Adding IQ5_KS - 5.25 bpw quants (#422)

* iq5_ks: basics

* iq5_ks: quantize

* iq5_ks: CUDA dequantize works

* iq5_ks: dot product works on CUDA

* iq5_ks: MMQ works

* iq5_ks: Zen4

* iq5_ks: AVX2

But is is not quite right, just like iq4_k, iq5_k, iq6_k, iq4_ks.
All these need fixing on AVX2.

* iq5_ks: NEON

* iq5_ks: Metal dequantize

* iq5_ks: Metal dot product

---------

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

1 files changed, 244 insertions, 0 deletions

diff --git a/ggml/src/iqk/iqk_quantize.cpp b/ggml/src/iqk/iqk_quantize.cpp
index ca5e008a..78b25525 100644
--- a/ggml/src/iqk/iqk_quantize.cpp
+++ b/ggml/src/iqk/iqk_quantize.cpp
@@ -3419,6 +3419,250 @@ void  vec_dot_iq4_ks_q8_k(int n, float * s, size_t bs, const void * vx, size_t b
 }
 
 namespace {
+static void quantize_row_iq5_ks_impl(const int super_block_size, const int block_size,
+        int n_per_row, const float * x, char * cy,
+        float * all_scales, float * weight,
+        const int8_t * values,
+        const float * quant_weights,
+        const int ntry) {
+
+    float * dptr = (float *)cy;
+    dptr[0] = 0;
+    block_iq5_ks * y = (block_iq5_ks *)(dptr + 1);
+
+    const int8_t * shifted_values = values + 32;
+
+    float amax_scale = 0;
+
+    for (int ibl = 0; ibl < n_per_row/super_block_size; ++ibl) {
+        memset(&y[ibl], 0, sizeof(block_iq5_ks));
+        const float * xbl = x + ibl*super_block_size;
+        auto scales = all_scales + ibl*(super_block_size/block_size);
+        float sigma2 = 0;
+        for (int j = 0; j < super_block_size; ++j) sigma2 += xbl[j]*xbl[j];
+        sigma2 *= 2.f/super_block_size;
+        for (int ib = 0; ib < super_block_size/block_size; ++ib) {
+            const float * xb = xbl + ib*block_size;
+            if (quant_weights) {
+                const float * qw = quant_weights + ibl*super_block_size + ib*block_size;
+                for (int j = 0; j < block_size; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]);
+            } else {
+                for (int j = 0; j < block_size; ++j) weight[j] = xb[j]*xb[j];
+            }
+            float amax = 0, max = 0;
+            for (int j = 0; j < block_size; ++j) {
+                float ax = fabsf(xb[j]);
+                if (ax > amax) {
+                    amax = ax; max = xb[j];
+                }
+            }
+            if (amax < 1e-15f) {
+                scales[ib] = 0;
+                continue;
+            }
+            float d = ntry > 0 ? -max/values[0] : max/values[0];
+            float id = 1/d;
+            float sumqx_p = 0, sumq2_p = 0;
+            float sumqx_m = 0, sumq2_m = 0;
+            for (int j = 0; j < block_size; ++j) {
+                float w = weight[j];
+                float al = id*xb[j];
+                int l = best_index_iq5nl(values, al);
+                float q = values[l];
+                sumqx_p += w*q*xb[j];
+                sumq2_p += w*q*q;
+                l = best_index_iq5nl(values, -al);
+                q = values[l];
+                sumqx_m += w*q*xb[j];
+                sumq2_m += w*q*q;
+            }
+            d = sumqx_p/sumq2_p;
+            bool is_shifted = false;
+            float best = d*sumqx_p;
+            if (sumq2_m > 0 && sumqx_m*sumqx_m > best*sumq2_m) {
+                d = sumqx_m/sumq2_m; best = d*sumqx_m;
+            }
+            for (int itry = -ntry; itry <= ntry; ++itry) {
+                id = (itry + values[0])/max;
+                sumqx_p = sumq2_p = 0;
+                sumqx_m = sumq2_m = 0;
+                for (int j = 0; j < block_size; ++j) {
+                    float w = weight[j];
+                    float al = id*xb[j];
+                    int l = best_index_iq5nl(values, al);
