Skip to content

Q4/Q8 Tiled Gemm Optimization. #16999

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Q4/Q8 Tiled Gemm Optimization. #16999

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
308 changes: 283 additions & 25 deletions ggml/src/ggml-cpu/llamafile/sgemm.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -119,6 +119,30 @@ inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vec_mul(x, y); }
#if defined(__MMA__)
typedef vector unsigned char vec_t;
typedef __vector_quad acc_t;

#include <pthread.h>

typedef vector unsigned char vec_t;
typedef __vector_quad acc_t;

static pthread_key_t t_data_key;
typedef struct {
vec_t* A_pack;
vec_t* B_pack;
int* comparray;
} thread_scratchpad_t;
void thread_cleanup(void* arg) {
thread_scratchpad_t* data = (thread_scratchpad_t*)arg;
if (data) {
delete[] data->A_pack;
delete[] data->B_pack;
delete[] data->comparray;

delete data;
}
}
static bool key_created = false;

Comment on lines +122 to +145
Copy link
Member

@ggerganov ggerganov Nov 5, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It would be better to avoid dynamic allocations - none of the code currently uses those. The mechanism for this is to use the wdata from ggml_compute_params to store scratch data. You'll need to reserve the worst-case wsize for your case.

Copy link
Contributor Author

@shalinib-ibm shalinib-ibm Nov 6, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@ggerganov Thank you for the input. I tried to avoid dynamic allocation from the code, but lost perforamce without this pthread based code. Below is the code Performance comparison after integrating thread-local scratchpad using wdata.

void matmul_tiled(const ggml_compute_params* params,
     int64_t m, int64_t n, int64_t mc, int64_t nc, int64_t kc) {  
     char* wdata = (char*) params->wdata;
    constexpr size_t ALIGN = 128;
    auto align_ptr = [&](char* ptr, size_t alignment) {
         return (char*)(((uintptr_t)ptr + alignment - 1) & ~(alignment - 1));
    };
    char* ptr = align_ptr(wdata, ALIGN);
    vec_t*  A_pack = (vec_t*)ptr;  ptr += sizeof(vec_t) * mc * kc * 2;
    vec_t*  B_pack = (vec_t*)ptr;  ptr += sizeof(vec_t) * nc * kc * 2;
    int* comparray = (int*)align_ptr(ptr, ALIGN);  // integer part aligned too
    ptr += sizeof(int) * mc * kc;
    // rest of your original matmul_tiled() code unchanged
}
Benchmark (llama-bench) Baseline pthread-based TLS ggml wdata-based TLS
pp128 69 t/s 89 t/s 36 t/s
pp256 69 t/s 94 t/s 36 t/s

This regression is likely due to:

  1. Loss of persistent per-thread cache locality — the previous pthread-based version reused buffers effectively across tiles.
  2. Higher memory initialization or shared buffer contention across threads.

I have also tried static allocation on stack by having just this code , But it has suffers from similar perf loss. ( 38 t/s)
vec_t A_pack [mckc2];
vec_t B_pack[nckc2];
int comparray[mc*kc];

Can you please suggest ?

Copy link
Member

@ggerganov ggerganov Nov 7, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

void matmul_tiled()

You wdata pointer has to take into account the thread ID. For example:

llama.cpp/ggml/src/ggml-cpu/ops.cpp

Lines 615 to 617 in 7f09a68

float * wdata = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;

From what I can tell from your codeblock, all threads are currently working on the same wdata range and most likely they are data racing. Unless I'm missing something 🤔

#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// VECTORIZED FUSED MULTIPLY ADD
Expand Down Expand Up @@ -621,6 +645,7 @@ class tinyBLAS_Q0_ARM {
const block_q8_0 *const B;
float *const C;
const int64_t k;
int64_t kc;
const int64_t lda;
const int64_t ldb;
const int64_t ldc;
Expand Down Expand Up @@ -1582,34 +1607,73 @@ class tinyBLAS_Q0_PPC {
float *C, int64_t ldc,
int ith, int nth)
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
kc = 64;
}

void matmul(int64_t m, int64_t n) {
mnpack(0, m, 0, n);
int mc = 64; int nc = 64;
const bool is_aligned = ((m & (mc - 1)) == 0) & ((n & (nc - 1)) == 0) & ((k & (kc - 1)) == 0);
if (is_aligned) {
matmul_tiled(m, n, mc, nc, kc);
} else {
mnpack(0, m, 0, n);
}
}

private:

inline void save_res(int ii, int jj, int idx, vector float* fin_res, int RM=4, int RN=4) {
for (int I = 0; I < RM; I++) {
for (int J = 0; J < RN; J++) {
*((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&fin_res[idx+I]+J);
}
}
for (int I = 0; I < RM; I++) {
for (int J = 0; J < RN; J++) {
*((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&fin_res[idx+I]+J);
}
}
}

