Q2_K_R4 (#136)

* q2_k_r4: Zen4

PP-512(LLaMA-3.1-8B) = 256 t/s

* q3_k_r4: AVX2

* q2_k_r4: AVX2

We get PP-512(LLaMA-3.1-8B) = 287 t/s.

Also cherry-picked the q3_k_r4 AVX2 adaptation that I somehow
forgot to push upstream.

* q2_k_r4: NEON

We get PP-512(LLaMA-3.1-8B) = 106.2 t/s.
TG-128 is 36.02 t/s, which is ~10% higher than q2_K_S.

* Make sure rows per thread are a multiple of 4

---------

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

1 files changed, 115 insertions, 0 deletions

diff --git a/ggml/src/iqk/iqk_quantize.cpp b/ggml/src/iqk/iqk_quantize.cpp
index 2e59fefe..49e2cf8e 100644
--- a/ggml/src/iqk/iqk_quantize.cpp
+++ b/ggml/src/iqk/iqk_quantize.cpp
@@ -4437,3 +4437,118 @@ void vec_dot_q3_k_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t b
     GGML_UNUSED(by);
 }
 
+//
+// ========================================= q2_k_r4
+//
+
+void quantize_row_q2_k_r4_ref(const float * x, block_q2_k_r4 * y, int64_t k) {
+    quantize_q3_k_r4(x, (void *)y, 4, k/4, nullptr);
+}
+
+void quantize_row_q2_k_r4(const float * x, void * y, int64_t k) {
+    quantize_q2_k_r4(x, y, 4, k/4, nullptr);
+}
+
+namespace {
+inline void convert_q2_k(const block_q2_K& x, uint8_t * L) {
+
+    const uint8_t * qs = x.qs;
+    for (int n = 0; n < QK_K; n += 128) {
+        for (int j = 0; j < 32; ++j) {
+            L[n + j +  0] = (qs[j] >> 0) & 0x3;
+            L[n + j + 32] = (qs[j] >> 2) & 0x3;
+            L[n + j + 64] = (qs[j] >> 4) & 0x3;
+            L[n + j + 96] = (qs[j] >> 6) & 0x3;
+        }
+        qs += 32;
+    }
+}
+}
+
+static void repack_q2_k(int nrows, int n_per_row, const block_q2_K * x, block_q2_k_r4 * y) {
+    GGML_ASSERT(nrows%4 == 0);
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    int nblock = n_per_row/QK_K;
+    const block_q2_K * x4[4];
+    uint8_t L[QK_K];
+    for (int row = 0; row < nrows; row += 4) {
+        for (int k = 0; k < 4; ++k) x4[k] = x + nblock*k;
+        for (int ibl = 0; ibl < nblock; ++ibl) {
+            for (int k = 0; k < 4; ++k) {
+                y[ibl].d[k+0] = x4[k][ibl].d;
+                y[ibl].d[k+4] = x4[k][ibl].dmin;
+                for (int ib = 0; ib < QK_K/16; ++ib) {
+                    y[ibl].scales[4*ib+k] = x4[k][ibl].scales[ib];
+                }
+                convert_q2_k(x4[k][ibl], L);
+                for (int ib = 0; ib < QK_K/32; ++ib) {
+                    for (int i = 0; i < 4; ++i) {
+                        y[ibl].qs[32*ib+4*k+i+ 0] = ((L[32*ib+i+ 0] & 0x3) << 0) | ((L[32*ib+i+ 4] & 0x3) << 2) | ((L[32*ib+i+ 8] & 0x3) << 4) | ((L[32*ib+i+12] & 0x3) << 6);
+                        y[ibl].qs[32*ib+4*k+i+16] = ((L[32*ib+i+16] & 0x3) << 0) | ((L[32*ib+i+20] & 0x3) << 2) | ((L[32*ib+i+24] & 0x3) << 4) | ((L[32*ib+i+28] & 0x3) << 6);
+                    }
