feat(draw): add unified NEON acceleration (#4860)

Signed-off-by: Peter Bee <bijunda1@xiaomi.com>
Co-authored-by: Gabor Kiss-Vamosi <kisvegabor@gmail.com>

lv_conf_template.h

@@ -17,7 +17,9 @@
 #ifndef LV_CONF_H
 #define LV_CONF_H
 
+#ifndef __ASSEMBLY__
 #include <stdint.h>
+#endif
 
 /*====================
    COLOR SETTINGS

lvgl.mk

@@ -1,8 +1,12 @@
 LVGL_PATH ?= ${shell pwd}/lvgl
 
+ASRCS += $(shell find $(LVGL_PATH)/src -type f -name '*.S')
 CSRCS += $(shell find $(LVGL_PATH)/src -type f -name '*.c')
 CSRCS += $(shell find $(LVGL_PATH)/demos -type f -name '*.c')
 CSRCS += $(shell find $(LVGL_PATH)/examples -type f -name '*.c')
-CXXSRCS += $(shell find $(LVGL_PATH)/src/libs/thorvg -type f -name '*.cpp')
+CXXEXT := .cpp
-CFLAGS += "-I$(LVGL_PATH)"
+CXXSRCS += $(shell find $(LVGL_PATH)/src/libs/thorvg -type f -name '*${CXXEXT}')
 
+AFLAGS += "-I$(LVGL_PATH)"
+CFLAGS += "-I$(LVGL_PATH)"
+CXXFLAGS += "-I$(LVGL_PATH)"

scripts/lv_conf_internal_gen.py

@@ -150,15 +150,19 @@ fout.write(
  * End of parsing lv_conf_template.h
  -----------------------------------*/
 
+#ifndef __ASSEMBLY__
 LV_EXPORT_CONST_INT(LV_DPI_DEF);
+#endif
 
 #undef _LV_KCONFIG_PRESENT
 
+#ifndef __ASSEMBLY__
 #if LV_USE_FLOAT
     typedef float lv_value_precise_t;
 #else
     typedef int32_t lv_value_precise_t;
 #endif
+#endif
 
 /*Set some defines if a dependency is disabled*/
 #if LV_USE_LOG == 0

src/draw/sw/blend/lv_draw_sw_blend_to_argb8888.c

@@ -16,6 +16,12 @@
 #include "../../../misc/lv_color.h"
 #include "../../../stdlib/lv_string.h"
 
+#if LV_USE_DRAW_SW_ASM == LV_DRAW_SW_ASM_NEON
+    #include "neon/lv_blend_neon.h"
+#elif LV_USE_DRAW_SW_ASM == LV_DRAW_SW_ASM_CUSTOM
+    #include LV_DRAW_SW_ASM_CUSTOM_INCLUDE
+#endif
+
 /*********************
  *      DEFINES
  *********************/
@@ -78,8 +84,20 @@ LV_ATTRIBUTE_FAST_MEM void lv_draw_sw_blend_color_to_argb8888(_lv_draw_sw_blend_
     int32_t x;
     int32_t y;
 
+    LV_UNUSED(w);
+    LV_UNUSED(h);
+    LV_UNUSED(x);
+    LV_UNUSED(y);
+    LV_UNUSED(opa);
+    LV_UNUSED(mask);
+    LV_UNUSED(mask_stride);
+    LV_UNUSED(dest_stride);
+
     /*Simple fill*/
     if(mask == NULL && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_COLOR_BLEND_TO_ARGB8888
+        LV_DRAW_SW_COLOR_BLEND_TO_ARGB8888(dsc);
+#else
         uint32_t color32 = lv_color_to_u32(dsc->color);
         uint32_t * dest_buf = dsc->dest_buf;
         for(y = 0; y < h; y++) {
@@ -110,9 +128,13 @@ LV_ATTRIBUTE_FAST_MEM void lv_draw_sw_blend_color_to_argb8888(_lv_draw_sw_blend_
 
             dest_buf = drawbuf_next_row(dest_buf, dest_stride);
         }
+#endif
     }
     /*Opacity only*/
     else if(mask == NULL && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_COLOR_BLEND_TO_ARGB8888_WITH_OPA
+        LV_DRAW_SW_COLOR_BLEND_TO_ARGB8888_WITH_OPA(dsc);
+#else
         lv_color32_t color_argb = lv_color_to_32(dsc->color, opa);
         lv_color32_t * dest_buf = dsc->dest_buf;
 
