Merge branch 'master' of https://git.exactas.uba.ar/orga2-puma-concolor/tp2

mediciones/Makefile

@@ -20,13 +20,17 @@ resoluciones=$(res_exp1) $(res_exp2)
 optimizaciones=O0 O1 O2 O3 Ofast
 filtros=asm-Cuadrados asm-Manchas asm-Offset asm-Sharpen #asm-Ruido
 filtros+=c-Cuadrados c-Manchas c-Offset c-Sharpen #c-Ruido
+filtros+=avx-Cuadrados mvec-Manchas simple-Manchas avx-Offset
 # Filtros que se usan para los gráficos simples (tienen implementación asm y c)
 filtros_simple=Cuadrados Manchas Offset Sharpen
+filtros_avx=Cuadrados Offset
 iteraciones?=50
 tmancha?=50
 retest=
 graficos_exp1= # Se rellenan automáticamente
 graficos_exp2= # Se rellenan automáticamente
+graficos_exp3= # Se rellenan automáticamente
+graficos_exp4= # Se rellenan automáticamente
 
 all: graficos
 
@@ -122,7 +126,45 @@ $(foreach foto,$(entradas), \
   $(foreach filtro,$(filtros_simple), \
     $(eval $(call RELACION_template,$(foto),$(filtro)))))
 
-graficos: $(graficos_exp1) $(graficos_exp2)
+# AVX_template foto filtro
+define AVX_template =
+graficos/avx.$(1).$(2).pdf: $(foreach res,$(res_exp1), \
+			      data/$(1).$(res).asm-$(2).O0.stats \
+			      data/$(1).$(res).avx-$(2).O0.stats \
+			      data/$(1).$(res).c-$(2).Ofast.stats)
+	for res in $(res_exp1); do \
+		echo $$$$res $$$$(cat data/$(1).$$$$res.asm-$(2).O0.stats) \
+		     $$$$(cat data/$(1).$$$$res.avx-$(2).O0.stats) \
+		     $$$$(cat data/$(1).$$$$res.c-$(2).Ofast.stats); \
+	done | ./grafico-avx.py $(2) $$@;
+
+graficos_exp3+=graficos/avx.$(1).$(2).pdf
+endef
+
+$(foreach foto,$(entradas), \
+  $(foreach filtro,$(filtros_avx), \
+    $(eval $(call AVX_template,$(foto),$(filtro)))))
+
+# MANCHAS_template foto filtro
+define MANCHAS_template =
+graficos/manchas.$(1).pdf: $(foreach res,$(res_exp1), \
+			     data/$(1).$(res).asm-Manchas.O0.stats \
+			     data/$(1).$(res).mvec-Manchas.O0.stats \
+			     data/$(1).$(res).simple-Manchas.O0.stats \
+			     data/$(1).$(res).c-Manchas.Ofast.stats)
+	for res in $(res_exp1); do \
+		echo $$$$res $$$$(cat data/$(1).$$$$res.asm-Manchas.O0.stats) \
+		     $$$$(cat data/$(1).$$$$res.mvec-Manchas.O0.stats) \
+		     $$$$(cat data/$(1).$$$$res.simple-Manchas.O0.stats) \
+		     $$$$(cat data/$(1).$$$$res.c-Manchas.Ofast.stats); \
+	done | ./grafico-manchas.py $$@;
+
+graficos_exp4+=graficos/manchas.$(1).$(2).pdf
+endef
+
+$(foreach foto,$(entradas), $(eval $(call MANCHAS_template,$(foto))))
+
+graficos: $(graficos_exp1) $(graficos_exp2) $(graficos_exp3) $(graficos_exp4)
 
 binarios/%/tp2:
 	$(MAKE) -C binarios $(@:binarios/%=%)

