From df9477dfa60ebb5d31bc142e58ce46535c17abce Mon Sep 17 00:00:00 2001 From: trav90 Date: Wed, 17 Oct 2018 05:59:08 -0500 Subject: Update aom to slightly newer commit ID --- third_party/aom/aom_dsp/aom_convolve.c | 267 ++++++++++++++++++++++++++++++++- 1 file changed, 259 insertions(+), 8 deletions(-) (limited to 'third_party/aom/aom_dsp/aom_convolve.c') diff --git a/third_party/aom/aom_dsp/aom_convolve.c b/third_party/aom/aom_dsp/aom_convolve.c index 74f4c00fb..4dac6aacc 100644 --- a/third_party/aom/aom_dsp/aom_convolve.c +++ b/third_party/aom/aom_dsp/aom_convolve.c @@ -337,14 +337,14 @@ static void convolve_add_src_horiz(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *x_filters, int x0_q4, int x_step_q4, int w, int h) { - int x, y; + int x, y, k; src -= SUBPEL_TAPS / 2 - 1; for (y = 0; y < h; ++y) { int x_q4 = x0_q4; for (x = 0; x < w; ++x) { const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; - int k, sum = 0; + int sum = 0; for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k]; dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS) + src_x[SUBPEL_TAPS / 2 - 1]); @@ -359,7 +359,7 @@ static void convolve_add_src_vert(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *y_filters, int y0_q4, int y_step_q4, int w, int h) { - int x, y; + int x, y, k; src -= src_stride * (SUBPEL_TAPS / 2 - 1); for (x = 0; x < w; ++x) { @@ -367,7 +367,7 @@ static void convolve_add_src_vert(const uint8_t *src, ptrdiff_t src_stride, for (y = 0; y < h; ++y) { const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; - int k, sum = 0; + int sum = 0; for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_y[k * src_stride] * y_filter[k]; dst[y * dst_stride] = @@ -446,6 +446,127 @@ void aom_convolve8_add_src_c(const uint8_t *src, ptrdiff_t src_stride, convolve_add_src(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4, filters_y, y0_q4, y_step_q4, w, h); } + +static void convolve_add_src_horiz_hip(const uint8_t *src, ptrdiff_t src_stride, + uint16_t *dst, ptrdiff_t dst_stride, + const InterpKernel *x_filters, int x0_q4, + int x_step_q4, int w, int h) { + const int bd = 8; + int x, y, k; + src -= SUBPEL_TAPS / 2 - 1; + for (y = 0; y < h; ++y) { + int x_q4 = x0_q4; + for (x = 0; x < w; ++x) { + const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; + const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; + int sum = ((int)src_x[SUBPEL_TAPS / 2 - 1] << FILTER_BITS) + + (1 << (bd + FILTER_BITS - 1)); + for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k]; + dst[x] = + (uint16_t)clamp(ROUND_POWER_OF_TWO(sum, FILTER_BITS - EXTRAPREC_BITS), + 0, EXTRAPREC_CLAMP_LIMIT(bd) - 1); + x_q4 += x_step_q4; + } + src += src_stride; + dst += dst_stride; + } +} + +static void convolve_add_src_vert_hip(const uint16_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *y_filters, int y0_q4, + int y_step_q4, int w, int h) { + const int bd = 8; + int x, y, k; + src -= src_stride * (SUBPEL_TAPS / 2 - 1); + + for (x = 0; x < w; ++x) { + int y_q4 = y0_q4; + for (y = 0; y < h; ++y) { + const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; + const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; + int sum = + ((int)src_y[(SUBPEL_TAPS / 2 - 1) * src_stride] << FILTER_BITS) - + (1 << (bd + FILTER_BITS + EXTRAPREC_BITS - 1)); + for (k = 0; k < SUBPEL_TAPS; ++k) + sum += src_y[k * src_stride] * y_filter[k]; + dst[y * dst_stride] = + clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS + EXTRAPREC_BITS)); + y_q4 += y_step_q4; + } + ++src; + ++dst; + } +} + +static void convolve_add_src_hip(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *const x_filters, int x0_q4, + int x_step_q4, + const InterpKernel *const y_filters, int y0_q4, + int y_step_q4, int w, int h) { + uint16_t temp[MAX_EXT_SIZE * MAX_SB_SIZE]; + int intermediate_height = + (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS; + + assert(w <= MAX_SB_SIZE); + assert(h <= MAX_SB_SIZE); + + assert(y_step_q4 <= 32); + assert(x_step_q4 <= 32); + + convolve_add_src_horiz_hip(src - src_stride * (SUBPEL_TAPS / 2 - 