IQ3_XXS_R4 (#153)

* iq3_xxs_r4: 1st shot on Zen4

PP-512: 107 t/s -> 137 t/s
TG-128(1 thread): 2.64 t/s -> 3.44 t/s

* iq4_xxs_r4: WIP

* iq4_xxs_r4: 1st shot at AVX2

Note: there is a bug in the AVX2 implementation for nrc_y = 1
for IQ quants with blocks of 32. I have fixed it for now by
using the nrc_y > 1 implementation (which works) also for nrc_y = 1.

* iq3_xxs_r4: 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 758ee018..21a82ff3 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_IQ3_XXS_R4] = {
+        .type_name                = "iq3_xxs_r4",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq3_xxs),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq3_xxs_r4,
+        .from_float               = quantize_row_iq3_xxs_r4,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_iq3_xxs_r4_ref,
+        .vec_dot                  = vec_dot_iq3_xxs_r4_q8_k,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+        .row_meta_size            = 0,
+    },
     [GGML_TYPE_IQ3_S] = {
         .type_name                = "iq3_s",
         .blck_size                = QK_K,
@@ -4200,6 +4213,7 @@ 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_XS:        wtype = GGML_TYPE_IQ2_XS;   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;
         case GGML_FTYPE_MOSTLY_IQ1_M:         wtype = GGML_TYPE_IQ1_M;    break;
         case GGML_FTYPE_MOSTLY_IQ1_BN:        wtype = GGML_TYPE_IQ1_BN;   break;
@@ -10741,6 +10755,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ1_BN:
@@ -11201,6 +11216,7 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ1_BN:
@@ -11358,6 +11374,7 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ1_BN:
@@ -14561,6 +14578,7 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ1_BN:
@@ -14958,6 +14976,7 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ1_BN:
@@ -15249,6 +15268,7 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ1_BN:
@@ -15869,6 +15889,7 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ1_BN:
@@ -22642,6 +22663,7 @@ void ggml_quantize_init(enum ggml_type type) {
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ1_M:   iq2xs_init_impl(type); break;
+        case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ3_XXS: iq3xs_init_impl(256); break;
         case GGML_TYPE_IQ3_S:   iq3xs_init_impl(512); break;
         default: // nothing
@@ -22715,6 +22737,7 @@ 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_XS:  result = quantize_iq2_xs (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;
         case GGML_TYPE_IQ2_S:   result = quantize_iq2_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ1_S:   result = quantize_iq1_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;