vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

iq3_s_mult: ifdef'd slow / fast versions

Browse source
This commit is contained in:
Iwan Kawrakow 2024-03-03 10:43:53 +02:00
parent 726aed307a
commit b6402fa757
2 changed files with 71 additions and 91 deletions
Download patch file Download diff file Expand all files Collapse all files

49
ggml-cuda.cu
View file

@ -2370,10 +2370,13 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
}
//#define IQ3S_MULTIPLIER 190842953LL
//#define IQ3S_MULTIPLIER 5718026
#define IQ3S_MULTIPLIER 898886
#ifdef IQ3S_SLOW_MULT
// Better (lower PPL), but requires more bit twidling, so slower
#define IQ3S_MULTIPLIER 190842953LL
#else
//#define IQ3S_MULTIPLIER 898886
#define IQ3S_MULTIPLIER 842866
#endif
template<typename dst_t>
static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
@ -2391,17 +2394,18 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
const int8_t * grid = (const int8_t *)aux32;
const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf));
const uint8_t signs = x[i].signs[4*ib + il];
//aux32[0] = ((qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256)) * IQ3S_MULTIPLIER) & 0x0f0f0f0f;
//aux32[1] = ((qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256)) * IQ3S_MULTIPLIER) & 0x0f0f0f0f;
#ifdef IQ3S_SLOW_MULT
aux32[0] = ((qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256)) * IQ3S_MULTIPLIER) & 0x0f0f0f0f;
aux32[1] = ((qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256)) * IQ3S_MULTIPLIER) & 0x0f0f0f0f;
#else
aux32[0] = (((qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256)) * IQ3S_MULTIPLIER) & 0x0f0f0f0f) | 0x01010101;
aux32[1] = (((qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256)) * IQ3S_MULTIPLIER) & 0x0f0f0f0f) | 0x01010101;
#endif
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
//aux32[0] = (((__vmaxs4(__vsub4(aux32[0], 0x01010101), 0x00000000) >> 1) & 0x07070707) << 1) | 0x01010101;
//aux32[1] = (((__vmaxs4(__vsub4(aux32[1], 0x01010101), 0x00000000) >> 1) & 0x07070707) << 1) | 0x01010101;
//aux32[0] = ((__vsub4(aux32[0], 0x01010101) >> 1) << 1) | 0x01010101;
//aux32[1] = ((__vsub4(aux32[1], 0x01010101) >> 1) << 1) | 0x01010101;
aux32[0] = ((aux32[0] >> 1) << 1) | 0x01010101;
aux32[1] = ((aux32[1] >> 1) << 1) | 0x01010101;
#ifdef IQ3S_SLOW_MULT
aux32[0] = (((__vmaxs4(__vsub4(aux32[0], 0x01010101), 0x00000000) >> 1) & 0x07070707) << 1) | 0x01010101;
aux32[1] = (((__vmaxs4(__vsub4(aux32[1], 0x01010101), 0x00000000) >> 1) & 0x07070707) << 1) | 0x01010101;
#endif
uint32_t signs0 = __vcmpeq4(((signs & 0xf) * 0x01010101) & 0x08040201, 0x08040201);
uint32_t signs1 = __vcmpeq4(((signs >> 4) * 0x01010101) & 0x08040201, 0x08040201);
aux32[0] = __vsub4(aux32[0] ^ signs0, signs0);
@ -2410,9 +2414,15 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
y[j] = d * grid[j];
}
#else
#ifdef IQ3S_SLOW_MULT
for (int j = 0; j < 8; ++j) {
y[j] = d * (2*((grid[j]-1)/2) + 1) * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
}
#else
for (int j = 0; j < 8; ++j) {
y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
}
#endif
#endif
#else
assert(false);
@ -5227,16 +5237,15 @@ static __device__ __forceinline__ float vec_dot_iq3_s_q8_1(
const int8_t * q8 = bq8_1[ib32].qs;
int sumi = 0;
for (int l = 0; l < 4; ++l) {
//aux32[0] = ((qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256)) * IQ3S_MULTIPLIER) & 0x0f0f0f0f;
//aux32[1] = ((qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256)) * IQ3S_MULTIPLIER) & 0x0f0f0f0f;
//aux32[0] = (((__vmaxs4(__vsub4(aux32[0], 0x01010101), 0) >> 1) & 0x07070707) << 1) | 0x01010101;
//aux32[1] = (((__vmaxs4(__vsub4(aux32[1], 0x01010101), 0) >> 1) & 0x07070707) << 1) | 0x01010101;
#ifdef IQ3S_SLOW_MULT
aux32[0] = ((qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256)) * IQ3S_MULTIPLIER) & 0x0f0f0f0f;
aux32[1] = ((qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256)) * IQ3S_MULTIPLIER) & 0x0f0f0f0f;
//aux32[0] = ((__vsub4(aux32[0], 0x01010101) >> 1) << 1) | 0x01010101;
//aux32[1] = ((__vsub4(aux32[1], 0x01010101) >> 1) << 1) | 0x01010101;
aux32[0] = ((aux32[0] >> 1) << 1) | 0x01010101;
aux32[1] = ((aux32[1] >> 1) << 1) | 0x01010101;
aux32[0] = (((__vmaxs4(__vsub4(aux32[0], 0x01010101), 0) >> 1) & 0x07070707) << 1) | 0x01010101;
aux32[1] = (((__vmaxs4(__vsub4(aux32[1], 0x01010101), 0) >> 1) & 0x07070707) << 1) | 0x01010101;
#else
aux32[0] = (((qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256)) * IQ3S_MULTIPLIER) & 0x0f0f0f0f) | 0x01010101;
aux32[1] = (((qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256)) * IQ3S_MULTIPLIER) & 0x0f0f0f0f) | 0x01010101;
#endif
uint32_t signs0 = __vcmpeq4(((bq2->signs[4*ib32+l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201);
uint32_t signs1 = __vcmpeq4(((bq2->signs[4*ib32+l] >> 4) * 0x01010101) & 0x08040201, 0x08040201);
const int grid_l = __vsub4(aux32[0] ^ signs0, signs0);