2048>(a12,b21,p2,ws);
U *out=(U*)C;
combine3_unpack<2048>(out+2048,out+2048*4096,out+2048*4096+2048,p6,p7,p5,p1,p4,p3,4096);
unpack_add<2048>(out,p2,p1,4096);_mm_sfence();
}
| Compilation | N/A | N/A | Compile OK | Score: N/A | 显示更多 |
| Testcase #1 | 563.39 ms | 159 MB + 364 KB | Accepted | Score: 100 | 显示更多 |