+                    float q = values[l];
+                    sumqx_p += w*q*xb[j];
+                    sumq2_p += w*q*q;
+                    l = best_index_iq5nl(values, -al);
+                    q = values[l];
+                    sumqx_m += w*q*xb[j];
+                    sumq2_m += w*q*q;
+                }
+                if (sumq2_p > 0 && sumqx_p*sumqx_p > best*sumq2_p) {
+                    d = sumqx_p/sumq2_p; best = d * sumqx_p; is_shifted = false;
+                }
+                if (sumq2_m > 0 && sumqx_m*sumqx_m > best*sumq2_m) {
+                    d = sumqx_m/sumq2_m; best = d * sumqx_m; is_shifted = false;
+                }
+                id = (itry + shifted_values[0])/max;
+                sumqx_p = sumq2_p = 0;
+                sumqx_m = sumq2_m = 0;
+                for (int j = 0; j < block_size; ++j) {
+                    float w = weight[j];
+                    float al = id*xb[j];
+                    int l = best_index_iq5nl(shifted_values, al);
+                    float q = shifted_values[l];
+                    sumqx_p += w*q*xb[j];
+                    sumq2_p += w*q*q;
+                    l = best_index_iq5nl(shifted_values, -al);
+                    q = shifted_values[l];
+                    sumqx_m += w*q*xb[j];
+                    sumq2_m += w*q*q;
+                }
+                if (sumq2_p > 0 && sumqx_p*sumqx_p > best*sumq2_p) {
+                    d = sumqx_p/sumq2_p; best = d * sumqx_p; is_shifted = true;
+                }
+                if (sumq2_m > 0 && sumqx_m*sumqx_m > best*sumq2_m) {
+                    d = sumqx_m/sumq2_m; best = d * sumqx_m; is_shifted = true;
+                }
+            }
+            if (is_shifted) y[ibl].scales[ib] = 0x01;
+            scales[ib] = d;
+            amax_scale = std::max(amax_scale, std::abs(d));
+        }
+    }
+    float d = amax_scale/127;
+    *dptr = d;
+    if (!d) return;
+    float id = d ? 1/d : 0.f;
+    float sumqx = 0, sumq2 = 0;
+    for (int ibl = 0; ibl < n_per_row/super_block_size; ++ibl) {
+        const float * xbl = x + ibl*super_block_size;
+        float sigma2 = 0;
+        for (int j = 0; j < super_block_size; ++j) sigma2 += xbl[j]*xbl[j];
+        sigma2 *= 2.f/super_block_size;
+        auto scales = all_scales + (super_block_size/block_size)*ibl;
+        for (int ib = 0; ib < super_block_size/block_size; ++ib) {
+            const int8_t * block_values = y[ibl].scales[ib] & 0x01 ? shifted_values : values;
+            int l = nearest_int(0.5f*(id*scales[ib]+127.f));
+            l = std::max(0, std::min(127, l)) << 1;
+            y[ibl].scales[ib] |= l;
+            l -= 127;
+            float dl = d * l;
+            float idl = dl ? 1/dl : 0.f;
+            const float * xb = xbl + ib*block_size;
+            if (quant_weights) {
+                const float * qw = quant_weights + ibl*super_block_size + ib*block_size;
+                for (int j = 0; j < block_size; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]);
+            } else {
+                for (int j = 0; j < block_size; ++j) weight[j] = xb[j]*xb[j];
+            }
+            for (int j = 0; j < block_size; ++j) {
+                uint8_t idx = best_index_iq5nl(block_values, idl*xb[j]);
+                y[ibl].qs[block_size*(ib/2) + j] |= ((idx & 0xf) << 4*(ib%2));
+                y[ibl].qh[j] |= ((idx >> 4) << ib);
+                float w = weight[j];
+                float q = block_values[idx]*l;
+                sumqx += w*q*xb[j];
+                sumq2 += w*q*q;
+            }
+        }
+    }
+    if (sumq2 > 0) *dptr = sumqx/sumq2;
+}
+}
+
+void quantize_row_iq5_ks_ref(const float * x, block_iq5_ks * y, int64_t k) {
+    quantize_iq5_ks(x, (void *)y, 1, k, nullptr);
+}
+
