Recover lost changes

src/main/resources/csrc/SimMemTrace.cc

@@ -1,50 +1,34 @@
+#ifndef NO_VPI
 #include <vpi_user.h>
 #include <svdpi.h>
-#include <memory>
+#endif
 #include <string>
-#include <fstream>
 #include <cstdio>
+#include <cassert>
 #include <unistd.h>
+#include "SimMemTrace.h"
 
-class MemTraceReader;
-
-// Global singleton instance of MemTraceReader
-static std::unique_ptr<MemTraceReader> reader;
-
-struct MemTraceLine {
-  bool valid = false;
-  unsigned long cycle = 0;
-  char loadstore[10];
-  int core_id = 0;
-  int thread_id = 0;
-  unsigned long address = 0;
-  unsigned long data = 0;
-  int data_size = 0;
-};
-
-class MemTraceReader {
-public:
-  MemTraceReader(const std::string &filename) {
-    char cwd[4096];
-    if (getcwd(cwd, sizeof(cwd))) {
-        printf("MemTraceReader: current working dir: %s\n", cwd);
-    }
-
-    infile.open(filename);
-    if (infile.fail()) {
-        fprintf(stderr, "failed to open file %s\n", filename);
-    }
+MemTraceReader::MemTraceReader(const std::string &filename) {
+  char cwd[4096];
+  if (getcwd(cwd, sizeof(cwd))) {
+    printf("MemTraceReader: current working dir: %s\n", cwd);
   }
-  ~MemTraceReader() {
-    infile.close();
-    printf("MemTraceReader destroyed\n");
+
+  infile.open(filename);
+  if (infile.fail()) {
+    fprintf(stderr, "failed to open file %s\n", filename.c_str());
   }
-  MemTraceLine tick();
+}
 
-  std::ifstream infile;
-};
+MemTraceReader::~MemTraceReader() {
+  infile.close();
+  printf("MemTraceReader destroyed\n");
+}
 
-MemTraceLine MemTraceReader::tick() {
+// Parse trace file in its entirety and store it into internal structure.
+// TODO: might block for a long time when the trace gets big, check if need to
+// be broken down
+void MemTraceReader::parse() {
   MemTraceLine line;
 
   printf("MemTraceReader: started parsing\n");
@@ -60,9 +44,9 @@ MemTraceLine MemTraceReader::tick() {
   printf("MemTraceReader: finished parsing\n");
 }
 
-// Try to read a memory request that might have happened at a given cycle, on a
-// given SIMD lane (= "thread").  In case no request happened at that point,
-// return an empty line with .valid = false.
+// Try to read a memory request that might have happened at a given cycle, on
+// given thread.  In case no request happened at that point, return an empty
+// line with .valid = false.
 MemTraceLine MemTraceReader::read_trace_at(const long cycle,
                                            const int thread_id) {
   MemTraceLine line;
@@ -75,16 +59,14 @@ MemTraceLine MemTraceReader::read_trace_at(const long cycle,
   }
 
   line = *read_pos;
-  // It should always be guaranteed that we consumed all of the past lines, and
-  // the next line is in the future.
+  // It should always be guaranteed that the next line is not read yet.
   if (line.cycle < cycle) {
     fprintf(stderr, "line.cycle=%ld, cycle=%ld\n", line.cycle, cycle);
     assert(false && "some trace lines are left unread in the past");
   }
 
   if (line.cycle > cycle) {
-    // We haven't reached the cycle mark specified in this line yet, so we don't
-    // read it right now.
+    // It's not ready to read this line yet.
     return MemTraceLine{};
   } else if (line.cycle == cycle) {
     printf("fire! cycle=%ld, valid=%d\n", cycle, line.valid);
@@ -99,11 +81,13 @@ MemTraceLine MemTraceReader::read_trace_at(const long cycle,
 }
 
 extern "C" void memtrace_init(const char *filename) {
-  reader = std::make_unique<MemTraceReader>(filename);
   printf("memtrace_init: filename=[%s]\n", filename);
+
+  reader = std::make_unique<MemTraceReader>(filename);
+  // parse file upfront
+  reader->parse();
 }
 