mediciones/grafico-avx.py

+#!/usr/bin/env python3
+import matplotlib.pyplot as plt
+import sys
+
+if len(sys.argv) != 3:
+    print('Uso: %s filtro salida' % (sys.argv[0],))
+    exit()
+
+lines = [line.split() for line in sys.stdin.readlines()]
+
+etiquetas = [line[0] for line in lines]
+medias_asm = [float(line[1]) for line in lines]
+medias_avx = [float(line[3]) for line in lines]
+medias_Ofast = [float(line[5]) for line in lines]
+errores_asm = [float(line[2]) for line in lines]
+errores_avx = [float(line[4]) for line in lines]
+errores_Ofast = [float(line[6]) for line in lines]
+
+filtro = sys.argv[1]
+salida = sys.argv[2]
+
+plt.xticks(range(0, len(etiquetas), 2), rotation=30, fontsize=8)
+plt.yscale('log')
+plt.errorbar(etiquetas, medias_asm, yerr=errores_asm, label='ASM')
+plt.errorbar(etiquetas, medias_avx, yerr=errores_avx, label='AVX')
+plt.errorbar(etiquetas, medias_Ofast, yerr=errores_Ofast, label='C (Ofast)')
+plt.xlabel('Tamaños')
+plt.ylabel('Ciclos del procesador')
+plt.title('Filtro ' + filtro)
+plt.legend()
+plt.savefig(fname=salida, format='pdf')

mediciones/grafico-manchas.py

+#!/usr/bin/env python3
+import matplotlib.pyplot as plt
+import sys
+
+if len(sys.argv) != 2:
+    print('Uso: %s salida' % (sys.argv[0],))
+    exit()
+
+lines = [line.split() for line in sys.stdin.readlines()]
+
+etiquetas = [line[0] for line in lines]
+medias_asm = [float(line[1]) for line in lines]
+medias_mvec = [float(line[3]) for line in lines]
+medias_simple = [float(line[5]) for line in lines]
+medias_Ofast = [float(line[7]) for line in lines]
+errores_asm = [float(line[2]) for line in lines]
+errores_mvec = [float(line[4]) for line in lines]
+errores_simple = [float(line[6]) for line in lines]
+errores_Ofast = [float(line[8]) for line in lines]
+
+salida = sys.argv[1]
+
+plt.xticks(range(0, len(etiquetas), 2), rotation=30, fontsize=8)
+plt.yscale('log')
+plt.errorbar(etiquetas, medias_asm, yerr=errores_asm, label='ASM (FPU+Tabla)')
+plt.errorbar(etiquetas, medias_simple, yerr=errores_simple, label='ASM (FPU)')
+plt.errorbar(etiquetas, medias_mvec, yerr=errores_mvec, label='ASM (libmvec)')
+plt.errorbar(etiquetas, medias_Ofast, yerr=errores_Ofast, label='C (Ofast)')
+plt.xlabel('Tamaños')
+plt.ylabel('Ciclos del procesador')
+plt.title('Diferentes implementaciones del filtro Manchas')
+plt.legend()
+plt.savefig(fname=salida, format='pdf')

src/Makefile

@@ -9,8 +9,9 @@ BIN = tp2
 
 FILTROS = Cuadrados Manchas Offset Ruido Sharpen
 
-FILTROS_OBJ = $(addsuffix .o, $(FILTROS)) $(addsuffix _asm.o, $(FILTROS))\
-	      $(addsuffix _c.o, $(FILTROS)) Manchas_mvec.o Manchas_simple.o
+FILTROS_OBJ = $(addsuffix .o, $(FILTROS)) $(addsuffix _asm.o, $(FILTROS)) \
+	      $(addsuffix _c.o, $(FILTROS)) Manchas_mvec.o Manchas_simple.o \
+	      Offset_avx.o Cuadrados_avx.o
 LIBS_OBJS   = libbmp.o imagenes.o utils.o
 MAIN_OBJS   = tp2.o cli.o
 MAIN_OBJS_CON_PATH = $(addprefix $(BUILD_DIR)/, $(MAIN_OBJS))

src/cli.c

@@ -132,7 +132,7 @@ void procesar_opciones(int argc, char **argv, configuracion_t *config) {
     } else if (strcmp(tipo_filtro, "asm") == 0) {
         config->tipo_filtro = FILTRO_ASM;
     } else if (strcmp(tipo_filtro, "simple") == 0 &&
-              (ES_FILTRO(config, Manchas) || ES_FILTRO(config, Offset))) {
+               ES_FILTRO(config, Manchas)) {
         config->tipo_filtro = FILTRO_SIMPLE;
     } else if (strcmp(tipo_filtro, "avx") == 0 &&
               (ES_FILTRO(config, Cuadrados) || ES_FILTRO(config, Offset))) {

src/filters/Cuadrados.c

@@ -10,8 +10,14 @@ void Cuadrados_asm (uint8_t *src, uint8_t *dst, int width, int height,
 void Cuadrados_c   (uint8_t *src, uint8_t *dst, int width, int height,
                       int src_row_size, int dst_row_size);
 
