Remove unused variables

8086_assembly_01/Makefile

@@ -1,2 +1,2 @@
 all:
-	clang++ -g dasm.cpp -o dasm
+	clang++ -g dasm.cc -o dasm

8086_assembly_02/Makefile

@@ -1,2 +1,2 @@
 all:
-	clang++ -g dasm.cpp -o dasm
+	clang++ -g dasm.cc -o dasm

8086_assembly_03/Makefile

@@ -1,2 +1,2 @@
 all:
-	clang++ -g dasm.cpp -o dasm
+	clang++ -g dasm.cc -o dasm

8086_sim/Makefile

@@ -1,7 +1,7 @@
 CC=clang++
 CFLAGS=-g -O0 -Wall -Wextra
 LIBS=-Wl,-rpath,./lib -L./lib -lsim86
-SRC=*.cpp
+SRC=*.cc
 OUT=sim86
 
 all:

haversine_01/Makefile

@@ -1,17 +1,17 @@
 mk_haversine_fscanf:
-	clang++ -g cpp/fscanf.cpp cpp/haversine.cpp -o cpp/haverscan
+	clang++ -g cpp/fscanf.cc cpp/haversine.cc -o cpp/haverscan
 
 run_haversine_fscanf:
 	cd ./cpp && ./haverscan
 
 mk_haversine_strtok:
-	clang++ -g cpp/strtok.cpp cpp/haversine.cpp -o cpp/haverstrtok
+	clang++ -g cpp/strtok.cc cpp/haversine.cc -o cpp/haverstrtok
 
 run_haversine_strtok:
 	cd ./cpp && ./haverstrtok
 
 mk_test:
-	clang++ -g -lpthread cpp/test.cpp cpp/haversine.cpp -o cpp/test
+	clang++ -g -lpthread cpp/test.cc cpp/haversine.cc -o cpp/test
 
 run_test:
 	cd ./cpp && ./test

haversine_02/.gitignore

@@ -4,6 +4,7 @@
 compile_commands.json
 count_and_distances
 pairs.json
+cache_test
 main
 genhavr
 prochavr

haversine_02/compile

@@ -4,6 +4,12 @@ CC=clang
 CXX=clang++
 CFLAGS="-Wall -Wextra -I$(realpath ./include) "
 
+ASM=nasm
+ASM_FLAGS="-f elf64 "
+
+AR=ar
+AR_FLAGS="rcs"
+
 # PARSE ARGUMENTS
 # From this StackOverflow answer https://stackoverflow.com/a/14203146
 while [[ $# > 0 ]];do
@@ -30,18 +36,18 @@ done
 
 # BUILD TYPE
 if [[ $RELEASE == true ]]; then
-	CFLAGS+="-O3"
+	CFLAGS+="-g -O1"
 else
 	CFLAGS+="-g"
 fi
 
 
 # GENERATOR
-GENSRC="./src/generator/gen_argparser.cpp \
+GENSRC="./src/generator/gen_argparser.cc \
-	./src/generator/generator.cpp \
+	./src/generator/generator.cc \
-	./src/haversine.cpp \
+	./src/haversine.cc \
-	./src/point_types.cpp \
+	./src/point_types.cc \
-	./src/generator/main.cpp"
+	./src/generator/main.cc"
 GENOUT=genhavr
 
 (set -x ; $CXX $CFLAGS $GENSRC -o $GENOUT)
@@ -58,10 +64,10 @@ JSONFLAGS="-c "
 JSON_BUILD_DIR=json_build
 
 PROCSRC="./$JSON_BUILD_DIR/*.o \
-	./src/haversine.cpp \
+	./src/haversine.cc \
-	./src/point_types.cpp \
+	./src/point_types.cc \
-	./src/processor/proc_argparser.cpp \
+	./src/processor/proc_argparser.cc \
-	./src/processor/main.cpp "
+	./src/processor/main.cc "
 PROCOUT=prochavr
 
