[software] Add f32 and f16 dotp app

pulp-platform · Sep 5, 2024 · c346c1c · c346c1c
1 parent e77933c
commit c346c1c
Show file tree

Hide file tree

Showing 10 changed files with 848 additions and 123 deletions.
diff --git a/software/apps/baremetal/dotp_f16/main.c b/software/apps/baremetal/dotp_f16/main.c
@@ -0,0 +1,78 @@
+// Copyright 2021 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+// Author: Marco Bertuletti, ETH Zurich
+
+#include <stdint.h>
+#include <string.h>
+
+#include "dma.h"
+#include "encoding.h"
+#include "printf.h"
+#include "runtime.h"
+#include "synchronization.h"
+
+#include "data_dotp_f16.h"
+#define NUM_BANKS (NUM_CORES * BANKING_FACTOR)
+// #define SINGLE_CORE_REDUCTION
+#define BINARY_REDUCTION
+
+// Vectors for kernel computation
+__fp16 l1_A[LEN] __attribute__((aligned(LEN), section(".l1_prio")));
+__fp16 l1_B[LEN] __attribute__((aligned(LEN), section(".l1_prio")));
+uint32_t red_barrier[NUM_BANKS]
+    __attribute__((aligned(NUM_BANKS), section(".l1_prio")));
+__fp16 sum[2 * NUM_BANKS]
+    __attribute__((aligned(NUM_BANKS), section(".l1_prio")));
+
+#include "baremetal/mempool_dotp_f16.h"
+
+int main() {
+
+  uint32_t core_id = mempool_get_core_id();
+  uint32_t num_cores = mempool_get_core_count();
+  uint32_t time_init, time_end;
+  mempool_barrier_init(core_id);
+
+  time_init = 0;
+  time_end = 0;
+  if (core_id == 0) {
+    dma_memcpy_blocking(l1_A, l2_A, LEN * sizeof(int16_t));
+    dma_memcpy_blocking(l1_B, l2_B, LEN * sizeof(int16_t));
+  }
+  for (uint32_t k = core_id; k < NUM_BANKS; k += num_cores) {
+    sum[k] = 0;
+    red_barrier[k] = 0;
+  }
+  mempool_barrier(num_cores);
+
+  //  // SINGLE-CORE
+  //  time_init = mempool_get_timer();
+  //  dotp_f16s(l1_A, l1_B, sum, LEN);
+  //  // dotp_f16s_unrolled4(l1_A, l1_B, sum, LEN);
+  //  time_end = mempool_get_timer();
+
+  //  // PARALLEL
+  //  time_init = mempool_get_timer();
+  //  dotp_f16vecp_unrolled4(l1_A, l1_B, sum, LEN, num_cores);
+  //  // dotp_f16p(l1_A, l1_B, sum, LEN, num_cores);
+  //  time_end = mempool_get_timer();
+
+  // PARALLEL, LOCAL ACCESSES
+  time_init = mempool_get_timer();
+  dotp_f16vecp_local_unrolled4(l1_A, l1_B, sum, LEN);
+  time_end = mempool_get_timer();
+
+  // Check results
+  mempool_barrier(num_cores);
+  if (core_id == 0) {
+    uint32_t clock_cycles = (time_end - time_init);
+    printf("\nKernel execution takes %d clock cycles\n", clock_cycles);
+    printf("Result ==> %x\n", *(uint32_t *)&sum[0]);
+    printf("Check  ==> %x\n\n", *(uint32_t *)&l2_C);
+  }
+  mempool_barrier(num_cores);
+
+  return 0;
+}
diff --git a/software/apps/baremetal/dotp_f32/main.c b/software/apps/baremetal/dotp_f32/main.c
@@ -0,0 +1,76 @@
+// Copyright 2021 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+// Author: Marco Bertuletti, ETH Zurich
+
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "dma.h"
+#include "encoding.h"
+#include "printf.h"
+#include "runtime.h"
+#include "synchronization.h"
+
+#include "data_dotp_f32.h"
+#define NUM_BANKS (NUM_CORES * BANKING_FACTOR)
+// #define SINGLE_CORE_REDUCTION
+#define BINARY_REDUCTION
+
+// Vectors for kernel computation
+float l1_A[LEN] __attribute__((aligned(LEN), section(".l1_prio")));
+float l1_B[LEN] __attribute__((aligned(LEN), section(".l1_prio")));
+uint32_t red_barrier[NUM_BANKS]
+    __attribute__((aligned(NUM_BANKS), section(".l1_prio")));
+float sum[NUM_BANKS] __attribute__((aligned(NUM_BANKS), section(".l1_prio")));
+
+#include "baremetal/mempool_dotp_f32.h"
+
+int main() {
+
+  uint32_t core_id = mempool_get_core_id();
+  uint32_t num_cores = mempool_get_core_count();
+  uint32_t time_init, time_end;
+  mempool_barrier_init(core_id);
+
+  time_init = 0;
+  time_end = 0;
+  if (core_id == 0) {
+    dma_memcpy_blocking(l1_A, l2_A, LEN * sizeof(int32_t));
+    dma_memcpy_blocking(l1_B, l2_B, LEN * sizeof(int32_t));
+  }
+  for (uint32_t k = core_id; k < NUM_BANKS; k += num_cores) {
+    sum[k] = 0;
+    red_barrier[k] = 0;
+  }
+  mempool_barrier(num_cores);
+
+  //    // SINGLE-CORE
+  //    time_init = mempool_get_timer();
+  //    dotp_f32s_unrolled4(l1_A, l1_B, sum, LEN);
+  //    time_end = mempool_get_timer();
+
+  //   // PARALLEL
+  //   time_init = mempool_get_timer();
+  //   dotp_f32p(l1_A, l1_B, sum, LEN, num_cores);
+  //   time_end = mempool_get_timer();
+
+  // PARALLEL, LOCAL ACCESSES
+  time_init = mempool_get_timer();
+  dotp_f32p_local_unrolled4(l1_A, l1_B, sum, LEN);
+  time_end = mempool_get_timer();
+
+  // Check results
+  mempool_barrier(num_cores);
+  if (core_id == 0) {
+    uint32_t clock_cycles = (time_end - time_init);
+    printf("\nKernel execution takes %d clock cycles\n", clock_cycles);
+    printf("Result ==> %d\n", sum[0]);
+    printf("Check  ==> %d\n\n", l2_C);
+  }
+  mempool_barrier(num_cores);
+
+  return 0;
+}
diff --git a/software/apps/baremetal/dotp_i32/main.c b/software/apps/baremetal/dotp_i32/main.c
@@ -28,8 +28,7 @@ uint32_t red_barrier[NUM_BANKS]
     __attribute__((aligned(NUM_BANKS), section(".l1_prio")));
 int32_t sum[NUM_BANKS] __attribute__((aligned(NUM_BANKS), section(".l1_prio")));
 
