diff --git a/CMakeLists.txt b/CMakeLists.txt
index 8c30b033..184dc807 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -11,6 +11,7 @@ set(app_sources "src/epdiy.c"
                 "src/output_common/line_queue.c"
                 "src/output_common/render_context.c"
                 "src/output_common/render_method.c"
+                "src/output_common/utils.c"
                 "src/font.c"
                 "src/displays.c"
                 "src/diff.S"
diff --git a/examples/fb_mode_test/main/main.c b/examples/fb_mode_test/main/main.c
index 567b3c66..5d3a93bc 100644
--- a/examples/fb_mode_test/main/main.c
+++ b/examples/fb_mode_test/main/main.c
@@ -48,9 +48,9 @@ static inline void checkError(enum EpdDrawError err) {
 /**
  * Clears the screen to white and resets the framebuffer.
  */
-void clear() {
+void clear(EpdRect area) {
     epd_poweron();
-    epd_clear();
+    epd_clear_area(area);
     epd_poweroff();
     memset(framebuffer, 0xFF, fb_size);
 }
@@ -189,13 +189,17 @@ void app_main() {
     fb_size = epd_width() * epd_height() / 2;
     framebuffer = heap_caps_aligned_alloc(16, fb_size, MALLOC_CAP_SPIRAM);
 
-    clear();
+    clear(epd_full_screen());
 
     test_8ppB();
 
     memset(framebuffer, 0xFF, fb_size);
 
     test_2ppB();
+    
+    // clear a small square to test selective clear.
+    EpdRect area = {.x = 250, .y = 150, .width=50, .height=50};
+    clear(area);
 
     printf("going to sleep...\n");
     epd_deinit();
diff --git a/src/output_common/lut.c b/src/output_common/lut.c
index 4ae87233..cad01700 100644
--- a/src/output_common/lut.c
+++ b/src/output_common/lut.c
@@ -96,33 +96,6 @@ static inline int max(int x, int y) {
 // status tracker for row skipping
 uint32_t skipping;
 
-__attribute__((optimize("O3"))) void IRAM_ATTR
-reorder_line_buffer(uint32_t* line_data, int buf_len) {
-    for (uint32_t i = 0; i < buf_len / 4; i++) {
-        uint32_t val = *line_data;
-        *(line_data++) = val >> 16 | ((val & 0x0000FFFF) << 16);
-    }
-}
-
-__attribute__((optimize("O3"))) void IRAM_ATTR
-bit_shift_buffer_right(uint8_t* buf, uint32_t len, int shift) {
-    uint8_t carry = 0xFF << (8 - shift);
-    for (uint32_t i = 0; i < len; i++) {
-        uint8_t val = buf[i];
-        buf[i] = (val >> shift) | carry;
-        carry = val << (8 - shift);
-    }
-}
-
-__attribute__((optimize("O3"))) void IRAM_ATTR
-nibble_shift_buffer_right(uint8_t* buf, uint32_t len) {
-    uint8_t carry = 0xF;
-    for (uint32_t i = 0; i < len; i++) {
-        uint8_t val = buf[i];
-        buf[i] = (val << 4) | carry;
-        carry = (val & 0xF0) >> 4;
-    }
-}
 
 __attribute__((optimize("O3"))) void IRAM_ATTR calc_epd_input_8ppB(
     const uint32_t* line_data, uint8_t* epd_input, const uint8_t* lut, uint32_t epd_width
diff --git a/src/output_common/lut.h b/src/output_common/lut.h
index 364636e4..c69f7713 100644
--- a/src/output_common/lut.h
+++ b/src/output_common/lut.h
@@ -31,17 +31,3 @@ typedef struct {
  */
 LutFunctionPair find_lut_functions(enum EpdDrawMode mode, uint32_t lut_size);
 
-/*
- * Reorder the output buffer to account for I2S FIFO order.
- */
-void reorder_line_buffer(uint32_t* line_data, int buf_len);
-
-/**
- * Apply a mask to a line buffer.
- * `len` must be divisible by 4.
- */
-void epd_apply_line_mask(uint8_t* buf, const uint8_t* mask, int len);
-
-// legacy functions
-void bit_shift_buffer_right(uint8_t* buf, uint32_t len, int shift);
-void nibble_shift_buffer_right(uint8_t* buf, uint32_t len);
diff --git a/src/output_i2s/render_i2s.c b/src/output_i2s/render_i2s.c
index 08db8106..8bb3b784 100644
--- a/src/output_i2s/render_i2s.c
+++ b/src/output_i2s/render_i2s.c
@@ -13,6 +13,7 @@
 
