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, 22 insertions, 0 deletions

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 45c873f2..526b1139 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1399,6 +1399,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .nrows                    = 1,
         .row_meta_size            = 0,
     },
+    [GGML_TYPE_IQ5_K_R4] = {
+        .type_name                = "iq5_k_r4",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq5_k),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq5_k_r4,
+        .from_float               = quantize_row_iq5_k_r4,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_iq5_k_r4_ref,
+        .vec_dot                  = vec_dot_iq5_k_r4_q8_k,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+        .row_meta_size            = 0,
+    },
     [GGML_TYPE_IQ6_K] = {
         .type_name                = "iq6_k",
         .blck_size                = QK_K,
@@ -4193,6 +4206,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_IQ3_K_R4:      wtype = GGML_TYPE_IQ3_K_R4; break;
         case GGML_FTYPE_MOSTLY_IQ4_K_R4:      wtype = GGML_TYPE_IQ4_K_R4; break;
         case GGML_FTYPE_MOSTLY_IQ5_K:         wtype = GGML_TYPE_IQ5_K;    break;
+        case GGML_FTYPE_MOSTLY_IQ5_K_R4:      wtype = GGML_TYPE_IQ5_K_R4; break;
         case GGML_FTYPE_MOSTLY_IQ6_K:         wtype = GGML_TYPE_IQ6_K;    break;
         case GGML_FTYPE_MOSTLY_IQ3_S:         wtype = GGML_TYPE_IQ3_S;    break;
         case GGML_FTYPE_MOSTLY_IQ2_S:         wtype = GGML_TYPE_IQ2_S;    break;
@@ -10732,6 +10746,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ3_K_R4:
         case GGML_TYPE_IQ4_K_R4:
         case GGML_TYPE_IQ5_K:
+        case GGML_TYPE_IQ5_K_R4:
         case GGML_TYPE_IQ6_K:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
@@ -11190,6 +11205,7 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ3_K_R4:
         case GGML_TYPE_IQ4_K_R4:
         case GGML_TYPE_IQ5_K:
+        case GGML_TYPE_IQ5_K_R4:
         case GGML_TYPE_IQ6_K:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
@@ -11345,6 +11361,7 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_IQ3_K_R4:
         case GGML_TYPE_IQ4_K_R4:
         case GGML_TYPE_IQ5_K:
+        case GGML_TYPE_IQ5_K_R4:
         case GGML_TYPE_IQ6_K:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
@@ -14546,6 +14563,7 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ3_K_R4:
         case GGML_TYPE_IQ4_K_R4:
         case GGML_TYPE_IQ5_K:
+        case GGML_TYPE_IQ5_K_R4:
         case GGML_TYPE_IQ6_K:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
@@ -14941,6 +14959,7 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_IQ3_K_R4:
         case GGML_TYPE_IQ4_K_R4:
         case GGML_TYPE_IQ5_K:
+        case GGML_TYPE_IQ5_K_R4:
         case GGML_TYPE_IQ6_K:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
@@ -15230,6 +15249,7 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ3_K_R4:
         case GGML_TYPE_IQ4_K_R4:
         case GGML_TYPE_IQ5_K:
+        case GGML_TYPE_IQ5_K_R4:
         case GGML_TYPE_IQ6_K:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
@@ -15848,6 +15868,7 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_IQ3_K_R4:
         case GGML_TYPE_IQ4_K_R4:
         case GGML_TYPE_IQ5_K:
+        case GGML_TYPE_IQ5_K_R4:
         case GGML_TYPE_IQ6_K:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
@@ -22694,6 +22715,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_IQ3_K_R4:result = quantize_iq3_k_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_K_R4:result = quantize_iq4_k_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ5_K:   result = quantize_iq5_k  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ5_K_R4:result = quantize_iq5_k_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ6_K:   result = quantize_iq6_k  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q4_0_4_4: result = quantize_q4_0_4x4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q4_0_4_8: result = quantize_q4_0_4x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;