libpayload: cbgfx: Replace bilinear resampling with Lanczos

This patch improves the image resampling (scaling) code in CBGFX to use
the Lanczos algorithm that is widely considered the "best" resampling
algorithm (e.g. also the first choice in Python's PIL library). It is of
course much more elaborate and therefore slower than bilinear
resampling, but a lot of the difference can be made up with
optimizations, and the resulting code was found to still produce
acceptable speeds for existing Chrome OS UI use cases (on an Arm
Cortex-A55 device, time to scale an image to 1101x593 went from ~88ms to
~275ms, a little over 3x slowdown). Nevertheless, if this should be too
slow for anyone there's also an option to tune it down a little, but
still much better than bilinear (same operation was ~170ms with this).

Example images (scaled up by a factor of 7):
Old (bilinear): https://i.imgur.com/ytr2n4Z.png
New (Lanczos a=3): https://i.imgur.com/f0vKluM.png

Signed-off-by: Julius Werner <jwerner@chromium.org>
Change-Id: Idde6f61865bfac2801ee4fff40ac64e4ebddff1a
Reviewed-on: https://review.coreboot.org/c/coreboot/+/42792
Tested-by: build bot (Jenkins) <no-reply@coreboot.org>
Reviewed-by: Yu-Ping Wu <yupingso@google.com>
Reviewed-by: Hung-Te Lin <hungte@chromium.org>

2 files changed, 297 insertions, 79 deletions

diff --git a/payloads/libpayload/Kconfig b/payloads/libpayload/Kconfig
index 7e506d9100..b5dc9a3c8b 100644
--- a/payloads/libpayload/Kconfig
+++ b/payloads/libpayload/Kconfig
@@ -334,6 +334,19 @@ config FONT_SCALE_FACTOR
 	  By default (value of 0), the scale factor is automatically
 	  calculated to ensure at least 130 columns (when possible).
 
+config CBGFX_FAST_RESAMPLE
+	bool "CBGFX: use faster (less pretty) image scaling"
+	default n
+	help
+	  When payloads use the CBGFX library to draw .BMPs on the screen,
+	  they will be resampled with an anti-aliasing filter to scale to the
+	  requested output size. The default implementation should normally be
+	  fast enough, but if desired this option can make it about 50-100%
+	  faster at the cost of quality. (It changes the 'a' parameter in the
+	  Lanczos resampling algorithm from 3 to 2.)
+
+	  Only affects .BMPs that aren't already provided at the right size.
+
 config PC_I8042
 	bool "A common PC i8042 driver"
 	default y if PC_KEYBOARD || PC_MOUSE
diff --git a/payloads/libpayload/drivers/video/graphics.c b/payloads/libpayload/drivers/video/graphics.c
index fa72c9b743..bb8467bf80 100644
--- a/payloads/libpayload/drivers/video/graphics.c
+++ b/payloads/libpayload/drivers/video/graphics.c
@@ -28,6 +28,7 @@
 
 #include <libpayload.h>
 #include <cbfs.h>
+#include <fpmath.h>
 #include <sysinfo.h>
 #include "bitmap.h"
 
@@ -468,33 +469,116 @@ int clear_screen(const struct rgb_color *rgb)
 	return CBGFX_SUCCESS;
 }
 