+void Cuadrados_avx (uint8_t *src, uint8_t *dst, int width, int height,
+                    int src_row_size, int dst_row_size);
+
 typedef void (Cuadrados_fn_t) (uint8_t*, uint8_t*, int, int, int, int);
 
+static Cuadrados_fn_t* implementaciones[] = {
+	Cuadrados_c, Cuadrados_asm, Cuadrados_avx
+};
 
 void leer_params_Cuadrados(configuracion_t *config, int argc, char *argv[]) {
 
@@ -19,7 +25,7 @@ void leer_params_Cuadrados(configuracion_t *config, int argc, char *argv[]) {
 
 void aplicar_Cuadrados(configuracion_t *config)
 {
-    Cuadrados_fn_t *Cuadrados = SWITCH_C_ASM( config, Cuadrados_c, Cuadrados_asm );
+    Cuadrados_fn_t *Cuadrados = implementaciones[config->tipo_filtro];
     buffer_info_t info = config->src;
     Cuadrados(info.bytes, config->dst.bytes, info.width, info.height, 
             info.row_size, config->dst.row_size);

src/filters/Cuadrados_asm.asm

-section .data
+section .rodata
 BLACK:	times	4	dd	0xFF000000
 
 

src/filters/Cuadrados_avx.asm

+section .rodata
+BLACK:	times	8	dd	0xFF000000
+
+section .text
+global Cuadrados_avx
+Cuadrados_avx:
+	; rdi = uint8_t *src
+	; rsi = uint8_t *dst
+	; rdx = int width
+	; rcx = int height
+	; r8  = int src_row_size
+	; r9  = int dst_row_size
+	; Ajusto por los marcos
+	sub	rdx,	8
+	shr	rdx,	3
+	; rax = int height
+	lea	rax,	[rcx - 8]
+	vmovdqu	ymm7,	[BLACK]
+	; r10 = r8 * 3
+	lea	r10,	[r8 + r8*2]
+black_start:
+	lea	rcx,	[rdx*4 + 4]
+	.x_loop:
+		vmovdqu	[rsi],	ymm7
+		lea	rsi,	[rsi + 32]
+		loop	.x_loop
+	lea	rdi,	[rdi + r8 * 4]
+y_loop:
+	movdqa	[rsi],	xmm7
+	lea	rdi,	[rdi + 16]
+	lea	rsi,	[rsi + 16]
+	mov	rcx,	rdx
+	.x_loop:
+		vmovdqu	ymm0,	[rdi]
+		vpmaxub	ymm0,	ymm0,	[rdi + 4]
+		vpmaxub	ymm0,	ymm0,	[rdi + 8]
+		vpmaxub	ymm0,	ymm0,	[rdi + 12]
+		vpmaxub	ymm0,	ymm0,	[rdi + r8]
+		vpmaxub	ymm0,	ymm0,	[rdi + r8 + 4]
+		vpmaxub	ymm0,	ymm0,	[rdi + r8 + 8]
+		vpmaxub	ymm0,	ymm0,	[rdi + r8 + 12]
+		vpmaxub	ymm0,	ymm0,	[rdi + r8*2]
+		vpmaxub	ymm0,	ymm0,	[rdi + r8*2 + 4]
+		vpmaxub	ymm0,	ymm0,	[rdi + r8*2 + 8]
+		vpmaxub	ymm0,	ymm0,	[rdi + r8*2 + 12]
+		vpmaxub	ymm0,	ymm0,	[rdi + r10]
+		vpmaxub	ymm0,	ymm0,	[rdi + r10 + 4]
+		vpmaxub	ymm0,	ymm0,	[rdi + r10 + 8]
+		vpmaxub	ymm0,	ymm0,	[rdi + r10 + 12]
+		vmovdqu	[rsi],	ymm0
+		lea	rdi,	[rdi + 32]
+		lea	rsi,	[rsi + 32]
+		loop	.x_loop
+	movdqa	[rsi],	xmm7
+	lea	rdi,	[rdi + 16]
+	lea	rsi,	[rsi + 16]
+	dec	rax
+	jnz	y_loop
+black_end:
+	lea	rcx,	[rdx*4 + 4]
+	.x_loop:
+		vmovdqu	[rsi],	ymm7
+		lea	rsi,	[rsi + 32]
+		loop	.x_loop
+	ret