 # MEMTESTER
@@ -71,6 +77,18 @@ MEMTESTOUT=memtest
 (set -x ; $CC $CFLAGS $MEMTESTSRC -o $MEMTESTOUT)
 echo
 
+# REPTEST ASSEMBLY
+ASM_BUILD_DIR=reptest_build
+ASM_SRC="./src/repetition_testing/reptest_functions.asm"
+ASM_OBJ="./$ASM_BUILD_DIR/funcs.o"
+ASM_LIB="./$ASM_BUILD_DIR/libfuncs.a"
+
+mkdir $ASM_BUILD_DIR
+
+(set -x ; $ASM $ASM_FLAGS $ASM_SRC -o $ASM_OBJ)
+(set -x ; $AR $AR_FLAGS $ASM_LIB $ASM_OBJ)
+echo
+
 if [[ $BASIC_PROFILING == true ]] || [[ $FULL_PROFILING == true ]]; then
 	if [[ $FULL_PROFILING == true ]]; then
 		JSONFLAGS+="-DFULL_PROFILING"
@@ -95,7 +113,7 @@ if [[ $BASIC_PROFILING == true ]] || [[ $FULL_PROFILING == true ]]; then
 	cd ../
 
 	# REPETITION TESTING
-	REPTESTSRC="./src/repetition_testing/*.cpp ./$PROF_BUILD_DIR/*.o"
+	REPTESTSRC="./src/repetition_testing/*.cc ./$PROF_BUILD_DIR/*.o $ASM_LIB"
 	REPTESTOUT=reptest
 
 	(set -x ; $CXX $CFLAGS $REPTESTFLAGS $REPTESTSRC -o $REPTESTOUT)
@@ -114,4 +132,4 @@ cd ../
 echo
 