 // output a row to the display.
 #include "../output_common/lut.h"
+#include "../output_common/utils.h"
 #include "../output_common/render_context.h"
 #include "i2s_data_bus.h"
 #include "rmt_pulse.h"
@@ -183,14 +184,11 @@ void i2s_do_update(RenderContext_t* ctx) {
 void IRAM_ATTR epd_push_pixels_i2s(RenderContext_t* ctx, EpdRect area, short time, int color) {
     int line_bytes = ctx->display_width / 4;
     uint8_t row[line_bytes];
-    memset(row, 0, line_bytes);
-
     const uint8_t color_choice[4] = { DARK_BYTE, CLEAR_BYTE, 0x00, 0xFF };
-    for (uint32_t i = 0; i < area.width; i++) {
-        uint32_t position = i + area.x % 4;
-        uint8_t mask = color_choice[color] & (0b00000011 << (2 * (position % 4)));
-        row[area.x / 4 + position / 4] |= mask;
-    }
+    memset(row, color_choice[color], line_bytes);
+
+    epd_populate_area_mask(ctx->line_mask, ctx->area, ctx->display_width);
+    epd_apply_line_mask(row, ctx->line_mask, ctx->display_width / 4);
     reorder_line_buffer((uint32_t*)row, line_bytes);
 
     i2s_start_frame();
diff --git a/src/output_lcd/lcd_driver.c b/src/output_lcd/lcd_driver.c
index 234d87c8..b5b2630e 100644
--- a/src/output_lcd/lcd_driver.c
+++ b/src/output_lcd/lcd_driver.c
@@ -450,7 +450,7 @@ static esp_err_t allocate_lcd_buffers() {
 
     // allocate bounce buffers
     for (int i = 0; i < 2; i++) {
-        lcd.bounce_buffer[i] = heap_caps_aligned_calloc(4, 1, lcd.bb_size, dma_flags);
+        lcd.bounce_buffer[i] = heap_caps_aligned_calloc(16, 1, lcd.bb_size, dma_flags);
         ESP_RETURN_ON_FALSE(lcd.bounce_buffer[i], ESP_ERR_NO_MEM, TAG, "install interrupt failed");
     }
 
diff --git a/src/output_lcd/render_lcd.c b/src/output_lcd/render_lcd.c
index 71aadabb..8b543cad 100644
--- a/src/output_lcd/render_lcd.c
+++ b/src/output_lcd/render_lcd.c
@@ -2,6 +2,7 @@
 #include <string.h>
 
 #include "../output_common/render_method.h"
+#include "../output_common/utils.h"
 
 #ifdef RENDER_METHOD_LCD
 
@@ -17,18 +18,39 @@
 #include "lcd_driver.h"
 #include "render_lcd.h"
 
+
+void epd_apply_line_mask_VE(uint8_t * line, const uint8_t* mask, int mask_len);
+
+
 static bool IRAM_ATTR fill_line_noop(RenderContext_t* ctx, uint8_t* line) {
     memset(line, 0x00, ctx->display_width / 4);
+    ctx->lines_consumed++;
     return false;
 }
 
 static bool IRAM_ATTR fill_line_white(RenderContext_t* ctx, uint8_t* line) {
+    // do a no-op if we're out of the draw area
+    if (ctx->lines_consumed < ctx->area.y || ctx->lines_consumed >= ctx->area.y + ctx->area.height) {
+        return fill_line_noop(ctx, line);
+    }
+
     memset(line, CLEAR_BYTE, ctx->display_width / 4);
+    // we use the non-VE version here, because the buffers are not extended to multiples of 16 bytes
+    epd_apply_line_mask(line, ctx->line_mask, ctx->display_width / 4);
+    ctx->lines_consumed++;
     return false;
 }
 
 static bool IRAM_ATTR fill_line_black(RenderContext_t* ctx, uint8_t* line) {
+    // do a no-op if we're out of the draw area
+    if (ctx->lines_consumed < ctx->area.y || ctx->lines_consumed >= ctx->area.y + ctx->area.height) {
+        return fill_line_noop(ctx, line);
+    }
+
     memset(line, DARK_BYTE, ctx->display_width / 4);
+    // we use the non-VE version here, because the buffers are not extended to multiples of 16 bytes
+    epd_apply_line_mask(line, ctx->line_mask, ctx->display_width / 4);
+    ctx->lines_consumed++;
     return false;
 }
 
@@ -100,6 +122,10 @@ void lcd_do_update(RenderContext_t* ctx) {
 void epd_push_pixels_lcd(RenderContext_t* ctx, short time, int color) {
     epd_set_mode(1);
     ctx->current_frame = 0;
+    ctx->lines_consumed = 0;
+
+    epd_populate_area_mask(ctx->line_mask, ctx->area, ctx->display_width);
+
     epd_lcd_frame_done_cb((frame_done_func_t)handle_lcd_frame_done, ctx);
     if (color == 0) {
         epd_lcd_line_source_cb((line_cb_func_t)&fill_line_black, ctx);
@@ -194,7 +220,6 @@ lcd_calculate_frame(RenderContext_t* ctx, int thread_id) {
         ctx->lut_lookup_func(lp, buf, ctx->conversion_lut, ctx->display_width);
 
         // apply the line mask
-        void epd_apply_line_mask_VE(uint8_t * line, const uint8_t* mask, int mask_len);
         epd_apply_line_mask_VE(buf, ctx->line_mask, ctx->display_width / 4);
 
         lq_commit(lq);