// ===== [s4zi-ROOT] **根层 pack+Strassen 融合**(对手 pdoom #112302 的新结构;本席移植进我方引擎)=====
// ===== REFERENCES =====
// [1] duck.ac 用户 pdoom,提交 **#112302** <https://duck.ac/submission/112302>(现行 T,判题机 570.335 ms)
// 及同作者相邻版本 #112273 / #112297(本题连续刷新)。站点提交正文按规则公开可见。
// `pack_root<Right>` 与根层调度**逐行转写自该提交**(仅改函数命名/承接到本引擎的 `pack`/`ld`/`st`)。
// [2] 本账号 saffah_cc_v41_agg1,现役最好件 #112300 <https://duck.ac/submission/112300>(580.129476 ms)
// —— 本件正文除 [1] 的根层结构外 = **#112300 正文逐字节**(含我方 micro24/micro16 的车道交错布局等自有改动)。
// 合规:duck.ac 提交正文按站点规则公开可见;原提交无独立许可证声明,按 RULES.md 第 3 节署名作者账号与原地址。
// ======================
// ===== 思路 =====
// 【结构】对手把"根层(4096)的 pack"与"根层 8 条 Strassen 面板组合"**融合成一次遍历**:
// 沿 2048 根的 64×64 瓦片网格递归,每片**同时**从根的四象限各取一片瓦(s、s+2048、s+2048*4096、
// s+2048*4096+2048),在寄存器里算出该根的 7 个操作数(A 侧 a11/a12/a22/s7/s5/s6/s3,
// B 侧 b11/b21/b22/t7/t5/t6/t4),NT 存直接写进 7 个 8.39 MB 面板(root_mem,15×RS)。
// ⇒ **不再需要**旧路径的"整矩阵 pack(33.5 MB 读+写 = 67 MB/侧)+ 根层 8 次 `combine<Sub>` 面板合并
// (每次 1 写 2 读 × 8 MB = 24 MB 流量)"。随后 7 次 `mul<2048>`(输出**原地**复用已消费的输入槽位:
// p7=pr、p5=ar+3R(s7)、p6=br+3R(t7)、p3=ar+4R(s5)、p4=br+4R(t5)、p1=ar+5R(s6)、p2=br+5R(t6)),
// 最后 `combine3_unpack<2048>` + `unpack_add<2048>` 把三个非对角象限与 (0,0) 象限写回 C。
// 【逐位等价(代数)】运算仍是 Z/(2^16) 逐元素加减;七乘十五加恒等式逐条对照一致:
// A 侧 s7=a11−a21 · s5=a21+a22 · s6=s5−a11 · s3=a12−s6;B 侧 t7=b22−b12 · t5=b12−b11 · t6=b22−t5 · t4=t6−b21。
// 【闸门((532) 先过)】本机 n=4096 三跑 ck **= `2d9bb5a397ce7d12`**(与基座逐位相同)✓ · `^int main`=0 ✓ · 编译 rc=0 ✓
// 【判题机读数(probe,单变量同通道背靠背)】基座 #112300 引擎 **659.308 ms** ↔ 本件 **651.842 ms** ⇒ **−7.466 ms(−1.13%)** ✓
// ★ (442)/(520) 重要实证:**本机同装置给出 +6.5%(反号)** ⇒ 该"省流量"类改动必须看判题机(本机 TSC 口径 1.54e9 ↔ 1.64e9)。
// 【上界算式(动工前)】旧根层流量 = 整 pack 67 MB + 8×24 MB = **≈326 MB**;新根层 = 33.5 读 + 7×8.39 写 = 92.2 MB/侧
// ⇒ **≈184 MB** ⇒ 省 ≈142 MB;在档 combine<Sub> 总份额 121.92 ms(20.95%)的根层部分 ⇒ 上界 6~12 ms ⇒ 实测 −7.47 ms 落在带内 ✓
// 【目的】真优化件(非测量件):落地对手新结构并叠加我方既有边(micro 车道交错/包次序)。
// ============================================================================================
#pragma GCC optimize("O3,unroll-loops")
#pragma GCC target("avx2")
#include <immintrin.h>
#include <stdint.h>
#include <string.h>
#ifndef BASE
#define BASE 64
#endif
using U = uint16_t;
static constexpr int MAXN = 4096;
alignas(4096) static U pa_[MAXN * MAXN + 8192], pb_[MAXN * MAXN + 8192];
alignas(4096) static U pc_[MAXN * MAXN + 8192], ws_[MAXN * MAXN + 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;
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+=2)
for(int r=0;r<4;++r) {
memcpy(d,s+(i+r)*stride+k,4); d+=2;
}
} else {
/* [s4zi-PKORD] k-outer / j-inner; destination `v = d + j*n + k*w`.
