ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 19:31:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq1_bn(Metal): 66.2 -> 67.1 t/s

Browse Source
This commit is contained in:
Iwan Kawrakow
2024-06-17 12:25:08 +02:00
parent d42e9e2922
commit 2f403d4c93
1 changed files with 5 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
ggml-metal.metal
View File

@@ -5046,8 +5046,6 @@ void kernel_mul_mv_iq1_bn_f32_impl(
d1bn[row] = scale.f; d1bn[row] = scale.f;
} }
//uint32_t aux32;
//thread const uint8_t * aux8 = (thread const uint8_t *)&aux32;
uint32_t aux32[2]; uint32_t aux32[2];
thread const uint8_t * aux8 = (thread const uint8_t *)aux32; thread const uint8_t * aux8 = (thread const uint8_t *)aux32;
@@ -5073,10 +5071,6 @@ void kernel_mul_mv_iq1_bn_f32_impl(
uint32_t v = iq1bn_grid_u16[ql[0] | ((qh[0] << (8 - 4*(ir%2))) & 0x0f00)]; uint32_t v = iq1bn_grid_u16[ql[0] | ((qh[0] << (8 - 4*(ir%2))) & 0x0f00)];
uint32_t v32 = v | (v << 12); uint32_t v32 = v | (v << 12);
//aux32 = v32 & 0x03030303;
//acc += yl[0] * aux8[0] + yl[4] * aux8[1] + yl[2]*aux8[2] + yl[6]*aux8[3];
//aux32 = v32 & 0x0c0c0c0c;
//acc += (yl[1] * aux8[0] + yl[5] * aux8[1] + yl[3]*aux8[2] + yl[7]*aux8[3]) * 0.25f;
aux32[0] = v32 & 0x03030303; aux32[1] = v32 & 0x0c0c0c0c; aux32[0] = v32 & 0x03030303; aux32[1] = v32 & 0x0c0c0c0c;
acc = yl[0] * aux8[0] + yl[4] * aux8[1] + yl[2]*aux8[2] + yl[6]*aux8[3]; acc = yl[0] * aux8[0] + yl[4] * aux8[1] + yl[2]*aux8[2] + yl[6]*aux8[3];
acc += (yl[1] * aux8[4] + yl[5] * aux8[5] + yl[3]*aux8[6] + yl[7]*aux8[7]) * 0.25f; acc += (yl[1] * aux8[4] + yl[5] * aux8[5] + yl[3]*aux8[6] + yl[7]*aux8[7]) * 0.25f;
@@ -5091,10 +5085,11 @@ void kernel_mul_mv_iq1_bn_f32_impl(
y4 += 32 * 8; y4 += 32 * 8;
} }
for (int row = 0; row < N_DST; ++row) { for (int row = 0; row < N_DST; row += 2) {
all_sum = simd_sum(sumf[row]); half2 r = {(half)sumf[row], (half)sumf[row+1]};
if (tiisg == 0) { r = simd_sum(r);
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * d1bn[row]; if (tiisg < 2) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row + tiisg] = r[tiisg] * d1bn[row + tiisg];
} }
} }
} }