#pragma GCC optimize("O3")
/* mmmc_k1: int8 GEMM C = A*B mod 256, EXACT, AVX2.
Split A = A0 + 128*A1 (A0 = A & 0x7F unsigned, A1 = top bit of each byte).
term1 = sum_k A0[i][k]*B[k][j] -> vpmaddubsw, no saturation (127*128*2 < 32768)
term2 = 128*(A1 . Bpar) mod 256 -> GF(2) matmul, 64 bits/word (Bpar = B&1)
Layout (lanes = columns): accum lane m is column j0+m.
x = vpbroadcastw of the 2-byte pair A0[i][2p..2p+1]
y = 32B load from BI[p] at 2*j0 ; BI[p][2j]=B[2p][j], BI[p][2j+1]=B[2p+1][j]
*/
#include <immintrin.h>
#include <xmmintrin.h>
#include <stdint.h>
#include <string.h>
#include <stdio.h>
#define AVX2 __attribute__((target("avx2")))
#define NMAX 4096
/* == mmmpad_fleet ARM: SCRATCH ROW-STRIDE PAD ==
A0b and PBb are PRIVATE scratch, so their row stride is not semantic -- but it is `n`,
i.e. 4096 B at n=4096, which is EXACTLY the L1 way-period (64 sets x 64 B). Every row of
both buffers then maps its line j into set j, on top of C's rows (also `n`) and the mikro's
two-row window: 6+ streams share one set out of 8 ways. LDA/LDB pad the two scratch
strides. The cell (row i, col j) is still the same value -- only WHERE it lives changes.
The mikro takes the scratch stride as an argument instead of deriving `A0+n`. */
#ifndef PADA
#define PADA 0
#endif
#define LDA ((int)(2*NMAX + PADA))
#ifndef PADB
#define PADB 0
#endif
#define LDB ((int)(NMAX + PADB))
static int8_t A0b[(size_t)NMAX*LDA] __attribute__((aligned(64)));
/* BW9: runtime scratch stride, sized to the ACTUAL m (not NMAX) */
static int GLDA = LDA, GLDB = LDB;
static uint64_t A1b[NMAX*NMAX/8] __attribute__((aligned(64)));
static uint64_t Bpb[NMAX*NMAX/8] __attribute__((aligned(64)));
static uint64_t Pbb[NMAX*NMAX/8] __attribute__((aligned(64)));
static uint8_t PBb[(size_t)NMAX*LDB+256] __attribute__((aligned(64)));
static void generic_mm(int n, const int8_t*A, const int8_t*B, int8_t*C){
for (int i=0;i<n;i++) for (int j=0;j<n;j++){
int s=0; for(int k=0;k<n;k++) s += (int)A[(size_t)i*n+k]*(int)B[(size_t)k*n+j];
C[(size_t)i*n+j]=(int8_t)s;
}
}
AVX2 static inline __m256i bcast16(const int8_t*p){
__m256i v; __asm__("vpbroadcastw %1, %0" : "=x"(v) : "m"(*(const short*)p)); return v;
}
AVX2 static inline __m256i packlo8(__m256i a, __m256i b){
const __m256i m=_mm256_set1_epi16(0x00FF);
return _mm256_permute4x64_epi64(
_mm256_packus_epi16(_mm256_and_si256(a,m), _mm256_and_si256(b,m)), 0xD8);
}
/* build_A0() DELETED by lane CH3: dead, and its A0b stride is the pre-duplication one. */
AVX2 static void build_A1bits(int n, const int8_t*A){
int W=n/64;
for (int i=0;i<n;i++){
const int8_t*row=A+(size_t)i*n; uint64_t*w=A1b+(size_t)i*W;
for (int k=0;k<n;k+=64){
uint32_t lo=(uint32_t)_mm256_movemask_epi8(_mm256_load_si256((const __m256i*)(row+k)));
uint32_t hi=(uint32_t)_mm256_movemask_epi8(_mm256_load_si256((const __m256i*)(row+k+32)));
w[k>>6]=(uint64_t)lo|((uint64_t)hi<<32);
}
}
}
AVX2 static void build_Bpbits(int n, const int8_t*B){
const __m256i one=_mm256_set1_epi8(1);
int W=n/64;
for (int k=0;k<n;k++){
const int8_t*row=B+(size_t)k*n; uint64_t*w=Bpb+(size_t)k*W;
for (int j=0;j<n;j+=64){
__m256i a=_mm256_slli_epi16(_mm256_and_si256(_mm256_load_si256((const __m256i*)(row+j)),one),7);
__m256i b=_mm256_slli_epi16(_mm256_and_si256(_mm256_load_si256((const __m256i*)(row+j+32)),one),7);
uint32_t lo=(uint32_t)_mm256_movemask_epi8(a);
uint32_t hi=(uint32_t)_mm256_movemask_epi8(b);
w[j>>6]=(uint64_t)lo|((uint64_t)hi<<32);
}
}
}
/* P[i] = XOR of Bp[k] over k with A1[i][k] set. NY = n/256 ymm per row. */
AVX2 static void corr_bits(int n){
int W=n/64, NY=W>>2;
for (int i=0;i<n;i++){
__m256i acc[16];
for(int t=0;t<NY;t++) acc[t]=_mm256_setzero_si256();
const uint64_t*row=A1b+(size_t)i*W;
for(int w=0;w<W;w++){
uint64_t b=row[w];
while(b){
int t=__builtin_ctzll(b); b&=b-1;
const uint64_t*br=Bpb+(size_t)(w*64+t)*W;
for(int q=0;q<NY;q++)
acc[q]=_mm256_xor_si256(acc[q],_mm256_load_si256((const __m256i*)(br+4*q)));
}
}
uint64_t*o=Pbb+(size_t)i*W;
for(int q=0;q<NY;q++) _mm256_store_si256((__m256i*)(o+4*q),acc[q]);
}
}
static uint64_t A1T[NMAX*NMAX/8] __attribute__((aligned(64)));
AVX2 static void transpose64(uint64_t *a){
int j,k; uint64_t m,t;
for (j=32,m=0xFFFFFFFF00000000ULL; j; j>>=1, m^=m>>j)
for (k=0;k<64;k=(k+j+1)&~j){
t=(a[k]^(a[k+j]<<j))&m;
a[k]^=t; a[k+j]^=(t>>j);
}
}
AVX2 static void build_A1T(int n){
int W=n/64;
for (int kb=0; kb<n; kb+=64)
for (int ib=0; ib<n; ib+=64){
uint64_t blk[64];
for (int i=0;i<64;i++) blk[i]=A1b[(size_t)(ib+i)*W+(kb>>6)];
transpose64(blk);
for (int k=0;k<64;k++) A1T[(size_t)(kb+k)*W+(ib>>6)]=blk[k];
}
}
/* P[i][j] = XOR over k with A1[i][k] set of Bp[k][j].