+                }
+            }
+        }
+        x += 4*nblock;
+        y += nblock;
+    }
+}
+
+size_t quantize_q2_k_r4(const float * src, void * dst, int64_t nrows, int64_t n_per_row, const float * imatrix) {
+    GGML_ASSERT(nrows%4 == 0);
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    char * qcur = (char *)dst;
+    auto row_size = ggml_row_size(GGML_TYPE_Q2_K, n_per_row);
+    std::vector<char> qtmp(4*row_size);
+    for (int row = 0; row < nrows; row += 4) {
+        quantize_q2_K(src, (void *)qtmp.data(), 4, n_per_row, imatrix);
+        repack_q2_k(4, n_per_row, (const block_q2_K *)qtmp.data(), (block_q2_k_r4 *)qcur);
+        qcur += 4*row_size;
+        src += 4*n_per_row;
+    }
+    return nrows*row_size;
+}
+
+void dequantize_row_q2_k_r4(const block_q2_k_r4 * x, float * y, int64_t k) {
+    auto n_per_row = k/4;
+    float * y4[4] = {y, y + n_per_row, y + 2*n_per_row, y + 3*n_per_row};
+    int nblock = n_per_row/QK_K;
+    for (int ibl = 0; ibl < nblock; ++ibl) {
+        for (int k = 0; k < 4; ++k) {
+            const float d = GGML_FP16_TO_FP32(x[ibl].d[k+0]);
+            const float m = GGML_FP16_TO_FP32(x[ibl].d[k+4]);
+            auto ql = x[ibl].qs;
+            for (int ib = 0; ib < QK_K/32; ++ib) {
+                float dl1 = d * (x[ibl].scales[8*ib + k + 0] & 0xf);
+                float ml1 = m * (x[ibl].scales[8*ib + k + 0] >>  4);
+                float dl2 = d * (x[ibl].scales[8*ib + k + 4] & 0xf);
+                float ml2 = m * (x[ibl].scales[8*ib + k + 4] >>  4);
+                for (int i = 0; i < 4; ++i) {
+                    y4[k][QK_K*ibl+32*ib+i+ 0] = dl1 * ((ql[4*k+i+ 0] >> 0) & 3) - ml1;
+                    y4[k][QK_K*ibl+32*ib+i+ 4] = dl1 * ((ql[4*k+i+ 0] >> 2) & 3) - ml1;
+                    y4[k][QK_K*ibl+32*ib+i+ 8] = dl1 * ((ql[4*k+i+ 0] >> 4) & 3) - ml1;
+                    y4[k][QK_K*ibl+32*ib+i+12] = dl1 * ((ql[4*k+i+ 0] >> 6) & 3) - ml1;
+                    y4[k][QK_K*ibl+32*ib+i+16] = dl2 * ((ql[4*k+i+16] >> 0) & 3) - ml2;
+                    y4[k][QK_K*ibl+32*ib+i+20] = dl2 * ((ql[4*k+i+16] >> 2) & 3) - ml2;
+                    y4[k][QK_K*ibl+32*ib+i+24] = dl2 * ((ql[4*k+i+16] >> 4) & 3) - ml2;
+                    y4[k][QK_K*ibl+32*ib+i+28] = dl2 * ((ql[4*k+i+16] >> 6) & 3) - ml2;
+                }
+                ql += 32;
+            }
+        }
+    }
+}
+
+void vec_dot_q2_k_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t bx, const void * vy, size_t by, int nrc) {
+#if GGML_USE_IQK_MULMAT
+    if (iqk_mul_mat(1, 1, n, GGML_TYPE_Q2_K_R4, vx, 0, GGML_TYPE_Q8_K, vy, 0, s, 0, 0, 1)) {
+        return;
+    }
+#endif
+    GGML_ASSERT(n%QK4_NL == 0);
+    GGML_ASSERT(nrc == 1);
+    GGML_UNUSED(bs);
+    GGML_UNUSED(bx);
+    GGML_UNUSED(by);
+}
+