src/filters/Makefile

@@ -19,7 +19,8 @@ BUILD_DIR = ../build
 FILTROS = Cuadrados Manchas Offset Ruido Sharpen
 
 FILTROS_OBJ = $(addsuffix .o, $(FILTROS)) $(addsuffix _asm.o, $(FILTROS)) \
-	      $(addsuffix _c.o, $(FILTROS)) Manchas_mvec.o Manchas_simple.o
+	      $(addsuffix _c.o, $(FILTROS)) Manchas_mvec.o Manchas_simple.o \
+	      Offset_avx.o Cuadrados_avx.o
 FILTROS_OBJ_CON_PATH = $(addprefix  $(BUILD_DIR)/, $(FILTROS_OBJ))
 
 .PHONY: filtros clean

src/filters/Manchas_asm.asm

-section .data
+section .rodata
 CINCUENTA: dd 50.0
 align 16
 VEINTICINCO:	times 4 dd 25

src/filters/Manchas_mvec.asm

-section .data
+section .rodata
 TAU: dd 6.28318530717958647692528676655900576839433879875021
 align 16
 VCINCO:    times 4 dd 25.0

src/filters/Manchas_simple.asm

-section .data
+section .rodata
 CTAU: dd 6.283185
 CINCUENTA: dd 50.0
 VCINCO: dd 25.0
@@ -278,7 +278,7 @@ Manchas_simple:
             jnz .x_loop
 
         ; Ya me movi todo un row => decremento los verticales
-        dec r15
+        dec r15d
         ; Si ya llegue abajo del todo salgo, si no repito
         jnz .y_loop
 

src/filters/Offset.c

@@ -10,6 +10,9 @@ void Offset_asm (uint8_t *src, uint8_t *dst, int width, int height,
 void Offset_c   (uint8_t *src, uint8_t *dst, int width, int height,
                       int src_row_size, int dst_row_size);
 
+void Offset_avx (uint8_t *src, uint8_t *dst, int width, int height,
+                 int src_row_size, int dst_row_size);
+
 typedef void (Offset_fn_t) (uint8_t*, uint8_t*, int, int, int, int);
 
 
@@ -17,9 +20,13 @@ void leer_params_Offset(configuracion_t *config, int argc, char *argv[]) {
 
 }
 
+static Offset_fn_t* implementaciones[] = {
+	Offset_c, Offset_asm, Offset_avx
+};
+
 void aplicar_Offset(configuracion_t *config)
 {
-    Offset_fn_t *Offset = SWITCH_C_ASM( config, Offset_c, Offset_asm );
+    Offset_fn_t *Offset = implementaciones[config->tipo_filtro];
     buffer_info_t info = config->src;
     Offset(info.bytes, config->dst.bytes, info.width, info.height, 
             info.row_size, config->dst.row_size);

src/filters/Offset_asm.asm

-section .data
+section .rodata
 align 16
 BLACK:	times	4	dd	0xFF000000
 MASK_B:	times	4	dd	0x000000FF