+static int pal_to_rgb(uint8_t index, const struct bitmap_palette_element_v3 *pal,
+		      size_t palcount, struct rgb_color *out)
+{
+	if (index >= palcount) {
+		LOG("Color index %d exceeds palette boundary\n", index);
+		return CBGFX_ERROR_BITMAP_DATA;
+	}
+
+	out->red = pal[index].red;
+	out->green = pal[index].green;
+	out->blue = pal[index].blue;
+	return CBGFX_SUCCESS;
+}
+
+/*
+ * We're using the Lanczos resampling algorithm to rescale images to a new size.
+ * Since output size is often not cleanly divisible by input size, an output
+ * pixel (ox,oy) corresponds to a point that lies in the middle between several
+ * input pixels (ix,iy), meaning that if you transformed the coordinates of the
+ * output pixel into the input image space, they would be fractional. To sample
+ * the color of this "virtual" pixel with fractional coordinates, we gather the
+ * 6x6 grid of nearest real input pixels in a sample array. Then we multiply the
+ * color values for each of those pixels (separately for red, green and blue)
+ * with a "weight" value that was calculated from the distance between that
+ * input pixel and the fractional output pixel coordinates. This is done for
+ * both X and Y dimensions separately. The combined weights for all 36 sample
+ * pixels add up to 1.0, so by adding up the multiplied color values we get the
+ * interpolated color for the output pixel.
+ *
+ * The CONFIG_LP_CBGFX_FAST_RESAMPLE option let's the user change the 'a'
+ * parameter from the Lanczos weight formula from 3 to 2, which effectively
+ * reduces the size of the sample array from 6x6 to 4x4. This is a bit faster
+ * but doesn't look as good. Most use cases should be fine without it.
+ */
+#if CONFIG(LP_CBGFX_FAST_RESAMPLE)
+#define LNCZ_A 2
+#else
+#define LNCZ_A 3
+#endif
+
+/*
+ * When walking the sample array we often need to start at a pixel close to our
+ * fractional output pixel (for convenience we choose the pixel on the top-left
+ * which corresponds to the integer parts of the output pixel coordinates) and
+ * then work our way outwards in both directions from there. Arrays in C must
+ * start at 0 but we'd really prefer indexes to go from -2 to 3 (for 6x6)
+ * instead, so that this "start pixel" could be 0. Since we cannot do that,
+ * define a constant for the index of that "0th" pixel instead.
+ */
+#define S0 (LNCZ_A - 1)
+
+/* The size of the sample array, which we need a lot. */
+#define SSZ (LNCZ_A * 2)
+
 /*
- * Bi-linear Interpolation
+ * This is implementing the Lanczos kernel according to:
+ * https://en.wikipedia.org/wiki/Lanczos_resampling
  *
- * It estimates the value of a middle point (tx, ty) using the values from four
- * adjacent points (q00, q01, q10, q11).
+ *         / 1							if x = 0
+ * L(x) = <  a * sin(pi * x) * sin(pi * x / a) / (pi^2 * x^2)	if -a < x <= a
+ *	   \ 0							otherwise
  */
-static uint32_t bli(uint32_t q00, uint32_t q10, uint32_t q01, uint32_t q11,
-		    struct fraction *tx, struct fraction *ty)
+static fpmath_t lanczos_weight(fpmath_t in, int off)
 {
-	uint32_t r0 = (tx->n * q10 + (tx->d - tx->n) * q00) / tx->d;
-	uint32_t r1 = (tx->n * q11 + (tx->d - tx->n) * q01) / tx->d;
-	uint32_t p = (ty->n * r1 + (ty->d - ty->n) * r0) / ty->d;
-	return p;
+	/*
+	 * |in| is the output pixel coordinate scaled into the input pixel
+	 * space. |off| is the offset in the sample array for the pixel whose
+	 * weight we're calculating. (off - S0) is the distance from that
+	 * sample pixel to the S0 pixel, and the fractional part of |in|
+	 * (in - floor(in)) is by definition the distance between S0 and the
+	 * output pixel.
+	 *
+	 * So (off - S0) - (in - floor(in)) is the distance from the sample
+	 * pixel to S0 minus the distance from S0 to the output pixel, aka
+	 * the distance from the sample pixel to the output pixel.
+	 */
+	fpmath_t x = fpisub(off - S0, fpsubi(in, fpfloor(in)));
+
+	if (fpequals(x, fp(0)))
+		return fp(1);
+
+	/* x * 2 / a can save some instructions if a == 2 */
+	fpmath_t x2a = x;
+	if (LNCZ_A != 2)
+		x2a = fpmul(x, fpfrac(2, LNCZ_A));
+
+	fpmath_t x_times_pi = fpmul(x, fppi());
+
+	/*
+	 * Rather than using sinr(pi*x), we leverage the "one-based" sine
+	 * function (see <fpmath.h>) with sin1(2*x) so that the pi is eliminated
+	 * since multiplication by an integer is a slightly faster operation.
+	 */
+	fpmath_t tmp = fpmuli(fpdiv(fpsin1(fpmuli(x, 2)), x_times_pi), LNCZ_A);
+	return fpdiv(fpmul(tmp, fpsin1(x2a)), x_times_pi);
 }
 
 static int draw_bitmap_v3(const struct vector *top_left,
-			  const struct scale *scale,
 			  const struct vector *dim,
 			  const struct vector *dim_org,
 			  const struct bitmap_header_v3 *header,
 			  const struct bitmap_palette_element_v3 *pal,
-			  const uint8_t *pixel_array,
-			  uint8_t invert)
+			  const uint8_t *pixel_array, uint8_t invert)
 {
 	const int bpp = header->bits_per_pixel;
 	int32_t dir;
 	struct vector p;
+	int32_t ox, oy;		/* output (resampled) pixel coordinates */
+	int32_t ix, iy;		/* input (source image) pixel coordinates */
+	int sx, sy;	/* index into |sample| (not ringbuffer adjusted) */
 