@@ -122,9 +144,13 @@ LV_ATTRIBUTE_FAST_MEM void lv_draw_sw_blend_color_to_argb8888(_lv_draw_sw_blend_
             }
             dest_buf = drawbuf_next_row(dest_buf, dest_stride);
         }
+#endif
     }
     /*Masked with full opacity*/
     else if(mask && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_COLOR_BLEND_TO_ARGB8888_WITH_MASK
+        LV_DRAW_SW_COLOR_BLEND_TO_ARGB8888_WITH_MASK(dsc);
+#else
         lv_color32_t color_argb = lv_color_to_32(dsc->color, 0xff);
         lv_color32_t * dest_buf = dsc->dest_buf;
         for(y = 0; y < h; y++) {
@@ -136,9 +162,13 @@ LV_ATTRIBUTE_FAST_MEM void lv_draw_sw_blend_color_to_argb8888(_lv_draw_sw_blend_
             dest_buf = drawbuf_next_row(dest_buf, dest_stride);
             mask += mask_stride;
         }
+#endif
     }
     /*Masked with opacity*/
     else {
+#ifdef LV_DRAW_SW_COLOR_BLEND_TO_ARGB8888_MIX_MASK_OPA
+        LV_DRAW_SW_COLOR_BLEND_TO_ARGB8888_MIX_MASK_OPA(dsc);
+#else
         lv_color32_t color_argb = lv_color_to_32(dsc->color, opa);
         lv_color32_t * dest_buf = dsc->dest_buf;
         for(y = 0; y < h; y++) {
@@ -149,6 +179,7 @@ LV_ATTRIBUTE_FAST_MEM void lv_draw_sw_blend_color_to_argb8888(_lv_draw_sw_blend_
             dest_buf = drawbuf_next_row(dest_buf, dest_stride);
             mask += mask_stride;
         }
+#endif
     }
 }
 
@@ -196,8 +227,18 @@ LV_ATTRIBUTE_FAST_MEM static void rgb565_image_blend(_lv_draw_sw_blend_image_dsc
     int32_t x;
     int32_t y;
 
+    LV_UNUSED(color_argb);
+
     if(dsc->blend_mode == LV_BLEND_MODE_NORMAL) {
         if(mask_buf == NULL) {
+#ifdef LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_ARGB8888
+            if(opa >= LV_OPA_MAX) {
+                LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_ARGB8888(dsc);
+            }
+            else {
+                LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_ARGB8888_WITH_OPA(dsc);
+            }
+#else
             color_argb.alpha = opa;
             for(y = 0; y < h; y++) {
                 for(x = 0; x < w; x++) {
@@ -209,8 +250,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb565_image_blend(_lv_draw_sw_blend_image_dsc
                 dest_buf_c32 = drawbuf_next_row(dest_buf_c32, dest_stride);
                 src_buf_c16 = drawbuf_next_row(src_buf_c16, src_stride);
             }
+#endif
         }
         else if(mask_buf && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_ARGB8888_WITH_MASK
+            LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_ARGB8888_WITH_MASK(dsc);
+#else
             for(y = 0; y < h; y++) {
                 for(x = 0; x < w; x++) {
                     color_argb.alpha = mask_buf[x];
@@ -223,8 +268,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb565_image_blend(_lv_draw_sw_blend_image_dsc
                 src_buf_c16 = drawbuf_next_row(src_buf_c16, src_stride);
                 mask_buf += mask_stride;
             }
+#endif
         }
         else {
+#ifdef LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_ARGB8888_MIX_MASK_OPA
+            LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_ARGB8888_MIX_MASK_OPA(dsc);
+#else
             for(y = 0; y < h; y++) {
                 for(x = 0; x < w; x++) {
                     color_argb.alpha = LV_OPA_MIX2(mask_buf[x], opa);
@@ -237,6 +286,7 @@ LV_ATTRIBUTE_FAST_MEM static void rgb565_image_blend(_lv_draw_sw_blend_image_dsc
                 src_buf_c16 = drawbuf_next_row(src_buf_c16, src_stride);
                 mask_buf += mask_stride;
             }
+#endif
         }
     }
     else {
@@ -278,9 +328,14 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
     int32_t src_x;
     int32_t y;
 
