ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-28 02:11:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

MMQ for iq4_k: working now

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-05-14 08:16:57 +03:00
parent f7802849b4
commit 5376413185
1 changed files with 104 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

141
ggml/src/ggml-cuda/mmq.cuh
View File

@@ -2426,69 +2426,135 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
int * x_qs = (int *) x_tile;
float * x_df = (float *) (x_qs + WARP_SIZE*2);
#else
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y);
constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16;
int * x_qs = (int *) x_tile;
float * x_df = (float *) (x_qs + txs.qs);
#endif // INT8_MMA_AVAILABLE
constexpr int qstep = 8;
const int kqsx = threadIdx.x % qstep;
uint32_t aux32[2];
const uint8_t * aux8 = (const uint8_t *)aux32;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
int i = i0 + threadIdx.y;
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/qstep) {
int i = i0 + threadIdx.y*(WARP_SIZE/qstep) + threadIdx.x/qstep;
if (need_check) {
i = min(i, i_max);
}
const block_iq4_k * bxi = (const block_iq4_k *)(x + i*stride) + kbx0;
const uint16_t extra = bxi->extra >> 2*kqsx;
const uint16_t extra = bxi->extra >> 2*(threadIdx.x/4);
auto values0 = iq4k_values + ((extra & 1) << 4);
auto values1 = iq4k_values + ((extra & 2) << 3);
const int q4 = get_int_b4(bxi->qs, threadIdx.x);
const int q40 = (q4 >> 0) & 0x0F0F0F0F;
const int q41 = (q4 >> 4) & 0x0F0F0F0F;
auto values_l = iq4k_values + ((extra & 1) << 4);
auto values_h = iq4k_values + ((extra & 2) << 3);
const int8_t * aux80 = (const int8_t *)&q40;
const char4 val0 = make_char4(values0[aux80[0]], values0[aux80[1]], values0[aux80[2]], values0[aux80[3]]);
const int8_t * aux81 = (const int8_t *)&q41;
const char4 val1 = make_char4(values1[aux80[1]], values1[aux81[1]], values1[aux81[2]], values1[aux81[3]]);
#pragma unroll
for (int l = 0; l < qstep/2; ++l) {
const int q4 = get_int_b4(bxi->qs, (qstep/2)*kqsx + l);
aux32[0] = (q4 >> 0) & 0x0f0f0f0f;
aux32[1] = (q4 >> 4) & 0x0f0f0f0f;
const char4 val0 = make_char4(values_l[aux8[0]], values_l[aux8[1]], values_l[aux8[2]], values_l[aux8[3]]);
const char4 val1 = make_char4(values_h[aux8[4]], values_h[aux8[5]], values_h[aux8[6]], values_h[aux8[7]]);
const int k0 = 8 * (threadIdx.x / 4) + threadIdx.x % 4;
#ifdef INT8_MMA_AVAILABLE
x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + k0 + 0] = *(const int *)&val0;
x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + k0 + 4] = *(const int *)&val1;
x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + l + 0] = *(const int *)&val0;
x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + l + 4] = *(const int *)&val1;
#else
x_qs[i*(2*WARP_SIZE + 1) + k0 + 0] = *(const int *)&val0;
x_qs[i*(2*WARP_SIZE + 1) + k0 + 4] = *(const int *)&val1;
x_qs[i*(2*WARP_SIZE + 1) + 8*kqsx + l + 0] = *(const int *)&val0;
x_qs[i*(2*WARP_SIZE + 1) + 8*kqsx + l + 4] = *(const int *)&val1;
#endif // INT8_MMA_AVAILABLE
}
const int ib32 = threadIdx.x % 8;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) {
int i = i0 + threadIdx.y * 4 + threadIdx.x / 8;
if (need_check) {
i = min(i, i_max);
}
const block_iq4_k * bxi = (const block_iq4_k *)(x + i*stride) + kbx0;
const uint8_t sh = bxi->scales_h[ib32/2] >> 4*(ib32%2);
const int ls1 = ((bxi->scales_l[ib32] & 0xf) | ((sh << 4) & 0x30)) - 32;
const int ls2 = ((bxi->scales_l[ib32] >> 4) | ((sh << 2) & 0x30)) - 32;
const uint8_t sh = bxi->scales_h[kqsx/2] >> 4*(kqsx%2);
const int ls1 = ((bxi->scales_l[kqsx] & 0xf) | ((sh << 4) & 0x30)) - 32;
const int ls2 = ((bxi->scales_l[kqsx] >> 4) | ((sh << 2) & 0x30)) - 32;
const float d = bxi->d;
#ifdef INT8_MMA_AVAILABLE
x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*ib32 + 0] = d * ls1;
