Misc. fixes/cleanups in superres upscaling
Change-Id: I97f7d390e580b6e1b2b471a6f1a95082323d39aa
diff --git a/av1/common/resize.c b/av1/common/resize.c
index db6b593..859ade1 100644
--- a/av1/common/resize.c
+++ b/av1/common/resize.c
@@ -451,19 +451,19 @@
#define UPSCALE_PROC_UNIT 64 // Source step (roughly)
#define UPSCALE_PROC_UNIT_SCALE (UPSCALE_PROC_UNIT / SCALE_NUMERATOR)
-static int get_upscale_convolve_params(int in_length, int out_length,
- int out_start, int32_t *x0_qn,
- int32_t *x_step_qn) {
- assert(out_length >= in_length);
- *x_step_qn =
- ((in_length << INTERP_PRECISION_BITS) + out_length / 2) / out_length;
- *x0_qn = (-((out_length - in_length) << (INTERP_PRECISION_BITS - 1)) +
- out_length / 2) /
- out_length +
- out_start * (*x_step_qn) + SUBPEL_INTERP_EXTRA_OFF;
- const int off = (*x0_qn >> INTERP_PRECISION_BITS);
- *x0_qn = (int32_t)((uint32_t)(*x0_qn) & INTERP_PRECISION_MASK);
- return off;
+static int32_t get_upscale_convolve_step(int in_length, int out_length) {
+ return ((in_length << INTERP_PRECISION_BITS) + out_length / 2) / out_length;
+}
+
+static int32_t get_upscale_convolve_x0(int in_length, int out_length,
+ int32_t x_step_qn) {
+ const int err = out_length * x_step_qn - (in_length << INTERP_PRECISION_BITS);
+ const int32_t x0 =
+ (-((out_length - in_length) << (INTERP_PRECISION_BITS - 1)) +
+ out_length / 2) /
+ out_length +
+ SUBPEL_INTERP_EXTRA_OFF - err / 2;
+ return (int32_t)((uint32_t)x0 & INTERP_PRECISION_MASK);
}
static void convolve_horiz_superres_c(const uint8_t *src, uint8_t *dst,
@@ -490,16 +490,24 @@
int superres_denom,
const int16_t *interp_filters,
int interp_taps) {
+ assert(in_length < out_length);
const int oproc_unit = UPSCALE_PROC_UNIT_SCALE * superres_denom;
- int32_t x_step_qn, x0_qn;
+ const int32_t x_step_qn = get_upscale_convolve_step(in_length, out_length);
+ const int32_t x0_qn =
+ get_upscale_convolve_x0(in_length, out_length, x_step_qn);
+ const uint8_t *srcp = src - 1;
+ int32_t x0 = x0_qn;
int olen = oproc_unit;
for (int op = 0; op < out_length; op += olen) {
- const uint8_t *srcp =
- src + get_upscale_convolve_params(in_length, out_length, op, &x0_qn,
- &x_step_qn);
+ // Note since we are upscaling, the first output sample is located before
+ // the first input sample. Hence srcp = src - 1 + ... below
olen = AOMMIN(oproc_unit, out_length - op);
- convolve_horiz_superres_c(srcp, dst + op, interp_filters, interp_taps,
- x0_qn, x_step_qn, olen);
+ convolve_horiz_superres_c(srcp, dst + op, interp_filters, interp_taps, x0,
+ x_step_qn, olen);
+ x0 += olen * x_step_qn;
+ // Note srcp may advance by UPSCALE_PROC_UNIT +/- 1
+ srcp += (x0 >> INTERP_PRECISION_BITS);
+ x0 &= INTERP_PRECISION_MASK;
}
}
@@ -507,7 +515,7 @@
uint8_t *output, int out_length,
int superres_denom) {
uint8_t *intbuf_alloc = (uint8_t *)aom_malloc(
- sizeof(uint8_t) * (in_length + UPSCALE_NORMATIVE_TAPS + 2));
+ sizeof(*input) * (in_length + UPSCALE_NORMATIVE_TAPS + 2));
uint8_t *intbuf = intbuf_alloc + UPSCALE_NORMATIVE_TAPS / 2 + 1;
memcpy(intbuf, input, sizeof(*intbuf) * in_length);
for (int k = 0; k < UPSCALE_NORMATIVE_TAPS / 2 + 1; ++k) {
@@ -903,16 +911,24 @@
int superres_denom, int bd,
const int16_t *interp_filters,