+    LV_UNUSED(color_argb);
+
     if(dsc->blend_mode == LV_BLEND_MODE_NORMAL) {
         /*Special case*/
         if(mask_buf == NULL && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_ARGB8888
+            LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_ARGB8888(dsc, src_px_size);
+#else
             if(src_px_size == 4) {
                 uint32_t line_in_bytes = w * 4;
                 for(y = 0; y < h; y++) {
@@ -301,8 +356,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
                     src_buf = drawbuf_next_row(src_buf, src_stride);
                 }
             }
+#endif
         }
         if(mask_buf == NULL && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_ARGB8888_WITH_OPA
+            LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_ARGB8888_WITH_OPA(dsc, src_px_size);
+#else
             color_argb.alpha = opa;
             for(y = 0; y < h; y++) {
                 for(dest_x = 0, src_x = 0; dest_x < w; dest_x++, src_x += src_px_size) {
@@ -314,8 +373,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
                 dest_buf_c32 = drawbuf_next_row(dest_buf_c32, dest_stride);
                 src_buf = drawbuf_next_row(src_buf, src_stride);
             }
+#endif
         }
         if(mask_buf && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_ARGB8888_WITH_MASK
+            LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_ARGB8888_WITH_MASK(dsc, src_px_size);
+#else
             for(y = 0; y < h; y++) {
                 for(dest_x = 0, src_x = 0; dest_x < w; dest_x++, src_x += src_px_size) {
                     color_argb.alpha = mask_buf[dest_x];
@@ -328,8 +391,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
                 src_buf = drawbuf_next_row(src_buf, src_stride);
                 mask_buf += mask_stride;
             }
+#endif
         }
         if(mask_buf && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_ARGB8888_MIX_MASK_OPA
+            LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_ARGB8888_MIX_MASK_OPA(dsc, src_px_size);
+#else
             for(y = 0; y < h; y++) {
                 for(dest_x = 0, src_x = 0; dest_x < w; dest_x++, src_x += src_px_size) {
                     color_argb.alpha = (opa * mask_buf[dest_x]) >> 8;
@@ -342,6 +409,7 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
                 src_buf = drawbuf_next_row(src_buf, src_stride);
                 mask_buf += mask_stride;
             }
+#endif
         }
     }
     else {
@@ -384,6 +452,9 @@ LV_ATTRIBUTE_FAST_MEM static void argb8888_image_blend(_lv_draw_sw_blend_image_d
 
     if(dsc->blend_mode == LV_BLEND_MODE_NORMAL) {
         if(mask_buf == NULL && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_ARGB8888
+            LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_ARGB8888(dsc);
+#else
             for(y = 0; y < h; y++) {
                 for(x = 0; x < w; x++) {
                     dest_buf_c32[x] = lv_color_32_32_mix(src_buf_c32[x], dest_buf_c32[x], &cache);
@@ -391,8 +462,12 @@ LV_ATTRIBUTE_FAST_MEM static void argb8888_image_blend(_lv_draw_sw_blend_image_d
                 dest_buf_c32 = drawbuf_next_row(dest_buf_c32, dest_stride);
                 src_buf_c32 = drawbuf_next_row(src_buf_c32, src_stride);
             }
+#endif
         }
         else if(mask_buf == NULL && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_ARGB8888_WITH_OPA
+            LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_ARGB8888_WITH_OPA(dsc);
+#else
             for(y = 0; y < h; y++) {
                 for(x = 0; x < w; x++) {
                     color_argb = src_buf_c32[x];
@@ -402,8 +477,12 @@ LV_ATTRIBUTE_FAST_MEM static void argb8888_image_blend(_lv_draw_sw_blend_image_d
                 dest_buf_c32 = drawbuf_next_row(dest_buf_c32, dest_stride);
                 src_buf_c32 = drawbuf_next_row(src_buf_c32, src_stride);
             }
+#endif
         }
         else if(mask_buf && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_ARGB8888_WITH_MASK
+            LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_ARGB8888_WITH_MASK(dsc);
+#else
             for(y = 0; y < h; y++) {
                 for(x = 0; x < w; x++) {
                     color_argb = src_buf_c32[x];
@@ -414,8 +493,12 @@ LV_ATTRIBUTE_FAST_MEM static void argb8888_image_blend(_lv_draw_sw_blend_image_d
                 src_buf_c32 = drawbuf_next_row(src_buf_c32, src_stride);
                 mask_buf += mask_stride;
             }
+#endif
         }
         else if(mask_buf && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_ARGB8888_MIX_MASK_OPA
+            LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_ARGB8888_MIX_MASK_OPA(dsc);
+#else
             for(y = 0; y < h; y++) {
                 for(x = 0; x < w; x++) {
                     color_argb = src_buf_c32[x];
@@ -426,6 +509,7 @@ LV_ATTRIBUTE_FAST_MEM static void argb8888_image_blend(_lv_draw_sw_blend_image_d
                 src_buf_c32 = drawbuf_next_row(src_buf_c32, src_stride);
                 mask_buf += mask_stride;
             }
+#endif
         }
     }
     else {

