ref: 976ddb61d34afc81cc65a56b91a9506ee28f47ef
parent: 83ba1880bf6596c922e67f99290a30b758b01379
author: Gabriel Marin <gmx@chromium.org>
date: Wed Dec 14 07:07:34 EST 2016
Add a vector form of routine vp9_model_rd_from_var_lapndz Add routine vp9_model_rd_from_var_lapndz_vec and call it from model_rd_for_sb to model the rate and distortion for MAX_MB_PLANE Laplacian sources in parallel. The caller ensures that all sources have non-zero variance. Measured a 18% to 25% reduction in retired instructions, and 17% to 24% reduction in instruction execution cost with different compilers for the Laplacian modeling. No change in behavior. TEST=Verified that encoded files match bit for bit, with and without this change. BUG=b/33678225 Change-Id: I6b76947f21c659a349adb896e13e99f6e3f951e6
--- a/vp9/encoder/vp9_rd.c
+++ b/vp9/encoder/vp9_rd.c
@@ -312,63 +312,62 @@
}
}
-static void model_rd_norm(int xsq_q10, int *r_q10, int *d_q10) {- // NOTE: The tables below must be of the same size.
+// NOTE: The tables below must be of the same size.
- // The functions described below are sampled at the four most significant
- // bits of x^2 + 8 / 256.
+// The functions described below are sampled at the four most significant
+// bits of x^2 + 8 / 256.
- // Normalized rate:
- // This table models the rate for a Laplacian source with given variance
- // when quantized with a uniform quantizer with given stepsize. The
- // closed form expression is:
- // Rn(x) = H(sqrt(r)) + sqrt(r)*[1 + H(r)/(1 - r)],
- // where r = exp(-sqrt(2) * x) and x = qpstep / sqrt(variance),
- // and H(x) is the binary entropy function.
- static const int rate_tab_q10[] = {- 65536, 6086, 5574, 5275, 5063, 4899, 4764, 4651, 4553, 4389, 4255, 4142,
- 4044, 3958, 3881, 3811, 3748, 3635, 3538, 3453, 3376, 3307, 3244, 3186,
- 3133, 3037, 2952, 2877, 2809, 2747, 2690, 2638, 2589, 2501, 2423, 2353,
- 2290, 2232, 2179, 2130, 2084, 2001, 1928, 1862, 1802, 1748, 1698, 1651,
- 1608, 1530, 1460, 1398, 1342, 1290, 1243, 1199, 1159, 1086, 1021, 963,
- 911, 864, 821, 781, 745, 680, 623, 574, 530, 490, 455, 424,
- 395, 345, 304, 269, 239, 213, 190, 171, 154, 126, 104, 87,
- 73, 61, 52, 44, 38, 28, 21, 16, 12, 10, 8, 6,
- 5, 3, 2, 1, 1, 1, 0, 0,
- };
+// Normalized rate:
+// This table models the rate for a Laplacian source with given variance
+// when quantized with a uniform quantizer with given stepsize. The
+// closed form expression is:
+// Rn(x) = H(sqrt(r)) + sqrt(r)*[1 + H(r)/(1 - r)],
+// where r = exp(-sqrt(2) * x) and x = qpstep / sqrt(variance),
+// and H(x) is the binary entropy function.
+static const int rate_tab_q10[] = {+ 65536, 6086, 5574, 5275, 5063, 4899, 4764, 4651, 4553, 4389, 4255, 4142, 4044,
+ 3958, 3881, 3811, 3748, 3635, 3538, 3453, 3376, 3307, 3244, 3186, 3133, 3037,
+ 2952, 2877, 2809, 2747, 2690, 2638, 2589, 2501, 2423, 2353, 2290, 2232, 2179,
+ 2130, 2084, 2001, 1928, 1862, 1802, 1748, 1698, 1651, 1608, 1530, 1460, 1398,
+ 1342, 1290, 1243, 1199, 1159, 1086, 1021, 963, 911, 864, 821, 781, 745,
+ 680, 623, 574, 530, 490, 455, 424, 395, 345, 304, 269, 239, 213,
+ 190, 171, 154, 126, 104, 87, 73, 61, 52, 44, 38, 28, 21,
+ 16, 12, 10, 8, 6, 5, 3, 2, 1, 1, 1, 0, 0,
+};
- // Normalized distortion:
- // This table models the normalized distortion for a Laplacian source
- // with given variance when quantized with a uniform quantizer
- // with given stepsize. The closed form expression is:
- // Dn(x) = 1 - 1/sqrt(2) * x / sinh(x/sqrt(2))
- // where x = qpstep / sqrt(variance).
