common,cpu: matmul: add fp4_e3m0 support

Author: mgouicem

src/common/matmul.cpp

@@ -284,14 +284,15 @@ status_t matmul_desc_init(matmul_desc_t *matmul_desc,
     // the memory to be byte aligned.
 
     // s4/u4/f4 weights requires n to be multiple of 2 to be byte aligned
-    VCHECK_MATMUL(
-            IMPLICATION(utils::one_of(weights_desc->data_type, data_type::s4,
-                                data_type::u4, data_type::f4_e2m1),
-                    weights_desc->dims[n_idx] % 2 == 0),
+    VCHECK_MATMUL(IMPLICATION(utils::one_of(weights_desc->data_type,
+                                      data_type::s4, data_type::u4,
+                                      data_type::f4_e2m1, data_type::f4_e3m0),
+                          weights_desc->dims[n_idx] % 2 == 0),
             VERBOSE_BAD_DIM, "weights", n_idx);
     // s4/u4/f4 src requires k to be multiple of 2 to be byte aligned
     VCHECK_MATMUL(IMPLICATION(utils::one_of(src_desc->data_type, data_type::s4,
-                                      data_type::u4, data_type::f4_e2m1),
+                                      data_type::u4, data_type::f4_e2m1,
+                                      data_type::f4_e3m0),
                           src_desc->dims[k_idx_src] % 2 == 0),
             VERBOSE_BAD_DIM, "src", n_idx);
 

src/cpu/matmul/ref_matmul.hpp

@@ -52,17 +52,17 @@ struct ref_matmul_t : public primitive_t {
             VDISPATCH_MATMUL(
                     is_dense_format_kind(), VERBOSE_UNSUPPORTED_SPARSE_CFG);
             VDISPATCH_MATMUL(utils::one_of(src_type, f32, bf16, f16, f8_e5m2,
-                                     f8_e4m3, f4_e2m1),
+                                     f8_e4m3, f4_e2m1, f4_e3m0),
                     VERBOSE_UNSUPPORTED_DT);
             VDISPATCH_MATMUL(utils::one_of(wei_type, f32, bf16, f16, f8_e5m2,
-                                     f8_e4m3, f4_e2m1, u8, s8, u4, s4),
+                                     f8_e4m3, f4_e2m1, f4_e3m0, u8, s8, u4, s4),
                     VERBOSE_UNSUPPORTED_DT);
             VDISPATCH_MATMUL(utils::one_of(dst_type, f32, bf16, f16, f8_e5m2,
-                                     f8_e4m3, f4_e2m1),
+                                     f8_e4m3, f4_e2m1, f4_e3m0),
                     VERBOSE_UNSUPPORTED_DT);
-            VDISPATCH_MATMUL(
-                    (src_type == wei_type
-                            || utils::one_of(wei_type, u8, s8, u4, s4)),
+            VDISPATCH_MATMUL((src_type == wei_type
+                                     || utils::one_of(wei_type, u8, s8, u4, s4,
+                                             f4_e3m0)),
                     VERBOSE_UNSUPPORTED_DT);
             /* int8 weights decompression support */
             VDISPATCH_MATMUL(IMPLICATION(utils::one_of(wei_type, u8, s8),