From aaa164773def027ed7cbc2c662c40f7deb4fb557 Mon Sep 17 00:00:00 2001
From: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Date: Sat, 21 Jun 2025 16:00:22 +0200
Subject: [PATCH] q5_1

64.2 t/s -> 114.9 t/s. There is no repacked variant.
---
 ggml/src/iqk/iqk_gemm_legacy_quants.cpp | 18 +++++++++++++++++-
 ggml/src/iqk/iqk_mul_mat.cpp            |  1 +
 2 files changed, 18 insertions(+), 1 deletion(-)

diff --git a/ggml/src/iqk/iqk_gemm_legacy_quants.cpp b/ggml/src/iqk/iqk_gemm_legacy_quants.cpp
index 3b598718..ab6eb130 100644
--- a/ggml/src/iqk/iqk_gemm_legacy_quants.cpp
+++ b/ggml/src/iqk/iqk_gemm_legacy_quants.cpp
@@ -2893,6 +2893,22 @@ struct DeqQ50 {
     const uint8x16_t mh = vdupq_n_u8(0xf0);
 };
 
+struct DeqQ51 {
+
+    inline int8x16x2_t dequant(const block_q5_1& x) const {
+        int8x16x2_t r;
+        bits.prepare1(x.qs, r.val);
+        auto qh = x.qh;
+        r.val[0] = vreinterpretq_s8_u8(vorrq_u8(vreinterpretq_u8_s8(r.val[0]), vandq_u8(mh, hbits.to_bytes(qh+0))));
+        r.val[1] = vreinterpretq_s8_u8(vorrq_u8(vreinterpretq_u8_s8(r.val[1]), vandq_u8(mh, hbits.to_bytes(qh+2))));
+        return r;
+    }
+
+    Q4LegacyBits bits;
+    HighBit5Legacy hbits;
+    const uint8x16_t mh = vdupq_n_u8(0x10);
+};
+
 struct DeqQ60 {
 
     inline int8x16x2_t dequant(const block_q6_0& x) const {
@@ -2992,7 +3008,7 @@ bool iqk_convert_legacy_quants_q8_r8(int type, int n, const void * vx, size_t bx
         case GGML_TYPE_Q4_0  : iqk_convert_qX_q80_r8<block_q4_0, DeqQ40>(n, vx, bx, vy, nrc_x); break;
         case GGML_TYPE_Q4_1  : iqk_convert_qX_1_q8_1_r8<block_q4_1, DeqQ41>(n, vx, bx, vy, nrc_x); break;
         case GGML_TYPE_Q5_0  : iqk_convert_qX_q80_r8<block_q5_0, DeqQ50>(n, vx, bx, vy, nrc_x); break;
-    //    case GGML_TYPE_Q5_1  : iqk_convert_qX_1_q8_1_r8<block_q5_1, Q5_1_Dequantizer<block_q5_1>>(n, vx, bx, vy, nrc_x); break;
+        case GGML_TYPE_Q5_1  : iqk_convert_qX_1_q8_1_r8<block_q5_1, DeqQ51>(n, vx, bx, vy, nrc_x); break;
         case GGML_TYPE_Q6_0  : iqk_convert_qX_q80_r8<block_q6_0, DeqQ60>(n, vx, bx, vy, nrc_x); break;
         case GGML_TYPE_IQ4_NL: iqk_convert_qX_q80_r8<block_iq4_nl, DeqIQ4NL>(n, vx, bx, vy, nrc_x); break;
         case GGML_TYPE_Q8_0  : iqk_convert_qX_q80_r8<block_q8_0, DeqQ80>(n, vx, bx, vy, nrc_x); break;
diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp
index 0f91c92a..d6753f0a 100644
--- a/ggml/src/iqk/iqk_mul_mat.cpp
+++ b/ggml/src/iqk/iqk_mul_mat.cpp
@@ -274,6 +274,7 @@ struct MulMat {
             case GGML_TYPE_Q4_0   : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
             case GGML_TYPE_Q4_1   : return nrc_y >= 32 ? GGML_TYPE_Q8_1    : type;
             case GGML_TYPE_Q5_0   : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
+            case GGML_TYPE_Q5_1   : return nrc_y >= 32 ? GGML_TYPE_Q8_1    : type;
             case GGML_TYPE_Q6_0   : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
             case GGML_TYPE_Q8_0   : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
             case GGML_TYPE_IQ4_NL : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;