- // Note the actual distortion is Dn * variance.
- static const int dist_tab_q10[] = {- 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 4, 5,
- 5, 6, 7, 7, 8, 9, 11, 12, 13, 15, 16, 17,
- 18, 21, 24, 26, 29, 31, 34, 36, 39, 44, 49, 54,
- 59, 64, 69, 73, 78, 88, 97, 106, 115, 124, 133, 142,
- 151, 167, 184, 200, 215, 231, 245, 260, 274, 301, 327, 351,
- 375, 397, 418, 439, 458, 495, 528, 559, 587, 613, 637, 659,
- 680, 717, 749, 777, 801, 823, 842, 859, 874, 899, 919, 936,
- 949, 960, 969, 977, 983, 994, 1001, 1006, 1010, 1013, 1015, 1017,
- 1018, 1020, 1022, 1022, 1023, 1023, 1023, 1024,
- };
- static const int xsq_iq_q10[] = {- 0, 4, 8, 12, 16, 20, 24, 28, 32,
- 40, 48, 56, 64, 72, 80, 88, 96, 112,
- 128, 144, 160, 176, 192, 208, 224, 256, 288,
- 320, 352, 384, 416, 448, 480, 544, 608, 672,
- 736, 800, 864, 928, 992, 1120, 1248, 1376, 1504,
- 1632, 1760, 1888, 2016, 2272, 2528, 2784, 3040, 3296,
- 3552, 3808, 4064, 4576, 5088, 5600, 6112, 6624, 7136,
- 7648, 8160, 9184, 10208, 11232, 12256, 13280, 14304, 15328,
- 16352, 18400, 20448, 22496, 24544, 26592, 28640, 30688, 32736,
- 36832, 40928, 45024, 49120, 53216, 57312, 61408, 65504, 73696,
- 81888, 90080, 98272, 106464, 114656, 122848, 131040, 147424, 163808,
- 180192, 196576, 212960, 229344, 245728,
- };
+// Normalized distortion:
+// This table models the normalized distortion for a Laplacian source
+// with given variance when quantized with a uniform quantizer
+// with given stepsize. The closed form expression is:
+// Dn(x) = 1 - 1/sqrt(2) * x / sinh(x/sqrt(2))
+// where x = qpstep / sqrt(variance).
+// Note the actual distortion is Dn * variance.
+static const int dist_tab_q10[] = {+ 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 4, 5, 5,
+ 6, 7, 7, 8, 9, 11, 12, 13, 15, 16, 17, 18, 21,
+ 24, 26, 29, 31, 34, 36, 39, 44, 49, 54, 59, 64, 69,
+ 73, 78, 88, 97, 106, 115, 124, 133, 142, 151, 167, 184, 200,
+ 215, 231, 245, 260, 274, 301, 327, 351, 375, 397, 418, 439, 458,
+ 495, 528, 559, 587, 613, 637, 659, 680, 717, 749, 777, 801, 823,
+ 842, 859, 874, 899, 919, 936, 949, 960, 969, 977, 983, 994, 1001,
+ 1006, 1010, 1013, 1015, 1017, 1018, 1020, 1022, 1022, 1023, 1023, 1023, 1024,
+};
+static const int xsq_iq_q10[] = {+ 0, 4, 8, 12, 16, 20, 24, 28, 32,