Per-output-row bit scan (dense 64-bit words -> predictable), K-BLOCKED so the
Bp slice stays in L2. P accumulated in memory across k-blocks. */
AVX2 static void corr_pass3(int n){
int W=n/64, NY=W>>2;
memset(Pbb,0,(size_t)n*W*8);
const int KB=256;
for (int kb=0; kb<n; kb+=KB){
int w0=kb>>6, w1=w0+(KB>>6);
for (int i=0;i<n;i++){
__m256i acc[16];
for (int q=0;q<NY;q++) acc[q]=_mm256_setzero_si256();
const uint64_t*row=A1b+(size_t)i*W;
for (int w=w0;w<w1;w++){
uint64_t b=row[w];
while(b){
int t=__builtin_ctzll(b); b&=b-1;
const uint64_t*br=Bpb+(size_t)(w*64+t)*W;
for (int q=0;q<NY;q++) acc[q]=_mm256_xor_si256(acc[q],_mm256_load_si256((const __m256i*)(br+4*q)));
}
}
uint64_t*o=Pbb+(size_t)i*W;
for (int q=0;q<NY;q++){
__m256i*po=(__m256i*)(o+4*q);
_mm256_store_si256(po,_mm256_xor_si256(_mm256_load_si256((const __m256i*)po),acc[q]));
}
}
}
}
/* specialized for n=4096 (W=64, NY=16): fully unrolled XOR loops */
/* corr16(): fully hand-written per-row GF(2) accumulate. Bit-identical output. */
#define AVX2BMI __attribute__((target("avx2,bmi")))
AVX2BMI static void corr16(void){
const int n=4096, W=64;
memset(Pbb,0,(size_t)n*W*8);
const int KB=64;
for (int kb=0; kb<n; kb+=KB){
int w0=kb>>6;
const char *bp0=(const char*)Bpb+(size_t)w0*32768;
const char *bp1=bp0; (void)bp1;
for (int i=0;i<n;i++){
const uint64_t*row=A1b+(size_t)i*W+w0;
uint64_t b0=row[0], b1=0;
uint64_t *o=Pbb+(size_t)i*W;
asm volatile(
"vpxor %%xmm0,%%xmm0,%%xmm0\n\t"
"vpxor %%xmm1,%%xmm1,%%xmm1\n\t"
"vpxor %%xmm2,%%xmm2,%%xmm2\n\t"
"vpxor %%xmm3,%%xmm3,%%xmm3\n\t"
"vpxor %%xmm4,%%xmm4,%%xmm4\n\t"
"vpxor %%xmm5,%%xmm5,%%xmm5\n\t"
"vpxor %%xmm6,%%xmm6,%%xmm6\n\t"
"vpxor %%xmm7,%%xmm7,%%xmm7\n\t"
"vpxor %%xmm8,%%xmm8,%%xmm8\n\t"
"vpxor %%xmm9,%%xmm9,%%xmm9\n\t"
"vpxor %%xmm10,%%xmm10,%%xmm10\n\t"
"vpxor %%xmm11,%%xmm11,%%xmm11\n\t"
"vpxor %%xmm12,%%xmm12,%%xmm12\n\t"
"vpxor %%xmm13,%%xmm13,%%xmm13\n\t"
"vpxor %%xmm14,%%xmm14,%%xmm14\n\t"
"vpxor %%xmm15,%%xmm15,%%xmm15\n\t"
"testq %[b0], %[b0]\n\t"
"jz 3%=f\n\t"
"1%=:\n\t"
"tzcntq %[b0], %%r11\n\t"
"shlq $9, %%r11\n\t"
"addq %[bp0], %%r11\n\t"
"vpxor 0(%%r11), %%ymm0, %%ymm0\n\t"
"vpxor 32(%%r11), %%ymm1, %%ymm1\n\t"
"vpxor 64(%%r11), %%ymm2, %%ymm2\n\t"
"vpxor 96(%%r11), %%ymm3, %%ymm3\n\t"
"vpxor 128(%%r11), %%ymm4, %%ymm4\n\t"
"vpxor 160(%%r11), %%ymm5, %%ymm5\n\t"
"vpxor 192(%%r11), %%ymm6, %%ymm6\n\t"
"vpxor 224(%%r11), %%ymm7, %%ymm7\n\t"
"vpxor 256(%%r11), %%ymm8, %%ymm8\n\t"
"vpxor 288(%%r11), %%ymm9, %%ymm9\n\t"
"vpxor 320(%%r11), %%ymm10, %%ymm10\n\t"
"vpxor 352(%%r11), %%ymm11, %%ymm11\n\t"
"vpxor 384(%%r11), %%ymm12, %%ymm12\n\t"
"vpxor 416(%%r11), %%ymm13, %%ymm13\n\t"
"vpxor 448(%%r11), %%ymm14, %%ymm14\n\t"
"vpxor 480(%%r11), %%ymm15, %%ymm15\n\t"
"blsrq %[b0], %[b0]\n\t"
"jnz 1%=b\n\t"
"3%=:\n\t"
"testq %[b1], %[b1]\n\t"
"jz 4%=f\n\t"
"2%=:\n\t"
"tzcntq %[b1], %%r11\n\t"
"shlq $9, %%r11\n\t"
"addq %[bp1], %%r11\n\t"
"vpxor 0(%%r11), %%ymm0, %%ymm0\n\t"
"vpxor 32(%%r11), %%ymm1, %%ymm1\n\t"
"vpxor 64(%%r11), %%ymm2, %%ymm2\n\t"
"vpxor 96(%%r11), %%ymm3, %%ymm3\n\t"
"vpxor 128(%%r11), %%ymm4, %%ymm4\n\t"
"vpxor 160(%%r11), %%ymm5, %%ymm5\n\t"
"vpxor 192(%%r11), %%ymm6, %%ymm6\n\t"
"vpxor 224(%%r11), %%ymm7, %%ymm7\n\t"
"vpxor 256(%%r11), %%ymm8, %%ymm8\n\t"
"vpxor 288(%%r11), %%ymm9, %%ymm9\n\t"
"vpxor 320(%%r11), %%ymm10, %%ymm10\n\t"
"vpxor 352(%%r11), %%ymm11, %%ymm11\n\t"
"vpxor 384(%%r11), %%ymm12, %%ymm12\n\t"
"vpxor 416(%%r11), %%ymm13, %%ymm13\n\t"
"vpxor 448(%%r11), %%ymm14, %%ymm14\n\t"
"vpxor 480(%%r11), %%ymm15, %%ymm15\n\t"
"blsrq %[b1], %[b1]\n\t"
"jnz 2%=b\n\t"
"4%=:\n\t"
"vpxor 0(%[o]), %%ymm0, %%ymm0\n\t"
"vmovdqu %%ymm0, 0(%[o])\n\t"
"vpxor 32(%[o]), %%ymm1, %%ymm1\n\t"
"vmovdqu %%ymm1, 32(%[o])\n\t"
"vpxor 64(%[o]), %%ymm2, %%ymm2\n\t"
"vmovdqu %%ymm2, 64(%[o])\n\t"
"vpxor 96(%[o]), %%ymm3, %%ymm3\n\t"
"vmovdqu %%ymm3, 96(%[o])\n\t"
"vpxor 128(%[o]), %%ymm4, %%ymm4\n\t"
"vmovdqu %%ymm4, 128(%[o])\n\t"
"vpxor 160(%[o]), %%ymm5, %%ymm5\n\t"
"vmovdqu %%ymm5, 160(%[o])\n\t"
"vpxor 192(%[o]), %%ymm6, %%ymm6\n\t"
"vmovdqu %%ymm6, 192(%[o])\n\t"
"vpxor 224(%[o]), %%ymm7, %%ymm7\n\t"
"vmovdqu %%ymm7, 224(%[o])\n\t"
"vpxor 256(%[o]), %%ymm8, %%ymm8\n\t"
"vmovdqu %%ymm8, 256(%[o])\n\t"
"vpxor 288(%[o]), %%ymm9, %%ymm9\n\t"
"vmovdqu %%ymm9, 288(%[o])\n\t"
"vpxor 320(%[o]), %%ymm10, %%ymm10\n\t"
"vmovdqu %%ymm10, 320(%[o])\n\t"
"vpxor 352(%[o]), %%ymm11, %%ymm11\n\t"
"vmovdqu %%ymm11, 352(%[o])\n\t"
"vpxor 384(%[o]), %%ymm12, %%ymm12\n\t"
"vmovdqu %%ymm12, 384(%[o])\n\t"
"vpxor 416(%[o]), %%ymm13, %%ymm13\n\t"
"vmovdqu %%ymm13, 416(%[o])\n\t"
"vpxor 448(%[o]), %%ymm14, %%ymm14\n\t"
"vmovdqu %%ymm14, 448(%[o])\n\t"
"vpxor 480(%[o]), %%ymm15, %%ymm15\n\t"
"vmovdqu %%ymm15, 480(%[o])\n\t"
: [b0]"+r"(b0), [b1]"+r"(b1)
: [bp0]"r"(bp0), [bp1]"r"(bp1), [o]"r"(o)
: "r11","cc","memory","ymm0","ymm1","ymm2","ymm3","ymm4","ymm5","ymm6","ymm7","ymm8","ymm9","ymm10","ymm11","ymm12","ymm13","ymm14","ymm15");
}
}
}
static void expand_bits(int n){
static uint8_t T[256*8]; static int ti=0;
if(!ti){ for(int v=0;v<256;v++) for(int b=0;b<8;b++) T[v*8+b]=((v>>b)&1)?0x80:0x00; ti=1; }
int W=n/64;
for(int i=0;i<n;i++){
const uint64_t*w=Pbb+(size_t)i*W; uint8_t*o=PBb+(size_t)i*GLDB;
for(int q=0;q<W;q++){
uint64_t v=w[q];
for(int s=0;s<8;s++) memcpy(o+(q*8+s)*8, T+(((uint8_t)(v>>(8*s)))*8), 8);
}
}
}
#define NIB 2
#define NJ 4
static const int16_t W0FF[16] __attribute__((aligned(32))) =
{0x00FF,0x00FF,0x00FF,0x00FF,0x00FF,0x00FF,0x00FF,0x00FF,
0x00FF,0x00FF,0x00FF,0x00FF,0x00FF,0x00FF,0x00FF,0x00FF};
static int8_t BI3b[NMAX*NMAX] __attribute__((aligned(64)));
AVX2 static void build_BI3(int n, const int8_t*B){
int ng=n/64;
for (int p=0;p<n/2;p++){