template<int size>
inline void compute(acc_t* ACC, int c_idx, int s_idx, std::array<int, size>& comparray, vector float* vs, vector float* fin_res) {
vector signed int vec_C[4];
vector float CA[4] = {0};
vector float res[4] = {0};
__builtin_mma_disassemble_acc(vec_C, ACC);
for (int i = 0; i < 4; i++) {
CA[i] = vec_splats((float)(((double)comparray[c_idx+i]) * -128.0));
res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
fin_res[s_idx+i] = vec_madd(res[i], vs[s_idx+i], fin_res[s_idx+i]);
}
inline void add_save_res(int ii, int jj, int idx, vector float* fin_res, int RM=4, int RN=4) {
for (int I = 0; I < RM; I++) {
for (int J = 0; J < RN; J++) {
float * c_ptr = (float *)(C+ii+((jj+J)*ldc)+I);
*c_ptr += *((float*)&fin_res[idx+I]+J);
}
}
}

template<typename ArrayType>
inline void compute(acc_t* ACC, int c_idx, int s_idx, ArrayType& comparray, vector float* vs, vector float* fin_res) {
vector signed int vec_C[4];
vector float CA[4] = {0};
vector float res[4] = {0};
__builtin_mma_disassemble_acc(vec_C, ACC);
for (int i = 0; i < 4; i++) {
CA[i] = vec_splats((float)(((double)comparray[c_idx+i]) * -128.0));
res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
fin_res[s_idx+i] = vec_madd(res[i], vs[s_idx+i], fin_res[s_idx+i]);
}
}

/*inline void compute_array(acc_t* ACC, int c_idx, int s_idx, int* comparray, vector float* vs, vector float* fin_res) {
vector signed int vec_C[4];
vector float CA[4] = {0};
vector float res[4] = {0};
__builtin_mma_disassemble_acc(vec_C, ACC);
for (int i = 0; i < 4; i++) {
CA[i] = vec_splats((float)(((double)comparray[c_idx+i]) * -128.0));
res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
fin_res[s_idx+i] = vec_madd(res[i], vs[s_idx+i], fin_res[s_idx+i]);
}
}*/

void compute_scale(int64_t ii, int64_t jj, int blk, vector float* vs){
for (int I = 0; I<8; I++) {
float a_scale = unhalf((A+((ii+I)*lda)+blk)->d);
for (int J = 0; J<4; J++) {
*((float*)&vs[I]+J) = (a_scale * unhalf((B+((jj+J)*ldb)+blk)->d));
*((float*)&vs[I+8]+J) = (a_scale * unhalf((B+((jj+J+4)*ldb)+blk)->d));
}
}
}

/* This function processes quantized data from block_q4_0 elements.
* First the we try to extract the two int4 values stored in single int8_t into two signed int8.
* And then we subtract each of the resultant element with 8, to convert signed int8 to unsigned int8.
Expand All @@ -1630,6 +1694,18 @@ class tinyBLAS_Q0_PPC {
*(ca) = vsum[0] + vsum[1] + vsum[2] + vsum[3];
}

inline void process_q8_elements(const int8_t *qs, int *ca) {
vector signed char c1 = vec_xl(0, qs);
vector signed char c2 = vec_xl(16, qs);
vector signed int vsum1 = {0};
vector signed int vsum2 = {0};
vsum1 = vec_sum4s(c1, vsum1);
vsum2 = vec_sum4s(c2, vsum2);
vector signed int vsum = vec_add(vsum1, vsum2);
*ca = vsum[0] + vsum[1] + vsum[2] + vsum[3];
}


