IQ2_XS_R4 (#155)

* iq2_xs_r4: Zen4

* iq2_xs_r4: AVX2

* iq2_xs_r4: slightly better matrix x vector on AVX2

* iq2_xs_r4: NEON - not much better than iq2_xs

* iq2_xs_r4: slightly better NEON

---------

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

1 files changed, 23 insertions, 0 deletions

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 27794cd3..0c3be11c 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1031,6 +1031,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .nrows                    = 1,
         .row_meta_size            = 0,
     },
+    [GGML_TYPE_IQ2_XS_R4] = {
+        .type_name                = "iq2_xs_r4",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq2_xs),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq2_xs_r4,
+        .from_float               = quantize_row_iq2_xs_r4,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_iq2_xs_r4_ref,
+        .vec_dot                  = vec_dot_iq2_xs_r4_q8_k,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+        .row_meta_size            = 0,
+    },
     [GGML_TYPE_IQ3_XXS] = {
         .type_name                = "iq3_xxs",
         .blck_size                = QK_K,
@@ -4226,6 +4239,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_IQ2_XXS:       wtype = GGML_TYPE_IQ2_XXS;  break;
         case GGML_FTYPE_MOSTLY_IQ2_XXS_R4:    wtype = GGML_TYPE_IQ2_XXS_R4;break;
         case GGML_FTYPE_MOSTLY_IQ2_XS:        wtype = GGML_TYPE_IQ2_XS;   break;
+        case GGML_FTYPE_MOSTLY_IQ2_XS_R4:     wtype = GGML_TYPE_IQ2_XS_R4;break;
         case GGML_FTYPE_MOSTLY_IQ3_XXS:       wtype = GGML_TYPE_IQ3_XXS;  break;
         case GGML_FTYPE_MOSTLY_IQ3_XXS_R4:    wtype = GGML_TYPE_IQ3_XXS_R4;break;
         case GGML_FTYPE_MOSTLY_IQ1_S:         wtype = GGML_TYPE_IQ1_S;    break;
@@ -10769,6 +10783,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XXS_R4:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_XS_R4:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ1_S:
@@ -11231,6 +11246,7 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XXS_R4:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_XS_R4:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ1_S:
@@ -11390,6 +11406,7 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XXS_R4:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_XS_R4:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ1_S:
@@ -14595,6 +14612,7 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XXS_R4:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_XS_R4:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ1_S:
@@ -14994,6 +15012,7 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XXS_R4:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_XS_R4:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ1_S:
@@ -15287,6 +15306,7 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XXS_R4:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_XS_R4:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ1_S:
@@ -15909,6 +15929,7 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XXS_R4:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_XS_R4:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ1_S:
@@ -22680,6 +22701,7 @@ void ggml_quantize_init(enum ggml_type type) {
 
     switch (type) {
         case GGML_TYPE_IQ2_XXS_R4: iq2xs_init_impl(GGML_TYPE_IQ2_XXS); break;
+        case GGML_TYPE_IQ2_XS_R4:  iq2xs_init_impl(GGML_TYPE_IQ2_XS);  break;
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ2_S:
@@ -22759,6 +22781,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_IQ2_XXS: result = quantize_iq2_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ2_XXS_R4:result = quantize_iq2_xxs_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ2_XS:  result = quantize_iq2_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ2_XS_R4:result = quantize_iq2_xs_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ3_XXS: result = quantize_iq3_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ3_XXS_R4:result = quantize_iq3_xxs_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ3_S:   result = quantize_iq3_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;