const int8_t *r0=B+(size_t)(2*p)*n, *r1=r0+n;
for (int g=0;g<ng;g++){
const int8_t *a=r0+64*g, *b=r1+64*g;
int8_t *o=BI3b+((size_t)g*(n/2)+p)*128;
for (int t=0;t<64;t+=32){
__m256i va=_mm256_load_si256((const __m256i*)(a+t));
__m256i vb=_mm256_load_si256((const __m256i*)(b+t));
__m256i lo=_mm256_unpacklo_epi8(va,vb), hi=_mm256_unpackhi_epi8(va,vb);
_mm256_store_si256((__m256i*)(o+2*t), _mm256_permute2x128_si256(lo,hi,0x20));
_mm256_store_si256((__m256i*)(o+2*t+32), _mm256_permute2x128_si256(lo,hi,0x31));
}
}
}
}
AVX2 static void mikro_u4(int n,const int8_t*A0,const int8_t*BI,int8_t*C,const uint8_t*PB,int jb,int lda,int ldb){
const int8_t *x0=A0;
long nn=(long)n;
const int8_t *x1=A0+lda;
const int8_t *bib=BI+(long)(n>>3)*512; long bidx=-((long)(n>>3)*512);
int8_t *C0=C+jb, *C1=C0+n;
const uint8_t *PB0=PB+jb, *PB1=PB0+ldb;
asm volatile(
"vpxor %%xmm8,%%xmm8,%%xmm8\n\t"
"vpxor %%xmm9,%%xmm9,%%xmm9\n\t"
"vpxor %%xmm10,%%xmm10,%%xmm10\n\t"
"vpxor %%xmm11,%%xmm11,%%xmm11\n\t"
"vpxor %%xmm12,%%xmm12,%%xmm12\n\t"
"vpxor %%xmm13,%%xmm13,%%xmm13\n\t"
"vpxor %%xmm14,%%xmm14,%%xmm14\n\t"
"vpxor %%xmm15,%%xmm15,%%xmm15\n\t"
".p2align 5\n\t"
"1:\n\t"
"vpbroadcastd 0(%[x0]), %%ymm0\n\t"
"vpbroadcastd 0(%[x1]), %%ymm1\n\t"
"vmovdqa 0(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm8, %%ymm8\n\t"
"vpaddw %%ymm6, %%ymm12, %%ymm12\n\t"
"vmovdqa 32(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm9, %%ymm9\n\t"
"vpaddw %%ymm6, %%ymm13, %%ymm13\n\t"
"vmovdqa 64(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm10, %%ymm10\n\t"
"vpaddw %%ymm6, %%ymm14, %%ymm14\n\t"
"vmovdqa 96(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm11, %%ymm11\n\t"
"vpaddw %%ymm6, %%ymm15, %%ymm15\n\t"
"vpbroadcastd 4(%[x0]), %%ymm0\n\t"
"vpbroadcastd 4(%[x1]), %%ymm1\n\t"
"vmovdqa 128(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm8, %%ymm8\n\t"
"vpaddw %%ymm6, %%ymm12, %%ymm12\n\t"
"vmovdqa 160(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm9, %%ymm9\n\t"
"vpaddw %%ymm6, %%ymm13, %%ymm13\n\t"
"vmovdqa 192(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm10, %%ymm10\n\t"
"vpaddw %%ymm6, %%ymm14, %%ymm14\n\t"
"vmovdqa 224(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm11, %%ymm11\n\t"
"vpaddw %%ymm6, %%ymm15, %%ymm15\n\t"
"vpbroadcastd 8(%[x0]), %%ymm0\n\t"
"vpbroadcastd 8(%[x1]), %%ymm1\n\t"
"vmovdqa 256(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm8, %%ymm8\n\t"
"vpaddw %%ymm6, %%ymm12, %%ymm12\n\t"
"vmovdqa 288(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm9, %%ymm9\n\t"
"vpaddw %%ymm6, %%ymm13, %%ymm13\n\t"
"vmovdqa 320(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm10, %%ymm10\n\t"
"vpaddw %%ymm6, %%ymm14, %%ymm14\n\t"
"vmovdqa 352(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm11, %%ymm11\n\t"
"vpaddw %%ymm6, %%ymm15, %%ymm15\n\t"
"vpbroadcastd 12(%[x0]), %%ymm0\n\t"
"vpbroadcastd 12(%[x1]), %%ymm1\n\t"
"vmovdqa 384(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm8, %%ymm8\n\t"
"vpaddw %%ymm6, %%ymm12, %%ymm12\n\t"
"vmovdqa 416(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm9, %%ymm9\n\t"
"vpaddw %%ymm6, %%ymm13, %%ymm13\n\t"
"vmovdqa 448(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm10, %%ymm10\n\t"
"vpaddw %%ymm6, %%ymm14, %%ymm14\n\t"
"vmovdqa 480(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm11, %%ymm11\n\t"
"vpaddw %%ymm6, %%ymm15, %%ymm15\n\t"
"addq $16, %[x0]\n\t"
"addq $16, %[x1]\n\t"
"addq $512, %[bidx]\n\t"
"jnz 1b\n\t"
"vpand %[mask], %%ymm8, %%ymm2\n\t"
"vpand %[mask], %%ymm9, %%ymm3\n\t"
"vpackuswb %%ymm3, %%ymm2, %%ymm2\n\t"
"vpermq $0xD8, %%ymm2, %%ymm2\n\t"
"vpxor 0(%[PB0]), %%ymm2, %%ymm2\n\t"
"vmovntdq %%ymm2, 0(%[C0])\n\t"
"vpand %[mask], %%ymm10, %%ymm2\n\t"
"vpand %[mask], %%ymm11, %%ymm3\n\t"
"vpackuswb %%ymm3, %%ymm2, %%ymm2\n\t"
"vpermq $0xD8, %%ymm2, %%ymm2\n\t"
"vpxor 32(%[PB0]), %%ymm2, %%ymm2\n\t"
"vmovntdq %%ymm2, 32(%[C0])\n\t"
"vpand %[mask], %%ymm12, %%ymm2\n\t"
"vpand %[mask], %%ymm13, %%ymm3\n\t"
"vpackuswb %%ymm3, %%ymm2, %%ymm2\n\t"
"vpermq $0xD8, %%ymm2, %%ymm2\n\t"
"vpxor 0(%[PB1]), %%ymm2, %%ymm2\n\t"
"vmovntdq %%ymm2, 0(%[C1])\n\t"
"vpand %[mask], %%ymm14, %%ymm2\n\t"
"vpand %[mask], %%ymm15, %%ymm3\n\t"
"vpackuswb %%ymm3, %%ymm2, %%ymm2\n\t"
"vpermq $0xD8, %%ymm2, %%ymm2\n\t"
"vpxor 32(%[PB1]), %%ymm2, %%ymm2\n\t"
"vmovntdq %%ymm2, 32(%[C1])\n\t"
: [x0]"+r"(x0), [bidx]"+r"(bidx), [x1]"+r"(x1)
: [C0]"r"(C0), [C1]"r"(C1), [PB0]"r"(PB0), [PB1]"r"(PB1), [nn]"r"(nn), [mask]"m"(W0FF), [bib]"r"(bib)
: "ymm0","ymm1","ymm2","ymm3","ymm4","ymm5","ymm6","ymm8","ymm9","ymm10","ymm11","ymm12","ymm13","ymm14","ymm15","cc","memory");
}
AVX2 static void mikro_u4r(int n,const int8_t*A0,const int8_t*BI,int8_t*C,const uint8_t*PB,int jb,int lda,int ldb){
const int8_t *x0=A0;
long nn=(long)n;
const int8_t *x1=A0+lda;
const int8_t *bib=BI; /* R1: slice START -- see mk_rev.py */ long bidx=((long)(n>>3)*512)-512; /* R1: start at the LAST 512-byte chunk */
int8_t *C0=C+jb, *C1=C0+n;
const uint8_t *PB0=PB+jb, *PB1=PB0+ldb;
asm volatile(
"vpxor %%xmm8,%%xmm8,%%xmm8\n\t"
"vpxor %%xmm9,%%xmm9,%%xmm9\n\t"
"vpxor %%xmm10,%%xmm10,%%xmm10\n\t"
"vpxor %%xmm11,%%xmm11,%%xmm11\n\t"
"vpxor %%xmm12,%%xmm12,%%xmm12\n\t"
"vpxor %%xmm13,%%xmm13,%%xmm13\n\t"
"vpxor %%xmm14,%%xmm14,%%xmm14\n\t"
"vpxor %%xmm15,%%xmm15,%%xmm15\n\t"
".p2align 5\n\t"
"1:\n\t"
"vpbroadcastd 0(%[x0]), %%ymm0\n\t"
"vpbroadcastd 0(%[x1]), %%ymm1\n\t"
"vmovdqa 0(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm8, %%ymm8\n\t"
"vpaddw %%ymm6, %%ymm12, %%ymm12\n\t"
"vmovdqa 32(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm9, %%ymm9\n\t"
"vpaddw %%ymm6, %%ymm13, %%ymm13\n\t"
"vmovdqa 64(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm10, %%ymm10\n\t"