 # CLEAR BUILD FILES
-rm -rvf $JSON_BUILD_DIR $PROF_BUILD_DIR
+rm -rvf $JSON_BUILD_DIR $PROF_BUILD_DIR $ASM_BUILD_DIR

haversine_02/src/repetition_testing/main.cpp

@@ -1,197 +0,0 @@
-#include "aliases.h"
-#include "profiler/timer.h"
-#include "repetition_testing/reptester.h"
-#include <stdint.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-#include <unistd.h>
-
-#define ARR_LEN(ARR) sizeof(ARR) / sizeof(*ARR)
-
-void test_fread(reptester *tester, alloc_type type);
-void test_read(reptester *tester, alloc_type type);
-void test_write(reptester *tester, alloc_type type);
-u64 get_file_length(FILE *fp);
-
-int main(int argc, char *argv[]) {
-  const char *filename = NULL;
-  u64 waves = 1;
-
-  switch (argc) {
-  case 3:
-    waves = atol(argv[2]);
-  // break left out intentionally
-  case 2:
-    filename = argv[1];
-    break;
-  default:
-    printf("Usage: reptest FILENAME [WAVE_COUNT]\n");
-    return -1;
-    break;
-  }
-
-  // clang-format off
-  reptester tester = {
-      {filename, NULL, 0, 0},     // params
-      get_cpu_freq(500),          // cpu_freq
-
-      10.0,                       // wait_time_secs
-      0.0,                        // test_time_secs
-      read_cpu_timer(),           // test_start_time
-
-      1,                          // current_run
-			{
-				UINT64_MAX,               // min_time
-				0,                        // max_time
-				0,                        // avg_time
-				0,                        // total_time
-			},
-			{
-				0,                        // min_faults
-				0,                        // max_faults
-				0,                        // avg_faults
-				0,                        // total_bytes
-				0,                        // total_faults
-			},
-		  {},                         // results
-  };
-  // clang-format on
-
-  FILE *fp = fopen(tester.params.filename, "rb");
-  if (!fp) {
-    return -1;
-  }
-
-  func_data funcs[] = {
-      {{"WRITE", "WRITE WITH MALLOC"}, test_write},
-      {{"READ", "READ WITH MALLOC"}, test_read},
-      {{"FREAD", "FREAD WITH MALLOC"}, test_fread},
-  };
-
-  tester.params.read_size = get_file_length(fp);
-  tester.params.read_count = 1;
-  tester.params.buffer = (char *)malloc(tester.params.read_size + 1);
-  memset(tester.params.buffer, 0, tester.params.read_size + 1);
-
-  for (u64 i = 0; i < waves; ++i) {
-    for (u64 j = 0; j < ARR_LEN(funcs); ++j) {
-      for (u64 k = 0; k < COUNT_ALLOC_TYPE; ++k) {
-        run_func_test(&tester, funcs[j].func, funcs[j].names[k], (alloc_type)k);
-      }
-    }
-  }
-
-  fclose(fp);
-
-  free(tester.params.buffer);
-
-  return 0;
-}
-
-void test_fread(reptester *tester, alloc_type type) {
-  FILE *fp = fopen(tester->params.filename, "rb");
-  if (!fp) {
-    return;
-  }
-
-  u64 start = read_cpu_timer();
-  u64 fault_count_start = page_fault_count();
-
-  handle_alloc(tester, type);
-
-  u64 obj_count = fread(tester->params.buffer, tester->params.read_size,
-                        tester->params.read_count, fp);
-
-  u64 fault_count_end = page_fault_count();
-  u64 end = read_cpu_timer();
-
-  u64 bytes_read = obj_count * tester->params.read_size;
-
-  u64 read_time = end - start;
-  u64 page_faults = fault_count_end - fault_count_start;
-
-  tester->results = {
-      bytes_read,
-      read_time,
-      page_faults,
-  };
-
-  handle_free(tester, type);
-
-  fclose(fp);
-}
-
-void test_read(reptester *tester, alloc_type type) {
-  FILE *fp = fopen(tester->params.filename, "rb");
-  if (!fp) {
-    return;
-  }
-
-  u64 start = read_cpu_timer();
-  u64 fault_count_start = page_fault_count();
-
-  handle_alloc(tester, type);
-
-  i32 fd = fileno(fp);
-
-  u64 bytes_read = read(fd, tester->params.buffer,
-                        tester->params.read_size * tester->params.read_count);
-
-  u64 fault_count_end = page_fault_count();
-  u64 end = read_cpu_timer();
-
-  u64 read_time = end - start;
-  u64 page_faults = fault_count_end - fault_count_start;
-
-  tester->results = {
-      bytes_read,
-      read_time,
-      page_faults,
-  };
-
-  handle_free(tester, type);
-
-  fclose(fp);
-}
-
-void test_write(reptester *tester, alloc_type type) {
-  u64 start = read_cpu_timer();
-  u64 fault_count_start = page_fault_count();
-
-  handle_alloc(tester, type);
-
-  u64 total_size = tester->params.read_size * tester->params.read_count;
-
-  for (u64 i = 0; i < total_size; ++i) {
-    tester->params.buffer[i] = '0';
-  }
-
-  u64 fault_count_end = page_fault_count();
-  u64 end = read_cpu_timer();
-
-  u64 read_time = end - start;
-  u64 page_faults = fault_count_end - fault_count_start;
-
-  tester->results = {
-      total_size,
-      read_time,
-      page_faults,
-  };
-
-  handle_free(tester, type);
-}
-
-u64 get_file_length(FILE *fp) {
-  if (!fp) {
-    return 0;
-  }
-
-  fseek(fp, 0, SEEK_END);
-
-  u64 length = ftell(fp);
-
-  fseek(fp, 0, SEEK_SET);
-
-  return length;
-}