+ 40, 48, 56, 64, 72, 80, 88, 96, 112,
+ 128, 144, 160, 176, 192, 208, 224, 256, 288,
+ 320, 352, 384, 416, 448, 480, 544, 608, 672,
+ 736, 800, 864, 928, 992, 1120, 1248, 1376, 1504,
+ 1632, 1760, 1888, 2016, 2272, 2528, 2784, 3040, 3296,
+ 3552, 3808, 4064, 4576, 5088, 5600, 6112, 6624, 7136,
+ 7648, 8160, 9184, 10208, 11232, 12256, 13280, 14304, 15328,
+ 16352, 18400, 20448, 22496, 24544, 26592, 28640, 30688, 32736,
+ 36832, 40928, 45024, 49120, 53216, 57312, 61408, 65504, 73696,
+ 81888, 90080, 98272, 106464, 114656, 122848, 131040, 147424, 163808,
+ 180192, 196576, 212960, 229344, 245728,
+};
+
+static void model_rd_norm(int xsq_q10, int *r_q10, int *d_q10) {const int tmp = (xsq_q10 >> 2) + 8;
const int k = get_msb(tmp) - 3;
const int xq = (k << 3) + ((tmp >> k) & 0x7);
@@ -379,6 +378,24 @@
*d_q10 = (dist_tab_q10[xq] * b_q10 + dist_tab_q10[xq + 1] * a_q10) >> 10;
}
+static void model_rd_norm_vec(int xsq_q10[MAX_MB_PLANE],
+ int r_q10[MAX_MB_PLANE],
+ int d_q10[MAX_MB_PLANE]) {+ int i;
+ const int one_q10 = 1 << 10;
+ for (i = 0; i < MAX_MB_PLANE; ++i) {+ const int tmp = (xsq_q10[i] >> 2) + 8;
+ const int k = get_msb(tmp) - 3;
+ const int xq = (k << 3) + ((tmp >> k) & 0x7);
+ const int a_q10 = ((xsq_q10[i] - xsq_iq_q10[xq]) << 10) >> (2 + k);
+ const int b_q10 = one_q10 - a_q10;
+ r_q10[i] = (rate_tab_q10[xq] * b_q10 + rate_tab_q10[xq + 1] * a_q10) >> 10;
+ d_q10[i] = (dist_tab_q10[xq] * b_q10 + dist_tab_q10[xq + 1] * a_q10) >> 10;
+ }
+}
+
+static const uint32_t MAX_XSQ_Q10 = 245727;
+
void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n_log2,
unsigned int qstep, int *rate,
int64_t *dist) {@@ -393,7 +410,6 @@
*dist = 0;
} else {int d_q10, r_q10;
- static const uint32_t MAX_XSQ_Q10 = 245727;
const uint64_t xsq_q10_64 =
(((uint64_t)qstep * qstep << (n_log2 + 10)) + (var >> 1)) / var;
const int xsq_q10 = (int)VPXMIN(xsq_q10_64, MAX_XSQ_Q10);
@@ -400,6 +416,30 @@
model_rd_norm(xsq_q10, &r_q10, &d_q10);
*rate = ROUND_POWER_OF_TWO(r_q10 << n_log2, 10 - VP9_PROB_COST_SHIFT);
*dist = (var * (int64_t)d_q10 + 512) >> 10;
+ }
+}
+
+// Implements a fixed length vector form of vp9_model_rd_from_var_lapndz where
+// vectors are of length MAX_MB_PLANE and all elements of var are non-zero.