"vpaddw %%ymm6, %%ymm14, %%ymm14\n\t"
"vmovdqa 96(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm11, %%ymm11\n\t"
"vpaddw %%ymm6, %%ymm15, %%ymm15\n\t"
"vpbroadcastd 4(%[x0]), %%ymm0\n\t"
"vpbroadcastd 4(%[x1]), %%ymm1\n\t"
"vmovdqa 128(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm8, %%ymm8\n\t"
"vpaddw %%ymm6, %%ymm12, %%ymm12\n\t"
"vmovdqa 160(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm9, %%ymm9\n\t"
"vpaddw %%ymm6, %%ymm13, %%ymm13\n\t"
"vmovdqa 192(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm10, %%ymm10\n\t"
"vpaddw %%ymm6, %%ymm14, %%ymm14\n\t"
"vmovdqa 224(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm11, %%ymm11\n\t"
"vpaddw %%ymm6, %%ymm15, %%ymm15\n\t"
"vpbroadcastd 8(%[x0]), %%ymm0\n\t"
"vpbroadcastd 8(%[x1]), %%ymm1\n\t"
"vmovdqa 256(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm8, %%ymm8\n\t"
"vpaddw %%ymm6, %%ymm12, %%ymm12\n\t"
"vmovdqa 288(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm9, %%ymm9\n\t"
"vpaddw %%ymm6, %%ymm13, %%ymm13\n\t"
"vmovdqa 320(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm10, %%ymm10\n\t"
"vpaddw %%ymm6, %%ymm14, %%ymm14\n\t"
"vmovdqa 352(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm11, %%ymm11\n\t"
"vpaddw %%ymm6, %%ymm15, %%ymm15\n\t"
"vpbroadcastd 12(%[x0]), %%ymm0\n\t"
"vpbroadcastd 12(%[x1]), %%ymm1\n\t"
"vmovdqa 384(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm8, %%ymm8\n\t"
"vpaddw %%ymm6, %%ymm12, %%ymm12\n\t"
"vmovdqa 416(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm9, %%ymm9\n\t"
"vpaddw %%ymm6, %%ymm13, %%ymm13\n\t"
"vmovdqa 448(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm10, %%ymm10\n\t"
"vpaddw %%ymm6, %%ymm14, %%ymm14\n\t"
"vmovdqa 480(%[bib],%[bidx]), %%ymm4\n\t"
"vpmaddubsw %%ymm4, %%ymm0, %%ymm5\n\t"
"vpmaddubsw %%ymm4, %%ymm1, %%ymm6\n\t"
"vpaddw %%ymm5, %%ymm11, %%ymm11\n\t"
"vpaddw %%ymm6, %%ymm15, %%ymm15\n\t"
"subq $16, %[x0]\n\t"
"subq $16, %[x1]\n\t"
"subq $512, %[bidx]\n\t"
"jns 1b\n\t"
"vpand %[mask], %%ymm8, %%ymm2\n\t"
"vpand %[mask], %%ymm9, %%ymm3\n\t"
"vpackuswb %%ymm3, %%ymm2, %%ymm2\n\t"
"vpermq $0xD8, %%ymm2, %%ymm2\n\t"
"vpxor 0(%[PB0]), %%ymm2, %%ymm2\n\t"
"vmovntdq %%ymm2, 0(%[C0])\n\t"
"vpand %[mask], %%ymm10, %%ymm2\n\t"
"vpand %[mask], %%ymm11, %%ymm3\n\t"
"vpackuswb %%ymm3, %%ymm2, %%ymm2\n\t"
"vpermq $0xD8, %%ymm2, %%ymm2\n\t"
"vpxor 32(%[PB0]), %%ymm2, %%ymm2\n\t"
"vmovntdq %%ymm2, 32(%[C0])\n\t"
"vpand %[mask], %%ymm12, %%ymm2\n\t"
"vpand %[mask], %%ymm13, %%ymm3\n\t"
"vpackuswb %%ymm3, %%ymm2, %%ymm2\n\t"
"vpermq $0xD8, %%ymm2, %%ymm2\n\t"
"vpxor 0(%[PB1]), %%ymm2, %%ymm2\n\t"
"vmovntdq %%ymm2, 0(%[C1])\n\t"
"vpand %[mask], %%ymm14, %%ymm2\n\t"
"vpand %[mask], %%ymm15, %%ymm3\n\t"
"vpackuswb %%ymm3, %%ymm2, %%ymm2\n\t"
"vpermq $0xD8, %%ymm2, %%ymm2\n\t"
"vpxor 32(%[PB1]), %%ymm2, %%ymm2\n\t"
"vmovntdq %%ymm2, 32(%[C1])\n\t"
: [x0]"+r"(x0), [bidx]"+r"(bidx), [x1]"+r"(x1)
: [C0]"r"(C0), [C1]"r"(C1), [PB0]"r"(PB0), [PB1]"r"(PB1), [nn]"r"(nn), [mask]"m"(W0FF), [bib]"r"(bib)
: "ymm0","ymm1","ymm2","ymm3","ymm4","ymm5","ymm6","ymm8","ymm9","ymm10","ymm11","ymm12","ymm13","ymm14","ymm15","cc","memory");
}
/* FUSED BUILDERS (ported from mmmc2k, pure-n): one pass over A, one over B */
AVX2 static void build_A0A1(int n, const int8_t*A){
const __m256i m=_mm256_set1_epi8(0x7F);
int W=n/64;
for (int i=0;i<n;i++){
const int8_t*row=A+(size_t)i*n;
int8_t*dst=A0b+(size_t)i*GLDA;
uint64_t*w1=A1b+(size_t)i*W;
for (int k=0;k<n;k+=64){
__m256i v0=_mm256_load_si256((const __m256i*)(row+k));
__m256i v1=_mm256_load_si256((const __m256i*)(row+k+32));
{ __m256i _w0=_mm256_and_si256(v0,m), _w1=_mm256_and_si256(v1,m);
__m256i _l0=_mm256_unpacklo_epi16(_w0,_w0), _h0=_mm256_unpackhi_epi16(_w0,_w0);
__m256i _l1=_mm256_unpacklo_epi16(_w1,_w1), _h1=_mm256_unpackhi_epi16(_w1,_w1);
_mm256_store_si256((__m256i*)(dst+2*k), _mm256_permute2x128_si256(_l0,_h0,0x20));
_mm256_store_si256((__m256i*)(dst+2*k+32), _mm256_permute2x128_si256(_l0,_h0,0x31));
_mm256_store_si256((__m256i*)(dst+2*k+64), _mm256_permute2x128_si256(_l1,_h1,0x20));
_mm256_store_si256((__m256i*)(dst+2*k+96), _mm256_permute2x128_si256(_l1,_h1,0x31)); }
w1[k>>6]=(uint64_t)(uint32_t)_mm256_movemask_epi8(v0)
|((uint64_t)(uint32_t)_mm256_movemask_epi8(v1)<<32);
}
}
}
AVX2 static void build_BI3Bp(int n, const int8_t*B){
const __m256i one=_mm256_set1_epi8(1);
int ng=n/64, W=n/64;
for (int p=0;p<n/2;p++){
const int8_t *r0=B+(size_t)(2*p)*n, *r1=r0+n;
uint64_t *w0p=Bpb+(size_t)(2*p)*W, *w1p=Bpb+(size_t)(2*p+1)*W;
for (int g=0;g<ng;g++){
const int8_t *a=r0+64*g, *b=r1+64*g;
int8_t *o=BI3b+((size_t)g*(n/2)+p)*128;
for (int t=0;t<64;t+=32){
__m256i va=_mm256_load_si256((const __m256i*)(a+t));
__m256i vb=_mm256_load_si256((const __m256i*)(b+t));
__m256i lo=_mm256_unpacklo_epi8(va,vb), hi=_mm256_unpackhi_epi8(va,vb);
_mm256_store_si256((__m256i*)(o+2*t), _mm256_permute2x128_si256(lo,hi,0x20));
_mm256_store_si256((__m256i*)(o+2*t+32), _mm256_permute2x128_si256(lo,hi,0x31));
uint32_t m0=(uint32_t)_mm256_movemask_epi8(_mm256_slli_epi16(_mm256_and_si256(va,one),7));
uint32_t m1=(uint32_t)_mm256_movemask_epi8(_mm256_slli_epi16(_mm256_and_si256(vb,one),7));
if(t==0) { w0p[g]=(uint64_t)m0; w1p[g]=(uint64_t)m1; }
else { w0p[g]|=(uint64_t)m0<<32; w1p[g]|=(uint64_t)m1<<32; }
}
}
}
}
/* ================= STRASSEN (exact over Z/256) =================
Every coefficient is +-1 and there is no division, so the identities hold in ANY commutative
ring; int8 arithmetic mod 256 is exactly that ring. The case table and the combine
coefficients are GENERATED from tables that mkstrassen4k.py's self-check brute-forces over
Z/256 before this file is written.
