[cheshire] Back-ref sw compilation flow for fmatmul

cheshire/README.md

@@ -8,7 +8,7 @@ Support for FPGA synthesis was added to Ara by integrating it into Cheshire. Sin
 
 1.  **Navigate to the Root Directory**
     Ensure you are in the root directory where the Makefile is located.
-    
+
 2.  **Set up environment**
     Set the `BACKREF_CHS_ROOT` variable to root directory of the Cheshire repository where you want to build the bitstream.
 
@@ -26,17 +26,17 @@ This command will:
 Here's how we use back-referencing in our setup:
 
 1.  **Generate Custom TCL File**:
-    
+
     -   We generate a custom `add_sources.vcu128.tcl` file using the `bender script vivado` command with our specific targets (`-t fpga -t cv64a6_imafdcv_sv39 -t cva6 -t vcu128 --define ARA`).
     -   This custom TCL file includes all the necessary sources and configurations required for the FPGA synthesis with Cheshire + Ara.
 
 2.  **Copy Custom TCL File**:
-    
+
     -   The generated custom TCL file is then copied into the Cheshire directory (`$(BACKREF_CHS_XIL_SCRIPTS)/add_sources.vcu128.tcl`).
 
 3.  **Invoke Cheshire Compile Flow**:
-    
+
     -   With the custom TCL file in place, we invoke the Cheshire compile flow by running `make -C $(BACKREF_CHS_ROOT) chs-xilinx-all`.
     -   The Cheshire compile flow target depends on the `add_sources.vcu128.tcl` file, and since we have provided our custom version, it will use ours for the synthesis process.
 
-This method ensures that we can extend and customize the compile flow for our specific needs without modifying the Cheshire repository directly.
+This method ensures that we can extend and customize the compile flow for our specific needs without modifying the Cheshire repository directly.

cheshire/sw/Makefile

@@ -6,22 +6,34 @@
 #
 # Copy and compile vector software on Cheshire
 
-CHS_ROOT ?= $(realpath ../../../../../..)
-ARA_SW   := $(dir $(realpath $(firstword $(MAKEFILE_LIST))))
+CHS_ROOT ?= $(dir $(realpath $(firstword $(MAKEFILE_LIST))))/../../../../../..
+ARA_ROOT := $(dir $(realpath $(firstword $(MAKEFILE_LIST))))/../..
 CHS_SW   := $(CHS_ROOT)/sw
-SRC      := $(wildcard $(ARA_SW)/*.c) $(wildcard $(ARA_SW)/*.h)
+ARA_SW   := $(ARA_ROOT)/cheshire/sw
+ARA_APPS := $(ARA_ROOT)/apps
+
+APPS   := $(patsubst $(ARA_APPS)/%/main.c,%,$(shell find $(ARA_APPS) -name "main.c"))
+SW_C   := $(wildcard $(ARA_SW)/src/*.c)
+DEPS_H := $(wildcard $(ARA_SW)/include/*.h)
+
+ARA_CONFIGURATION ?= 2_lanes
+include $(ARA_ROOT)/config/$(ARA_CONFIGURATION).mk
 
 # Get the original compiler options and add the support for vector extension
 CHS_SW_FLAGS ?= $(shell grep "^CHS_SW_FLAGS\s\+?=\s\+" -- $(CHS_SW)/sw.mk | sed 's/^.*?= //' | sed s/rv64gc/rv64gcv/)
+# Tweak the compilation to include Cheshire-related headers and files
+CHS_SW_FLAGS += -DCHESHIRE -DNR_LANES=$(nr_lanes) -DVLEN=$(vlen)
 
-.PHONY: chs-sw-all copy_vector_sw
+.PHONY: chs-sw-all copy_vector_sw copy-vector-deps
 
 # Forward build command to the main Cheshire makefile and attach the correct -march
-# Rename the .c vector files not to break the cheshire vanilla flow
-chs-sw-all: copy-vector-sw
+chs-sw-all: copy-vector-sw copy-vector-deps
 	make -C $(CHS_ROOT) $@ CHS_SW_FLAGS="$(CHS_SW_FLAGS)"
-	for f in $(filter %.c, $(SRC)); do mv $(CHS_SW)/tests/$f $(CHS_SW)/tests/$f.bkp; done
 
-# Copy the vector programs to cheshire
-copy-vector-sw:
-	cp $(SRC) $(CHS_SW)/tests
+# Copy the dependencies from this folder to Cheshire
+copy-vector-deps: $(DEPS_H)
+	cp $^ $(CHS_SW)/tests
+
+# Copy the vector programs from the src folder to cheshire
+copy-vector-sw: $(SW_C)
+	cp $^ $(CHS_SW)/tests

cheshire/sw/README.md

@@ -1,9 +1,11 @@
 # Build software for Cheshire Ara
 
-Compile the `.c` programs in this folder with:
+## Compile the vector code for Cheshire
+
+Compile the source files with the vector extension support enable:
 
 ```bash
 make chs-sw-all
 ```
 
-This command will copy the necessary source files into Cheshire's `sw/tests` directory and compile them with the support for vector extension.
+This command will also copy the necessary dependencies to `sw/tests` and enable the vector extension at compile time.

cheshire/sw/cheshire_util.h → cheshire/sw/include/cheshire_util.h

@@ -11,14 +11,14 @@
 
 #include "printf.h"
 
-inline void cheshire_start() {
+void cheshire_start() {
   // Initialize Cheshire's UART
   uint32_t rtc_freq = *reg32(&__base_regs, CHESHIRE_RTC_FREQ_REG_OFFSET);
   uint64_t reset_freq = clint_get_core_freq(rtc_freq, 2500);
   uart_init(&__base_uart, reset_freq, __BOOT_BAUDRATE);
 }
 
