Adding IQ2_KL (#602)

* Experiments for 2.6875 bpw quants

At least according to rmse, this is significantly better than
q2_K, while using only 1/16 more bits per weight.

* iq2_kl: basics

* iq2_kl: CUDA dequantize

* iq2_kl: small improvement in PPL

Also check the two neighbouring values for the block scale
and use the one that minimizes RMSE.

* iq2_kl: MMQ

Quite good: PP-512(L3-8B) = 8472 t/s.

* iq2_kl: MMVQ

We get PP-128(L3-8B) = 162 t/s.
Which means that this is not quite as good as it should be as
(almost) same bpq q2_K is at 170 t/s.

* iq2_kl: Zen4 GEMM/GEMV

Not particularly fast. I may need to think about rearranging the bits.

* iq2_kl: better Zen4

* iq2_kl: convert/repack to q8_k_r8 (AVX2)

* iq2_kl: AVX2 GEMM/GEMV

* iq2_kl: WIP NEON

The compiler started crashing!!!

* iq2_kl: NEON

Had to work around a compiler crash when using vzip2q_u8 using
vqtbl2q_u8.

* iq2_kl: convert/repack to q8_k_r8 (NEON)

* iq2_kl: Metal dequantize

* iq2_kl: Metal GEMV - pretty slow

* iq2_kl: Metal GEMV - slightly better (40 t/s -> 44.5 t/s)

* iq2_kl: Metal GEMV - slightly better (44.5 t/s -> 46.5 t/s)

* iq2_kl: Metal GEMV - slightly better (46.5 t/s -> 47.2 t/s)

* iq2_kl: slightly better Metal dequantize

PP-512 goes to 476 t/s up from 466 t/s.

* iq2_kl: slightly better Metal dequantize

PP-512 goes to 492 t/s up from 476 t/s.

* Add iq2_kl to constants.py

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

1 files changed, 70 insertions, 0 deletions

diff --git a/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_kl.cu b/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_kl.cu
new file mode 100644
index 00000000..a5c22879
--- /dev/null
+++ b/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_kl.cu
@@ -0,0 +1,70 @@
+#include "../mmq.cuh"
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_kl(
+    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_IQ4_XS, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x/4;
+
+    uint32_t aux32[2];
+    const uint8_t * a8 = (const uint8_t *)aux32;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += 4*nwarps) {
+        int i = i0 + 4*threadIdx.y + threadIdx.x%4;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const half * dptr = (const half *)(x + i*stride);
+        const float d = *dptr;
+        const block_iq2_kl * bxi = (const block_iq2_kl *)(dptr + 1) + kbx0;
+
+        #pragma unroll
+        for (int j = 0; j < 2; ++j) {
+            auto ql = get_int_b2(bxi->qs, 4*(kqsx/2) + 2*(kqsx%2) + j);
+            auto qh = get_int_b2(bxi->qh, 2*(kqsx%2) + j) >> 2*(kqsx/2);
+            aux32[0] = ((ql >> 0) & 0x0f0f0f0f) | ((qh << 4) & 0x10101010);
+            aux32[1] = ((ql >> 4) & 0x0f0f0f0f) | ((qh << 3) & 0x10101010);
+            #pragma unroll
+            for (int l = 0; l < 2; ++l) {
+                int val1 = iq2kl_values[a8[2*l+0]] | (iq2kl_values[a8[2*l+1]] << 16);
+                int val2 = iq2kl_values[a8[2*l+4]] | (iq2kl_values[a8[2*l+5]] << 16);
+#ifdef INT8_MMA_AVAILABLE
+                x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 16*(kqsx/2) + 4*(kqsx%2) + 2*j + l + 0] = val1;
+                x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 16*(kqsx/2) + 4*(kqsx%2) + 2*j + l + 8] = val2;
+#else
+                x_qs[i*(2*WARP_SIZE + 1)     + 16*(kqsx/2) + 4*(kqsx%2) + 2*j + l + 0] = val1;
+                x_qs[i*(2*WARP_SIZE + 1)     + 16*(kqsx/2) + 4*(kqsx%2) + 2*j + l + 8] = val2;
+#endif
+            }
+        }
+
+        int ls = int(((bxi->scales_l[kqsx%4] >> 4*(kqsx/4)) & 0xf) | (((bxi->scales_h >> 2*kqsx) & 3) << 4)) - 32;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = d * ls;
+#else
+        x_df[i*(WARP_SIZE/4) + i/4   + kqsx] = d * ls;
+#endif
+    }
+
+}
+
+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_IQ2_KL> {
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq2_kl<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+DECL_MMQ_CASE(GGML_TYPE_IQ2_KL);