============================================================== */
#define SM 2048
static int8_t SA[SM*SM] __attribute__((aligned(64)));
static int8_t SB[SM*SM] __attribute__((aligned(64)));
static int8_t MMS[7][1024*1024+256] __attribute__((aligned(64)));
/* ============ FUSED QUADRANT PACKERS (--fuse) ============
The unfused path builds S2A/S2B with blk() and hands them to build_A0A1/build_BI3Bp, which
read them straight back: 2 MB written and 2 MB re-read per sub-product, x49. These write the
PACKED forms DIRECTLY from the parent's quadrants. Every stored value is identical by
construction, so this is a pure traffic deletion, not a re-derivation.
buildA_q/buildB_q(int h,int n,X,r0,c0,Y,r1,c1,op) == pack(X quad op Y quad)
========================================================= */
AVX2 static void buildA_q(int h,int n,const int8_t*X,int r0,int c0,
const int8_t*Y,int r1,int c1,int op){
const __m256i mx=_mm256_set1_epi8(0x7F);
int W=h/64;
for (int i=0;i<h;i++){
const int8_t*xa=X+(size_t)(r0+i)*n+c0, *ya=Y+(size_t)(r1+i)*n+c1;
int8_t*dst=A0b+(size_t)i*GLDA;
uint64_t*w1=A1b+(size_t)i*W;
for (int k=0;k<h;k+=64){
__m256i v0=_mm256_load_si256((const __m256i*)(xa+k));
__m256i v1=_mm256_load_si256((const __m256i*)(xa+k+32));
if (op){ __m256i u0=_mm256_load_si256((const __m256i*)(ya+k));
__m256i u1=_mm256_load_si256((const __m256i*)(ya+k+32));
if(op==1){ v0=_mm256_add_epi8(v0,u0); v1=_mm256_add_epi8(v1,u1); }
else { v0=_mm256_sub_epi8(v0,u0); v1=_mm256_sub_epi8(v1,u1); } }
{ __m256i _w0=_mm256_and_si256(v0,mx), _w1=_mm256_and_si256(v1,mx);
__m256i _l0=_mm256_unpacklo_epi16(_w0,_w0), _h0=_mm256_unpackhi_epi16(_w0,_w0);
__m256i _l1=_mm256_unpacklo_epi16(_w1,_w1), _h1=_mm256_unpackhi_epi16(_w1,_w1);
_mm256_store_si256((__m256i*)(dst+2*k), _mm256_permute2x128_si256(_l0,_h0,0x20));
_mm256_store_si256((__m256i*)(dst+2*k+32), _mm256_permute2x128_si256(_l0,_h0,0x31));
_mm256_store_si256((__m256i*)(dst+2*k+64), _mm256_permute2x128_si256(_l1,_h1,0x20));
_mm256_store_si256((__m256i*)(dst+2*k+96), _mm256_permute2x128_si256(_l1,_h1,0x31)); }
w1[k>>6]=(uint64_t)(uint32_t)_mm256_movemask_epi8(v0)
|((uint64_t)(uint32_t)_mm256_movemask_epi8(v1)<<32);
}
}
}
AVX2 static void buildB_q(int h,int n,const int8_t*X,int r0,int c0,
const int8_t*Y,int r1,int c1,int op){
const __m256i one=_mm256_set1_epi8(1);
int ng=h/64, W=h/64;
for (int p=0;p<h/2;p++){
const int8_t *ra=X+(size_t)(r0+2*p)*n+c0, *rb=X+(size_t)(r0+2*p+1)*n+c0;
const int8_t *sa=Y+(size_t)(r1+2*p)*n+c1, *sb=Y+(size_t)(r1+2*p+1)*n+c1;
uint64_t *w0p=Bpb+(size_t)(2*p)*W, *w1p=Bpb+(size_t)(2*p+1)*W;
for (int g=0;g<ng;g++){
for (int t=0;t<64;t+=32){
__m256i va=_mm256_load_si256((const __m256i*)(ra+64*g+t));
__m256i vb=_mm256_load_si256((const __m256i*)(rb+64*g+t));
if (op){ __m256i pa=_mm256_load_si256((const __m256i*)(sa+64*g+t));
__m256i pb=_mm256_load_si256((const __m256i*)(sb+64*g+t));
if(op==1){ va=_mm256_add_epi8(va,pa); vb=_mm256_add_epi8(vb,pb); }
else { va=_mm256_sub_epi8(va,pa); vb=_mm256_sub_epi8(vb,pb); } }
__m256i lo=_mm256_unpacklo_epi8(va,vb), hi=_mm256_unpackhi_epi8(va,vb);
int8_t *o=BI3b+((size_t)g*(h/2)+p)*128;
_mm256_store_si256((__m256i*)(o+2*t), _mm256_permute2x128_si256(lo,hi,0x20));
_mm256_store_si256((__m256i*)(o+2*t+32), _mm256_permute2x128_si256(lo,hi,0x31));
uint32_t mm0=(uint32_t)_mm256_movemask_epi8(_mm256_slli_epi16(_mm256_and_si256(va,one),7));
uint32_t mm1=(uint32_t)_mm256_movemask_epi8(_mm256_slli_epi16(_mm256_and_si256(vb,one),7));
if(t==0){ w0p[g]=(uint64_t)mm0; w1p[g]=(uint64_t)mm1; }
else { w0p[g]|=(uint64_t)mm0<<32; w1p[g]|=(uint64_t)mm1<<32; }
}
}
}
}
/* ---- LEVEL 2 scratch: the m x m sub-product is itself Strassen-decomposed into 7 of h=m/2 ---- */
static int8_t S2M[7][512*512+256] __attribute__((aligned(64)));
/* Cblk += sum_k s_k * M_k (up to 4 terms; NULL skips a term). dst stride is `n`, block is m x m. */
AVX2 static void cmb(int8_t*C,int n,int r0,int c0,int m,
const int8_t*M0,int s0,const int8_t*M1,int s1,
const int8_t*M2,int s2,const int8_t*M3,int s3){
for (int i=0;i<m;i++){
int8_t *d=C+(size_t)(r0+i)*n+c0;
const int8_t *p0=M0?M0+(size_t)i*m:0, *p1=M1?M1+(size_t)i*m:0;
const int8_t *p2=M2?M2+(size_t)i*m:0, *p3=M3?M3+(size_t)i*m:0;
for (int j=0;j<m;j+=32){
__m256i v;
if (p0){ __m256i t=_mm256_load_si256((const __m256i*)(p0+j));
v = s0>0?t:_mm256_sub_epi8(_mm256_setzero_si256(),t); }
else v = _mm256_setzero_si256();
if (p1){ __m256i t=_mm256_load_si256((const __m256i*)(p1+j));
