Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-28 15:18:26 +01:00
Code Issues Packages Projects Releases Wiki Activity

iq3_s_mult: Metal works - slower than lookup

Browse Source
This commit is contained in:
Iwan Kawrakow 2024-03-03 06:41:58 +02:00
parent bf90920fb2
commit 3000e0ac9e
1 changed files with 34 additions and 81 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

115
ggml-metal.metal
View File

@ -2546,6 +2546,7 @@ typedef struct {
uint8_t signs[QK_K/8];
uint8_t scales[IQ3S_N_SCALE];
} block_iq3_s;
#define IQ3S_MULTIPLIER 190842953
typedef struct {
half d;
@ -4083,73 +4084,6 @@ constexpr constant static uint32_t iq3xxs_grid[256] = {
0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
};
constexpr constant static uint32_t iq3xs_grid[512] = {
0x04040404, 0x0404040c, 0x04040414, 0x0404042c, 0x0404043e, 0x04040c04, 0x04040c0c, 0x04040c14,
0x04040c24, 0x04040c34, 0x04041404, 0x0404140c, 0x0404142c, 0x04041c1c, 0x04042404, 0x04042414,
0x0404242c, 0x0404243e, 0x04042c0c, 0x04042c1c, 0x04043404, 0x04043414, 0x04043e0c, 0x04043e24,
0x04043e3e, 0x040c0404, 0x040c040c, 0x040c0414, 0x040c0424, 0x040c0c04, 0x040c0c0c, 0x040c0c2c,
0x040c1404, 0x040c141c, 0x040c143e, 0x040c1c0c, 0x040c1c2c, 0x040c2424, 0x040c340c, 0x040c342c,
0x040c3e14, 0x04140404, 0x0414040c, 0x0414042c, 0x0414043e, 0x04140c04, 0x04140c1c, 0x04140c34,
0x0414140c, 0x0414142c, 0x04141c04, 0x04141c24, 0x04142414, 0x0414242c, 0x0414243e, 0x04142c0c,
0x04142c1c, 0x04143e04, 0x04143e1c, 0x041c041c, 0x041c0c0c, 0x041c0c2c, 0x041c1404, 0x041c1414,
0x041c1c0c, 0x041c1c1c, 0x041c1c34, 0x041c2424, 0x041c2c04, 0x041c2c14, 0x041c343e, 0x041c3e0c,
0x041c3e2c, 0x04240404, 0x04240c1c, 0x04240c3e, 0x0424140c, 0x04241424, 0x04241c14, 0x04242404,
0x0424241c, 0x04242c0c, 0x04243e04, 0x042c0414, 0x042c0424, 0x042c1404, 0x042c1414, 0x042c1434,
0x042c1c1c, 0x042c240c, 0x042c242c, 0x042c243e, 0x042c3434, 0x042c3e1c, 0x04340434, 0x04340c0c,
0x04340c1c, 0x04341c0c, 0x04342c14, 0x04343e0c, 0x043e0404, 0x043e0414, 0x043e0424, 0x043e1404,
0x043e1414, 0x043e1434, 0x043e1c1c, 0x043e2c04, 0x043e2c24, 0x0c040404, 0x0c04040c, 0x0c040414,
0x0c040424, 0x0c040c04, 0x0c040c0c, 0x0c040c1c, 0x0c040c2c, 0x0c040c3e, 0x0c041404, 0x0c041414,
0x0c041c0c, 0x0c041c24, 0x0c041c34, 0x0c042c24, 0x0c042c34, 0x0c04340c, 0x0c043e14, 0x0c0c0404,
0x0c0c040c, 0x0c0c041c, 0x0c0c0434, 0x0c0c0c04, 0x0c0c0c24, 0x0c0c140c, 0x0c0c1c04, 0x0c0c1c1c,
0x0c0c240c, 0x0c0c2c04, 0x0c0c2c14, 0x0c0c3e04, 0x0c0c3e34, 0x0c140404, 0x0c140c14, 0x0c140c2c,
0x0c140c3e, 0x0c141404, 0x0c141424, 0x0c141c14, 0x0c142404, 0x0c14241c, 0x0c142c2c, 0x0c143404,
0x0c143e14, 0x0c1c040c, 0x0c1c0424, 0x0c1c043e, 0x0c1c0c04, 0x0c1c0c1c, 0x0c1c140c, 0x0c1c143e,
0x0c1c1c04, 0x0c1c1c24, 0x0c1c240c, 0x0c1c3414, 0x0c1c3e04, 0x0c24041c, 0x0c24042c, 0x0c240c14,