src/filters/Offset_avx.asm

+section .rodata
+align 32
+BLACK:	times	8	dd	0xFF000000
+MASK_B:	times	8	dd	0x000000FF
+MASK_R:	times	8	dd	0x00FF0000
+
+section .text
+global Offset_avx
+Offset_avx:
+	; rdi = uint8_t *src
+	; rsi = uint8_t *dst
+	; rdx = int width
+	; rcx = int height
+	; r8  = int src_row_size
+	; r9  = int dst_row_size
+	sub	rdx,	16
+	shr	rdx,	3
+	sub	rcx,	16
+	vmovdqa	ymm7,	[BLACK]
+	vmovdqa	ymm6,	[MASK_B]
+	vmovdqa	ymm5,	[MASK_R]
+	; rax = int height
+	mov	rax,	rcx
+black_start:
+	lea	rcx,	[rdx*8 + 16]
+	.x_loop:
+		vmovdqu	[rsi],	ymm7
+		lea	rsi,	[rsi + 32]
+		loop	.x_loop
+	lea	rdi,	[rdi + r8*8]
+y_loop:
+	vmovdqu	[rsi],	ymm7
+	lea	rdi,	[rdi + 32]
+	lea	rsi,	[rsi + 32]
+	mov	rcx,	rdx
+	.x_loop:
+		vmovdqu		ymm1,	[rdi + 32]
+		vpblendvb	ymm1,	ymm1,	[rdi + r8*8],		ymm6
+		vpblendvb	ymm1,	ymm1,	[rdi + r8*8 + 32],	ymm5
+		vmovdqu		[rsi],	ymm1
+		lea	rdi,	[rdi + 32]
+		lea	rsi,	[rsi + 32]
+		loop	.x_loop
+	vmovdqu	[rsi],	ymm7
+	lea	rdi,	[rdi + 32]
+	lea	rsi,	[rsi + 32]
+	dec	rax
+	jnz	y_loop
+black_end:
+	lea	rcx,	[rdx*8 + 16]
+	.x_loop:
+		vmovdqu	[rsi],	ymm7
+		lea	rsi,	[rsi + 32]
+		loop	.x_loop
+	ret

src/filters/Sharpen_asm.asm

-section .data
+section .rodata
 align 16
 %define BLACK	0xFF000000
 BLACK_BORDER: dd 0xFF000000, 0xFF000000, 0xFF000000, 0xFF000000

src/tests/2_test_diff_cat_asm.py

@@ -13,6 +13,22 @@ for corrida in corridas:
         ok = verificar(corrida['filtro'], corrida['params'], corrida['tolerancia'], 'asm', imagen)
         todos_ok = todos_ok and ok
 
+for imagen in archivos:
+    ok = verificar('Offset', '', 5, 'avx', imagen)
+    todos_ok = todos_ok and ok
+
+for imagen in archivos:
+    ok = verificar('Cuadrados', '', 5, 'avx', imagen)
+    todos_ok = todos_ok and ok
+
+for imagen in archivos:
+    ok = verificar('Manchas', '100', 5, 'mvec', imagen)
+    todos_ok = todos_ok and ok
+
+for imagen in archivos:
+    ok = verificar('Manchas', '100', 5, 'simple', imagen)
+    todos_ok = todos_ok and ok
+
 if todos_ok:
     print(colored("Test de filtros finalizados correctamente", 'green'))
 else:

src/tests/3_correr_test_mem.sh

@@ -43,41 +43,56 @@ function run_test {
     ret=0; return;
 }
 
-for imp in asm; do
-
-  # Cuadrados
-  for s in ${SIZESMEM[*]}; do
-    run_test "$TP2ALU" "Cuadrados" "$imp" "$TESTINDIR/$img1.$s.bmp" ""
-    if [ $ret -ne 0 ]; then exit -1; fi
-  done
-
-  # Manchas
-  for s in ${SIZESMEM[*]}; do
-    v=0.555
-    run_test "$TP2ALU" "Manchas" "$imp" "$TESTINDIR/$img1.$s.bmp" "100"
-    if [ $ret -ne 0 ]; then exit -1; fi
-  done
-
-  # Offset
-  for s in ${SIZESMEM[*]}; do
-    run_test "$TP2ALU" "Offset" "$imp" "$TESTINDIR/$img1.$s.bmp" ""
-    if [ $ret -ne 0 ]; then exit -1; fi
-  done
-
-  # Sharpen
-  for s in ${SIZESMEM[*]}; do
-    run_test "$TP2ALU" "Sharpen" "$imp" "$TESTINDIR/$img1.$s.bmp" ""
-    if [ $ret -ne 0 ]; then exit -1; fi
-  done
-  
-  # Ruido
-#  for s in ${SIZESMEM[*]}; do
-#    run_test "$TP2ALU" "Ruido" "$imp" "$TESTINDIR/$img1.$s.bmp" ""
-#    if [ $ret -ne 0 ]; then exit -1; fi
-#  done
-  
+# Cuadrados
+for s in ${SIZESMEM[*]}; do
+  run_test "$TP2ALU" Cuadrados asm "$TESTINDIR/$img1.$s.bmp" ""
+  if [ $ret -ne 0 ]; then exit -1; fi
 done
 