 	if (header->compression) {
 		LOG("Compressed bitmaps are not supported\n");
@@ -508,10 +592,6 @@ static int draw_bitmap_v3(const struct vector *top_left,
 		LOG("Unsupported bits per pixel: %d\n", bpp);
 		return CBGFX_ERROR_BITMAP_FORMAT;
 	}
-	if (scale->x.n == 0 || scale->y.n == 0) {
-		LOG("Scaling out of range\n");
-		return CBGFX_ERROR_SCALE_OUT_OF_RANGE;
-	}
 
 	const int32_t y_stride = ROUNDUP(dim_org->width * bpp / 8, 4);
 	/*
@@ -530,63 +610,202 @@ static int draw_bitmap_v3(const struct vector *top_left,
 		p.y += dim->height - 1;
 		dir = -1;
 	}
+
+	/* Don't waste time resampling when the scale is 1:1. */
+	if (dim_org->width == dim->width && dim_org->height == dim->height) {
+		for (oy = 0; oy < dim->height; oy++, p.y += dir) {
+			p.x = top_left->x;
+			for (ox = 0; ox < dim->width; ox++, p.x++) {
+				struct rgb_color rgb;
+				if (pal_to_rgb(pixel_array[oy * y_stride + ox],
+					       pal, header->colors_used, &rgb))
+					return CBGFX_ERROR_BITMAP_DATA;
+				set_pixel(&p, calculate_color(&rgb, invert));
+			}
+		}
+		return CBGFX_SUCCESS;
+	}
+
+	/* Precalculate the X-weights for every possible ox so that we only have
+	   to multiply weights together in the end. */
+	fpmath_t (*weight_x)[SSZ] = malloc(sizeof(fpmath_t) * SSZ * dim->width);
+	if (!weight_x)
+		return CBGFX_ERROR_UNKNOWN;
+	for (ox = 0; ox < dim->width; ox++) {
+		for (sx = 0; sx < SSZ; sx++) {
+			fpmath_t ixfp = fpfrac(ox * dim_org->width, dim->width);
+			weight_x[ox][sx] = lanczos_weight(ixfp, sx);
+		}
+	}
+
 	/*
-	 * Plot pixels scaled by the bilinear interpolation. We scan over the
-	 * image on canvas (using d) and find the corresponding pixel in the
-	 * bitmap data (using s0, s1).
-	 *
-	 * When d hits the right bottom corner, s0 also hits the right bottom
-	 * corner of the pixel array because that's how scale->x and scale->y
-	 * have been set. Since the pixel array size is already validated in
-	 * parse_bitmap_header_v3, s0 is guaranteed not to exceed pixel array
-	 * boundary.
+	 * For every sy in the sample array, we directly cache a pointer into
+	 * the .BMP pixel array for the start of the corresponding line. On the
+	 * edges of the image (where we don't have any real pixels to fill all
+	 * lines in the sample array), we just reuse the last valid lines inside
+	 * the image for all lines that would lie outside.
 	 */
-	struct vector s0, s1, d;
-	struct fraction tx, ty;
-	for (d.y = 0; d.y < dim->height; d.y++, p.y += dir) {
-		s0.y = d.y * scale->y.d / scale->y.n;
-		s1.y = s0.y;
-		if (s1.y + 1 < dim_org->height)
-			s1.y++;
-		ty.d = scale->y.n;
-		ty.n = (d.y * scale->y.d) % scale->y.n;
-		const uint8_t *data0 = pixel_array + s0.y * y_stride;
-		const uint8_t *data1 = pixel_array + s1.y * y_stride;
+	const uint8_t *ypix[SSZ];
+	for (sy = 0; sy < SSZ; sy++) {
+		if (sy <= S0)
+			ypix[sy] = pixel_array;
+		else if (sy - S0 >= dim_org->height)
+			ypix[sy] = ypix[sy - 1];
+		else
+			ypix[sy] = &pixel_array[y_stride * (sy - S0)];
+	}
+
+	/* iy and ix track the input pixel corresponding to sample[S0][S0]. */
+	iy = 0;
+	for (oy = 0; oy < dim->height; oy++, p.y += dir) {
+		struct rgb_color sample[SSZ][SSZ];
+
+		/* Like with X weights, we also cache all Y weights. */
+		fpmath_t iyfp = fpfrac(oy * dim_org->height, dim->height);
+		fpmath_t weight_y[SSZ];
+		for (sy = 0; sy < SSZ; sy++)
+			weight_y[sy] = lanczos_weight(iyfp, sy);
+
+		/*
+		 * If we have a new input pixel line between the last oy and
+		 * this one, we have to adjust iy forward. When upscaling, this
+		 * is not always the case for each new output line. When
+		 * downscaling, we may even cross more than one line per output
+		 * pixel.
+		 */
+		while (fpfloor(iyfp) > iy) {
+			iy++;
+
+			/* Shift ypix array up to center around next iy line. */
+			for (sy = 0; sy < SSZ - 1; sy++)
+				ypix[sy] = ypix[sy + 1];
+
+			/* Calculate the last ypix that is being shifted in,
+			   but beware of reaching the end of the input image. */
+			if (iy + LNCZ_A < dim_org->height)
+				ypix[SSZ - 1] = &pixel_array[y_stride *
+							     (iy + LNCZ_A)];
+		}
+
+		/*
+		 * Initialize the sample array for this line. For pixels to the
