Minor fix

kaist-cp · Nov 19, 2024 · 5d554d4 · 5d554d4
1 parent 9a35402
commit 5d554d4
Show file tree

Hide file tree

Showing 4 changed files with 43 additions and 110 deletions.
diff --git a/hazardflow-designs/src/gemmini/execute/mod.rs b/hazardflow-designs/src/gemmini/execute/mod.rs
@@ -906,7 +906,7 @@ fn spad_write_req(resp: (MeshRespExtended, (Dataflow, U<3>, U<16>)), bank_i: U<2
         .mesh_resp
         .data
         .map(|e| {
-            let e_clipped = clip_with_saturation(U::from(e)); // Lower 8 bits
+            let e_clipped = clip_with_saturation(U::from(e[0])); // Lower 8 bits
             if activation == 1.into_u() {
                 // Check MSB for signedness.
                 if e_clipped[8 - 1] {
@@ -940,7 +940,7 @@ fn acc_write_req(resp: (MeshRespExtended, (Dataflow, U<3>, U<16>)), bank_i: U<1>
     let write_signals = compute_write_signal(resp);
     let resp = resp.0;
 
-    let wdata = resp.mesh_resp.data.map(|v| U::from(v.sext::<32>()));
+    let wdata = resp.mesh_resp.data.map(|v| U::from(v[0].sext::<32>()));
     let wmask = write_signals.w_mask.map(|v| v.repeat::<4>()).concat();
 
     if write_signals.start_array_outputting

diff --git a/hazardflow-designs/src/gemmini/execute/systolic_array/mesh_with_delays.rs b/hazardflow-designs/src/gemmini/execute/systolic_array/mesh_with_delays.rs
@@ -19,6 +19,8 @@ const FIFO_LENGTH: usize = MAX_SIMULTANEOUS_MATMULS + 1;
 pub type A = Array<Array<S<INPUT_BITS>, TILE_ROWS>, MESH_ROWS>;
 /// Type of data of `b`.
 pub type B = Array<Array<S<INPUT_BITS>, TILE_COLS>, MESH_COLS>;
+/// Type of data of `c`.
+pub type C = Array<Array<S<OUTPUT_BITS>, TILE_COLS>, MESH_COLS>;
 /// Type of data of `d`.
 pub type D = Array<Array<S<INPUT_BITS>, TILE_COLS>, MESH_COLS>;
 
@@ -88,8 +90,8 @@ pub struct MeshResp {
     pub tag: MeshTag,
     /// Indicates that the row represents the last row.
     pub last: bool,
-    /// Data.
-    pub data: Array<S<OUTPUT_BITS>, MESH_COLS>,
+    /// Output data.
+    pub data: C,
 }
 
 /// Matmul operation configuration.

diff --git a/hazardflow-designs/src/gemmini/execute/systolic_array/utils.rs b/hazardflow-designs/src/gemmini/execute/systolic_array/utils.rs
@@ -146,34 +146,29 @@ pub fn preprocess_type((data, req): (Valid<(A, B, D)>, Valid<(ReqExtended, bool)
 }
 
