3 files changed, 76 insertions, 64 deletions
diff --git a/TRNS/benchmark-scripts/ccmcc25-sim.sh b/TRNS/benchmark-scripts/ccmcc25-sim.sh
new file mode 100755
index 0000000..a7aa79c
--- /dev/null
+++ b/TRNS/benchmark-scripts/ccmcc25-sim.sh
@@ -0,0 +1,53 @@
+#!/bin/bash
+
+mkdir -p log/$(hostname)
+
+# Args: -m m -n n -o M_ -p N_
+#
+# Input: (M_ * m) × (N_ * n) matrix
+# Output: (N_* n) × (M_ * m) matrix
+# Step 1: transpose (M_ * m) × N_ matrix that consists of tiles of size n
+#   CPU version: explicit
+#   DPU version: implicit (M_ * m write operations of #DPUs * n elements to DPUs)
+# Step 2: transpose m × n matrix; this happens N_ * M_ times.
+#   DPU version: Each tasklet transposes a single m × n matrix / tile.
+#   (16 × 8 tile takes up 1 KiB WRAM)
+# Step 3: Transpose M_ × n matrix that consists of tiles of size m.
+#
+# Note for DPU version: if M_ > #DPUs, steps 1 through 3 are repeated.
+# Number of repetitions == ceil(M_ / #DPUS)
+# For Hetsim benchmarks, we set M_ == #DPUs to simplify the task graph (no repetitions that depend on the number of available DPUs).
+# Just in case, there is also a configuration with M_ == 2048 independent of #DPUs
+#
+# input size: uint64(DPU)/double(CPU) * M_ * m * N_ * n
+# output size: uint64(DPU)/double(CPU) * M_ * m * N_ * n -- on DPU only; CPU version operates in-place
+# Upstream DPU version uses int64_t, -p 2048 -o 12288 -x 1 [implicit -m 16 -n 8]
+# Upstream CPU version uses double,  -p 2556 -o 4096 -m 16 -n 8 and fails with -o 12288 (allocation error)
+#
+# -p         2048 -o 2048 -m 16 -n 8 -> matrix size: 4 GiB
+# -p [64 .. 2304] -o 2048 -m 16 -n 8 -> matrix size: 128 MiB .. 4.5 GiB
+
+run_benchmark_nmc() {
+	local "$@"
+	set -e
+	make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets}  \
+		aspectc=1 aspectc_timing=1 dfatool_timing=0
+	bin/host_code -w 0 -e 2 -p ${cols} -o ${rows} -m ${tile_rows} -n ${tile_cols}
+}
+
+export -f run_benchmark_nmc
+
+fn=log/$(hostname)/ccmcc25-sdk${sdk}-sim
+
+source ~/lib/local/upmem/upmem-2025.1.0-Linux-x86_64/upmem_env.sh simulator
+
+echo "prim-benchmarks  TRNS  $(git describe --all --long)  $(git rev-parse HEAD)  $(date -R)" >> ${fn}.txt
+
+parallel -j1 --eta --joblog ${fn}.joblog --resume --header : \
+	run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 cols={cols} rows={rows} tile_cols={tile_cols} tile_rows={tile_rows} \
+	::: nr_dpus 1 4 16 32 64 \
+	::: rows 64 128 256 512 \
+	::: cols 64 128 256 512 \
+	::: tile_rows 16 \
+	::: tile_cols 8 \
+>> ${fn}.txt
diff --git a/TRNS/benchmark-scripts/ccmcc25.sh b/TRNS/benchmark-scripts/ccmcc25.sh
index 751566b..7c66306 100755
--- a/TRNS/benchmark-scripts/ccmcc25.sh
+++ b/TRNS/benchmark-scripts/ccmcc25.sh
@@ -1,73 +1,32 @@
 #!/bin/bash
 
-mkdir -p log/$(hostname) baselines/cpu/log/$(hostname)
-fn=log/$(hostname)/ccmcc25
-
-source /opt/upmem/upmem-2025.2.0-Linux-x86_64/upmem_env.sh
-
-# Args: -m m -n n -o M_ -p N_
-#
-# Input: (M_ * m) × (N_ * n) matrix
-# Output: (N_* n) × (M_ * m) matrix
-# Step 1: transpose (M_ * m) × N_ matrix that consists of tiles of size n
-#   CPU version: explicit
-#   DPU version: implicit (M_ * m write operations of #DPUs * n elements to DPUs)
-# Step 2: transpose m × n matrix; this happens N_ * M_ times.
-#   DPU version: Each tasklet transposes a single m × n matrix / tile.
-#   (16 × 8 tile takes up 1 KiB WRAM)
-# Step 3: Transpose M_ × n matrix that consists of tiles of size m.
-#
-# Note for DPU version: if M_ > #DPUs, steps 1 through 3 are repeated.
-# Number of repetitions == ceil(M_ / #DPUS)
-# For Hetsim benchmarks, we set M_ == #DPUs to simplify the task graph (no repetitions that depend on the number of available DPUs).
-# Just in case, there is also a configuration with M_ == 2048 independent of #DPUs
-#
-# input size: uint64(DPU)/double(CPU) * M_ * m * N_ * n
-# output size: uint64(DPU)/double(CPU) * M_ * m * N_ * n -- on DPU only; CPU version operates in-place
-# Upstream DPU version uses int64_t, -p 2048 -o 12288 -x 1 [implicit -m 16 -n 8]
-# Upstream CPU version uses double,  -p 2556 -o 4096 -m 16 -n 8 and fails with -o 12288 (allocation error)
-#
-# -p         2048 -o 2048 -m 16 -n 8 -> matrix size: 4 GiB
-# -p [64 .. 2304] -o 2048 -m 16 -n 8 -> matrix size: 128 MiB .. 4.5 GiB
+mkdir -p log/$(hostname)
 