+		 * left of S0 there are no corresponding input pixels so just
+		 * copy the S0 values over.
+		 *
+		 * Also initialize the equals counter, which counts how many of
+		 * the latest pixels were exactly equal. We know the columns
+		 * left of S0 must be equal to S0, so start with that number.
+		 */
+		int equals = S0 * SSZ;
+		uint8_t last_equal = ypix[0][0];
+		for (sy = 0; sy < SSZ; sy++) {
+			for (sx = S0; sx < SSZ; sx++) {
+				if (sx >= dim_org->width) {
+					sample[sx][sy] = sample[sx - 1][sy];
+					equals++;
+					continue;
+				}
+				uint8_t i = ypix[sy][sx - S0];
+				if (pal_to_rgb(i, pal, header->colors_used,
+					       &sample[sx][sy]))
+					goto bitmap_error;
+				if (i == last_equal) {
+					equals++;
+				} else {
+					last_equal = i;
+					equals = 1;
+				}
+			}
+			for (sx = S0 - 1; sx >= 0; sx--)
+				sample[sx][sy] = sample[S0][sy];
+		}
+
+		ix = 0;
 		p.x = top_left->x;
-		for (d.x = 0; d.x < dim->width; d.x++, p.x++) {
-			s0.x = d.x * scale->x.d / scale->x.n;
-			s1.x = s0.x;
-			if (s1.x + 1 < dim_org->width)
-				s1.x++;
-			tx.d = scale->x.n;
-			tx.n = (d.x * scale->x.d) % scale->x.n;
-			uint8_t c00 = data0[s0.x];
-			uint8_t c10 = data0[s1.x];
-			uint8_t c01 = data1[s0.x];
-			uint8_t c11 = data1[s1.x];
-			if (c00 >= header->colors_used
-					|| c10 >= header->colors_used
-					|| c01 >= header->colors_used
-					|| c11 >= header->colors_used) {
-				LOG("Color index exceeds palette boundary\n");
-				return CBGFX_ERROR_BITMAP_DATA;
+		for (ox = 0; ox < dim->width; ox++, p.x++) {
+			/* Adjust ix forward, same as iy above. */
+			fpmath_t ixfp = fpfrac(ox * dim_org->width, dim->width);
+			while (fpfloor(ixfp) > ix) {
+				ix++;
+
+				/*
+				 * We want to reuse the sample columns we
+				 * already have, but we don't want to copy them
+				 * all around for every new column either.
+				 * Instead, treat the X dimension of the sample
+				 * array like a ring buffer indexed by ix. rx is
+				 * the ringbuffer-adjusted offset of the new
+				 * column in sample (the rightmost one) we're
+				 * trying to fill.
+				 */
+				int rx = (SSZ - 1 + ix) % SSZ;
+				for (sy = 0; sy < SSZ; sy++) {
+					if (ix + LNCZ_A >= dim_org->width) {
+						sample[rx][sy] = sample[(SSZ - 2
+							+ ix) % SSZ][sy];
+						equals++;
+						continue;
+					}
+					uint8_t i = ypix[sy][ix + LNCZ_A];
+					if (i == last_equal) {
+						if (equals++ >= (SSZ * SSZ))
+							continue;
+					} else {
+						last_equal = i;
+						equals = 1;
+					}
+					if (pal_to_rgb(i, pal,
+						       header->colors_used,
+						       &sample[rx][sy]))
+						goto bitmap_error;
+				}
+			}
+
+			/* If all pixels in sample are equal, fast path. */
+			if (equals >= (SSZ * SSZ)) {
+				set_pixel(&p, calculate_color(&sample[0][0],
+							      invert));
+				continue;
+			}
+
+			fpmath_t red = fp(0);
+			fpmath_t green = fp(0);
+			fpmath_t blue = fp(0);
+			for (sy = 0; sy < SSZ; sy++) {
+				for (sx = 0; sx < SSZ; sx++) {
+					int rx = (sx + ix) % SSZ;
+					fpmath_t weight = fpmul(weight_x[ox][sx],
+								weight_y[sy]);
+					red = fpadd(red, fpmuli(weight,
+						sample[rx][sy].red));
+					green = fpadd(green, fpmuli(weight,
+						sample[rx][sy].green));
+					blue = fpadd(blue, fpmuli(weight,
+						sample[rx][sy].blue));
+				}
 			}
-			const struct rgb_color rgb = {
-				.red = bli(pal[c00].red, pal[c10].red,
-					   pal[c01].red, pal[c11].red,
-					   &tx, &ty),
-				.green = bli(pal[c00].green, pal[c10].green,
-					     pal[c01].green, pal[c11].green,
-					     &tx, &ty),
-				.blue = bli(pal[c00].blue, pal[c10].blue,
-					    pal[c01].blue, pal[c11].blue,
-					    &tx, &ty),
+
+			/*
+			 * Weights *should* sum up to 1.0 (making this not
+			 * necessary) but just to hedge against rounding errors
+			 * we should clamp color values to their legal limits.
+			 */
+			struct rgb_color rgb = {
+				.red = MAX(0, MIN(UINT8_MAX, fpround(red))),
+				.green = MAX(0, MIN(UINT8_MAX, fpround(green))),
+				.blue = MAX(0, MIN(UINT8_MAX, fpround(blue))),
 			};
+
 			set_pixel(&p, calculate_color(&rgb, invert));
 		}
 	}
 