1), + src_stride, temp, MAX_SB_SIZE, x_filters, x0_q4, + x_step_q4, w, intermediate_height); + convolve_add_src_vert_hip(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1), + MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4, + y_step_q4, w, h); +} + +void aom_convolve8_add_src_horiz_hip_c(const uint8_t *src, ptrdiff_t src_stride, + uint16_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, + int w, int h) { + const InterpKernel *const filters_x = get_filter_base(filter_x); + const int x0_q4 = get_filter_offset(filter_x, filters_x); + + (void)filter_y; + (void)y_step_q4; + + convolve_add_src_horiz_hip(src, src_stride, dst, dst_stride, filters_x, x0_q4, + x_step_q4, w, h); +} + +void aom_convolve8_add_src_vert_hip_c(const uint16_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, + int w, int h) { + const InterpKernel *const filters_y = get_filter_base(filter_y); + const int y0_q4 = get_filter_offset(filter_y, filters_y); + + (void)filter_x; + (void)x_step_q4; + + convolve_add_src_vert_hip(src, src_stride, dst, dst_stride, filters_y, y0_q4, + y_step_q4, w, h); +} + +void aom_convolve8_add_src_hip_c(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, + int h) { + const InterpKernel *const filters_x = get_filter_base(filter_x); + const int x0_q4 = get_filter_offset(filter_x, filters_x); + + const InterpKernel *const filters_y = get_filter_base(filter_y); + const int y0_q4 = get_filter_offset(filter_y, filters_y); + + convolve_add_src_hip(src, src_stride, dst, dst_stride, filters_x, x0_q4, + x_step_q4, filters_y, y0_q4, y_step_q4, w, h); +} #endif // CONFIG_LOOP_RESTORATION #if CONFIG_HIGHBITDEPTH @@ -721,7 +842,7 @@ static void highbd_convolve_add_src_horiz(const uint8_t *src8, const InterpKernel *x_filters, int x0_q4, int x_step_q4, int w, int h, int bd) { - int x, y; + int x, y, k; uint16_t *src = CONVERT_TO_SHORTPTR(src8); uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); src -= SUBPEL_TAPS / 2 - 1; @@ -730,7 +851,7 @@ static void highbd_convolve_add_src_horiz(const uint8_t *src8, for (x = 0; x < w; ++x) { const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; - int k, sum = 0; + int sum = 0; for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k]; dst[x] = clip_pixel_highbd( ROUND_POWER_OF_TWO(sum, FILTER_BITS) + src_x[SUBPEL_TAPS / 2 - 1], @@ -748,7 +869,7 @@ static void highbd_convolve_add_src_vert(const uint8_t *src8, const InterpKernel *y_filters, int y0_q4, int y_step_q4, int w, int h, int bd) { - int x, y; + int x, y, k; uint16_t *src = CONVERT_TO_SHORTPTR(src8); uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); src -= src_stride * (SUBPEL_TAPS / 2 - 1); @@ -757,7 +878,7 @@ static void highbd_convolve_add_src_vert(const uint8_t *src8, for (y = 0; y < h; ++y) { const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; - int k, sum = 0; + int sum = 0; for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_y[k * src_stride] * y_filter[k]; dst[y * dst_stride] = @@ -850,5 +971,135 @@ void aom_highbd_convolve8_add_src_c(const uint8_t *src, ptrdiff_t src_stride, highbd_convolve_add_src(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4, filters_y, y0_q4, y_step_q4, w, h, bd); } + +static void highbd_convolve_add_src_horiz_hip( + const uint8_t *src8, ptrdiff_t src_stride, uint16_t *dst, + ptrdiff_t dst_stride, const InterpKernel *x_filters, int x0_q4, + int x_step_q4, int w, int h, int bd) { + const int extraprec_clamp_limit = EXTRAPREC_CLAMP_LIMIT(bd); + int x, y, k; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + src -= SUBPEL_TAPS / 2 - 1; + for (y = 0; y < h; ++y) { + int x_q4 = x0_q4; + for (x = 0; x < w; ++x) { + const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; + const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; + int sum = ((int)src_x[SUBPEL_TAPS / 2 - 1] << FILTER_BITS) + + (1 << (bd + FILTER_BITS - 1)); + for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k]; + dst[x] = + (uint16_t)clamp(ROUND_POWER_OF_TWO(sum, FILTER_BITS - EXTRAPREC_BITS), + 0, extraprec_clamp_limit - 1); + x_q4 += x_step_q4; + } + src += src_stride; + dst += dst_stride; + } +} + +static