template <typename V1, typename V2>
inline void vector_permute_store(V2 &s1, V2 &s2, V2 &s3, V2 &s4, V1 *vecOffset, bool flip) {
vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
Expand Down Expand Up @@ -1877,6 +1953,98 @@ class tinyBLAS_Q0_PPC {
}
}
}
template<typename VA, typename VB>
void packNormal_large(const block_q8_0* a, int64_t lda, int rows, int cols, VA* vec, bool flip, int* comparray=nullptr) {
int64_t i, j;
block_q8_0 *aoffset = NULL;
VA *vecOffset = NULL;
block_q8_0* aoffsets[8];
__vector_pair arr[8];
VB c[8][2] = {0};
VB c1[8] = {0}; VB c2[8] = {0};
aoffset = const_cast<block_q8_0*>(a);
vecOffset = vec;
j = (rows >> 3);
int index = 0;
if (j > 0) {
do {
for (int it = 0; it < 8; it++)
aoffsets[it] = aoffset + it*lda;
aoffset += 8 * lda;
for (int blk = 0; blk < kc; blk++) {
for (int it = 0; it < 8; it++) {
arr[it] = __builtin_vsx_lxvp(0, (__vector_pair*)(aoffsets[it]+blk)->qs);
__builtin_vsx_disassemble_pair(c[it], &arr[it]);
c1[it] = c[it][0];
c2[it] = c[it][1];
if (comparray){
process_q8_elements((aoffsets[it]+ blk)->qs, &comparray[index + 8*blk + it]);
}
}
vector_permute_store<VA, VB>(c1[0], c1[1], c1[2], c1[3], vecOffset, flip);
vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset+64, flip);
vector_permute_store<VA, VB>(c1[4], c1[5], c1[6], c1[7], vecOffset+128, flip);
vector_permute_store<VA, VB>(c2[4], c2[5], c2[6], c2[7], vecOffset+192, flip);
vecOffset += 256;
}
j--;
index += 8*kc;
} while(j > 0);
}
}

void packNormalInt4_large(const TA* a, int64_t lda, int rows, int cols, int8_t* vec, int*comparray) {
int64_t i, j;
TA *aoffset = NULL;
int8_t *vecOffset = NULL;
TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
vector signed char c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2] = {0};
vector signed char c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2] = {0};
aoffset = const_cast<TA*>(a);
vecOffset = vec;
int index = 0;
j = (rows >> 3);
if (j > 0) {
do {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
aoffset4 = aoffset3 + lda;
aoffset5 = aoffset4 + lda;
aoffset6 = aoffset5 + lda;
aoffset7 = aoffset6 + lda;
aoffset8 = aoffset7 + lda;
aoffset += 8 * lda;
for (int blk = 0; blk < kc; blk++) {
c1[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset1+blk)->qs));
c2[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset2+blk)->qs));
c3[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset3+blk)->qs));
c4[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset4+blk)->qs));
c5[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset5+blk)->qs));
c6[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset6+blk)->qs));
c7[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset7+blk)->qs));
c8[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset8+blk)->qs));

process_q4_elements(c1, &comparray[index + 8*blk+0]);
process_q4_elements(c2, &comparray[index + 8*blk+1]);
process_q4_elements(c3, &comparray[index + 8*blk+2]);
process_q4_elements(c4, &comparray[index + 8*blk+3]);
process_q4_elements(c5, &comparray[index + 8*blk+4]);
process_q4_elements(c6, &comparray[index + 8*blk+5]);
process_q4_elements(c7, &comparray[index + 8*blk+6]);
process_q4_elements(c8, &comparray[index + 8*blk+7]);
vector_permute_store<int8_t, vector signed char>(c1[0], c2[0], c3[0], c4[0], vecOffset, false);
vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset+64, false);
vector_permute_store<int8_t, vector signed char>(c5[0], c6[0], c7[0], c8[0], vecOffset+128, false);
vector_permute_store<int8_t, vector signed char>(c5[1], c6[1], c7[1], c8[1], vecOffset+192, false);
vecOffset += 256;
}
j--;
index += 8*kc;
} while (j > 0);
}
}

void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int m_rem = MIN(m - m0, 16);
Expand Down Expand Up @@ -1953,8 +2121,8 @@ class tinyBLAS_Q0_PPC {
aoffset += lda;
}
}
compute<4>(&acc_0, 0, 0, comparray, vs, fin_res);
compute<4>(&acc_1, 0, 4, comparray, vs, fin_res);
compute(&acc_0, 0, 0, comparray, vs, fin_res);
compute(&acc_1, 0, 4, comparray, vs, fin_res);
}
save_res(ii, jj, 0, fin_res);
save_res(ii, jj+4, 4, fin_res);
Expand Down Expand Up @@ -1997,8 +2165,8 @@ class tinyBLAS_Q0_PPC {
aoffset += lda;
}
}
compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
compute(&acc_0, 0, 0, comparray, vs, fin_res);
compute(&acc_1, 4, 4, comparray, vs, fin_res);
}
save_res(ii, jj, 0, fin_res);
save_res(ii+4, jj, 4, fin_res);
Expand Down Expand Up @@ -2046,17 +2214,106 @@ class tinyBLAS_Q0_PPC {
aoffset += lda;
}
}
compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
compute<8>(&acc_2, 0, 8, comparray, vs, fin_res);
compute<8>(&acc_3, 4, 12, comparray, vs, fin_res);
compute(&acc_0, 0, 0, comparray, vs, fin_res);
compute(&acc_1, 4, 4, comparray, vs, fin_res);
compute(&acc_2, 0, 8, comparray, vs, fin_res);
compute(&acc_3, 4, 12, comparray, vs, fin_res);
}
save_res(ii, jj, 0, fin_res);
save_res(ii+4, jj, 4, fin_res);
save_res(ii, jj+4, 8, fin_res);
save_res(ii+4, jj+4, 12, fin_res);
}