src/draw/sw/blend/lv_draw_sw_blend_to_rgb565.c

@@ -1,5 +1,5 @@
 /**
- * @file lv_draw_sw_blend.c
+ * @file lv_draw_sw_blend_to_rgb565.c
  *
  */
 
@@ -16,6 +16,12 @@
 #include "../../../misc/lv_color.h"
 #include "../../../stdlib/lv_string.h"
 
+#if LV_USE_DRAW_SW_ASM == LV_DRAW_SW_ASM_NEON
+    #include "neon/lv_blend_neon.h"
+#elif LV_USE_DRAW_SW_ASM == LV_DRAW_SW_ASM_CUSTOM
+    #include LV_DRAW_SW_ASM_CUSTOM_INCLUDE
+#endif
+
 /*********************
  *      DEFINES
  *********************/
@@ -77,8 +83,22 @@ LV_ATTRIBUTE_FAST_MEM void lv_draw_sw_blend_color_to_rgb565(_lv_draw_sw_blend_fi
     int32_t x;
     int32_t y;
 
+    LV_UNUSED(w);
+    LV_UNUSED(h);
+    LV_UNUSED(x);
+    LV_UNUSED(y);
+    LV_UNUSED(opa);
+    LV_UNUSED(mask);
+    LV_UNUSED(color16);
+    LV_UNUSED(mask_stride);
+    LV_UNUSED(dest_stride);
+    LV_UNUSED(dest_buf_u16);
+
     /*Simple fill*/
     if(mask == NULL && opa >= LV_OPA_MAX)  {
+#ifdef LV_DRAW_SW_COLOR_BLEND_TO_RGB565
+        LV_DRAW_SW_COLOR_BLEND_TO_RGB565(dsc);
+#else
         for(y = 0; y < h; y++) {
             uint16_t * dest_end_final = dest_buf_u16 + w;
             uint32_t * dest_end_mid = (uint32_t *)((uint16_t *) dest_buf_u16 + ((w - 1) & ~(0xF)));
@@ -111,9 +131,13 @@ LV_ATTRIBUTE_FAST_MEM void lv_draw_sw_blend_color_to_rgb565(_lv_draw_sw_blend_fi
             dest_buf_u16 = drawbuf_next_row(dest_buf_u16, dest_stride);
             dest_buf_u16 -= w;
         }
+#endif
     }
     /*Opacity only*/
     else if(mask == NULL && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_COLOR_BLEND_TO_RGB565_WITH_OPA
+        LV_DRAW_SW_COLOR_BLEND_TO_RGB565_WITH_OPA(dsc);
+#else
         uint32_t last_dest32_color = dest_buf_u16[0] + 1; /*Set to value which is not equal to the first pixel*/
         uint32_t last_res32_color = 0;
 
@@ -150,10 +174,14 @@ LV_ATTRIBUTE_FAST_MEM void lv_draw_sw_blend_color_to_rgb565(_lv_draw_sw_blend_fi
             }
             dest_buf_u16 = drawbuf_next_row(dest_buf_u16, dest_stride);
         }
+#endif
     }
 