-// TODO: accept core_id as well
 extern "C" void memtrace_query(unsigned char trace_read_ready,
                                unsigned long trace_read_cycle,
                                int trace_read_thread_id,
@@ -113,9 +97,16 @@ extern "C" void memtrace_query(unsigned char trace_read_ready,
   printf("memtrace_query(cycle=%ld, tid=%d)\n", trace_read_cycle,
          trace_read_thread_id);
 
+  if (!trace_read_ready) {
+    return;
+  }
+
+  auto line = reader->read_trace_at(trace_read_cycle, trace_read_thread_id);
   *trace_read_valid = line.valid;
-  *trace_read_cycle = line.cycle;
   *trace_read_address = line.address;
+  // This means finished and valid will go up at the same cycle.  Need to
+  // handle this without skipping the last line.
+  *trace_read_finished = reader->finished();
 
   return;
 }

src/main/resources/csrc/SimMemTrace.h

@@ -0,0 +1,40 @@
+#include <vector>
+#include <memory>
+#include <fstream>
+
+class MemTraceReader;
+
+// Global singleton instance of MemTraceReader
+static std::unique_ptr<MemTraceReader> reader;
+
+struct MemTraceLine {
+  bool valid = false;
+  long cycle = 0;
+  char loadstore[10];
+  int core_id = 0;
+  int thread_id = 0;
+  unsigned long address = 0;
+  unsigned long data = 0;
+  int data_size = 0;
+};
+
+class MemTraceReader {
+public:
+  MemTraceReader(const std::string &filename);
+  ~MemTraceReader();
+  void parse();
+  MemTraceLine read_trace_at(const long cycle, const int thread_id);
+  bool finished() const { return read_pos == trace.cend(); }
+
+  std::ifstream infile;
+  std::vector<MemTraceLine> trace;
+  std::vector<MemTraceLine>::const_iterator read_pos;
+};
+
+extern "C" void memtrace_init(const char *filename);
+extern "C" void memtrace_query(unsigned char trace_read_ready,
+                               unsigned long trace_read_cycle,
+                               int trace_read_thread_id,
+                               unsigned char *trace_read_valid,
+                               unsigned long *trace_read_address,
+                               unsigned char *trace_read_finished);

src/main/resources/vsrc/SimMemTrace.v

@@ -1,65 +1,103 @@
 `define DATA_WIDTH 64
+`define MAX_NUM_THREADS 32
 
 import "DPI-C" function void memtrace_init(
-    input  string  filename
+  input  string  filename
 );
 
-import "DPI-C" function void memtrace_tick
+// Make sure to sync the parameters for:
+// (1) import "DPI-C" declaration
+// (2) C function declaration
+// (3) DPI function calls inside initial/always blocks
+import "DPI-C" function void memtrace_query
 (
-    output bit     trace_read_valid,
-    input  bit     trace_read_ready,
-    output longint trace_read_cycle,
-    output longint trace_read_address
+  input  bit     trace_read_ready,
+  input  longint trace_read_cycle,
+  input  int     trace_read_tid,
+  output bit     trace_read_valid,
+  output longint trace_read_address,
+  output bit     trace_read_finished
 );
 
-module SimMemTrace  (
-    input              clock,
-    input              reset,
+module SimMemTrace #(parameter NUM_THREADS = 4) (
+  input              clock,
+  input              reset,
 
-    output                   trace_read_valid,
-    input                    trace_read_ready,
-    output [`DATA_WIDTH-1:0] trace_read_cycle,
-    output [`DATA_WIDTH-1:0] trace_read_address
+  // These have to match the IO port of the Chisel wrapper module.
+  input                    trace_read_ready,
+  output [NUM_THREADS-1:0] trace_read_valid,
+  output [`DATA_WIDTH*NUM_THREADS-1:0] trace_read_address,
+  output                   trace_read_finished
 );
+  bit __in_valid[NUM_THREADS-1:0];
+  longint __in_address[NUM_THREADS-1:0];
+  bit __in_finished;
+  string __uartlog;
 