int interp_taps) {
+ assert(in_length < out_length);
const int oproc_unit = UPSCALE_PROC_UNIT_SCALE * superres_denom;
- int32_t x_step_qn, x0_qn;
+ const int32_t x_step_qn = get_upscale_convolve_step(in_length, out_length);
+ const int32_t x0_qn =
+ get_upscale_convolve_x0(in_length, out_length, x_step_qn);
+ const uint16_t *srcp = src - 1;
+ int32_t x0 = x0_qn;
int olen = oproc_unit;
for (int op = 0; op < out_length; op += olen) {
- const uint16_t *srcp =
- src + get_upscale_convolve_params(in_length, out_length, op, &x0_qn,
- &x_step_qn);
+ // Note since we are upscaling, the first output sample is located before
+ // the first input sample. Hence srcp = src - 1 + ... below
olen = AOMMIN(oproc_unit, out_length - op);
highbd_convolve_horiz_superres_c(srcp, dst + op, interp_filters,
- interp_taps, x0_qn, x_step_qn, olen, bd);
+ interp_taps, x0, x_step_qn, olen, bd);
+ x0 += olen * x_step_qn;
+ // Note srcp may advance by UPSCALE_PROC_UNIT +/- 1
+ srcp += (x0 >> INTERP_PRECISION_BITS);
+ x0 &= INTERP_PRECISION_MASK;
}
}
@@ -921,10 +937,10 @@
int out_length, int bd,
int superres_denom) {
uint16_t *intbuf_alloc = (uint16_t *)aom_malloc(
- sizeof(uint16_t) * (in_length + UPSCALE_NORMATIVE_TAPS));
- uint16_t *intbuf = intbuf_alloc + UPSCALE_NORMATIVE_TAPS / 2;
+ sizeof(*input) * (in_length + UPSCALE_NORMATIVE_TAPS + 2));
+ uint16_t *intbuf = intbuf_alloc + UPSCALE_NORMATIVE_TAPS / 2 + 1;
memcpy(intbuf, input, sizeof(*intbuf) * in_length);
- for (int k = 0; k < UPSCALE_NORMATIVE_TAPS / 2; ++k) {
+ for (int k = 0; k < UPSCALE_NORMATIVE_TAPS / 2 + 1; ++k) {
intbuf[-k - 1] = intbuf[0];
intbuf[in_length + k] = intbuf[in_length - 1];
}
@@ -1366,27 +1382,28 @@
// Calculates scaled dimensions given original dimensions and the scale
// denominator. If 'scale_height' is 1, both width and height are scaled;
// otherwise, only the width is scaled.
-static void calculate_scaled_size_helper(int *width, int *height, int denom,
- int scale_height) {
+static void calculate_scaled_size_helper(int *dim, int denom) {
if (denom != SCALE_NUMERATOR) {
- *width = *width * SCALE_NUMERATOR / denom;
- *width += *width & 1; // Make it even.
- if (scale_height) {
- *height = *height * SCALE_NUMERATOR / denom;
- *height += *height & 1; // Make it even.
- }
+ // Use this version if we need *dim to be even
+ // *width = (*width * SCALE_NUMERATOR + denom) / (2 * denom);
+ // *width <<= 1;
+ *dim = (*dim * SCALE_NUMERATOR + denom / 2) / (denom);
}
}
void av1_calculate_scaled_size(int *width, int *height, int resize_denom) {
- calculate_scaled_size_helper(width, height, resize_denom, 1);
+ calculate_scaled_size_helper(width, resize_denom);
+ calculate_scaled_size_helper(height, resize_denom);
}
#if CONFIG_FRAME_SUPERRES
void av1_calculate_scaled_superres_size(int *width, int *height,
int superres_denom) {
- calculate_scaled_size_helper(width, height, superres_denom,
- !CONFIG_HORZONLY_FRAME_SUPERRES);
+ (void)height;
+ calculate_scaled_size_helper(width, superres_denom);
+#if !CONFIG_HORZONLY_FRAME_SUPERRES
+ calculate_scaled_size_helper(height, superres_denom);
+#endif // !CONFIG_HORZONLY_FRAME_SUPERRES
}
void av1_calculate_unscaled_superres_size(int *width, int *height, int denom) {