     /*Masked with full opacity*/
     else if(mask && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_COLOR_BLEND_TO_RGB565_WITH_MASK
+        LV_DRAW_SW_COLOR_BLEND_TO_RGB565_WITH_MASK(dsc);
+#else
         uint32_t c32 = color16 + ((uint32_t)color16 << 16);
         for(y = 0; y < h; y++) {
             for(x = 0; x < w && ((lv_uintptr_t)(mask) & 0x3); x++) {
@@ -189,9 +217,13 @@ LV_ATTRIBUTE_FAST_MEM void lv_draw_sw_blend_color_to_rgb565(_lv_draw_sw_blend_fi
             dest_buf_u16 = drawbuf_next_row(dest_buf_u16, dest_stride);
             mask += mask_stride;
         }
+#endif
     }
     /*Masked with opacity*/
     else if(mask && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_COLOR_BLEND_TO_RGB565_MIX_MASK_OPA
+        LV_DRAW_SW_COLOR_BLEND_TO_RGB565_MIX_MASK_OPA(dsc);
+#else
         for(y = 0; y < h; y++) {
             for(x = 0; x < w; x++) {
                 dest_buf_u16[x] = lv_color_16_16_mix(color16, dest_buf_u16[x], LV_OPA_MIX2(mask[x], opa));
@@ -199,6 +231,7 @@ LV_ATTRIBUTE_FAST_MEM void lv_draw_sw_blend_color_to_rgb565(_lv_draw_sw_blend_fi
             dest_buf_u16 = drawbuf_next_row(dest_buf_u16, dest_stride);
             mask += mask_stride;
         }
+#endif
     }
 }
 
@@ -244,14 +277,21 @@ LV_ATTRIBUTE_FAST_MEM static void rgb565_image_blend(_lv_draw_sw_blend_image_dsc
 
     if(dsc->blend_mode == LV_BLEND_MODE_NORMAL) {
         if(mask_buf == NULL && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB565
+            LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB565(dsc);
+#else
             uint32_t line_in_bytes = w * 2;
             for(y = 0; y < h; y++) {
                 lv_memcpy(dest_buf_u16, src_buf_u16, line_in_bytes);
                 dest_buf_u16 = drawbuf_next_row(dest_buf_u16, dest_stride);
                 src_buf_u16 = drawbuf_next_row(src_buf_u16, src_stride);
             }
+#endif
         }
         else if(mask_buf == NULL && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB565_WITH_OPA
+            LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB565_WITH_OPA(dsc);
+#else
             for(y = 0; y < h; y++) {
                 for(x = 0; x < w; x++) {
                     dest_buf_u16[x] = lv_color_16_16_mix(src_buf_u16[x], dest_buf_u16[x], opa);
@@ -259,8 +299,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb565_image_blend(_lv_draw_sw_blend_image_dsc
                 dest_buf_u16 = drawbuf_next_row(dest_buf_u16, dest_stride);
                 src_buf_u16 = drawbuf_next_row(src_buf_u16, src_stride);
             }
+#endif
         }
         else if(mask_buf && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB565_WITH_MASK
+            LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB565_WITH_MASK(dsc);
+#else
             for(y = 0; y < h; y++) {
                 for(x = 0; x < w; x++) {
                     dest_buf_u16[x] = lv_color_16_16_mix(src_buf_u16[x], dest_buf_u16[x], mask_buf[x]);
@@ -269,8 +313,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb565_image_blend(_lv_draw_sw_blend_image_dsc
                 src_buf_u16 = drawbuf_next_row(src_buf_u16, src_stride);
                 mask_buf += mask_stride;
             }
+#endif
         }
         else {
+#ifdef LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB565_MIX_MASK_OPA
+            LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB565_MIX_MASK_OPA(dsc);
+#else
             for(y = 0; y < h; y++) {
                 for(x = 0; x < w; x++) {
                     dest_buf_u16[x] = lv_color_16_16_mix(src_buf_u16[x], dest_buf_u16[x], LV_OPA_MIX2(mask_buf[x], opa));
@@ -279,6 +327,7 @@ LV_ATTRIBUTE_FAST_MEM static void rgb565_image_blend(_lv_draw_sw_blend_image_dsc
                 src_buf_u16 = drawbuf_next_row(src_buf_u16, src_stride);
                 mask_buf += mask_stride;
             }
+#endif
         }
     }
     else {
@@ -345,6 +394,9 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
 