+for s in ${SIZESMEM[*]}; do
+  run_test "$TP2ALU" Cuadrados avx "$TESTINDIR/$img1.$s.bmp" ""
+  if [ $ret -ne 0 ]; then exit -1; fi
+done
+
+# Manchas
+for s in ${SIZESMEM[*]}; do
+  run_test "$TP2ALU" Manchas asm "$TESTINDIR/$img1.$s.bmp" "100"
+  if [ $ret -ne 0 ]; then exit -1; fi
+done
+
+for s in ${SIZESMEM[*]}; do
+  run_test "$TP2ALU" Manchas simple "$TESTINDIR/$img1.$s.bmp" "100"
+  if [ $ret -ne 0 ]; then exit -1; fi
+done
+
+for s in ${SIZESMEM[*]}; do
+  run_test "$TP2ALU" Manchas mvec "$TESTINDIR/$img1.$s.bmp" "100"
+  if [ $ret -ne 0 ]; then exit -1; fi
+done
+
+# Offset
+for s in ${SIZESMEM[*]}; do
+  run_test "$TP2ALU" Offset asm "$TESTINDIR/$img1.$s.bmp" ""
+  if [ $ret -ne 0 ]; then exit -1; fi
+done
+
+for s in ${SIZESMEM[*]}; do
+  run_test "$TP2ALU" Offset avx "$TESTINDIR/$img1.$s.bmp" ""
+  if [ $ret -ne 0 ]; then exit -1; fi
+done
+
+# Sharpen
+for s in ${SIZESMEM[*]}; do
+  run_test "$TP2ALU" Sharpen asm "$TESTINDIR/$img1.$s.bmp" ""
+  if [ $ret -ne 0 ]; then exit -1; fi
+done
+
+# Ruido
+#for s in ${SIZESMEM[*]}; do
+#  run_test "$TP2ALU" "Ruido" "$imp" "$TESTINDIR/$img1.$s.bmp" ""
+#  if [ $ret -ne 0 ]; then exit -1; fi
+#done
+
 echo ""
 echo -e "$VERDE Felicitaciones los test de MEMORIA finalizaron correctamente $DEFAULT"
 

src/tp2.h

@@ -10,7 +10,6 @@
 #define FILTRO_AVX 2
 /* Manchas */
 #define FILTRO_MVEC 2
-/* Manchas, Offset */
 #define FILTRO_SIMPLE 3
 
 #define ES_FILTRO(config, nombre) (strcmp(config->nombre_filtro, #nombre) == 0)
@@ -73,7 +72,16 @@ typedef struct configuracion_t {
 } configuracion_t;
 
 #define SWITCH_C_ASM(config,c_ver,asm_ver) ( config->tipo_filtro == FILTRO_C ? c_ver : asm_ver )
-#define C_ASM(config) ( SWITCH_C_ASM(config, "C", "ASM") )
+#define C_ASM(config) (config->tipo_filtro == FILTRO_C ? "C" \
+                     : config->tipo_filtro == FILTRO_ASM ? "ASM" \
+                     : config->tipo_filtro == FILTRO_MVEC \
+                       && ES_FILTRO(config, Manchas) ? "MVEC" \
+                     : config->tipo_filtro == FILTRO_SIMPLE \
+                       && ES_FILTRO(config, Manchas) ? "SIMPLE" \
+                     : config->tipo_filtro == FILTRO_AVX \
+                       && (ES_FILTRO(config, Cuadrados) \
+                           || ES_FILTRO(config, Offset)) ? "AVX" \
+                     : "Desconocido")
 
 typedef void (lector_params_fn_t) (configuracion_t *config, int, char *[]);
 typedef void (aplicador_fn_t) (configuracion_t*);