Merge branch 'ae' into ae-hopper

kernels/sgemm_tcore/sgemm_impl.hpp

@@ -6,7 +6,7 @@
 #include "include/gemmini.h"
 #include "gemmini_mmio.h"
 
-#define FP_SIZE 32
+#define FP_SIZE 16
 
 // "fake" fp16 type that only has the correct data width.
 using float16_t = uint16_t;
@@ -19,7 +19,7 @@ using float_type = float16_t;
 
 // Generate kernel for the Hopper-style SMEM-decoupled tensor core.  This uses
 // asynchronous HGMMA and HGMMA_WAIT instructions.
-#define TENSOR_HOPPER 0
+#define TENSOR_HOPPER 1
 
 // Constraints on parameters:
 // * Memory:

lib/gemmini/include/gemmini_params.h

@@ -1 +1 @@
-gemmini_params.dim8fp32.h
+gemmini_params.dim16fp16.h

lib/include/gemmini_mmio.h

@@ -12,9 +12,9 @@
 // 64KB
 // #define SMEM_SIZE 0x10000
 // 128KB (FP16 GEMM)
-// #define SMEM_SIZE 0x20000
+#define SMEM_SIZE 0x20000
 // 256KB (FlashAttention)
-#define SMEM_SIZE 0x40000
+// #define SMEM_SIZE 0x40000
 
 #define SMEM_MASK (SMEM_SIZE - 1)
 #define SMEM_ADDR_END (SMEM_BASE + SMEM_SIZE)