Address set is IDENTICAL to the running-`d` form (each j-block
advances by exactly its width w, so sum_{j'<j} w_j' == j). */
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) {
__m128i x0=_mm_load_si128((const __m128i*)(s+k*stride+j+t));
__m128i y0=_mm_load_si128((const __m128i*)(s+(k+1)*stride+j+t));
if(t+16<=w) {
__m128i x1=_mm_load_si128((const __m128i*)(s+k*stride+j+t+8));
__m128i y1=_mm_load_si128((const __m128i*)(s+(k+1)*stride+j+t+8));
_mm_store_si128((__m128i*)v, _mm_unpacklo_epi16(x0,y0));
_mm_store_si128((__m128i*)(v+8), _mm_unpacklo_epi16(x1,y1));
_mm_store_si128((__m128i*)(v+16), _mm_unpackhi_epi16(x0,y0));
_mm_store_si128((__m128i*)(v+24), _mm_unpackhi_epi16(x1,y1));
v+=32;
} else {
_mm_store_si128((__m128i*)v, _mm_unpacklo_epi16(x0,y0));
_mm_store_si128((__m128i*)(v+8), _mm_unpackhi_epi16(x0,y0));
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;
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)));
}
}
}
static inline void micro24(const U *a,const U *b,U *c,int n) {
const U *end=a+4*n; const U *end2=end-64; 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"
"add $64, %[a]\n\t"
"add $384, %[b]\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"
"add $64, %[a]\n\t"
"add $384, %[b]\n\t"
"Lloop_%=:\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 %[end2], %[a]\n\t"
"jbe Lloop_%=\n\t"
"Ldone_%=:\n\t"
"vpcmpeqd %%ymm13, %%ymm13, %%ymm13\n\t"
"vpsrld $16, %%ymm13, %%ymm13\n\t"
"vpand %%ymm13, %%ymm0, %%ymm0\n\t"
"vpand %%ymm13, %%ymm1, %%ymm1\n\t"
"vpand %%ymm13, %%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 %%ymm13, %%ymm3, %%ymm3\n\t"
"vpand %%ymm13, %%ymm4, %%ymm4\n\t"
"vpand %%ymm13, %%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 %%ymm13, %%ymm6, %%ymm6\n\t"
"vpand %%ymm13, %%ymm7, %%ymm7\n\t"
"vpand %%ymm13, %%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 %%ymm13, %%ymm9, %%ymm9\n\t"
"vpand %%ymm13, %%ymm10, %%ymm10\n\t"
"vpand %%ymm13, %%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), [end2] "r"(end2)
: "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; const U *end2=end-32; 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"
"vpbroadcastd 4(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm2\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm3\n\t"
"vpbroadcastd 8(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm4\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm5\n\t"
"vpbroadcastd 12(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm6\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm7\n\t"
"vmovdqa 64(%[b]), %%ymm12\n\t"
"vmovdqa 96(%[b]), %%ymm13\n\t"
"vpbroadcastd 16(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 20(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 24(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 28(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"vmovdqa 128(%[b]), %%ymm12\n\t"
"vmovdqa 160(%[b]), %%ymm13\n\t"
"vpbroadcastd 32(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 36(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 40(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 44(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"vmovdqa 192(%[b]), %%ymm12\n\t"
"vmovdqa 224(%[b]), %%ymm13\n\t"
"vpbroadcastd 48(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 52(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 56(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 60(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"add $64, %[a]\n\t"
"add $256, %[b]\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"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 4(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 8(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 12(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"vmovdqa 64(%[b]), %%ymm12\n\t"
"vmovdqa 96(%[b]), %%ymm13\n\t"
"vpbroadcastd 16(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 20(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 24(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 28(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"vmovdqa 128(%[b]), %%ymm12\n\t"
"vmovdqa 160(%[b]), %%ymm13\n\t"
"vpbroadcastd 32(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 36(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 40(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 44(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"vmovdqa 192(%[b]), %%ymm12\n\t"
"vmovdqa 224(%[b]), %%ymm13\n\t"
"vpbroadcastd 48(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 52(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 56(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 60(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"add $64, %[a]\n\t"
"add $256, %[b]\n\t"