haversine_02/src/repetition_testing/reptest_functions.asm

@@ -0,0 +1,529 @@
+global mov_all_bytes_asm
+global nop_all_bytes_asm
+global nop_1x3_all_bytes_asm
+global nop_1x9_all_bytes_asm
+global inc_all_bytes_asm
+global dec_all_bytes_asm
+global align64_loop
+global align1_loop
+global align15_loop
+global align31_loop
+global align63_loop
+global align75_loop
+global align90_loop
+global align112_loop
+global rat_add
+global rat_mov_add
+global read_1
+global read_2
+global read_3
+global read_4
+global read_8
+global write_1
+global write_2
+global write_3
+global write_4
+global write_8
+global read_1x2_low
+global read_1x2_high
+global read_2x2
+global read_4x2
+global read_8x2
+global read_4x2_simd
+global read_8x2_simd
+global read_16x2_simd
+global read_32x2_simd_offset
+global read_32x2_simd_no_offset
+global read_16x4_simd
+global read_32x4_simd
+global cache_test ; Expects 3 inputs (pointer, read_count, mask)
+global cache_test_unaligned ; Expects 3 inputs (pointer, read_count, mask)
+
+mov_all_bytes_asm:
+  xor rax, rax
+
+  .loop:
+    mov BYTE [rdi + rax * 1], al
+    inc rax
+    cmp rsi, rax
+    jne .loop
+
+  ret
+
+nop_all_bytes_asm:
+  xor rax, rax
+
+  .loop:
+    db 0x0f, 0x1f, 0x00
+    inc rax
+    cmp rdi, rax
+    jne .loop
+
+  ret
+
+nop_1x3_all_bytes_asm:
+  xor rax, rax
+
+  .loop:
+    nop
+    nop
+    nop
+    inc rax
+    cmp rdi, rax
+    jne .loop
+
+  ret
+
+nop_1x9_all_bytes_asm:
+  xor rax, rax
+
+  .loop:
+    nop
+    nop
+    nop
+    nop
+    nop
+    nop
+    nop
+    nop
+    nop
+    inc rax
+    cmp rdi, rax
+    jne .loop
+
+  ret
+
+inc_all_bytes_asm:
+  xor rax, rax
+
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
+
+  ret
+
+dec_all_bytes_asm:
+  .loop:
+    dec rdi
+    jnz .loop
+
+  ret
+
+align64_loop:
+  xor rax, rax
+
+  align 64
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
+
+  ret
+
+align1_loop:
+  xor rax, rax
+
+  align 64
+  nop
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
+
+  ret
+
+align15_loop:
+  xor rax, rax
+
+  align 64
+  %rep 15
+  nop
+  %endrep
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
+
+  ret
+
+align31_loop:
+  xor rax, rax
+
+  align 64
+  %rep 31
+  nop
+  %endrep
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
+
+  ret
+
+align63_loop:
+  xor rax, rax
+
+  align 64
+  %rep 63
+  nop
+  %endrep
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
+
+  ret
+
+align75_loop:
+  xor rax, rax
+
+  align 64
+  %rep 75
+  nop
+  %endrep
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
+
+  ret
+
+align90_loop:
+  xor rax, rax
+
+  align 64
+  %rep 90
+  nop
+  %endrep
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
+
+  ret
+
+align112_loop:
+  xor rax, rax
+
+  align 64
+  %rep 112
+  nop
+  %endrep
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
+
+  ret
+
+rat_add:
+  mov rax, rdi
+
+  .loop:
+    add rcx, 1
+    add rcx, 1
+    dec rax
+    jnz .loop
+  ret
+
+rat_mov_add:
+  mov rax, rdi
+
+  .loop:
+    mov rcx, rax
+    add rcx, 1
+    mov rcx, rax
+    add rcx, 1
+    dec rax
+    jnz .loop
+  ret
+
+read_1:
+  align 64
+  .loop:
+    mov rax, [rdi]
+    sub rsi, 1
+    jnle .loop
+
+  ret
+
+read_2:
+  align 64
+  .loop:
+    %rep 2
+    mov rax, [rdi]
+    %endrep
+    sub rsi, 2
+    jnle .loop
+
+  ret
+
+read_3:
+  align 64
+  .loop:
+    %rep 3
+    mov rax, [rdi]
+    %endrep
+    sub rsi, 3
+    jnle .loop
+
+  ret
+
+read_4:
+  align 64
+  .loop:
+    %rep 4
+    mov rax, [rdi]
+    %endrep
+    sub rsi, 4
+    jnle .loop
+
+  ret
+
+read_8:
+  align 64
+  .loop:
