Q5_0_R4 (#121)

* Adding q5_0_r4

We get PP-512(LLaMA-3.1-8B) = 256.7 t/s on a Ryzen-7950X.
We even get TG-128 improvement to 11.7 t/s from 11.1 t/s.

* q5_0_r4: NEON

We get PP-512(LLaMA-3.1-8B) = 99.6 t/s on M2-Max,
up from 71.0 t/s for Q5_0. The difference to mainline llama.cpp
is no longer funny: they get 26.5 t/s for Q5_0.

For TG, we are nor able to fully saturate memory bandwidth
and arrive at 22.1 t/s @ 8 threads. Mainline llama.cpp gets
20.6 t/s for Q5_0.

---------

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

1 files changed, 14 insertions, 0 deletions

diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h
index 2af3323d..fb87a602 100644
--- a/ggml/src/ggml-common.h
+++ b/ggml/src/ggml-common.h
@@ -182,6 +182,13 @@ typedef struct {
 } block_q5_0;
 static_assert(sizeof(block_q5_0) == sizeof(ggml_half) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
 
+typedef struct {
+    ggml_half d[4];        // delta
+    uint8_t qh[QK5_0/2];   // 5-th bit of quants
+    uint8_t qs[QK5_0*2];   // nibbles / quants
+} block_q5_0_r4;
+static_assert(sizeof(block_q5_0_r4) == 4*sizeof(ggml_half) + QK5_0*2 + QK5_0/2, "wrong q5_0_r4 block size/padding");
+
 #define QK5_1 32
 typedef struct {
     GGML_SCALE_TYPE1(m, dm);
@@ -198,6 +205,13 @@ typedef struct {
 } block_q6_0;
 static_assert(sizeof(block_q6_0) == sizeof(ggml_half) + QK6_0/2 + QK6_0/4, "wrong q6_0 block size/padding");
 
+typedef struct {
+    ggml_half d[4];      // delta
+    uint8_t qh[QK6_0];   // 5+6-th bit of quants
+    uint8_t qs[QK6_0*2]; // nibbles / quants
+} block_q6_0_r4;
+static_assert(sizeof(block_q6_0_r4) == 4*sizeof(ggml_half) + QK6_0*2 + QK6_0, "wrong q6_0_r4 block size/padding");
+
 #define QK8_0 32
 typedef struct {
     ggml_half d;       // delta