     if(dsc->blend_mode == LV_BLEND_MODE_NORMAL) {
         if(mask_buf == NULL && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB565
+            LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB565(dsc, src_px_size);
+#else
             for(y = 0; y < h; y++) {
                 for(dest_x = 0, src_x = 0; dest_x < w; dest_x++, src_x += src_px_size) {
                     dest_buf_u16[dest_x]  = ((src_buf_u8[src_x + 2] & 0xF8) << 8) +
@@ -354,8 +406,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
                 dest_buf_u16 = drawbuf_next_row(dest_buf_u16, dest_stride);
                 src_buf_u8 += src_stride;
             }
+#endif
         }
         else if(mask_buf == NULL && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB565_WITH_OPA
+            LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB565_WITH_OPA(dsc, src_px_size);
+#else
             for(y = 0; y < h; y++) {
                 for(dest_x = 0, src_x = 0; dest_x < w; dest_x++, src_x += src_px_size) {
                     dest_buf_u16[dest_x] = lv_color_24_16_mix(&src_buf_u8[src_x], dest_buf_u16[dest_x], opa);
@@ -363,8 +419,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
                 dest_buf_u16 = drawbuf_next_row(dest_buf_u16, dest_stride);
                 src_buf_u8 += src_stride;
             }
+#endif
         }
         if(mask_buf && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB565_WITH_MASK
+            LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB565_WITH_MASK(dsc, src_px_size);
+#else
             for(y = 0; y < h; y++) {
                 for(dest_x = 0, src_x = 0; dest_x < w; dest_x++, src_x += src_px_size) {
                     dest_buf_u16[dest_x] = lv_color_24_16_mix(&src_buf_u8[src_x], dest_buf_u16[dest_x], mask_buf[dest_x]);
@@ -373,8 +433,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
                 src_buf_u8 += src_stride;
                 mask_buf += mask_stride;
             }
+#endif
         }
         if(mask_buf && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB565_MIX_MASK_OPA
+            LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB565_MIX_MASK_OPA(dsc, src_px_size);
+#else
             for(y = 0; y < h; y++) {
                 for(dest_x = 0, src_x = 0; dest_x < w; dest_x++, src_x += src_px_size) {
                     dest_buf_u16[dest_x] = lv_color_24_16_mix(&src_buf_u8[src_x], dest_buf_u16[dest_x], LV_OPA_MIX2(mask_buf[dest_x], opa));
@@ -383,6 +447,7 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
                 src_buf_u8 += src_stride;
                 mask_buf += mask_stride;
             }
+#endif
         }
     }
     else {
@@ -445,6 +510,9 @@ LV_ATTRIBUTE_FAST_MEM static void argb8888_image_blend(_lv_draw_sw_blend_image_d
 
     if(dsc->blend_mode == LV_BLEND_MODE_NORMAL) {
         if(mask_buf == NULL && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_RGB565
+            LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_RGB565(dsc);
+#else
             for(y = 0; y < h; y++) {
                 for(dest_x = 0, src_x = 0; dest_x < w; dest_x++, src_x += 4) {
                     dest_buf_u16[dest_x] = lv_color_24_16_mix(&src_buf_u8[src_x], dest_buf_u16[dest_x], src_buf_u8[src_x + 3]);
@@ -452,8 +520,12 @@ LV_ATTRIBUTE_FAST_MEM static void argb8888_image_blend(_lv_draw_sw_blend_image_d
                 dest_buf_u16 = drawbuf_next_row(dest_buf_u16, dest_stride);
                 src_buf_u8 += src_stride;
             }
+#endif
         }
         else if(mask_buf == NULL && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_RGB565_WITH_OPA
+            LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_RGB565_WITH_OPA(dsc);
+#else
             for(y = 0; y < h; y++) {
                 for(dest_x = 0, src_x = 0; dest_x < w; dest_x++, src_x += 4) {
                     dest_buf_u16[dest_x] = lv_color_24_16_mix(&src_buf_u8[src_x], dest_buf_u16[dest_x], LV_OPA_MIX2(src_buf_u8[src_x + 3],
@@ -462,8 +534,12 @@ LV_ATTRIBUTE_FAST_MEM static void argb8888_image_blend(_lv_draw_sw_blend_image_d
                 dest_buf_u16 = drawbuf_next_row(dest_buf_u16, dest_stride);
                 src_buf_u8 += src_stride;
             }
+#endif
         }
         else if(mask_buf && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_RGB565_WITH_MASK
+            LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_RGB565_WITH_MASK(dsc);
+#else
             for(y = 0; y < h; y++) {
                 for(dest_x = 0, src_x = 0; dest_x < w; dest_x++, src_x += 4) {
                     dest_buf_u16[dest_x] = lv_color_24_16_mix(&src_buf_u8[src_x], dest_buf_u16[dest_x],
@@ -473,8 +549,12 @@ LV_ATTRIBUTE_FAST_MEM static void argb8888_image_blend(_lv_draw_sw_blend_image_d
                 src_buf_u8 += src_stride;
                 mask_buf += mask_stride;
             }
+#endif
         }
         else if(mask_buf && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_RGB565_MIX_MASK_OPA
+            LV_DRAW_SW_ARGB8888_BLEND_NORMAL_TO_RGB565_MIX_MASK_OPA(dsc);
+#else
             for(y = 0; y < h; y++) {
                 for(dest_x = 0, src_x = 0; dest_x < w; dest_x++, src_x += 4) {
                     dest_buf_u16[dest_x] = lv_color_24_16_mix(&src_buf_u8[src_x], dest_buf_u16[dest_x],
@@ -484,6 +564,7 @@ LV_ATTRIBUTE_FAST_MEM static void argb8888_image_blend(_lv_draw_sw_blend_image_d
                 src_buf_u8 += src_stride;
                 mask_buf += mask_stride;
             }
+#endif
         }
     }
     else {