 run_benchmark_nmc() {
 	local "$@"
 	set -e
 	sudo limit_ranks_to_numa_node ${numa_rank}
 	make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} dfatool_timing=0 aspectc=1 aspectc_timing=1
-	bin/host_code -w 0 -e 40 -p ${p} -o 2048 -m 16 -n 8
+	bin/host_code -w 0 -e 4 -p ${cols} -o ${rows} -m ${tile_rows} -n ${tile_cols}
 }
 
 export -f run_benchmark_nmc
 
-(
-
-echo "NMC single-node operation (1/4)" >&2
-
-parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
-	run_benchmark_nmc p={nr_dpus} nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \
-	::: numa_rank 0 1 \
-	::: nr_dpus 64 128 256 512 768 1024
-
-echo "NMC multi-node operation (2/4)" >&2
-
-parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
-	run_benchmark_nmc p={nr_dpus} nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \
-	::: numa_rank any \
-	::: nr_dpus 1536 2048 2304
+for sdk in 2023.2.0 2024.1.0 2024.2.0 2025.1.0; do
 
-echo "NMC single-node operation (3/4)" >&2
+	fn=log/$(hostname)/ccmcc25-sdk${sdk}
 
-parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \
-	run_benchmark_nmc p=2048 nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \
-	::: numa_rank 0 1 \
-	::: nr_dpus 64 128 256 512 768 1024
+	source /opt/upmem/upmem-${sdk}-Linux-x86_64/upmem_env.sh
 
-echo "NMC multi-node operation (4/4)" >&2
+	echo "prim-benchmarks  TRNS  $(git describe --all --long)  $(git rev-parse HEAD)  $(date -R)" >> ${fn}.txt
 
-parallel -j1 --eta --joblog ${fn}.4.joblog --resume --header : \
-	run_benchmark_nmc p=2048 nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \
-	::: numa_rank any \
-	::: nr_dpus 1536 2048 2304
+	parallel -j1 --eta --joblog ${fn}.joblog --resume --header : \
+		run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank=any cols={cols} rows={rows} tile_cols={tile_cols} tile_rows={tile_rows} \
+		::: nr_dpus 64 128 256 512 768 1024 1536 2048 2304 \
+		::: rows 1024 2048 4096 \
+		::: cols 64 128 256 512 768 1024 1536 2048 2304 \
+		::: tile_rows 16 \
+		::: tile_cols 8 \
+	>> ${fn}.txt
 
-) >> ${fn}.txt
+done
diff --git a/TRNS/include/dfatool_host.ah b/TRNS/include/dfatool_host.ah
index d9065d8..72978cc 100644
--- a/TRNS/include/dfatool_host.ah
+++ b/TRNS/include/dfatool_host.ah
@@ -5,7 +5,7 @@
 
 aspect DfatoolHostTiming : public DfatoolHostDPUTiming {
 
-	unsigned int n_rows_f1, n_rows_f2, n_cols_f1, n_cols_f2;
+	unsigned int n_rows_outer, n_rows_tile, n_cols_outer, n_cols_tile;
 	unsigned int element_size;
 
 	virtual int getKernel() { return kernel; }
@@ -19,18 +19,18 @@ aspect DfatoolHostTiming : public DfatoolHostDPUTiming {
 		/*
 		 * Input: (M_ * m) × (N_ * n) matrix
 		 */
-		n_rows_f1 = p->M_;
-		n_rows_f2 = p->m;
-		n_cols_f1 = p->N_;
-		n_cols_f2 = p->n;
-		printf("[>>] TRNS | n_dpus=%u n_rows_f1=%u n_rows_f2=%u n_cols_f1=%u n_cols_f2=%u\n", NR_DPUS, n_rows_f1, n_rows_f2, n_cols_f1, n_cols_f2);
+		n_rows_outer = p->M_;
+		n_rows_tile = p->m;
+		n_cols_outer = p->N_;
+		n_cols_tile = p->n;
+		printf("[>>] TRNS | n_dpus=%u n_rows_outer=%u n_rows_tile=%u n_cols_outer=%u n_cols_tile=%u\n", NR_DPUS, n_rows_outer, n_rows_tile, n_cols_outer, n_cols_tile);
 	}
 
 	advice call("% trns_host(...)") : after() {
-		printf("[--] TRNS | n_dpus=%u n_rows_f1=%u n_rows_f2=%u n_cols_f1=%u n_cols_f2=%u\n", NR_DPUS, n_rows_f1, n_rows_f2, n_cols_f1, n_cols_f2);
+		printf("[--] TRNS | n_dpus=%u n_rows_outer=%u n_rows_tile=%u n_cols_outer=%u n_cols_tile=%u\n", NR_DPUS, n_rows_outer, n_rows_tile, n_cols_outer, n_cols_tile);
 	}
 
 	advice execution("% main(...)") : after() {
-		printf("[<<] TRNS | n_dpus=%u n_rows_f1=%u n_rows_f2=%u n_cols_f1=%u n_cols_f2=%u\n", NR_DPUS, n_rows_f1, n_rows_f2, n_cols_f1, n_cols_f2);
+		printf("[<<] TRNS | n_dpus=%u n_rows_outer=%u n_rows_tile=%u n_cols_outer=%u n_cols_tile=%u\n", NR_DPUS, n_rows_outer, n_rows_tile, n_cols_outer, n_cols_tile);
 	}
 };