+    %rep 8
+    mov rax, [rdi]
+    %endrep
+    sub rsi, 8
+    jnle .loop
+
+  ret
+
+write_1:
+  align 64
+  .loop:
+    mov QWORD [rdi], 0
+    sub rsi, 1
+    jnle .loop
+
+  ret
+
+write_2:
+  align 64
+  .loop:
+    %rep 2
+    mov QWORD [rdi], 0
+    %endrep
+    sub rsi, 2
+    jnle .loop
+
+  ret
+
+write_3:
+  align 64
+  .loop:
+    %rep 3
+    mov QWORD [rdi], 0
+    %endrep
+    sub rsi, 3
+    jnle .loop
+
+  ret
+
+write_4:
+  align 64
+  .loop:
+    %rep 4
+    mov QWORD [rdi], 0
+    %endrep
+    sub rsi, 4
+    jnle .loop
+
+  ret
+
+write_8:
+  align 64
+  .loop:
+    %rep 8
+    mov QWORD [rdi], 0
+    %endrep
+    sub rsi, 8
+    jnle .loop
+
+  ret
+
+read_1x2_low:
+  align 64
+  .loop:
+    %rep 2
+    mov al, [rdi]
+    %endrep
+    sub rsi, 2
+    jnle .loop
+
+  ret
+
+read_1x2_high:
+  align 64
+  .loop:
+    %rep 2
+    mov ah, [rdi]
+    %endrep
+    sub rsi, 2
+    jnle .loop
+
+  ret
+
+read_2x2:
+  align 64
+  .loop:
+    %rep 2
+    mov ax, [rdi]
+    %endrep
+    sub rsi, 2
+    jnle .loop
+
+  ret
+
+read_4x2:
+  align 64
+  .loop:
+    %rep 2
+    mov eax, [rdi]
+    %endrep
+    sub rsi, 2
+    jnle .loop
+
+  ret
+
+read_8x2:
+  align 64
+  .loop:
+    %rep 2
+    mov rax, [rdi]
+    %endrep
+    sub rsi, 2
+    jnle .loop
+
+  ret
+
+read_4x2_simd:
+  xor rax, rax
+
+  align 64
+  .loop:
+    mov r8d, [rdi]
+    mov r8d, [rdi + 4]
+    add rax, 8
+    cmp rax, rsi
+    jb .loop
+
+  ret
+
+read_8x2_simd:
+  xor rax, rax
+
+  align 64
+  .loop:
+    mov r8, [rdi]
+    mov r8, [rdi + 8]
+    add rax, 16
+    cmp rax, rsi
+    jb .loop
+
+  ret
+
+read_16x2_simd:
+  xor rax, rax
+
+  align 64
+  .loop:
+    vmovdqu xmm0, [rdi]
+    vmovdqu xmm0, [rdi + 16]
+    add rax, 32
+    cmp rax, rsi
+    jb .loop
+
+  ret
+
+read_32x2_simd_offset:
+  xor rax, rax
+
+  align 64
+  .loop:
+    vmovdqu ymm0, [rdi]
+    vmovdqu ymm0, [rdi + 32]
+    add rax, 64
+    cmp rax, rsi
+    jb .loop
+
+  ret
+
+read_32x2_simd_no_offset:
+  xor rax, rax
+
+  align 64
+  .loop:
+    vmovdqu ymm0, [rdi]
+    vmovdqu ymm0, [rdi]
+    add rax, 64
+    cmp rax, rsi
+    jb .loop
+
+  ret
+
+read_16x4_simd:
+  xor rax, rax
+
+  align 64
+  .loop:
+    %rep 2
+    vmovdqu xmm0, [rdi]
+    vmovdqu xmm0, [rdi + 16]
+    %endrep
+    add rax, 64
+    cmp rax, rsi
+    jb .loop
+
+  ret
+
+read_32x4_simd:
+  xor rax, rax
+
+  align 64
+  .loop:
+    %rep 2
+    vmovdqu ymm0, [rdi]
+    vmovdqu ymm0, [rdi]
+    %endrep
+    add rax, 128
+    cmp rax, rsi
+    jb .loop
+
+  ret
+
+cache_test:
+  xor r10, r10           ; Zero loop counter
+  mov rbx, rdi           ; Save original pointer
+  .loop:
+    add rdi, r10         ; Advance the pointer
+    add r10, 128         ; Increment loop counter
+    and r10, rdx         ; Mask offset
+    vmovdqu ymm0, [rdi + 0]
+    vmovdqu ymm1, [rdi + 32]
+    vmovdqu ymm2, [rdi + 64]
+    vmovdqu ymm3, [rdi + 96]
+    mov rdi, rbx         ; Restore original pointer
+    sub rsi, 128         ; Decrement count
+    ja .loop
+  ret
+
+cache_test_unaligned:
+  xor r10, r10           ; Zero loop counter
+	add rdi, 5             ; Unalign pointer
+  mov rbx, rdi           ; Save original pointer
+  .loop:
+    add rdi, r10         ; Advance the pointer
+    add r10, 128         ; Increment loop counter
+    and r10, rdx         ; Mask offset
+    vmovdqu ymm0, [rdi + 0]
+    vmovdqu ymm1, [rdi + 32]
+    vmovdqu ymm2, [rdi + 64]
+    vmovdqu ymm3, [rdi + 96]
+    mov rdi, rbx         ; Restore original pointer
+    sub rsi, 128         ; Decrement count
+    ja .loop
+  ret

