IQ4_NL_X4 (#118)

* Adding iq4_nl_x4

Looks very promising - I get PP-512(LLaMA-3.1-8B) = 230 t/s
on the Ryzen-7950X! This is faster than any other quant and
~40% faster than iq4_nl.

* iq4_nl_x4: getting amazing

This Zen4 variant gets us to PP-512(LLaMA-3.1-8B) = 263 t/s!

* iq4_nl_x4: AVX2

Here we gain only 25% compared to iq4_nl

* iq4_nl_x4: NEON

On M2-Max we get PP-512(LLaMA-3.1-8B) = 109.7 t/s, up from
82.4 t/s for iq4_nl.

* iq4_nl_x4: minor NEON improvement and cleanup

This gets us to 110.3 t/s. In comparison,
IQ4_NL_4_4 in mainline llama.cpp achieves 92.3 t/s.

* iq4_nl_x4: NEON specialization for matrix x vector

---------

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

1 files changed, 26 insertions, 0 deletions

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 39218ff4..c975212e 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1245,6 +1245,23 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .nrows                    = 1,
         .row_meta_size            = 0,
     },
+    [GGML_TYPE_IQ4_NL_X4] = {
+        .type_name                = "iq4_nl_x4",
+        .blck_size                = QK4_NL,
+        .type_size                = sizeof(block_iq4_nl),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq4_nl_x4,
+        .from_float               = quantize_row_iq4_nl_x4,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_iq4_nl_x4_ref,
+        .vec_dot                  = vec_dot_iq4_nl_x4_q8_0,
+#if GGML_USE_IQK_MULMAT && defined __AVX2__
+        .vec_dot_type             = GGML_TYPE_Q8_1,
+#else
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+#endif
+        .nrows                    = 1,
+        .row_meta_size            = 0,
+    },
 };
 
 // For internal test use
@@ -3903,6 +3920,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_IQ1_BN:        wtype = GGML_TYPE_IQ1_BN;   break;
         case GGML_FTYPE_MOSTLY_IQ2_BN:        wtype = GGML_TYPE_IQ2_BN;   break;
         case GGML_FTYPE_MOSTLY_IQ4_NL:        wtype = GGML_TYPE_IQ4_NL;   break;
+        case GGML_FTYPE_MOSTLY_IQ4_NL_X4:     wtype = GGML_TYPE_IQ4_NL_X4;break;
         case GGML_FTYPE_MOSTLY_IQ4_XS:        wtype = GGML_TYPE_IQ4_XS;   break;
         case GGML_FTYPE_MOSTLY_IQ4_KS:        wtype = GGML_TYPE_IQ4_KS;   break;
         case GGML_FTYPE_MOSTLY_IQ4_KSS:       wtype = GGML_TYPE_IQ4_KSS;  break;
@@ -10426,6 +10444,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ1_BN:
         case GGML_TYPE_IQ2_BN:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_NL_X4:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
         case GGML_TYPE_IQ4_KSS:
@@ -10868,6 +10887,7 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ1_BN:
         case GGML_TYPE_IQ2_BN:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_NL_X4:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
         case GGML_TYPE_IQ4_KSS:
@@ -11007,6 +11027,7 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_IQ1_BN:
         case GGML_TYPE_IQ2_BN:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_NL_X4:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
         case GGML_TYPE_IQ4_KSS:
@@ -14192,6 +14213,7 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ1_BN:
         case GGML_TYPE_IQ2_BN:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_NL_X4:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
         case GGML_TYPE_IQ4_KSS:
@@ -14571,6 +14593,7 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_IQ1_BN:
         case GGML_TYPE_IQ2_BN:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_NL_X4:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
         case GGML_TYPE_IQ4_KSS:
@@ -14844,6 +14867,7 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ1_BN:
         case GGML_TYPE_IQ2_BN:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_NL_X4:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
         case GGML_TYPE_IQ4_KSS:
@@ -15444,6 +15468,7 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_IQ1_BN:
         case GGML_TYPE_IQ2_BN:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_NL_X4:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
         case GGML_TYPE_IQ4_KSS:
@@ -22270,6 +22295,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_IQ1_BN:  result = quantize_iq1_bn (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ2_BN:  result = quantize_iq2_bn (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_NL:  result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ4_NL_X4: result = quantize_iq4_nl_x4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_XS:  result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_KS:  result = quantize_iq4_ks (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_KSS: result = quantize_iq4_kss(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;