 /// Interface type conversion.
-pub fn postprocess_type(
-    (out_row, out_col): (MeshRowData, MeshColData),
-) -> Valid<(Array<S<OUTPUT_BITS>, MESH_COLS>, PeColControl)> {
+pub fn postprocess_type((out_row, out_col): (MeshRowData, MeshColData)) -> Valid<(C, PeColControl)> {
     // # Safety
     //
     // All the input and output interfaces are `Valid` type.
     unsafe {
-        (out_row, out_col).fsm::<Valid<(Array<S<OUTPUT_BITS>, MESH_COLS>, PeColControl)>, ()>(
-            (),
-            |(_, col_data), _, ()| {
-                let out_valid = col_data[0][0].0.is_some();
-                let dataflow_os = col_data[0][0].1.is_some_and(|v| matches!(v.control.dataflow, Dataflow::OS));
-
-                let out_b = col_data
-                    .map(|tile_r| tile_r.map(|(data, _)| data.map_or(0.into_u(), |v| U::from(v.b))).concat())
-                    .map(S::from);
-                let out_c = col_data
-                    .map(|tile_r| tile_r.map(|(data, _)| data.map_or(0.into_u(), |v| U::from(v.d))).concat())
-                    .map(S::from);
-
-                let matmul_result = if dataflow_os { out_c } else { out_b };
-
-                let ep = if out_valid { Some((matmul_result, col_data[0][0].1.unwrap())) } else { None };
-                let ir0 = ().repeat::<1>().repeat::<MESH_COLS>();
-                let ir1 = ((), ()).repeat::<1>().repeat::<MESH_COLS>();
-
-                (ep, (ir0, ir1), ())
-            },
-        )
+        (out_row, out_col).fsm::<Valid<(C, PeColControl)>, ()>((), |(_, col_data), _, ()| {
+            let out_valid = col_data[0][0].0.is_some();
+            let dataflow_os = col_data[0][0].1.is_some_and(|v| matches!(v.control.dataflow, Dataflow::OS));
+
+            let out_b = col_data
+                .map(|tile_r| tile_r.map(|(data, _)| data.map_or(0.into_u(), |v| U::from(v.b))).concat())
+                .map(|v| S::from(v).repeat::<TILE_COLS>());
+            let out_c = col_data
+                .map(|tile_r| tile_r.map(|(data, _)| data.map_or(0.into_u(), |v| U::from(v.d))).concat())
+                .map(|v| S::from(v).repeat::<TILE_COLS>());
+
+            let matmul_result = if dataflow_os { out_c } else { out_b };
+
+            let ep = if out_valid { Some((matmul_result, col_data[0][0].1.unwrap())) } else { None };
+            let ir0 = ().repeat::<1>().repeat::<MESH_COLS>();
+            let ir1 = ((), ()).repeat::<1>().repeat::<MESH_COLS>();
+
+            (ep, (ir0, ir1), ())
+        })
     }
 }
diff --git a/scripts/gemmini/unit_tests/mesh_with_delays/test_mesh_with_delays.py b/scripts/gemmini/unit_tests/mesh_with_delays/test_mesh_with_delays.py
@@ -313,38 +313,18 @@ async def ws_no_transpose(dut):
     await tb.in_ctrl_req.send(req_with_rob_id(WS, False, False, REG2))
 
     for i in range(16):
-        await tb.in_a_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=0,
-            )
-        )
-        await tb.in_b_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=0,
-            )
-        )
+        await tb.in_a_data_req.send(InpDataTransaction(payload_Some_0_0=0))
+        await tb.in_b_data_req.send(InpDataTransaction(payload_Some_0_0=0))
         # When running WS dataflow, weight should be sent in reverse order
         # due to the way the weight is preloaded in the PEs.
-        await tb.in_d_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=concatenate_data(weight[15 - i], 8),
-            )
-        )
+        await tb.in_d_data_req.send(InpDataTransaction(payload_Some_0_0=concatenate_data(weight[15 - i], 8)))
 
     # 2. Send activation and bias data
     await tb.in_ctrl_req.send(req_with_none_rob_id(WS, False, False, REG1))
 
     for i in range(16):
-        await tb.in_a_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=concatenate_data(activation[i], 8),
-            )
-        )
-        await tb.in_b_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=concatenate_data(bias[i], 8),
-            )
-        )
+        await tb.in_a_data_req.send(InpDataTransaction(payload_Some_0_0=concatenate_data(activation[i], 8)))
+        await tb.in_b_data_req.send(InpDataTransaction(payload_Some_0_0=concatenate_data(bias[i], 8)))
         await tb.in_d_data_req.send(InpDataTransaction(payload_Some_0_0=0))
 
     output_data = []
@@ -386,30 +366,18 @@ async def ws_transpose_a(dut):
     # 1. Preload weight data
     await tb.in_ctrl_req.send(req_with_rob_id(WS, True, False, REG2))
     for i in range(16):
-        await tb.in_a_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=concatenate_data(activation[i], 8),
-            )
-        )
+        await tb.in_a_data_req.send(InpDataTransaction(payload_Some_0_0=concatenate_data(activation[i], 8)))
         await tb.in_b_data_req.send(InpDataTransaction(payload_Some_0_0=0))
         # When running WS dataflow, weight should be sent in reverse order
         # due to the way the weight is preloaded in the PEs.
-        await tb.in_d_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=concatenate_data(weight[15 - i], 8),
-            )
-        )
+        await tb.in_d_data_req.send(InpDataTransaction(payload_Some_0_0=concatenate_data(weight[15 - i], 8)))
 