v = _mm256_add_epi8(v, s1>0?t:_mm256_sub_epi8(_mm256_setzero_si256(),t)); }
if (p2){ __m256i t=_mm256_load_si256((const __m256i*)(p2+j));
v = _mm256_add_epi8(v, s2>0?t:_mm256_sub_epi8(_mm256_setzero_si256(),t)); }
if (p3){ __m256i t=_mm256_load_si256((const __m256i*)(p3+j));
v = _mm256_add_epi8(v, s3>0?t:_mm256_sub_epi8(_mm256_setzero_si256(),t)); }
_mm256_stream_si256((__m256i*)(d+j), v);
}
}
}
/* ===== lane c2k3: FUSED LEVEL-2 COMBINE. Was 4 cmb() calls per k reading 12 x 256 KB for the
7 leaves (1.75 MB) -- 8.75 MB of DRAM re-reads over the 7 k's. One row-wise pass reads each
leaf row exactly ONCE and writes all four quadrants. Identical arithmetic to the 4 cmb calls:
q00 = S0+S3-S4+S6, q01 = S2+S4, q10 = S1+S3, q11 = S0+S2-S1+S5. */
AVX2 static void cmb_s2f(int8_t*Cp,int m,const int8_t*S0,const int8_t*S1,const int8_t*S2,
const int8_t*S3,const int8_t*S4,const int8_t*S5,const int8_t*S6){
const int h=m>>1;
for (int i=0;i<h;i++){
const int8_t *r0=S0+(size_t)i*h,*r1=S1+(size_t)i*h,*r2=S2+(size_t)i*h,*r3=S3+(size_t)i*h;
const int8_t *r4=S4+(size_t)i*h,*r5=S5+(size_t)i*h,*r6=S6+(size_t)i*h;
int8_t *d0=Cp+(size_t)i*m, *d1=d0+h, *d2=Cp+(size_t)(h+i)*m, *d3=d2+h;
for (int j=0;j<h;j+=32){
__m256i a0=_mm256_load_si256((const __m256i*)(r0+j));
__m256i a1=_mm256_load_si256((const __m256i*)(r1+j));
__m256i a2=_mm256_load_si256((const __m256i*)(r2+j));
__m256i a3=_mm256_load_si256((const __m256i*)(r3+j));
__m256i a4=_mm256_load_si256((const __m256i*)(r4+j));
__m256i a5=_mm256_load_si256((const __m256i*)(r5+j));
__m256i a6=_mm256_load_si256((const __m256i*)(r6+j));
__m256i q0=_mm256_add_epi8(_mm256_sub_epi8(_mm256_add_epi8(a0,a3),a4),a6);
__m256i q1=_mm256_add_epi8(a2,a4);
__m256i q2=_mm256_add_epi8(a1,a3);
__m256i q3=_mm256_add_epi8(_mm256_sub_epi8(_mm256_add_epi8(a0,a2),a1),a5);
_mm256_stream_si256((__m256i*)(d0+j),q0);
_mm256_stream_si256((__m256i*)(d1+j),q1);
_mm256_stream_si256((__m256i*)(d2+j),q2);
_mm256_stream_si256((__m256i*)(d3+j),q3);
}
}
}
/* ===== lane c2k3: FUSED LEVEL-1 COMBINE. Same construction one level up: C's four quadrants
written in ONE pass so each of the 7 MMS blocks is read EXACTLY ONCE (was 12 MB of reads for
7 MB of data). q00 = M0+M3-M4+M6, q01 = M2+M4, q10 = M1+M3, q11 = M0+M2-M1+M5. */
AVX2 static void cmb_l1f(int8_t*C,int n,int m){
for (int i=0;i<m;i++){
const int8_t *r0=MMS[0]+(size_t)i*m,*r1=MMS[1]+(size_t)i*m,*r2=MMS[2]+(size_t)i*m;
const int8_t *r3=MMS[3]+(size_t)i*m,*r4=MMS[4]+(size_t)i*m,*r5=MMS[5]+(size_t)i*m;
const int8_t *r6=MMS[6]+(size_t)i*m;
int8_t *d0=C+(size_t)i*n, *d1=d0+m, *d2=C+(size_t)(m+i)*n, *d3=d2+m;
for (int j=0;j<m;j+=32){
__m256i a0=_mm256_load_si256((const __m256i*)(r0+j));
__m256i a1=_mm256_load_si256((const __m256i*)(r1+j));
__m256i a2=_mm256_load_si256((const __m256i*)(r2+j));
__m256i a3=_mm256_load_si256((const __m256i*)(r3+j));
__m256i a4=_mm256_load_si256((const __m256i*)(r4+j));
__m256i a5=_mm256_load_si256((const __m256i*)(r5+j));
__m256i a6=_mm256_load_si256((const __m256i*)(r6+j));
__m256i q0=_mm256_add_epi8(_mm256_sub_epi8(_mm256_add_epi8(a0,a3),a4),a6);
__m256i q1=_mm256_add_epi8(a2,a4);
__m256i q2=_mm256_add_epi8(a1,a3);
__m256i q3=_mm256_add_epi8(_mm256_sub_epi8(_mm256_add_epi8(a0,a2),a1),a5);
_mm256_stream_si256((__m256i*)(d0+j),q0);
_mm256_stream_si256((__m256i*)(d1+j),q1);
_mm256_stream_si256((__m256i*)(d2+j),q2);
_mm256_stream_si256((__m256i*)(d3+j),q3);
}
}
}
/* dst[i*h+j] = X[(r0+i)*n+(c0+j)] op Y[(r1+i)*n+(c1+j)], op: 0 copy 1 add 2 sub (mod 256) */
AVX2 static void blk(int8_t*dst,const int8_t*X,int n,int r0,int c0,int h,
const int8_t*Y,int r1,int c1,int op){
for (int i=0;i<h;i++){
const int8_t *xa=X+(size_t)(r0+i)*n+c0, *ya=Y+(size_t)(r1+i)*n+c1;
int8_t *d=dst+(size_t)i*h;
int j=0;
if (op==0) for(;j+32<=h;j+=32) _mm256_store_si256((__m256i*)(d+j),
_mm256_load_si256((const __m256i*)(xa+j)));
else if (op==1) for(;j+32<=h;j+=32) _mm256_store_si256((__m256i*)(d+j),
_mm256_add_epi8(_mm256_load_si256((const __m256i*)(xa+j)),
_mm256_load_si256((const __m256i*)(ya+j))));
else for(;j+32<=h;j+=32) _mm256_store_si256((__m256i*)(d+j),
_mm256_sub_epi8(_mm256_load_si256((const __m256i*)(xa+j)),
_mm256_load_si256((const __m256i*)(ya+j))));
if (op) for(;j<h;j++) d[j]=(int8_t)(xa[j]+(op==1?ya[j]:-ya[j]));
}
}
/* The GF(2) + combine tail of a sub-product, i.e. everything after the operand packers. Split
out so the FUSED path can reach it without materialising S2A/S2B first. */
AVX2BMI static void corr1024(void);
AVX2BMI static void corr512(void);
AVX2BMI static void corr256(void);
