Added separate block sizes for zen5 and zen4 in dlp classic. (#508)

Added a new block size map macro for Zen5 (DLP_GEMM_BLKSZ_MAP_ZEN5) and moved the 
previous Zen4 block size map to a new macro (DLP_GEMM_BLKSZ_MAP_ZEN4). 

This ensures that Zen5 and Zen4 architectures use their own optimized block sizes.

Author: MithunMohanKadavil

classic/config/dlp_gemm_blksz_map.h

@@ -33,7 +33,7 @@
 // ID,MC,NC,KC,MR,NR,PACKA_RS,PACKA_CS,PACKB_RS,PACKB_CS: ID = One of the
 // AOCL_DLP_OPERATION_TYPE enum.
 
-#define DLP_GEMM_BLKSZ_MAP_ZEN4                                                \
+#define DLP_GEMM_BLKSZ_MAP_ZEN5                                                \
     XMACRO(U8S8S32OS32, 144, 1024, 2048, 6, 64, 4, 24, 4 * 64, 64)             \
     XMACRO(F32F32F32OF32, 192, 8064, 1024, 6, 64, 1, 6, 64, 1)                 \
     XMACRO(BF16BF16F32OF32, 144, 1024, 4096, 6, 64, 0, 0, 2 * 64, 64 / 2)      \
@@ -45,6 +45,18 @@
     XMACRO(F32OBF16, 144, 1024, 4096, 6, 64, 0, 0, 2 * 64, 64 / 2)             \
     XMACRO(BF16U4F32OF32, 144, 1024, 4096, 6, 64, 0, 0, 2 * 64, 64 / 2)
 
+#define DLP_GEMM_BLKSZ_MAP_ZEN4                                                \
+    XMACRO(U8S8S32OS32, 144, 1024, 2048, 6, 64, 4, 24, 4 * 64, 64)             \
+    XMACRO(F32F32F32OF32, 192, 8064, 512, 6, 64, 1, 6, 64, 1)                  \
+    XMACRO(BF16BF16F32OF32, 144, 1024, 4096, 6, 64, 0, 0, 2 * 64, 64 / 2)      \
+    XMACRO(BF16S4F32OF32, 144, 1024, 4096, 6, 64, 0, 0, 2 * 64, 64 / 2)        \
+    XMACRO(F16F16F16OF16, 288, 1024, 2048, 6, 128, 1, 12, 128, 1)              \
+    XMACRO(F32F16F32OF32, 192, 4096, 2048, 6, 64, 1, 6, 64, 1)                 \
+    XMACRO(S8S8S32OS32, 144, 1024, 2048, 6, 64, 4, 24, 4 * 64, 64)             \
+    XMACRO(U8S4S32OS32, 144, 1024, 2048, 6, 64, 4, 24, 4 * 64, 64)             \
+    XMACRO(F32OBF16, 144, 1024, 4096, 6, 64, 0, 0, 2 * 64, 64 / 2)             \
+    XMACRO(BF16U4F32OF32, 144, 1024, 4096, 6, 64, 0, 0, 2 * 64, 64 / 2)
+
 #define DLP_GEMM_BLKSZ_MAP_ZEN                                                 \
     XMACRO(U8S8S32OS32, 144, 1024, 2048, 6, 64, 4, 24, 4 * 64, 64)             \
     XMACRO(F32F32F32OF32, 144, 8160, 512, 6, 16, 1, 6, 16, 1)                  \

classic/config/dlp_gemm_config.c

@@ -289,6 +289,15 @@ _dlp_gemm_cntx_init_func_map()
 
         _DLP_GEMM_CNTX_UPD_FUNC_MAP_FOR_CONFIGURED_ARCH()
 #endif
+    } else if (dlp_cpuid_is_avx512_supported() == TRUE) {
+#ifdef DLP_KERNELS_ZEN4
+        DLP_GEMM_KERN_FUNC_MAP_AVX512
+        DLP_GEMM_PACKA_FUNC_MAP_AVX512
+        DLP_GEMM_PACKB_FUNC_MAP_AVX512
+        DLP_GEMM_PACKBMXP_FUNC_MAP_AVX512
+
+        _DLP_GEMM_CNTX_UPD_FUNC_MAP_FOR_CONFIGURED_ARCH()
+#endif
     } else if (dlp_cpuid_is_avx2fma3_supported() == TRUE) {
 #ifdef DLP_KERNELS_ZEN3
         DLP_GEMM_KERN_FUNC_MAP_AVX2
@@ -376,7 +385,15 @@ _dlp_gemm_cntx_init_blksz_map()
     // the blocksize for a particular version of zen id is generalized
     // for all machines that support the ISA supported by that particular
     // zen id.
-    if (dlp_cpuid_is_avx512vnni_supported() == TRUE) {
+    if (dlp_cpuid_is_similar_zen5_arch() == TRUE) {
+        DLP_GEMM_BLKSZ_MAP_ZEN5
+
+        // Fallback to zen3 blocksizes has the same logic for both
+        // zen5 and zen4.
+        if (global_dlp_gemmenable_arch == DLP_ARCH_ZEN3) {
+            DLP_GEMM_BLKSZ_UPD_MAP_ZEN4_TO_ZEN
+        }
+    } else if (dlp_cpuid_is_avx512_supported() == TRUE) {
         DLP_GEMM_BLKSZ_MAP_ZEN4
 
         if (global_dlp_gemmenable_arch == DLP_ARCH_ZEN3) {