0x0c240c24, 0x0c241c0c, 0x0c241c1c, 0x0c242414, 0x0c242434, 0x0c242c04, 0x0c242c24, 0x0c2c040c,
0x0c2c0c04, 0x0c2c0c1c, 0x0c2c140c, 0x0c2c1c04, 0x0c2c1c14, 0x0c2c2c0c, 0x0c341404, 0x0c341424,
0x0c34143e, 0x0c342424, 0x0c342434, 0x0c3e040c, 0x0c3e041c, 0x0c3e0c04, 0x0c3e0c14, 0x0c3e140c,
0x0c3e1c2c, 0x0c3e240c, 0x0c3e3414, 0x0c3e3e04, 0x14040404, 0x1404040c, 0x1404041c, 0x1404042c,
0x1404043e, 0x14040c04, 0x14040c14, 0x14040c24, 0x14040c34, 0x1404140c, 0x1404141c, 0x1404143e,
0x14041c04, 0x14041c14, 0x1404240c, 0x1404241c, 0x1404242c, 0x14042c04, 0x14042c14, 0x1404343e,
0x14043e04, 0x14043e1c, 0x14043e2c, 0x140c0404, 0x140c0414, 0x140c0c04, 0x140c0c1c, 0x140c0c3e,
0x140c1414, 0x140c142c, 0x140c1c0c, 0x140c1c24, 0x140c2414, 0x140c2c0c, 0x1414040c, 0x14140424,
0x1414043e, 0x1414140c, 0x1414141c, 0x14141c04, 0x14141c3e, 0x1414240c, 0x14142c1c, 0x14142c3e,
0x14143e0c, 0x14143e24, 0x141c0404, 0x141c0414, 0x141c042c, 0x141c0c0c, 0x141c1414, 0x141c1424,
0x141c1c0c, 0x141c1c1c, 0x141c2414, 0x141c2c04, 0x141c3434, 0x1424040c, 0x1424043e, 0x14241404,
0x1424141c, 0x14241c14, 0x14241c2c, 0x1424240c, 0x14243e14, 0x14243e2c, 0x142c0424, 0x142c0c0c,
0x142c1414, 0x142c1c3e, 0x142c2404, 0x142c2c1c, 0x142c3e04, 0x14340404, 0x14340414, 0x1434043e,
0x1434140c, 0x14342c2c, 0x1434340c, 0x143e042c, 0x143e0c0c, 0x143e1434, 0x143e1c04, 0x143e241c,
0x143e2c04, 0x1c040414, 0x1c040c0c, 0x1c040c1c, 0x1c040c2c, 0x1c040c3e, 0x1c041414, 0x1c041c0c,
0x1c041c1c, 0x1c041c2c, 0x1c042414, 0x1c042424, 0x1c04243e, 0x1c042c0c, 0x1c04341c, 0x1c043e0c,
0x1c0c040c, 0x1c0c041c, 0x1c0c042c, 0x1c0c0c24, 0x1c0c140c, 0x1c0c141c, 0x1c0c2404, 0x1c0c3404,
0x1c0c3e14, 0x1c0c3e34, 0x1c140404, 0x1c140c14, 0x1c141404, 0x1c141c14, 0x1c141c24, 0x1c142c04,
0x1c1c040c, 0x1c1c0c04, 0x1c1c0c24, 0x1c1c140c, 0x1c1c141c, 0x1c1c143e, 0x1c1c1c04, 0x1c1c240c,
0x1c1c241c, 0x1c1c243e, 0x1c1c2c2c, 0x1c1c3e1c, 0x1c24041c, 0x1c240c0c, 0x1c240c34, 0x1c241414,
0x1c241c0c, 0x1c242c14, 0x1c243404, 0x1c243424, 0x1c2c040c, 0x1c2c0c04, 0x1c2c0c14, 0x1c2c142c,
0x1c2c1c14, 0x1c2c2424, 0x1c2c2c34, 0x1c2c3e1c, 0x1c340c34, 0x1c34240c, 0x1c3e040c, 0x1c3e041c,
0x1c3e1404, 0x1c3e1414, 0x1c3e1c2c, 0x24040404, 0x24040424, 0x24040c14, 0x24041404, 0x24041424,
0x2404143e, 0x24041c14, 0x2404240c, 0x24042c04, 0x24043e04, 0x240c0414, 0x240c043e, 0x240c0c0c,
0x240c0c1c, 0x240c1414, 0x240c1c04, 0x240c1c2c, 0x240c241c, 0x240c2c0c, 0x240c2c2c, 0x2414040c,
0x2414041c, 0x24140c04, 0x24140c2c, 0x2414140c, 0x24141c1c, 0x24142404, 0x24142c3e, 0x24143414,
0x24143e04, 0x241c0424, 0x241c0c0c, 0x241c0c1c, 0x241c1404, 0x241c1414, 0x241c1c0c, 0x241c1c2c,
0x24240404, 0x24240414, 0x24241424, 0x24241c3e, 0x24242404, 0x24243e0c, 0x242c042c, 0x242c043e,
0x242c140c, 0x242c3414, 0x24340c1c, 0x24341c24, 0x24343404, 0x243e0c04, 0x243e0c2c, 0x243e1c04,
0x243e241c, 0x243e2c0c, 0x2c040414, 0x2c040c04, 0x2c040c24, 0x2c041414, 0x2c042404, 0x2c042424,
