Merge branch 'new-cisc' into kernels-asplos-ae

tests/regression/sgemm_gemmini_dma/compile_ampere.sh

@@ -0,0 +1,11 @@
+rm kernel.radiance.elf
+rm -rf binaries
+mkdir binaries
+for a in args/*; do
+    cp -f $a args.bin
+    aa=$(basename "$a")
+    cp -f input.a/"$aa" input.a.bin
+    cp -f input.b/"$aa" input.b.bin
+    make > /dev/null
+    mv kernel.radiance.elf binaries/gemmini_fp16dma"$aa".elf
+done

tests/regression/sgemm_gemmini_dma/compile_hopper.sh

@@ -0,0 +1,11 @@
+rm kernel.radiance.elf
+rm -rf binaries
+mkdir binaries
+for a in args/*; do
+    cp -f $a args.bin
+    aa=$(basename "$a")
+    cp -f input.a/"$aa" input.a.bin
+    cp -f input.b/"$aa" input.b.bin
+    make > /dev/null
+    mv kernel.radiance.elf binaries/gemmini_hopper_dma"$aa".elf
+done

tests/regression/sgemm_gemmini_dma/generate_operands.py

@@ -15,20 +15,20 @@ def truncated_matrix_multiplication(matrix_a, matrix_b, size):
     result = np.matmul(truncated_a, truncated_b)
     return result.astype(np.float16)
 
-# Generate the 512x512 matrices
-size = 512
-matrix_a = generate_fp16_matrix(size)
-matrix_b = generate_fp16_matrix(size)
-
-# Save the operand matrices to binary files
-save_matrix_to_bin("input.a.bin", matrix_a)
-save_matrix_to_bin("input.b.bin", matrix_b)
-
 # Generate and save the reference matrices for 128x128, 256x256, and 512x512 sizes
-sizes = [128, 256, 512]
+sizes = [128, 256, 512, 1024]
 for s in sizes:
+    np.random.seed(0)
+    matrix_a = generate_fp16_matrix(s)
+    matrix_b = generate_fp16_matrix(s)
+    
+    # Save the operand matrices to binary files
+    save_matrix_to_bin("input.a.bin", matrix_a)
+    save_matrix_to_bin(f"input.a/{s}", matrix_a)
+    save_matrix_to_bin("input.b.bin", matrix_b)
+    save_matrix_to_bin(f"input.b/{s}", matrix_b)
+
     ref_matrix = truncated_matrix_multiplication(matrix_a, matrix_b, s)
-    print(ref_matrix)
     save_matrix_to_bin(f"ref{s}.bin", ref_matrix)
 
 print("All files generated successfully.")

tests/regression/sgemm_gemmini_dma/kernel.cpp

@@ -107,7 +107,7 @@ void thread_block_matmul_gemmini(kernel_arg_t *__UNIFORM__ arg,
       for (uint32_t tile_j = 0; tile_j < num_tiles_n; tile_j += 1) {
         for (uint32_t tile_k = 0; tile_k < num_tiles_k; tile_k += 1) {
           uint32_t a_hexadecile = (tile_k & 1) << 2;
-          uint32_t b_hexadecile = a_hexadecile + 8;
+          uint32_t b_hexadecile = a_hexadecile + 11;
           gemmini_tile_load_ab(A, B,
               a_hexadecile, b_hexadecile, tile_i, tile_j, tile_k,
               dim_m, dim_n, dim_k, TILE_M, TILE_N, TILE_K);