/* lane BW9: unrolled mirror of corr1024 for the m=512 leaf (W=8) */
AVX2BMI static void corr512(void){
const int n=512, W=8;
memset(Pbb,0,(size_t)n*W*8);
const int KB=64;
for (int kb=0; kb<n; kb+=KB){
int w0=kb>>6;
const char *bp0=(const char*)Bpb+(size_t)w0*4096;
for (int i=0;i<n;i++){
const uint64_t*row=A1b+(size_t)i*W+w0;
uint64_t b0=row[0], b1=0;
uint64_t *o=Pbb+(size_t)i*W;
asm volatile(
"vpxor %%xmm0,%%xmm0,%%xmm0\n\t"
"vpxor %%xmm1,%%xmm1,%%xmm1\n\t"
"testq %[b0], %[b0]\n\t"
"jz 3%=f\n\t"
"1%=:\n\t"
"tzcntq %[b0], %%r11\n\t"
"shlq $6, %%r11\n\t"
"addq %[bp0], %%r11\n\t"
"vpxor 0(%%r11), %%ymm0, %%ymm0\n\t"
"vpxor 32(%%r11), %%ymm1, %%ymm1\n\t"
"blsrq %[b0], %[b0]\n\t"
"jnz 1%=b\n\t"
"3%=:\n\t"
"testq %[b1], %[b1]\n\t"
"jz 4%=f\n\t"
"2%=:\n\t"
"tzcntq %[b1], %%r11\n\t"
"shlq $6, %%r11\n\t"
"addq %[bp0], %%r11\n\t"
"vpxor 0(%%r11), %%ymm0, %%ymm0\n\t"
"vpxor 32(%%r11), %%ymm1, %%ymm1\n\t"
"blsrq %[b1], %[b1]\n\t"
"jnz 2%=b\n\t"
"4%=:\n\t"
"vpxor 0(%[o]), %%ymm0, %%ymm0\n\t"
"vmovdqu %%ymm0, 0(%[o])\n\t"
"vpxor 32(%[o]), %%ymm1, %%ymm1\n\t"
"vmovdqu %%ymm1, 32(%[o])\n\t"
: [b0]"+r"(b0), [b1]"+r"(b1)
: [bp0]"r"(bp0), [o]"r"(o)
: "r11","cc","memory","ymm0","ymm1");
}
}
}
/* lane BW9: unrolled mirror of corr1024 for the m=256 leaf (W=4) */
AVX2BMI static void corr256(void){
const int n=256, W=4;
memset(Pbb,0,(size_t)n*W*8);
const int KB=64;
for (int kb=0; kb<n; kb+=KB){
int w0=kb>>6;
const char *bp0=(const char*)Bpb+(size_t)w0*2048;
for (int i=0;i<n;i++){
const uint64_t*row=A1b+(size_t)i*W+w0;
uint64_t b0=row[0], b1=0;
uint64_t *o=Pbb+(size_t)i*W;
asm volatile(
"vpxor %%xmm0,%%xmm0,%%xmm0\n\t"
"testq %[b0], %[b0]\n\t"
"jz 3%=f\n\t"
"1%=:\n\t"
"tzcntq %[b0], %%r11\n\t"
"shlq $5, %%r11\n\t"
"addq %[bp0], %%r11\n\t"
"vpxor 0(%%r11), %%ymm0, %%ymm0\n\t"
"blsrq %[b0], %[b0]\n\t"
"jnz 1%=b\n\t"
"3%=:\n\t"
"testq %[b1], %[b1]\n\t"
"jz 4%=f\n\t"
"2%=:\n\t"
"tzcntq %[b1], %%r11\n\t"
"shlq $5, %%r11\n\t"
"addq %[bp0], %%r11\n\t"
"vpxor 0(%%r11), %%ymm0, %%ymm0\n\t"
"blsrq %[b1], %[b1]\n\t"
"jnz 2%=b\n\t"
"4%=:\n\t"
"vpxor 0(%[o]), %%ymm0, %%ymm0\n\t"
"vmovdqu %%ymm0, 0(%[o])\n\t"
: [b0]"+r"(b0), [b1]"+r"(b1)
: [bp0]"r"(bp0), [o]"r"(o)
: "r11","cc","memory","ymm0");
}
}
}
/* corr1024(): a hand-unrolled mirror of the SHIPPED corr16 for n=1024 (W=16, NY=4), which
corr16's own `if (n==4096)` dispatch never reaches. corr_pass3's inner `for(q<NY)` is
runtime-bounded, so GCC cannot unroll it; corr16 exists because the unrolled form pays.
Only the constants differ from corr16: 4 accumulators instead of 16, Bpb row stride 128 B
instead of 512 B (shlq $7), and offsets 0/32/64/96 instead of 0..480. */
AVX2BMI static void corr1024(void){
const int n=1024, W=16;
memset(Pbb,0,(size_t)n*W*8);
const int KB=64;
for (int kb=0; kb<n; kb+=KB){
int w0=kb>>6;
const char *bp0=(const char*)Bpb+(size_t)w0*8192;
for (int i=0;i<n;i++){
const uint64_t*row=A1b+(size_t)i*W+w0;
uint64_t b0=row[0], b1=0;
uint64_t *o=Pbb+(size_t)i*W;
asm volatile(
"vpxor %%xmm0,%%xmm0,%%xmm0\n\t"
"vpxor %%xmm1,%%xmm1,%%xmm1\n\t"
"vpxor %%xmm2,%%xmm2,%%xmm2\n\t"
"vpxor %%xmm3,%%xmm3,%%xmm3\n\t"
"testq %[b0], %[b0]\n\t"
"jz 3%=f\n\t"
"1%=:\n\t"
"tzcntq %[b0], %%r11\n\t"
"shlq $7, %%r11\n\t"
"addq %[bp0], %%r11\n\t"
"vpxor 0(%%r11), %%ymm0, %%ymm0\n\t"
"vpxor 32(%%r11), %%ymm1, %%ymm1\n\t"
"vpxor 64(%%r11), %%ymm2, %%ymm2\n\t"
"vpxor 96(%%r11), %%ymm3, %%ymm3\n\t"
"blsrq %[b0], %[b0]\n\t"
"jnz 1%=b\n\t"
"3%=:\n\t"
"testq %[b1], %[b1]\n\t"
"jz 4%=f\n\t"
"2%=:\n\t"
"tzcntq %[b1], %%r11\n\t"
"shlq $7, %%r11\n\t"
"addq %[bp0], %%r11\n\t"
"vpxor 0(%%r11), %%ymm0, %%ymm0\n\t"
"vpxor 32(%%r11), %%ymm1, %%ymm1\n\t"
"vpxor 64(%%r11), %%ymm2, %%ymm2\n\t"
"vpxor 96(%%r11), %%ymm3, %%ymm3\n\t"
"blsrq %[b1], %[b1]\n\t"
"jnz 2%=b\n\t"
"4%=:\n\t"
"vpxor 0(%[o]), %%ymm0, %%ymm0\n\t"
"vmovdqu %%ymm0, 0(%[o])\n\t"
"vpxor 32(%[o]), %%ymm1, %%ymm1\n\t"
"vmovdqu %%ymm1, 32(%[o])\n\t"
"vpxor 64(%[o]), %%ymm2, %%ymm2\n\t"
"vmovdqu %%ymm2, 64(%[o])\n\t"
"vpxor 96(%[o]), %%ymm3, %%ymm3\n\t"
"vmovdqu %%ymm3, 96(%[o])\n\t"
: [b0]"+r"(b0), [b1]"+r"(b1)
: [bp0]"r"(bp0), [o]"r"(o)
: "r11","cc","memory","ymm0","ymm1","ymm2","ymm3");
}
}
}
/* ===================== lane CH3: NIBBLE-TABLE GF(2) PRODUCT =====================
P[i][:] = XOR_{k: A1[i][k]=1} Bpb[k][:] -- the same contract as corr512/corr1024,
verified bit-identical against them on the real judge (work/lane_ch3_mmmc/rig/).
Rationale and measurements: work/lane_ch3_mmmc/mk_nibcorr.py docstring.
