happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Cleanup, consistent style

This commit is contained in:
Alberto Cabrera 2025-12-16 12:33:26 +00:00
parent 815123c472
commit 26ef6770d0
1 changed files with 25 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

51
ggml/src/ggml-cpu/repack.cpp
View File

@ -751,6 +751,7 @@ void ggml_gemv_q8_0_4x8_q8_0_generic(int n,
const int ncols_interleaved = 4; const int ncols_interleaved = 4;
const int blocklen = 8; const int blocklen = 8;
assert(nr == 1);
assert(n % qk == 0); assert(n % qk == 0);
assert(nc % ncols_interleaved == 0); assert(nc % ncols_interleaved == 0);
@ -1328,38 +1329,36 @@ void ggml_gemm_q8_0_4x4_q8_0_generic(int n,
assert(nr % 4 == 0); assert(nr % 4 == 0);
assert(nc % ncols_interleaved == 0); assert(nc % ncols_interleaved == 0);
{ float sumf[4][4];
float sumf[4][4]; int sumi;
int sumi;
for (int y = 0; y < nr / 4; y++) { for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) { for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb); const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb);
for (int m = 0; m < 4; m++) { for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) { for (int j = 0; j < ncols_interleaved; j++) {
sumf[m][j] = 0.0; sumf[m][j] = 0.0;
}
} }
for (int l = 0; l < nb; l++) { }
for (int k = 0; k < (qk / blocklen); k++) { for (int l = 0; l < nb; l++) {
for (int m = 0; m < 4; m++) { for (int k = 0; k < (qk / blocklen); k++) {
for (int j = 0; j < ncols_interleaved; j++) { for (int m = 0; m < 4; m++) {
sumi = 0; for (int j = 0; j < ncols_interleaved; j++) {
for (int i = 0; i < blocklen; ++i) { sumi = 0;
const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i]; for (int i = 0; i < blocklen; ++i) {
sumi += v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]; const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i];
} sumi += v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i];
sumf[m][j] +=
sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
} }
sumf[m][j] +=
sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
} }
} }
} }
for (int m = 0; m < 4; m++) { }
for (int j = 0; j < ncols_interleaved; j++) { for (int m = 0; m < 4; m++) {
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; for (int j = 0; j < ncols_interleaved; j++) {
} s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
} }
} }
} }