void KERNEL_Q0(int64_t ii, int64_t jj, int64_t mc, int64_t nc, int64_t kc, int64_t l, vec_t *vec_A, vec_t *vec_B, int *comparray) {
acc_t acc[4];
for (int i = 0; i < mc ; i += 8) {
for (int j = 0; j < nc; j += 8) {
vector float fin_res[16] = {0};
vector float vs[16] = {0};
for (int64_t kk = 0; kk < kc; kk++) {
for (int x = 0; x < 4; x++) {
__builtin_mma_xxsetaccz(&acc[x]);
}
int A_block_idx = (i/8)*(16*kc) + kk*16;
int B_block_idx = (j/8)*(16*kc)+ kk*16;
vec_t *A_block = &vec_A[A_block_idx];
vec_t *B_block = &vec_B[B_block_idx];
for (int x = 0; x < 8; x++) {
__builtin_mma_xvi8ger4pp(&acc[0], A_block[x], B_block[x]);
__builtin_mma_xvi8ger4pp(&acc[1], A_block[x + 8], B_block[x]);
__builtin_mma_xvi8ger4pp(&acc[2], A_block[x], B_block[x+8]);
__builtin_mma_xvi8ger4pp(&acc[3], A_block[x+8], B_block[x+8]);
}
compute_scale(ii+i, jj+j, l+kk, vs);
int c_index = (i/8)*(8*kc)+ kk*8;
int* c_block = &comparray[c_index];
compute(&acc[0], 0, 0, c_block, vs, fin_res);
compute(&acc[1], 4, 4, c_block, vs, fin_res);
compute(&acc[2], 0, 8, c_block, vs, fin_res);
compute(&acc[3], 4, 12, c_block, vs, fin_res);
}
if (l == 0) {
save_res(ii+i, jj+j, 0, fin_res);
save_res(ii+i+4, jj+j, 4, fin_res);
save_res(ii+i, jj+j+4, 8, fin_res);
save_res(ii+i+4, jj+j+4, 12, fin_res);
} else {
add_save_res(ii+i, jj+j, 0, fin_res);
add_save_res(ii+i+4, jj+j, 4, fin_res);
add_save_res(ii+i, jj+j+4, 8, fin_res);
add_save_res(ii+i+4, jj+j+4, 12, fin_res);
}
}
}
}

void matmul_tiled(int64_t m, int64_t n, int64_t mc, int64_t nc, int64_t kc) {
if (!key_created) {
if (pthread_key_create(&t_data_key, thread_cleanup) == 0) {
key_created = true;
} else {
return;
}
}
thread_scratchpad_t* t_data = (thread_scratchpad_t*)pthread_getspecific(t_data_key);
if (t_data == nullptr) {
t_data = new thread_scratchpad_t;
t_data->A_pack = new vec_t[mc * kc * 2];
t_data->B_pack = new vec_t[nc * kc * 2];
t_data->comparray = new int[mc * kc];
pthread_setspecific(t_data_key, t_data);
}
vec_t* A_pack = t_data->A_pack;
vec_t* B_pack = t_data->B_pack;
int* comparray = t_data->comparray;

constexpr bool is_Ablock_q4 = std::is_same_v<TA, block_q4_0>;
int64_t ytiles = m / mc;
int64_t xtiles = n / nc;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (end > tiles) {
end = tiles;
}
for (int64_t job = start; job < end; ++job) {
int64_t ii = (job / xtiles) * mc;
int64_t jj = (job % xtiles) * nc;
for (int64_t kk = 0; kk < k; kk += kc) {
if constexpr(is_Ablock_q4) {
packNormalInt4_large(A + ii*lda + kk, lda, mc, 4, (int8_t*)A_pack, comparray);
} else {
packNormal_large<int8_t, vector signed char>(A + ii*lda + kk, lda, mc, 8, (int8_t*)A_pack, false, comparray);
}
packNormal_large<uint8_t, vector unsigned char>(B + jj*ldb + kk, ldb, nc, 8, (uint8_t*)B_pack, true);
KERNEL_Q0(ii, jj, mc, nc, kc, kk, A_pack, B_pack, comparray);
}
}
}


void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
Expand Down Expand Up @@ -2159,6 +2416,7 @@ class tinyBLAS_Q0_PPC {
const block_q8_0 *const B;
float *C;
const int64_t k;
int64_t kc;
const int64_t lda;
const int64_t ldb;
const int64_t ldc;
Expand Down
Loading