+	free(weight_x);
 	return CBGFX_SUCCESS;
+
+bitmap_error:
+	free(weight_x);
+	return CBGFX_ERROR_BITMAP_DATA;
 }
 
 static int get_bitmap_file_header(const void *bitmap, size_t size,
@@ -780,7 +999,6 @@ int draw_bitmap(const void *bitmap, size_t size,
 	const struct bitmap_palette_element_v3 *palette;
 	const uint8_t *pixel_array;
 	struct vector top_left, dim, dim_org;
-	struct scale scale;
 	int rv;
 	const uint8_t pivot = flags & PIVOT_MASK;
 	const uint8_t invert = (flags & INVERT_COLORS) >> INVERT_SHIFT;
@@ -799,12 +1017,6 @@ int draw_bitmap(const void *bitmap, size_t size,
 	if (rv)
 		return rv;
 
-	/* Calculate self scale */
-	scale.x.n = dim.width;
-	scale.x.d = dim_org.width;
-	scale.y.n = dim.height;
-	scale.y.d = dim_org.height;
-
 	/* Calculate coordinate */
 	rv = calculate_position(&dim, pos_rel, pivot, &top_left);
 	if (rv)
@@ -816,7 +1028,7 @@ int draw_bitmap(const void *bitmap, size_t size,
 		return rv;
 	}
 
-	return draw_bitmap_v3(&top_left, &scale, &dim, &dim_org,
+	return draw_bitmap_v3(&top_left, &dim, &dim_org,
 			      &header, palette, pixel_array, invert);
 }
 
@@ -827,7 +1039,6 @@ int draw_bitmap_direct(const void *bitmap, size_t size,
 	const struct bitmap_palette_element_v3 *palette;
 	const uint8_t *pixel_array;
 	struct vector dim;
-	struct scale scale;
 	int rv;
 
 	if (cbgfx_init())
@@ -839,19 +1050,13 @@ int draw_bitmap_direct(const void *bitmap, size_t size,
 	if (rv)
 		return rv;
 
-	/* Calculate self scale */
-	scale.x.n = 1;
-	scale.x.d = 1;
-	scale.y.n = 1;
-	scale.y.d = 1;
-
 	rv = check_boundary(top_left, &dim, &screen);
 	if (rv) {
 		LOG("Bitmap image exceeds screen boundary\n");
 		return rv;
 	}
 
-	return draw_bitmap_v3(top_left, &scale, &dim, &dim,
+	return draw_bitmap_v3(top_left, &dim, &dim,
 			      &header, palette, pixel_array, 0);
 }