+void quantize_row_iq5_ks(const float * x, void * y, int64_t k) {
+    quantize_iq5_ks(x, (void *)y, 1, k, nullptr);
+}
+
+size_t quantize_iq5_ks(const float * src, void * dst, int64_t nrows, int64_t n_per_row, const float * imatrix) {
+    constexpr int kBlockSize = 32;
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    auto row_size = ggml_row_size(GGML_TYPE_IQ5_KS, n_per_row);
+    char * qrow = (char *)dst;
+    float weight[kBlockSize];
+    std::vector<float> all_scales(n_per_row/kBlockSize);
+    for (int64_t row = 0; row < nrows; ++row) {
+        quantize_row_iq5_ks_impl(QK_K, kBlockSize, n_per_row, src, qrow, all_scales.data(), weight, iq5nl_values, imatrix, 5);
+        src += n_per_row;
+        qrow += row_size;
+    }
+    return nrows * row_size;
+}
+
+void dequantize_row_iq5_ks(const block_iq5_ks * x, float * y, int64_t k) {
+    constexpr int kBlockSize = 32;
+    GGML_ASSERT(k%QK_K == 0);
+    const float * dptr = (const float *)x;
+    float d = *dptr;
+    x = (const block_iq5_ks *)(dptr + 1);
+    int nblock = k/QK_K;
+    for (int ibl = 0; ibl < nblock; ++ibl) {
+        auto qs = x[ibl].qs;
+        auto qh = x[ibl].qh;
+        for (int ib64 = 0; ib64 < QK_K/(2*kBlockSize); ++ib64) {
+            float dl1 = d * ((int)(x[ibl].scales[2*ib64+0] & 254) - 127);
+            float dl2 = d * ((int)(x[ibl].scales[2*ib64+1] & 254) - 127);
+            const int8_t * values1 = iq5nl_values + ((x[ibl].scales[2*ib64+0] & 1) << 5);
+            const int8_t * values2 = iq5nl_values + ((x[ibl].scales[2*ib64+1] & 1) << 5);
+            for (int j = 0; j < kBlockSize; ++j) {
+                y[j           ] = dl1 * values1[(qs[j] & 0xf) | (((qh[j] >> (2*ib64+0)) & 1) << 4)];
+                y[j+kBlockSize] = dl2 * values2[(qs[j] >>  4) | (((qh[j] >> (2*ib64+1)) & 1) << 4)];
+            }
+            y  += 2*kBlockSize;
+            qs += kBlockSize;
+        }
+    }
+}
+
+void  vec_dot_iq5_ks_q8_k(int n, float * s, size_t bs, const void * vx, size_t bx, const void * vy, size_t by, int nrc) {
+    constexpr int kBlockSize = 32;
+#if GGML_USE_IQK_MULMAT
+    if (iqk_mul_mat(1, 1, n, GGML_TYPE_IQ5_KS, vx, 0, GGML_TYPE_Q8_K, vy, 0, s, 0, 0, 1)) {
+        return;
+    }
+#endif
+    GGML_ASSERT(n%QK_K == 0);
+    GGML_ASSERT(nrc == 1);
+    GGML_UNUSED(bs);
+    GGML_UNUSED(bx);
+    GGML_UNUSED(by);
+    const float * dptr = (const float *)vx;
+    const float d = *dptr;
+    const block_iq5_ks * x = (const block_iq5_ks *)(dptr + 1);
+    const block_q8_K   * y = (const block_q8_K    *)vy;
+    int nblock = n/QK_K;
+    float sumf = 0;
+    for (int ibl = 0; ibl < nblock; ++ibl) {
+        auto qy = y[ibl].qs;
+        auto qs = x[ibl].qs;
+        auto qh = x[ibl].qh;
+        float db = d * y[ibl].d;
+        for (int ib64 = 0; ib64 < QK_K/(2*kBlockSize); ++ib64) {
+            float dl1 = db * ((int)(x[ibl].scales[2*ib64+0] & 254) - 127);
+            float dl2 = db * ((int)(x[ibl].scales[2*ib64+1] & 254) - 127);
+            const int8_t * values1 = iq5nl_values + ((x[ibl].scales[2*ib64+0] & 1) << 5);
+            const int8_t * values2 = iq5nl_values + ((x[ibl].scales[2*ib64+1] & 1) << 5);
+            int suml1 = 0;
+            int suml2 = 0;
+            for (int j = 0; j < kBlockSize; ++j) {
+                suml1 += qy[j           ] * values1[(qs[j] & 0xf) | (((qh[j] >> (2*ib64+0)) & 1) << 4)];
+                suml2 += qy[j+kBlockSize] * values2[(qs[j] >>  4) | (((qh[j] >> (2*ib64+1)) & 1) << 4)];
+            }
+            sumf += dl1*suml1 + dl2*suml2;
+            y  += 2*kBlockSize;
+            qs += kBlockSize;
+        }
+    }
+    *s = sumf;
+}
+
+namespace {
 const uint16_t * scramble_table() {
     static std::mutex mutex;
     static std::vector<uint16_t> table;