"Lloop_%=:\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"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 4(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 8(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 12(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"vmovdqa 64(%[b]), %%ymm12\n\t"
"vmovdqa 96(%[b]), %%ymm13\n\t"
"vpbroadcastd 16(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 20(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 24(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 28(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"vmovdqa 128(%[b]), %%ymm12\n\t"
"vmovdqa 160(%[b]), %%ymm13\n\t"
"vpbroadcastd 32(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 36(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 40(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 44(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"vmovdqa 192(%[b]), %%ymm12\n\t"
"vmovdqa 224(%[b]), %%ymm13\n\t"
"vpbroadcastd 48(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 52(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 56(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 60(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"vmovdqa 256(%[b]), %%ymm12\n\t"
"vmovdqa 288(%[b]), %%ymm13\n\t"
"vpbroadcastd 64(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 68(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 72(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 76(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"vmovdqa 320(%[b]), %%ymm12\n\t"
"vmovdqa 352(%[b]), %%ymm13\n\t"
"vpbroadcastd 80(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 84(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 88(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 92(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"vmovdqa 384(%[b]), %%ymm12\n\t"
"vmovdqa 416(%[b]), %%ymm13\n\t"
"vpbroadcastd 96(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 100(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 104(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 108(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"vmovdqa 448(%[b]), %%ymm12\n\t"
"vmovdqa 480(%[b]), %%ymm13\n\t"
"vpbroadcastd 112(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm0, %%ymm0\n\t"
"vpaddd %%ymm14, %%ymm1, %%ymm1\n\t"
"vpbroadcastd 116(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm2, %%ymm2\n\t"
"vpaddd %%ymm14, %%ymm3, %%ymm3\n\t"
"vpbroadcastd 120(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm4, %%ymm4\n\t"
"vpaddd %%ymm14, %%ymm5, %%ymm5\n\t"
"vpbroadcastd 124(%[a]), %%ymm14\n\t"
"vpmaddwd %%ymm12, %%ymm14, %%ymm15\n\t"
"vpmaddwd %%ymm13, %%ymm14, %%ymm14\n\t"
"vpaddd %%ymm15, %%ymm6, %%ymm6\n\t"
"vpaddd %%ymm14, %%ymm7, %%ymm7\n\t"
"add $128, %[a]\n\t"
"add $512, %[b]\n\t"
"cmp %[end2], %[a]\n\t"
"jbe Lloop_%=\n\t"
"Ldone_%=:\n\t"
"vpcmpeqd %%ymm13, %%ymm13, %%ymm13\n\t"
"vpsrld $16, %%ymm13, %%ymm13\n\t"
"vpand %%ymm13, %%ymm0, %%ymm0\n\t"
"vpand %%ymm13, %%ymm1, %%ymm1\n\t"
"vpackusdw %%ymm1, %%ymm0, %%ymm0\n\t"
"vmovdqu %%ymm0, 0(%[c])\n\t"
"vpand %%ymm13, %%ymm2, %%ymm2\n\t"
"vpand %%ymm13, %%ymm3, %%ymm3\n\t"
"vpackusdw %%ymm3, %%ymm2, %%ymm2\n\t"
"vmovdqu %%ymm2, 0(%[c],%[s],1)\n\t"
"vpand %%ymm13, %%ymm4, %%ymm4\n\t"
"vpand %%ymm13, %%ymm5, %%ymm5\n\t"
"vpackusdw %%ymm5, %%ymm4, %%ymm4\n\t"
"vmovdqu %%ymm4, 0(%[c],%[s],2)\n\t"
"vpand %%ymm13, %%ymm6, %%ymm6\n\t"
"vpand %%ymm13, %%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), [end2] "r"(end2)
: "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"
"add $64, %[a]\n\t"
"add $128, %[b]\n\t"
"cmp %[end], %[a]\n\t"
"jae 2f\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"
"add $64, %[a]\n\t"
"add $128, %[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) {
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;
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;
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;
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,*next=work+2*q;
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);
}
}
#define ROOT_PAD 1024
static constexpr int RQ=2048*2048,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;
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+=16){
__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 x7,x5,x6,x34;
if constexpr(Right){x7=_mm256_sub_epi16(d,b);x5=_mm256_sub_epi16(b,a);x6=_mm256_sub_epi16(d,x5);x34=_mm256_sub_epi16(x6,c);}
else{x7=_mm256_sub_epi16(a,c);x5=_mm256_add_epi16(c,d);x6=_mm256_sub_epi16(x5,a);x34=_mm256_sub_epi16(b,x6);}
_mm256_stream_si256((__m256i*)(r0+i),a);
_mm256_stream_si256((__m256i*)(r1+i),Right?c:b);
_mm256_stream_si256((__m256i*)(r2+i),d);
_mm256_stream_si256((__m256i*)(v7+i),x7);
_mm256_stream_si256((__m256i*)(v5+i),x5);
_mm256_stream_si256((__m256i*)(v6+i),x6);
_mm256_stream_si256((__m256i*)(v34+i),x34);
}
}
}
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 | 572.895 ms | 157 MB + 412 KB | Accepted | Score: 100 | 显示更多 |