-#include "baremetal/mempool_dotp_i32p.h"
-#include "baremetal/mempool_dotp_i32s.h"
+#include "baremetal/mempool_dotp_i32.h"
 
 int main() {
 

diff --git a/software/data/data_dotp_f16.h.tpl b/software/data/data_dotp_f16.h.tpl
@@ -0,0 +1,24 @@
+// Copyright 2022 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+\
+<% def array_to_cstr(array):
+    out = '{'
+    i = 0
+    out += '\n'
+    for a in array:
+        out += '(__fp16){:.4f}, '.format(a)
+        i += 1
+        if i % 8 == 0:
+            out += '\n'
+    out = out[:-2] + '}'
+    return out
+%> \
+
+#define LEN (${Len})
+
+__fp16 __attribute__((aligned(sizeof(int32_t)), section(".l2"))) l2_A[${Len}] = ${array_to_cstr(A)};
+
+__fp16 __attribute__((aligned(sizeof(int32_t)), section(".l2"))) l2_B[${Len}] = ${array_to_cstr(B)};
+
+__fp16 __attribute__((aligned(sizeof(int32_t)), section(".l2"))) l2_C = (__fp16)${C}f;
diff --git a/software/data/data_dotp_f32.h.tpl b/software/data/data_dotp_f32.h.tpl
@@ -0,0 +1,24 @@
+// Copyright 2022 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+\
+<% def array_to_cstr(array):
+    out = '{'
+    i = 0
+    out += '\n'
+    for a in array:
+        out += '{}f, '.format(a)
+        i += 1
+        if i % 8 == 0:
+            out += '\n'
+    out = out[:-2] + '}'
+    return out
+%> \
+
+#define LEN (${Len})
+
+float __attribute__((aligned(sizeof(int32_t)), section(".l2"))) l2_A[${Len}] = ${array_to_cstr(A)};
+
+float __attribute__((aligned(sizeof(int32_t)), section(".l2"))) l2_B[${Len}] = ${array_to_cstr(B)};
+
+float __attribute__((aligned(sizeof(int32_t)), section(".l2"))) l2_C = ${C}f;
diff --git a/software/data/generate_dotp.py b/software/data/generate_dotp.py
@@ -22,6 +22,24 @@ def generate_dotp_i32(Len):
     C = np.dot(A, B)
     return A, B, C
 
+
+def generate_dotp_f32(Len):
+
+    # Create matrix
+    A = np.random.rand(Len).astype(np.float32)
+    B = np.random.rand(Len).astype(np.float32)
+    C = (np.dot(A, B)).astype(np.float32)
+    return A, B, C
+
+
+def generate_dotp_f16(Len):
+
+    # Create matrix
+    A = np.random.rand(Len).astype(np.float16)
+    B = np.random.rand(Len).astype(np.float16)
+    C = (np.dot(A, B)).astype(np.float16)
+    return A, B, C
+
 ##################
 # compute_result #
 ##################
@@ -72,6 +90,26 @@ def main():
         'Len': Len}
     gen_data_header_file(args.outdir, tpl, **kwargs)
 
+    A, B, C = generate_dotp_f32(Len)
+    tpl = pathlib.Path(__file__).parent.absolute() / "data_dotp_f32.h.tpl"
+    kwargs = {
+        'name': 'data_dotp_f32',
+        'A': A,
+        'B': B,
+        'C': C,
+        'Len': Len}
+    gen_data_header_file(args.outdir, tpl, **kwargs)
+
+    A, B, C = generate_dotp_f16(Len)
+    tpl = pathlib.Path(__file__).parent.absolute() / "data_dotp_f16.h.tpl"
+    kwargs = {
+        'name': 'data_dotp_f16',
+        'A': A,
+        'B': B,
+        'C': C,
+        'Len': Len}
+    gen_data_header_file(args.outdir, tpl, **kwargs)
+
 
 if __name__ == "__main__":
     main()