-inline void cheshire_finish() {
+void cheshire_end() {
   // Flush teh UART
   uart_write_flush(&__base_uart);
 }

cheshire/sw/include/encoding.h

@@ -0,0 +1 @@
+../../../apps/common/encoding.h

cheshire/sw/include/fmatmul.c.h

@@ -0,0 +1 @@
+../../../apps/fmatmul/kernel/fmatmul.c

cheshire/sw/include/fmatmul.h

@@ -0,0 +1 @@
+../../../apps/fmatmul/kernel/fmatmul.h

cheshire/sw/vector_util.h → cheshire/sw/include/vector_util.h

@@ -13,9 +13,31 @@
 #include <riscv_vector.h>
 #include "encoding.h"
 
+#define start_timer()
+#define stop_timer()
+#define get_timer() 0
+
+#define FABS(x) ((x < 0) ? -x : x)
+
 inline void enable_rvv() {
   asm volatile ("li t0, %0" :: "i"(MSTATUS_VS));
   asm volatile ("csrs mstatus, t0" );
 }
 
+inline int similarity_check(double a, double b, double threshold) {
+  double diff = a - b;
+  if (FABS(diff) > threshold)
+    return 0;
+  else
+    return 1;
+}
+
+inline int similarity_check_32b(float a, float b, float threshold) {
+  float diff = a - b;
+  if (FABS(diff) > threshold)
+    return 0;
+  else
+    return 1;
+}
+
 #endif

cheshire/sw/src/fmatmul.c

@@ -0,0 +1,126 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Matteo Perotti <[email protected]>
+//
+// fmatmul wrapper for Cheshire
+
+#include "regs/cheshire.h"
+#include "dif/clint.h"
+#include "dif/uart.h"
+#include "params.h"
+#include "util.h"
+
+#include "cheshire_util.h"
+#include "vector_util.h"
+
+#include "fmatmul.c.h"
+
+#ifndef _MM_SIZE_
+#define _MM_SIZE_ 32
+#endif
+
+// Define Matrix dimensions:
+// C = AB with A=[MxN], B=[NxP], C=[MxP]
+uint64_t M = _MM_SIZE_;
+uint64_t N = _MM_SIZE_;
+uint64_t P = _MM_SIZE_;
+
+// Max matrix size: 256x256
+double a[_MM_SIZE_*_MM_SIZE_] __attribute__((aligned(32 * NR_LANES)));
+double b[_MM_SIZE_*_MM_SIZE_] __attribute__((aligned(32 * NR_LANES)));
+double c[_MM_SIZE_*_MM_SIZE_] __attribute__((aligned(32 * NR_LANES)));
+// Gold results
+double g[_MM_SIZE_*_MM_SIZE_] __attribute__((aligned(32 * NR_LANES)));
+
+#define THRESHOLD 0.001
+
+// Verify the matrix
+int verify_matrix(double *result, double *gold, size_t R, size_t C,
+                  double threshold) {
+  for (uint64_t i = 0; i < R; ++i) {
+    for (uint64_t j = 0; j < C; ++j) {
+      int idx = i * C + j;
+      if (!similarity_check(result[idx], gold[idx], threshold)) {
+        return (i + j) == 0 ? -1 : idx;
+      }
+    }
+  }
+  return 0;
+}
+
+int main() {
+  printf("\n");
+  printf("=============\n");
+  printf("=  FMATMUL  =\n");
+  printf("=============\n");
+  printf("\n");
+  printf("------------------------------------------------------------\n");
+  printf("Calculating a (%d x %d) x (%d x %d) matrix multiplication...\n", M,
+         N, N, P);
+  printf("------------------------------------------------------------\n");
+  printf("\n");
+
+  cheshire_start();
+  enable_rvv();
+
+  unsigned int s = M;
+
+  // Initialize matrices
+  for (unsigned int i = 0; i < s; ++i) {
+    for (unsigned int k = 0; k < s; ++k) {
+      a[k + i*s] = (double) (i + k);
+    }
+  }
+  for (unsigned int k = 0; k < s; ++k) {
+    for (unsigned int j = 0; j < s; ++j) {
+      b[j + k*s] = (double) (k - j);
+    }
+  }
+
+  // Run scalar check
+  printf("Calculating fmatmul on scalar core...\n");
+  for (unsigned int i = 0; i < s; ++i) {
+    for (unsigned int j = 0; j < s; ++j) {
+      double sum = 0;
+      for (unsigned int k = 0; k < s; ++k) {
+        sum += a[k + i * s] * b[j + k * s];
+      }
+      g[i + j*s] = sum;
+    }
+  }
+
+  // Run vector kernel
+  printf("Calculating fmatmul on vector core...\n");
+  start_timer();
+  fmatmul(c, a, b, s, s, s);
+  stop_timer();
+
+  // Metrics
+  int64_t runtime = get_timer();
+  float performance = 2.0 * s * s * s / runtime;
+  float utilization = 100 * performance / (2.0 * NR_LANES);
+
+  printf("The execution took %d cycles.\n", runtime);
+  printf("The performance is %f FLOP/cycle (%f%% utilization).\n",
+         performance, utilization);
+
+  // Verify the result only for s == M (to keep it simple)
+  if (s == M) {
+    printf("Verifying result...\n");
+    int error = verify_matrix(c, g, s, s, THRESHOLD);
+    if (error != 0) {
+      printf("Error code %d\n", error);
+      printf("c[%d]=%d\n", error, c[error]);
+      return error;
+    } else {
+      printf("Passed.\n");
+    }
+  }
+
+
+  cheshire_end();
+
+  return 0;
+}