src/draw/sw/blend/lv_draw_sw_blend_to_rgb888.c

@@ -1,5 +1,5 @@
 /**
- * @file lv_draw_sw_blend_rgb888.c
+ * @file lv_draw_sw_blend_to_rgb888.c
  *
  */
 
@@ -17,7 +17,7 @@
 #include "../../../stdlib/lv_string.h"
 
 #if LV_USE_DRAW_SW_ASM == LV_DRAW_SW_ASM_NEON
-    #include "neon/lv_blend_to_rgb888_neon.h"
+    #include "neon/lv_blend_neon.h"
 #elif LV_USE_DRAW_SW_ASM == LV_DRAW_SW_ASM_CUSTOM
     #include LV_DRAW_SW_ASM_CUSTOM_INCLUDE
 #endif
@@ -237,6 +237,9 @@ LV_ATTRIBUTE_FAST_MEM static void rgb565_image_blend(_lv_draw_sw_blend_image_dsc
 
     if(dsc->blend_mode == LV_BLEND_MODE_NORMAL) {
         if(mask_buf == NULL && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB888
+            LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB888(dsc, dest_px_size);
+#else
             for(y = 0; y < h; y++) {
                 for(src_x = 0, dest_x = 0; src_x < w; dest_x += dest_px_size, src_x++) {
                     dest_buf_u8[dest_x + 2] = (src_buf_c16[src_x].red * 2106) >> 8;  /*To make it rounded*/
@@ -246,8 +249,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb565_image_blend(_lv_draw_sw_blend_image_dsc
                 dest_buf_u8 += dest_stride;
                 src_buf_c16 = drawbuf_next_row(src_buf_c16, src_stride);
             }
+#endif
         }
         else if(mask_buf == NULL && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB888_WITH_OPA
+            LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB888_WITH_OPA(dsc, dest_px_size);
+#else
             uint8_t res[3];
             for(y = 0; y < h; y++) {
                 for(src_x = 0, dest_x = 0; src_x < w; dest_x += dest_px_size, src_x++) {
@@ -259,8 +266,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb565_image_blend(_lv_draw_sw_blend_image_dsc
                 dest_buf_u8 += dest_stride;
                 src_buf_c16 = drawbuf_next_row(src_buf_c16, src_stride);
             }
+#endif
         }
         else if(mask_buf && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB888_WITH_MASK
+            LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB888_WITH_MASK(dsc, dest_px_size);
+#else
             uint8_t res[3];
             for(y = 0; y < h; y++) {
                 for(src_x = 0, dest_x = 0; src_x < w; dest_x += dest_px_size, src_x++) {
@@ -273,8 +284,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb565_image_blend(_lv_draw_sw_blend_image_dsc
                 src_buf_c16 = drawbuf_next_row(src_buf_c16, src_stride);
                 mask_buf += mask_stride;
             }
+#endif
         }
         else {
+#ifdef LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB888_MIX_MASK_OPA
+            LV_DRAW_SW_RGB565_BLEND_NORMAL_TO_RGB888_MIX_MASK_OPA(dsc, dest_px_size);
+#else
             uint8_t res[3];
             for(y = 0; y < h; y++) {
                 for(src_x = 0, dest_x = 0; src_x < w; dest_x += dest_px_size, src_x++) {
@@ -287,6 +302,7 @@ LV_ATTRIBUTE_FAST_MEM static void rgb565_image_blend(_lv_draw_sw_blend_image_dsc
                 src_buf_c16 = drawbuf_next_row(src_buf_c16, src_stride);
                 mask_buf += mask_stride;
             }
+#endif
         }
     }
     else {
@@ -327,6 +343,9 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
     if(dsc->blend_mode == LV_BLEND_MODE_NORMAL) {
         /*Special case*/
         if(mask_buf == NULL && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB888
+            LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB888(dsc, dest_px_size, src_px_size);
