IQ5_K_R4 (#149)

* iq5_k_r4: Zen4

Much slower than the others.

* iq5_k_r5: WIP

* Minor

* iq5_k_r4: fix AVX2 nrc_y = 1 case

* iq5_k_r4: better Zen4

But TG is still slower than iq5_k

* iq5_k_r4: slightly better AVX2

* iq5_k_r4: NEON

---------

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

1 files changed, 10 insertions, 0 deletions

diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h
index 03cc3460..0af461c7 100644
--- a/ggml/src/ggml-common.h
+++ b/ggml/src/ggml-common.h
@@ -585,6 +585,16 @@ typedef struct {
 static_assert(sizeof(block_iq5_k) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/2 + QK_K/8 + 3*QK_K/64, "wrong iq5_k block size/padding");
 
 typedef struct {
+    ggml_half d[4];
+    uint8_t  extra[8];
+    uint8_t  scales_h[QK_K/16];
+    uint8_t  scales_l[QK_K/8 ];
+    uint8_t  qs[QK_K*2];
+    uint8_t  qh[QK_K/2];
+} block_iq5_k_r4;
+static_assert(sizeof(block_iq5_k_r4) == 4*sizeof(block_iq5_k), "wrong iq5_k_r4 block size/padding");
+
+typedef struct {
     ggml_half d;
     uint16_t extra;
     int8_t   scales[QK_K/16];