GROUPS = 16 groups (= one 64-bit A1 word) per k-block; table = 16*16*W words. */
static uint64_t TB[((size_t)NMAX/64)*256] __attribute__((aligned(64)));
#define DEF_NIBCORR(TAG, M) \
AVX2BMI static void nibcorr##TAG(void){ \
const int n=M, W=n/64, G=n>>2, GB=16, NY=W>>2; \
memset(Pbb,0,(size_t)n*W*8); \
for (int g0=0; g0<G; g0+=GB){ \
for (int t=0;t<GB;t++){ \
const int g=g0+t; \
uint64_t *T=TB+(size_t)t*16*W; \
const uint64_t *r=Bpb+(size_t)(g*4)*W; \
for (int w=0;w<W;w++){ \
uint64_t a1=r[w], a2=r[W+w], a4=r[2*W+w], a8=r[3*W+w]; \
uint64_t x12=a1^a2, x14=a1^a4, x18=a1^a8; \
T[ 0*W+w]=0; \
T[ 1*W+w]=a1; T[ 2*W+w]=a2; T[ 4*W+w]=a4; T[ 8*W+w]=a8; \
T[ 3*W+w]=x12; T[ 5*W+w]=x14; T[ 6*W+w]=a2^a4; T[ 7*W+w]=x12^a4; \
T[ 9*W+w]=x18; T[10*W+w]=a2^a8; T[12*W+w]=a4^a8; \
T[11*W+w]=a2^x18; T[13*W+w]=a4^x18; T[14*W+w]=a2^a4^a8; \
T[15*W+w]=x12^a4^a8; \
} \
} \
for (int i=0;i<n;i++){ \
const uint64_t *arow=A1b+(size_t)i*W+(g0>>4); \
__m256i acc[NY]; \
for (int q=0;q<NY;q++) acc[q]=_mm256_setzero_si256(); \
const uint64_t *T=TB; \
for (int wb=0; wb<(GB>>4); wb++){ \
const uint64_t word=arow[wb]; \
_Pragma("GCC unroll 16") \
for (int b=0;b<16;b++, T+=(size_t)16*W){ \
const uint64_t *e=T+(size_t)(((word>>(4*b))&15u)*W); \
for (int q=0;q<NY;q++) \
acc[q]=_mm256_xor_si256(acc[q],_mm256_load_si256((const __m256i*)(e+4*q))); \
} \
} \
uint64_t *o=Pbb+(size_t)i*W; \
for (int q=0;q<NY;q++){ \
__m256i *po=(__m256i*)(o+4*q); \
_mm256_store_si256(po,_mm256_xor_si256(_mm256_load_si256((const __m256i*)po),acc[q])); \
} \
} \
} \
}
DEF_NIBCORR(512,512)
DEF_NIBCORR(1024,1024)
AVX2 static void submm_tail(int m,int8_t*Cp){
GLDA = 2*(m) + PADA; GLDB = m + PADB;
if (m==1024) corr1024(); else if (m==512) nibcorr512(); else if (m==256) corr256();
else corr_pass3(m); expand_bits(m);
const int GB=64, IB=128;
for (int ib=0; ib<m; ib+=IB){
const int ihi = (ib+IB<m)? ib+IB : m;
for (int gb=0; gb<m; gb+=GB){
const int ghi = (gb+GB<m)? gb+GB : m;
for (int g=gb; g<ghi; g+=64){
const int8_t *bi = BI3b+(size_t)(g>>6)*(m/2)*128;
{ int rev=0;
for (int i=ib; i<ihi; i+=2){
/* R1 boustrophedon: every other call walks A and BI BACKWARD, so it starts
where the previous call stopped -- that data is still in L1. */
const int8_t *ap=A0b+(size_t)i*GLDA;
if (rev) mikro_u4r(m, ap+(2*m)-16, bi,
Cp+(size_t)i*m+g, PBb+(size_t)i*GLDB+g, 0, GLDA, GLDB);
else mikro_u4 (m, ap, bi,
Cp+(size_t)i*m+g, PBb+(size_t)i*GLDB+g, 0, GLDA, GLDB);
rev^=1; }
}
}
}
}
}
/* one m x m sub-product through the SHIPPED pipeline, NT-stored to Cp (m = 2048 in the
shipped dispatch; `m` is a parameter so the SAME code path gates at smaller m). */
AVX2 static void submm_plain(int m,const int8_t*Ap,const int8_t*Bp,int8_t*Cp){
GLDA = 2*(m) + PADA; GLDB = m + PADB;
build_A0A1(m,Ap); build_BI3Bp(m,Bp); submm_tail(m,Cp);
}
AVX2 static void submm(int m,const int8_t*Ap,const int8_t*Bp,int8_t*Cp);
#define S2MIN 1024
/* level 2: 7 sub-products of h=m/2, recombined by the SAME verified identity.
⛔ THE SUB-PRODUCT IS COMPUTED BY `submm_plain`, **NOT** BY `submm`.
Calling `submm(h,...)` re-enters this function at h, and the inner call **overwrites the
parent's own `S2A`/`S2B`/`S2M` scratch** -- so every `S2M[k]` except the last is destroyed
before the combine reads it. That is a 100 %-wrong build, and the 6-second A=I probe at
n=4096 found it (87 % of bytes wrong, all four blocks). If a deeper level is ever wanted,
it needs its own scratch, not the shared pool. */
AVX2 static void submm_s2(int m,const int8_t*Ap,const int8_t*Bp,int8_t*Cp){
GLDA = 2*((m>>1)) + PADA; GLDB = (m>>1) + PADB;
const int h=m>>1, o=h;
for (int k=0;k<7;k++){
switch(k){
case 0: buildA_q(h,m,Ap,0,0,Ap,o,o,1); buildB_q(h,m,Bp,0,0,Bp,o,o,1); submm_tail(h,S2M[0]); break;
case 1: buildA_q(h,m,Ap,o,0,Ap,o,o,1); buildB_q(h,m,Bp,0,0,Bp,0,0,0); submm_tail(h,S2M[1]); break;
case 2: buildA_q(h,m,Ap,0,0,Ap,0,0,0); buildB_q(h,m,Bp,0,o,Bp,o,o,2); submm_tail(h,S2M[2]); break;
case 3: buildA_q(h,m,Ap,o,o,Ap,o,o,0); buildB_q(h,m,Bp,o,0,Bp,0,0,2); submm_tail(h,S2M[3]); break;
case 4: buildA_q(h,m,Ap,0,0,Ap,0,o,1); buildB_q(h,m,Bp,o,o,Bp,o,o,0); submm_tail(h,S2M[4]); break;
case 5: buildA_q(h,m,Ap,o,0,Ap,0,0,2); buildB_q(h,m,Bp,0,0,Bp,0,o,1); submm_tail(h,S2M[5]); break;
case 6: buildA_q(h,m,Ap,0,o,Ap,o,o,2); buildB_q(h,m,Bp,o,0,Bp,o,o,1); submm_tail(h,S2M[6]); break;
}
}
_mm_sfence();
cmb_s2f(Cp,m, S2M[0],S2M[1],S2M[2],S2M[3],S2M[4],S2M[5],S2M[6]);
_mm_sfence();
}
AVX2 static void submm(int m,const int8_t*Ap,const int8_t*Bp,int8_t*Cp){
if (m>=S2MIN) { submm_s2(m,Ap,Bp,Cp); return; }
submm_plain(m,Ap,Bp,Cp);
}
AVX2 static void strassen1(int n,const int8_t*A,const int8_t*B,int8_t*C){
const int m=n>>1, o=m;
for (int k=0;k<7;k++){
switch(k){
case 0: blk(SA,A,n,0,0,m,A,o,o,1); blk(SB,B,n,0,0,m,B,o,o,1); break;
case 1: blk(SA,A,n,o,0,m,A,o,o,1); blk(SB,B,n,0,0,m,B,0,0,0); break;
case 2: blk(SA,A,n,0,0,m,A,0,0,0); blk(SB,B,n,0,o,m,B,o,o,2); break;
case 3: blk(SA,A,n,o,o,m,A,o,o,0); blk(SB,B,n,o,0,m,B,0,0,2); break;
case 4: blk(SA,A,n,0,0,m,A,0,o,1); blk(SB,B,n,o,o,m,B,o,o,0); break;
case 5: blk(SA,A,n,o,0,m,A,0,0,2); blk(SB,B,n,0,0,m,B,0,o,1); break;
case 6: blk(SA,A,n,0,o,m,A,o,o,2); blk(SB,B,n,o,0,m,B,o,o,1); break;
}
submm(m,SA,SB,MMS[k]);
}
_mm_sfence();
cmb_l1f(C,n,m);
_mm_sfence();
}
AVX2 void matrix_multiply(int n, const int8_t *A, const int8_t *B, int8_t *C){
if (n<=0 || (n&255) || n>NMAX) { generic_mm(n,A,B,C); return; }
if (n==2048) { strassen1(n,A,B,C); return; }
/* x4: A1T is written by build_A1T and NEVER READ -- dead 16.78 MB + its transpose pass. */
/* FUSED: one pass over A (was two), one over B (was two) */
GLDA = 2*(n) + PADA; GLDB = n + PADB;
build_A0A1(n,A); build_BI3Bp(n,B);
if (n==4096) corr16(); else if (n==1024) corr1024();
else corr_pass3(n); expand_bits(n);
for (int g=0; g<n; g+=64)
for (int i=0; i<n; i+=2)
mikro_u4(n, A0b+(size_t)i*GLDA, BI3b+(size_t)(g>>6)*(n/2)*128,
C+(size_t)i*n+g, PBb+(size_t)i*GLDB+g, 0, GLDA, GLDB);
_mm_sfence();
}
| Compilation | N/A | N/A | Compile OK | Score: N/A | 显示更多 |
| Testcase #1 | 53.05 ms | 15 MB + 928 KB | Accepted | Score: 100 | 显示更多 |