x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*ib32 + 1] = d * ls2;
x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*kqsx+0] = d * ls1;
x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*kqsx+1] = d * ls2;
#else
// TODO
x_df[i*(WARP_SIZE/4) + i/4 + threadIdx.x % 8] = d * ls1;
x_df[i*(2*WARP_SIZE*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = d * ls1;
x_df[i*(2*WARP_SIZE*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = d * ls2;
#endif // INT8_MMA_AVAILABLE
}
}
//template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq4_k(
// const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
//
//#ifdef INT8_MMA_AVAILABLE
// int * x_qs = (int *) x_tile;
// float * x_df = (float *) (x_qs + WARP_SIZE*2);
//#else
// constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y);
// int * x_qs = (int *) x_tile;
// float * x_df = (float *) (x_qs + txs.qs);
//#endif // INT8_MMA_AVAILABLE
//
//#pragma unroll
// for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
// int i = i0 + threadIdx.y;
//
// if (need_check) {
// i = min(i, i_max);
// }
//
// const block_iq4_k * bxi = (const block_iq4_k *)(x + i*stride) + kbx0;
//
// const uint16_t extra = bxi->extra >> 2*(threadIdx.x/4);
// auto values0 = iq4k_values + ((extra & 1) << 4);
// auto values1 = iq4k_values + ((extra & 2) << 3);
// const int q4 = get_int_b4(bxi->qs, threadIdx.x);
// const int q40 = (q4 >> 0) & 0x0F0F0F0F;
// const int q41 = (q4 >> 4) & 0x0F0F0F0F;
//
// const int8_t * aux80 = (const int8_t *)&q40;
// const char4 val0 = make_char4(values0[aux80[0]], values0[aux80[1]], values0[aux80[2]], values0[aux80[3]]);
// const int8_t * aux81 = (const int8_t *)&q41;
// const char4 val1 = make_char4(values1[aux80[1]], values1[aux81[1]], values1[aux81[2]], values1[aux81[3]]);
//
// const int k0 = 8 * (threadIdx.x / 4) + threadIdx.x % 4;
//#ifdef INT8_MMA_AVAILABLE
// x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + k0 + 0] = *(const int *)&val0;
// x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + k0 + 4] = *(const int *)&val1;
//#else
// x_qs[i*(2*WARP_SIZE + 1) + k0 + 0] = *(const int *)&val0;
// x_qs[i*(2*WARP_SIZE + 1) + k0 + 4] = *(const int *)&val1;
//#endif // INT8_MMA_AVAILABLE
// }
//
// const int ib32 = threadIdx.x % 8;
//#pragma unroll
// for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) {
// int i = i0 + threadIdx.y * 4 + threadIdx.x / 8;
//
// if (need_check) {
// i = min(i, i_max);
// }
//
// const block_iq4_k * bxi = (const block_iq4_k *)(x + i*stride) + kbx0;
// const uint8_t sh = bxi->scales_h[ib32/2] >> 4*(ib32%2);
// const int ls1 = ((bxi->scales_l[ib32] & 0xf) | ((sh << 4) & 0x30)) - 32;
// const int ls2 = ((bxi->scales_l[ib32] >> 4) | ((sh << 2) & 0x30)) - 32;
//
// const float d = bxi->d;
//#ifdef INT8_MMA_AVAILABLE
// x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*ib32 + 0] = d * ls1;
// x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*ib32 + 1] = d * ls2;
//#else
// // TODO
// x_df[i*(WARP_SIZE/4) + i/4 + threadIdx.x % 8] = d * ls1;
//#endif // INT8_MMA_AVAILABLE
// }
//
//}
template<int mmq_x, int mmq_y, int nwarps, bool need_check>
static __device__ __forceinline__ void mmq_write_back_dp4a(
const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) {
@@ -2712,7 +2778,8 @@ struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ4_XS> {
template <int mmq_x, int mmq_y, int nwarps, bool need_check>
struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ4_K> {
static constexpr int vdr = VDR_IQ4_XS_Q8_1_MMQ;
static constexpr int vdr = VDR_IQ2_XS_Q8_1_MMQ;
//static constexpr int vdr = VDR_IQ4_XS_Q8_1_MMQ;
static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq4_k<mmq_y, nwarps, need_check>;
static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_16_q8_1_mma<mmq_x, mmq_y, nwarps>;
static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_16_q8_1_dp4a<mmq_x, mmq_y, nwarps>;