+void vp9_model_rd_from_var_lapndz_vec(unsigned int var[MAX_MB_PLANE],
+ unsigned int n_log2[MAX_MB_PLANE],
+ unsigned int qstep[MAX_MB_PLANE],
+ int64_t *rate_sum, int64_t *dist_sum) {+ int i;
+ int xsq_q10[MAX_MB_PLANE], d_q10[MAX_MB_PLANE], r_q10[MAX_MB_PLANE];
+ for (i = 0; i < MAX_MB_PLANE; ++i) {+ const uint64_t xsq_q10_64 =
+ (((uint64_t)qstep[i] * qstep[i] << (n_log2[i] + 10)) + (var[i] >> 1)) /
+ var[i];
+ xsq_q10[i] = (int)VPXMIN(xsq_q10_64, MAX_XSQ_Q10);
+ }
+ model_rd_norm_vec(xsq_q10, r_q10, d_q10);
+ for (i = 0; i < MAX_MB_PLANE; ++i) {+ int rate =
+ ROUND_POWER_OF_TWO(r_q10[i] << n_log2[i], 10 - VP9_PROB_COST_SHIFT);
+ int64_t dist = (var[i] * (int64_t)d_q10[i] + 512) >> 10;
+ *rate_sum += rate;
+ *dist_sum += dist;
}
}
--- a/vp9/encoder/vp9_rd.h
+++ b/vp9/encoder/vp9_rd.h
@@ -140,6 +140,11 @@
void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n,
unsigned int qstep, int *rate, int64_t *dist);
+void vp9_model_rd_from_var_lapndz_vec(unsigned int var[MAX_MB_PLANE],
+ unsigned int n_log2[MAX_MB_PLANE],
+ unsigned int qstep[MAX_MB_PLANE],
+ int64_t *rate_sum, int64_t *dist_sum);
+
int vp9_get_switchable_rate(const struct VP9_COMP *cpi,
const MACROBLOCKD *const xd);
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -164,11 +164,9 @@
const int ref = xd->mi[0]->ref_frame[0];
unsigned int sse;
unsigned int var = 0;
- unsigned int sum_sse = 0;
int64_t total_sse = 0;
int skip_flag = 1;
const int shift = 6;
- int rate;
int64_t dist;
const int dequant_shift =
#if CONFIG_VP9_HIGHBITDEPTH
@@ -175,6 +173,10 @@
(xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd - 5 :
#endif // CONFIG_VP9_HIGHBITDEPTH
3;
+ unsigned int qstep_vec[MAX_MB_PLANE];
+ unsigned int nlog2_vec[MAX_MB_PLANE];
+ unsigned int sum_sse_vec[MAX_MB_PLANE];
+ int any_zero_sum_sse = 0;
x->pred_sse[ref] = 0;
@@ -186,6 +188,7 @@
const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
const int64_t dc_thr = p->quant_thred[0] >> shift;
const int64_t ac_thr = p->quant_thred[1] >> shift;
+ unsigned int sum_sse = 0;
// The low thresholds are used to measure if the prediction errors are
// low enough so that we can skip the mode search.
const int64_t low_dc_thr = VPXMIN(50, dc_thr >> 2);
@@ -196,8 +199,6 @@
int lw = b_width_log2_lookup[unit_size] + 2;
int lh = b_height_log2_lookup[unit_size] + 2;
- sum_sse = 0;
-
for (idy = 0; idy < bh; ++idy) { for (idx = 0; idx < bw; ++idx) {uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
@@ -233,12 +234,18 @@
}
total_sse += sum_sse;
+ sum_sse_vec[i] = sum_sse;
+ any_zero_sum_sse = any_zero_sum_sse || (sum_sse == 0);
+ qstep_vec[i] = pd->dequant[1] >> dequant_shift;
+ nlog2_vec[i] = num_pels_log2_lookup[bs];
+ }
- // Fast approximate the modelling function.
- if (cpi->sf.simple_model_rd_from_var) {+ // Fast approximate the modelling function.
+ if (cpi->sf.simple_model_rd_from_var) {+ for (i = 0; i < MAX_MB_PLANE; ++i) {int64_t rate;
- const int64_t square_error = sum_sse;
- int quantizer = (pd->dequant[1] >> dequant_shift);
+ const int64_t square_error = sum_sse_vec[i];
+ int quantizer = qstep_vec[i];
if (quantizer < 120)
rate = (square_error * (280 - quantizer)) >> (16 - VP9_PROB_COST_SHIFT);
@@ -247,12 +254,19 @@
dist = (square_error * quantizer) >> 8;
rate_sum += rate;
dist_sum += dist;
+ }
+ } else {+ if (any_zero_sum_sse) {+ for (i = 0; i < MAX_MB_PLANE; ++i) {+ int rate;
+ vp9_model_rd_from_var_lapndz(sum_sse_vec[i], nlog2_vec[i], qstep_vec[i],
+ &rate, &dist);
+ rate_sum += rate;
+ dist_sum += dist;
+ }
} else {- vp9_model_rd_from_var_lapndz(sum_sse, num_pels_log2_lookup[bs],
- pd->dequant[1] >> dequant_shift, &rate,
- &dist);
- rate_sum += rate;
- dist_sum += dist;
+ vp9_model_rd_from_var_lapndz_vec(sum_sse_vec, nlog2_vec, qstep_vec,
+ &rate_sum, &dist_sum);
}
}