0x2c04243e, 0x2c042c14, 0x2c043434, 0x2c043e24, 0x2c0c040c, 0x2c0c041c, 0x2c0c042c, 0x2c0c0c14,
0x2c0c140c, 0x2c0c1c14, 0x2c0c3e14, 0x2c140404, 0x2c140c0c, 0x2c14141c, 0x2c141c04, 0x2c141c34,
0x2c142c1c, 0x2c1c0414, 0x2c1c043e, 0x2c1c0c04, 0x2c1c143e, 0x2c1c2424, 0x2c1c2c0c, 0x2c1c342c,
0x2c1c3e1c, 0x2c24040c, 0x2c240424, 0x2c241404, 0x2c241c14, 0x2c242434, 0x2c2c0c14, 0x2c2c1434,
0x2c2c2c0c, 0x2c2c2c1c, 0x2c342414, 0x2c3e0414, 0x2c3e0424, 0x2c3e1414, 0x34040c0c, 0x34040c1c,
0x34040c2c, 0x34041c0c, 0x34041c1c, 0x34043404, 0x340c0404, 0x340c1404, 0x340c143e, 0x340c3424,
0x34140c14, 0x34141c24, 0x34142414, 0x34142c2c, 0x34143414, 0x34143e04, 0x341c0404, 0x341c0c24,
0x341c140c, 0x341c2404, 0x3424142c, 0x3424241c, 0x34243414, 0x342c0404, 0x342c041c, 0x342c1c24,
0x342c3404, 0x3434042c, 0x34342404, 0x343e0c0c, 0x343e0c1c, 0x3e040404, 0x3e040424, 0x3e04043e,
0x3e041404, 0x3e041414, 0x3e041c34, 0x3e042404, 0x3e042c24, 0x3e043414, 0x3e0c0414, 0x3e0c0c0c,
0x3e0c1424, 0x3e0c241c, 0x3e0c242c, 0x3e14040c, 0x3e140424, 0x3e140c04, 0x3e140c34, 0x3e14140c,
0x3e141c04, 0x3e142c0c, 0x3e1c0414, 0x3e1c1c14, 0x3e1c1c2c, 0x3e1c2c1c, 0x3e24040c, 0x3e24042c,
0x3e240c1c, 0x3e241404, 0x3e242c04, 0x3e2c1414, 0x3e2c2414, 0x3e340414, 0x3e341c0c, 0x3e3e0404,
};
#define NGRID_IQ1S 512
constexpr constant static uint64_t iq1s_grid[NGRID_IQ1S] = {
0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000,
@ -4757,14 +4691,23 @@ void kernel_mul_mv_iq3_s_f32_impl(
threadgroup uint32_t * values = (threadgroup uint32_t *)shared_values;
{
uint32_t aux32;
thread int8_t * q = (thread int8_t *)&aux32;
int nval = 8;
int pos = (32*sgitg + tiisg)*nval;
for (int i = 0; i < nval; ++i) values[pos + i] = iq3xs_grid[pos + i];
for (int i = 0; i < nval; ++i) {
aux32 = (IQ3S_MULTIPLIER * (pos + i)) & 0x0f0f0f0f;
for (int k = 0; k < 4; ++k) q[k] = 2*((q[k]-1)/2) + 1;
values[pos + i] = aux32;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
}
const int ix = tiisg;
uint32_t aux32[2];
thread const int8_t * grid = (thread const int8_t *)aux32;
device const float * y4 = y + 32 * ix;
for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
@ -4786,13 +4729,21 @@ void kernel_mul_mv_iq3_s_f32_impl(
for (int row = 0; row < N_DST; row++) {
const float db = dh[0];
const float d = db * (0.5f + ((sc[0] >> 4*(ib%2)) & 0xf));
const float d = db * (1 + 2*((sc[0] >> 4*(ib%2)) & 0xf));
float2 sum = {0};
for (int l = 0; l < 4; ++l) {
const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + (qs[2*l+0] | ((qh[0] << (8-2*l)) & 256)));
const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + (qs[2*l+1] | ((qh[0] << (7-2*l)) & 256)));
//aux32[0] = (IQ3S_MULTIPLIER * (qs[2*l+0] | ((qh[0] << (8-2*l)) & 256))) & 0x0f0f0f0f;
//aux32[1] = (IQ3S_MULTIPLIER * (qs[2*l+1] | ((qh[0] << (7-2*l)) & 256))) & 0x0f0f0f0f;