-    bit __in_valid;
-    longint __in_cycle;
-    longint __in_address;
-    string __uartlog;
-    int __uartno;
+  // Cycle counter that is used to query C parser whether we have a request
+  // coming in at the current cycle.
+  reg [`DATA_WIDTH-1:0] cycle_counter;
+  wire [`DATA_WIDTH-1:0] next_cycle_counter;
+  assign next_cycle_counter = cycle_counter + 1'b1;
 
-    initial begin
-        $value$plusargs("uartlog=%s", __uartlog);
-        memtrace_init("vecadd.core1.thread4.trace");
+  // registers that stage outputs of the C parser
+  reg [NUM_THREADS-1:0] __in_valid_reg;
+  reg [`DATA_WIDTH-1:0] __in_address_reg [NUM_THREADS-1:0];
+  reg __in_finished_reg;
+
+  genvar g;
+
+  generate
+    for (g = 0; g < NUM_THREADS; g = g + 1) begin
+      assign trace_read_valid[g] = __in_valid_reg[g];
+      assign trace_read_address[`DATA_WIDTH*(g+1)-1:`DATA_WIDTH*g]  = __in_address_reg[g];
     end
+  endgenerate
+  assign trace_read_finished = __in_finished_reg;
 
-    reg __in_valid_reg;
-    reg [`DATA_WIDTH-1:0] __in_cycle_reg;
-    reg [`DATA_WIDTH-1:0] __in_address_reg;
+  initial begin
+      /* $value$plusargs("uartlog=%s", __uartlog); */
+      memtrace_init("vecadd.core1.thread4.trace");
+  end
 
-    assign trace_read_valid   = __in_valid_reg;
-    assign trace_read_cycle   = __in_cycle_reg;
-    assign trace_read_address = __in_address_reg;
+  // Evaluate the signals on the positive edge
+  always @(posedge clock) begin
+    if (reset) begin
+      for (integer tid = 0; tid < NUM_THREADS; tid = tid + 1) begin
+        __in_valid[tid] = 1'b0;
+        __in_address[tid] = `DATA_WIDTH'b0;
+      end
+      __in_finished = 1'b0;
 
-    // Evaluate the signals on the positive edge
-    always @(posedge clock) begin
-        if (reset) begin
-            __in_valid = 1'b0;
+      cycle_counter <= `DATA_WIDTH'b0;
 
-            __in_valid_reg <= 1'b0;
-            __in_cycle_reg <= `DATA_WIDTH'b0;
-        end else begin
-            memtrace_tick(
-                __in_valid,
-                trace_read_ready,
-                __in_cycle,
-                __in_address
-            );
+      for (integer tid = 0; tid < NUM_THREADS; tid = tid + 1) begin
+        __in_valid_reg[tid] <= 1'b0;
+        __in_address_reg[tid] <= `DATA_WIDTH'b0;
+      end
+      __in_finished_reg <= 1'b0;
+    end else begin
+      cycle_counter <= next_cycle_counter;
 
-            __in_valid_reg   <= __in_valid;
-            __in_cycle_reg   <= __in_cycle;
-            __in_address_reg <= __in_address;
-        end
+      for (integer tid = 0; tid < NUM_THREADS; tid = tid + 1) begin
+        memtrace_query(
+          trace_read_ready,
+          // Since parsed results are latched to the output on the next
+          // cycle due to staging registers, we need to pass in the next cycle
+          // to sync up.
+          next_cycle_counter,
+          tid,
+          __in_valid[tid],
+          __in_address[tid],
+          __in_finished
+        );
+      end
+
+      for (integer tid = 0; tid < NUM_THREADS; tid = tid + 1) begin
+        __in_valid_reg[tid]   <= __in_valid[tid];
+        __in_address_reg[tid] <= __in_address[tid];
+      end
+      __in_finished_reg <= __in_finished;
     end
-
+  end
 endmodule

src/main/scala/tilelink/Coalescing.scala

@@ -159,6 +159,7 @@ class SimMemTrace(val trace_file: String, num_threads: Int) extends BlackBox(Map
 
   addResource("/vsrc/SimMemTrace.v")
   addResource("/csrc/SimMemTrace.cc")
+  addResource("/csrc/SimMemTrace.h")
 }