add read widths tests

Makefile

@@ -1,8 +1,10 @@
 CFLAGS=-lm -g -Wall -O1
 
-build/main: src/main.c src/repetition_tester.c build/multi_nop_loop.o build/write_loops.o build/load_uop.o build/store_uop.o build/short_load_uop.o src/rprof.h
+# TODO: Move this to a build.zig. This is becoming a mess, I don't want to deal with Makefiles anymore
+
+build/main: src/main.c src/repetition_tester.c build/multi_nop_loop.o build/write_loops.o build/load_uop.o build/store_uop.o build/short_load_uop.o src/rprof.h build/read_widths.o
 	mkdir -p build
-	gcc -o build/main src/main.c build/multi_nop_loop.o build/write_loops.o build/load_uop.o build/store_uop.o build/short_load_uop.o $(CFLAGS)
+	gcc -o build/main src/main.c build/multi_nop_loop.o build/write_loops.o build/load_uop.o build/store_uop.o build/short_load_uop.o build/read_widths.o $(CFLAGS)
 
 build/short_load_uop.o: src/short_load_uop.asm
 	nasm -g -f elf64 -o build/short_load_uop.o src/short_load_uop.asm
@@ -19,6 +21,9 @@ build/multi_nop_loop.o: src/multi_nop_loop.asm
 build/store_uop.o: src/store_uop.asm
 	nasm -g -f elf64 -o build/store_uop.o src/store_uop.asm
 
+build/read_widths.o: src/read_widths.asm
+	nasm -g -f elf64 -o build/read_widths.o src/read_widths.asm
+
 run: ./build/main
 	./build/main $(TEST_NAME)
 

src/main.c

@@ -17,6 +17,7 @@
 #include "main_load_uop.c"
 #include "main_store_uop.c"
 #include "main_short_load_uop.c"
+#include "main_read_widths.c"
 
 int main(int argc, char **argv) {
 	if (argc < 2) {
@@ -28,6 +29,8 @@ int main(int argc, char **argv) {
 
 	if (!strncmp(test_name, "write_bytes_asm", sizeof("write_bytes_asm"))) {
 		return main_test_write_bytes_asm();
+	} else if (!strncmp(test_name, "read_widths", sizeof("read_widths"))) {
+		return main_test_read_widths();
 	} else if (!strncmp(test_name, "write_all_bytes", sizeof("write_bytes"))) {
 		return main_test_write_all_bytes();
 	} else if (!strncmp(test_name, "load_uop", sizeof("load_uop"))) {

src/main_read_widths.c

@@ -0,0 +1,39 @@
+#include "repetition_tester.c"
+#include "read_widths.h"
+
+int main_test_read_widths() {
+	typedef void (*test_cb)(uint8_t *buffer, uint64_t byte_count);
+	struct testcase {
+		char *name;
+		test_cb cb;
+	};
+
+	struct testcase cases[] = {
+		{ .name = "read_width_4x2()", .cb = read_width_4x2 },
+		{ .name = "read_width_8x2()", .cb = read_width_8x2 },
+		{ .name = "read_width_16x2()", .cb = read_width_16x2 },
+		{ .name = "read_width_32x2()", .cb = read_width_32x2 },
+	};
+
+	struct repetitor repetitor = {};
+	repetitor_init(&repetitor);
+	printf("CPU Frequency: %ldHz (~%.2fGHz)\n", repetitor.cpu_freq, (float)repetitor.cpu_freq/(1000*1000*1000));
+
+	uint64_t byte_count = 4096 * 1024;
+	uint8_t buffer[byte_count];
+
+	for (int i = 0; i < ARRAY_LEN(cases); i++) {
+		struct testcase *testcase = &cases[i];
+		repetitor_clear(&repetitor);
+		while (repetitor_repeat(&repetitor, 2)) {
+			repetitor_start(&repetitor);
+			repetitor_measure_start(&repetitor);
+			testcase->cb(buffer, byte_count);
+			repetitor_measure_stop(&repetitor, byte_count);
+			repetitor_stop(&repetitor);
+		}
+		repetitor_print_results_label(&repetitor, testcase->name);
+	}
+
+	return 0;
+}

src/read_widths.asm

@@ -0,0 +1,58 @@
+global read_width_4x2
+global read_width_8x2
+global read_width_16x2
+global read_width_32x2
+
+section .text
+
+; rsi - byte_count
+; rdi - buffer
+read_width_4x2:
+	xor rcx, rcx
+	align 64
+.loop:
+	mov r8d, [rdi]
+	mov r8d, [rdi + 4]
+	add rcx, 8
+	cmp rcx, rsi
+	jb .loop
+	ret
+
+; rsi - byte_count
+; rdi - buffer
+read_width_8x2:
+	xor rcx, rcx
+	align 64
+.loop:
+	mov r8, [rdi]
+	mov r8, [rdi + 8]
+	add rcx, 16
+	cmp rcx, rsi
+	jb .loop
+	ret
+
+; rsi - byte_count
+; rdi - buffer
+read_width_16x2:
+	xor rcx, rcx
+	align 64
+.loop:
+	vmovdqu xmm0, [rdi]
+	vmovdqu xmm0, [rdi + 16]
+	add rcx, 32
+	cmp rcx, rsi
+	jb .loop
+	ret
+
+; rsi - byte_count
+; rdi - buffer
+read_width_32x2:
+	xor rcx, rcx
+	align 64
+.loop:
+	vmovdqu ymm0, [rdi]
+	vmovdqu ymm0, [rdi + 32]
+	add rcx, 64
+	cmp rcx, rsi
+	jb .loop
+	ret

src/read_widths.h

@@ -0,0 +1,6 @@
+#include <stdint.h>
+
+void read_width_4x2(uint8_t *buffer, uint64_t byte_count);
+void read_width_8x2(uint8_t *buffer, uint64_t byte_count);
+void read_width_16x2(uint8_t *buffer, uint64_t byte_count);
+void read_width_32x2(uint8_t *buffer, uint64_t byte_count);