     # 2. Send activation and bias data
     await tb.in_ctrl_req.send(req_with_none_rob_id(WS, True, False, REG1))
 
     for i in range(16):
         await tb.in_a_data_req.send(InpDataTransaction(payload_Some_0_0=0))
-        await tb.in_b_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=concatenate_data(bias[i], 8),
-            )
-        )
+        await tb.in_b_data_req.send(InpDataTransaction(payload_Some_0_0=concatenate_data(bias[i], 8)))
         await tb.in_d_data_req.send(InpDataTransaction(payload_Some_0_0=0))
 
     output_data = []
@@ -537,11 +505,7 @@ async def os_no_transpose(dut):
     await tb.in_ctrl_req.send(req_with_none_rob_id(OS, False, False, REG1))
     for i in range(16):
         await tb.in_a_data_req.send(InpDataTransaction(payload_Some_0_0=0))
-        await tb.in_b_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=concatenate_data(weight[i], 8),
-            )
-        )
+        await tb.in_b_data_req.send(InpDataTransaction(payload_Some_0_0=concatenate_data(weight[i], 8)))
         await tb.in_d_data_req.send(InpDataTransaction(payload_Some_0_0=0))
 
     await tb.in_ctrl_req.send(os_flush_request(REG1, rnd_shift))
@@ -592,24 +556,12 @@ async def os_transpose_a(dut):
     for i in range(16):
         await tb.in_a_data_req.send(InpDataTransaction(payload_Some_0_0=0))
         await tb.in_b_data_req.send(InpDataTransaction(payload_Some_0_0=0))
-        await tb.in_d_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=concatenate_data(bias[15 - i], 8),
-            )
-        )
+        await tb.in_d_data_req.send(InpDataTransaction(payload_Some_0_0=concatenate_data(bias[15 - i], 8)))
 
     await tb.in_ctrl_req.send(req_with_none_rob_id(OS, True, False, REG1))
     for i in range(16):
-        await tb.in_a_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=concatenate_data(activation[i], 8),
-            )
-        )
-        await tb.in_b_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=concatenate_data(weight[i], 8),
-            )
-        )
+        await tb.in_a_data_req.send(InpDataTransaction(payload_Some_0_0=concatenate_data(activation[i], 8)))
+        await tb.in_b_data_req.send(InpDataTransaction(payload_Some_0_0=concatenate_data(weight[i], 8)))
         await tb.in_d_data_req.send(InpDataTransaction(payload_Some_0_0=0))
 
     await tb.in_ctrl_req.send(os_flush_request(REG1, rnd_shift))
@@ -646,9 +598,7 @@ async def os_transpose_both(dut):
     await ClockCycles(dut.clk, 10)
 
     # Input data
-    activation, weight, bias, expected_output, rnd_shift = generate_os_test_data(
-        True, True
-    )
+    activation, weight, bias, expected_output, rnd_shift = generate_os_test_data(True, True)
 
     tb.log.info(f"[Ouptut-Stationary] Transpose A and B, Shift: {rnd_shift}")
     tb.log.info(f"Activation:\n{activation}")
@@ -661,33 +611,19 @@ async def os_transpose_both(dut):
     for i in range(16):
         await tb.in_a_data_req.send(InpDataTransaction(payload_Some_0_0=0))
         await tb.in_b_data_req.send(InpDataTransaction(payload_Some_0_0=0))
-        await tb.in_d_data_req.send(
-            InpDataTransaction(payload_Some_0_0=concatenate_data(bias[15 - i], 8))
-        )
+        await tb.in_d_data_req.send(InpDataTransaction(payload_Some_0_0=concatenate_data(bias[15 - i], 8)))
 
     # 1. Send weight data
     await tb.in_ctrl_req.send(req_with_rob_id(OS, True, True, REG2))
     for i in range(16):
         await tb.in_a_data_req.send(InpDataTransaction(payload_Some_0_0=0))
-        await tb.in_b_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=concatenate_data(weight[i], 8),
-            )
-        )
-        await tb.in_d_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=concatenate_data(bias[15 - i], 8),
-            )
-        )
+        await tb.in_b_data_req.send(InpDataTransaction(payload_Some_0_0=concatenate_data(weight[i], 8)))
+        await tb.in_d_data_req.send(InpDataTransaction(payload_Some_0_0=concatenate_data(bias[15 - i], 8)))
 
     # 2. Send activation data
     await tb.in_ctrl_req.send(req_with_none_rob_id(OS, True, True, REG1))
     for i in range(16):
-        await tb.in_a_data_req.send(
-            InpDataTransaction(
-                payload_Some_0_0=concatenate_data(activation[i], 8),
-            )
-        )
+        await tb.in_a_data_req.send(InpDataTransaction(payload_Some_0_0=concatenate_data(activation[i], 8)))
         await tb.in_b_data_req.send(InpDataTransaction(payload_Some_0_0=0))
         await tb.in_d_data_req.send(InpDataTransaction(payload_Some_0_0=0))