void highbd_convolve_add_src_vert_hip( + const uint16_t *src, ptrdiff_t src_stride, uint8_t *dst8, + ptrdiff_t dst_stride, const InterpKernel *y_filters, int y0_q4, + int y_step_q4, int w, int h, int bd) { + int x, y, k; + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); + src -= src_stride * (SUBPEL_TAPS / 2 - 1); + for (x = 0; x < w; ++x) { + int y_q4 = y0_q4; + for (y = 0; y < h; ++y) { + const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; + const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; + int sum = + ((int)src_y[(SUBPEL_TAPS / 2 - 1) * src_stride] << FILTER_BITS) - + (1 << (bd + FILTER_BITS + EXTRAPREC_BITS - 1)); + for (k = 0; k < SUBPEL_TAPS; ++k) + sum += src_y[k * src_stride] * y_filter[k]; + dst[y * dst_stride] = clip_pixel_highbd( + ROUND_POWER_OF_TWO(sum, FILTER_BITS + EXTRAPREC_BITS), bd); + y_q4 += y_step_q4; + } + ++src; + ++dst; + } +} + +static void highbd_convolve_add_src_hip( + const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, + ptrdiff_t dst_stride, const InterpKernel *const x_filters, int x0_q4, + int x_step_q4, const InterpKernel *const y_filters, int y0_q4, + int y_step_q4, int w, int h, int bd) { + // Note: Fixed size intermediate buffer, temp, places limits on parameters. + // 2d filtering proceeds in 2 steps: + // (1) Interpolate horizontally into an intermediate buffer, temp. + // (2) Interpolate temp vertically to derive the sub-pixel result. + // Deriving the maximum number of rows in the temp buffer (135): + // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative). + // --Largest block size is 64x64 pixels. + // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the + // original frame (in 1/16th pixel units). + // --Must round-up because block may be located at sub-pixel position. + // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails. + // --((64 - 1) * 32 + 15) >> 4 + 8 = 135. + uint16_t temp[MAX_EXT_SIZE * MAX_SB_SIZE]; + int intermediate_height = + (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS; + + assert(w <= MAX_SB_SIZE); + assert(h <= MAX_SB_SIZE); + assert(y_step_q4 <= 32); + assert(x_step_q4 <= 32); + + highbd_convolve_add_src_horiz_hip( + src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp, MAX_SB_SIZE, + x_filters, x0_q4, x_step_q4, w, intermediate_height, bd); + highbd_convolve_add_src_vert_hip(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1), + MAX_SB_SIZE, dst, dst_stride, y_filters, + y0_q4, y_step_q4, w, h, bd); +} + +void aom_highbd_convolve8_add_src_horiz_hip_c( + const uint8_t *src, ptrdiff_t src_stride, uint16_t *dst, + ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, int h, int bd) { + const InterpKernel *const filters_x = get_filter_base(filter_x); + const int x0_q4 = get_filter_offset(filter_x, filters_x); + (void)filter_y; + (void)y_step_q4; + + highbd_convolve_add_src_horiz_hip(src, src_stride, dst, dst_stride, filters_x, + x0_q4, x_step_q4, w, h, bd); +} + +void aom_highbd_convolve8_add_src_vert_hip_c( + const uint16_t *src, ptrdiff_t src_stride, uint8_t *dst, + ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, int h, int bd) { + const InterpKernel *const filters_y = get_filter_base(filter_y); + const int y0_q4 = get_filter_offset(filter_y, filters_y); + (void)filter_x; + (void)x_step_q4; + + highbd_convolve_add_src_vert_hip(src, src_stride, dst, dst_stride, filters_y, + y0_q4, y_step_q4, w, h, bd); +} + +void aom_highbd_convolve8_add_src_hip_c(const uint8_t *src, + ptrdiff_t src_stride, uint8_t *dst, + ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, + int w, int h, int bd) { + const InterpKernel *const filters_x = get_filter_base(filter_x); + const int x0_q4 = get_filter_offset(filter_x, filters_x); + + const InterpKernel *const filters_y = get_filter_base(filter_y); + const int y0_q4 = get_filter_offset(filter_y, filters_y); + + highbd_convolve_add_src_hip(src, src_stride, dst, dst_stride, filters_x, + x0_q4, x_step_q4, filters_y, y0_q4, y_step_q4, w, + h, bd); +} + #endif // CONFIG_LOOP_RESTORATION #endif // CONFIG_HIGHBITDEPTH -- cgit v1.2.3