//threadgroup const uint8_t * grid1 = (threadgroup const uint8_t *)(values + (qs[2*l+0] | ((qh[0] << (8-2*l)) & 256)));
//threadgroup const uint8_t * grid2 = (threadgroup const uint8_t *)(values + (qs[2*l+1] | ((qh[0] << (7-2*l)) & 256)));
threadgroup const uint8_t * grid1 = (threadgroup const uint8_t *)(values + qs[2*l+0] +
select(0, 256, qh[0] & kmask_iq2xs[2*l+0]));
threadgroup const uint8_t * grid2 = (threadgroup const uint8_t *)(values + qs[2*l+1] +
select(0, 256, qh[0] & kmask_iq2xs[2*l+1]));
for (int j = 0; j < 4; ++j) {
//sum[0] += yl[8*l + j + 0] * (2*((grid[j+0] - 1)/2) + 1) * select(1, -1, signs[l] & kmask_iq2xs[j+0]);
//sum[1] += yl[8*l + j + 4] * (2*((grid[j+4] - 1)/2) + 1) * select(1, -1, signs[l] & kmask_iq2xs[j+4]);
sum[0] += yl[8*l + j + 0] * grid1[j] * select(1, -1, signs[l] & kmask_iq2xs[j+0]);
sum[1] += yl[8*l + j + 4] * grid2[j] * select(1, -1, signs[l] & kmask_iq2xs[j+4]);
}
@ -4812,7 +4763,7 @@ void kernel_mul_mv_iq3_s_f32_impl(
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.5f;
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
}
}
}
@ -5703,18 +5654,20 @@ void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 &
device const uint8_t * qs = xb->qs + 8*ib32;
device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
const uint8_t qh = xb->qh[ib32] >> 4*il;
const float dl = d * (0.5f + ((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * 0.5f;
constant uint8_t * grid1 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+0] | ((qh << 8) & 256)));
constant uint8_t * grid2 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+1] | ((qh << 7) & 256)));
const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
uint32_t aux32[2];
thread const int8_t * grid = (thread const int8_t *)aux32;
aux32[0] = (IQ3S_MULTIPLIER * (qs[4*il+0] | ((qh << 8) & 256))) & 0x0f0f0f0f;
aux32[1] = (IQ3S_MULTIPLIER * (qs[4*il+1] | ((qh << 7) & 256))) & 0x0f0f0f0f;
for (int i = 0; i < 4; ++i) {
reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
reg[0][i] = dl * (2*((grid[i+0]-1)/2)+1) * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
reg[1][i] = dl * (2*((grid[i+4]-1)/2)+1) * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
}
grid1 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+2] | ((qh << 6) & 256)));
grid2 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+3] | ((qh << 5) & 256)));
aux32[0] = (IQ3S_MULTIPLIER * (qs[4*il+2] | ((qh << 6) & 256))) & 0x0f0f0f0f;
aux32[1] = (IQ3S_MULTIPLIER * (qs[4*il+3] | ((qh << 5) & 256))) & 0x0f0f0f0f;
for (int i = 0; i < 4; ++i) {
reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
reg[2][i] = dl * (2*((grid[i+0]-1)/2)+1) * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
reg[3][i] = dl * (2*((grid[i+4]-1)/2)+1) * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
}
}