haversine_02/src/repetition_testing/reptester.cc

@@ -3,6 +3,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <sys/mman.h>
 #include <sys/resource.h>
 #include <sys/time.h>
 
@@ -10,7 +11,9 @@ void handle_alloc(reptester *tester, alloc_type type) {
   switch (type) {
   case ALLOC_TYPE_WITH_MALLOC:
     if (!(tester->params.buffer)) {
-      tester->params.buffer = (char *)malloc(tester->params.read_size + 1);
+      tester->params.buffer = (char *)mmap(
+          NULL, tester->params.read_size + 1, PROT_READ | PROT_WRITE,
+          MAP_ANON | MAP_SHARED | MAP_NORESERVE, -1, 0);
       memset(tester->params.buffer, 0, tester->params.read_size + 1);
     }
 
@@ -24,7 +27,7 @@ void handle_free(reptester *tester, alloc_type type) {
   switch (type) {
   case ALLOC_TYPE_WITH_MALLOC:
     if (tester->params.buffer) {
-      free(tester->params.buffer);
+      munmap(tester->params.buffer, tester->params.read_size + 1);
       tester->params.buffer = NULL;
     }
 
@@ -58,7 +61,9 @@ void run_func_test(reptester *tester, reptest_func func, const char *func_name,
 
   if (type == ALLOC_TYPE_WITH_MALLOC) {
     buffer = tester->params.buffer;
-    tester->params.buffer = (char *)malloc(tester->params.read_size + 1);
+    tester->params.buffer =
+        (char *)mmap(NULL, tester->params.read_size + 1, PROT_READ | PROT_WRITE,
+                     MAP_ANON | MAP_SHARED | MAP_NORESERVE, -1, 0);
     memset(tester->params.buffer, 0, tester->params.read_size + 1);
   }
 
@@ -100,7 +105,7 @@ void run_func_test(reptester *tester, reptest_func func, const char *func_name,
   }
 
   if (type == ALLOC_TYPE_WITH_MALLOC) {
-    free(tester->params.buffer);
+    munmap(tester->params.buffer, tester->params.read_size + 1);
     tester->params.buffer = buffer;
   }