+#else
             if(src_px_size == dest_px_size) {
                 for(y = 0; y < h; y++) {
                     lv_memcpy(dest_buf, src_buf, w);
@@ -345,8 +364,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
                     src_buf += src_stride;
                 }
             }
+#endif
         }
         if(mask_buf == NULL && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB888_WITH_OPA
+            LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB888_WITH_OPA(dsc, dest_px_size, src_px_size);
+#else
             for(y = 0; y < h; y++) {
                 for(dest_x = 0, src_x = 0; dest_x < w; dest_x += dest_px_size, src_x += src_px_size) {
                     lv_color_24_24_mix(&src_buf[src_x], &dest_buf[dest_x], opa);
@@ -354,8 +377,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
                 dest_buf += dest_stride;
                 src_buf += src_stride;
             }
+#endif
         }
         if(mask_buf && opa >= LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB888_WITH_MASK
+            LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB888_WITH_MASK(dsc, dest_px_size, src_px_size);
+#else
             uint32_t mask_x;
             for(y = 0; y < h; y++) {
                 for(mask_x = 0, dest_x = 0, src_x = 0; dest_x < w; mask_x++, dest_x += dest_px_size, src_x += src_px_size) {
@@ -365,8 +392,12 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
                 src_buf += src_stride;
                 mask_buf += mask_stride;
             }
+#endif
         }
         if(mask_buf && opa < LV_OPA_MAX) {
+#ifdef LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB888_MIX_MASK_OPA
+            LV_DRAW_SW_RGB888_BLEND_NORMAL_TO_RGB888_MIX_MASK_OPA(dsc, dest_px_size, src_px_size);
+#else
             uint32_t mask_x;
             for(y = 0; y < h; y++) {
                 for(mask_x = 0, dest_x = 0, src_x = 0; dest_x < w; mask_x++, dest_x += dest_px_size, src_x += src_px_size) {
@@ -376,6 +407,7 @@ LV_ATTRIBUTE_FAST_MEM static void rgb888_image_blend(_lv_draw_sw_blend_image_dsc
                 src_buf += src_stride;
                 mask_buf += mask_stride;
             }
+#endif
         }
     }
     else {

src/lv_conf_internal.h

@@ -55,7 +55,9 @@
  * Start parsing lv_conf_template.h
  -----------------------------------*/
 
+#ifndef __ASSEMBLY__
 #include <stdint.h>
+#endif
 
 /*====================
    COLOR SETTINGS
@@ -2890,15 +2892,19 @@
  * End of parsing lv_conf_template.h
  -----------------------------------*/
 
+#ifndef __ASSEMBLY__
 LV_EXPORT_CONST_INT(LV_DPI_DEF);
+#endif
 
 #undef _LV_KCONFIG_PRESENT
 
+#ifndef __ASSEMBLY__
 #if LV_USE_FLOAT
     typedef float lv_value_precise_t;
 #else
     typedef int32_t lv_value_precise_t;
 #endif
+#endif
 
 /*Set some defines if a dependency is disabled*/
 #if LV_USE_LOG == 0

src/misc/lv_types.h

@@ -13,7 +13,10 @@ extern "C" {
 /*********************
  *      INCLUDES
  *********************/
+
+#ifndef __ASSEMBLY__
 #include <stdint.h>
+#endif
 
 /*********************
  *      DEFINES
@@ -55,6 +58,9 @@ extern "C" {
  *      TYPEDEFS
  **********************/
 
+/* Exclude C enum and struct definitions when included by assembly code */
+#ifndef __ASSEMBLY__
+
 /**
  * LVGL error codes.
  */
@@ -88,6 +94,8 @@ typedef int32_t lv_intptr_t;
 
 #endif
 
+#endif /*__ASSEMBLY__*/
+
 /**********************
  * GLOBAL PROTOTYPES
  **********************/