ref: bc1c18e18c6001dc3db233c4dcb34c4e54923ce7
parent: 0fefc6873a7f11032e52900aa1d8adf9ad8dbeed
author: Linfeng Zhang <linfengz@google.com>
date: Tue Feb 7 05:06:21 EST 2017
Add vpx_idct16x16_38_add_neon() The RunQuantCheck() test on it exposes 16-bit overflow in stage 7 of pass 2. Change to use saturating add/sub for both vpx_idct16x16_38_add_neon() and vpx_idct16x16_256_add_neon() for high bitdepth. Change-Id: Ibf4c107a887553a52852cc582e28d38a5a5a2712
--- a/test/partial_idct_test.cc
+++ b/test/partial_idct_test.cc
@@ -501,6 +501,8 @@
make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_256_add_c>,
&wrapper<vpx_idct16x16_256_add_neon>, TX_16X16, 256, 8, 1),
make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_256_add_c>,
+ &wrapper<vpx_idct16x16_38_add_neon>, TX_16X16, 38, 8, 1),
+ make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_256_add_c>,
&wrapper<vpx_idct16x16_10_add_neon>, TX_16X16, 10, 8, 1),
make_tuple(&vpx_fdct16x16_c, &wrapper<vpx_idct16x16_256_add_c>,
&wrapper<vpx_idct16x16_1_add_neon>, TX_16X16, 1, 8, 1),
--- a/vpx_dsp/arm/idct16x16_add_neon.c
+++ b/vpx_dsp/arm/idct16x16_add_neon.c
@@ -158,8 +158,105 @@
wrap_low_4x2(t32, d0, d1);
}
+static INLINE void idct16x16_add_stage7(const int16x8_t *const step2,
+ int16x8_t *const out) {+#if CONFIG_VP9_HIGHBITDEPTH
+ // Use saturating add/sub to avoid overflow in 2nd pass
+ out[0] = vqaddq_s16(step2[0], step2[15]);
+ out[1] = vqaddq_s16(step2[1], step2[14]);
+ out[2] = vqaddq_s16(step2[2], step2[13]);
+ out[3] = vqaddq_s16(step2[3], step2[12]);
+ out[4] = vqaddq_s16(step2[4], step2[11]);
+ out[5] = vqaddq_s16(step2[5], step2[10]);
+ out[6] = vqaddq_s16(step2[6], step2[9]);
+ out[7] = vqaddq_s16(step2[7], step2[8]);
+ out[8] = vqsubq_s16(step2[7], step2[8]);
+ out[9] = vqsubq_s16(step2[6], step2[9]);
+ out[10] = vqsubq_s16(step2[5], step2[10]);
+ out[11] = vqsubq_s16(step2[4], step2[11]);
+ out[12] = vqsubq_s16(step2[3], step2[12]);
+ out[13] = vqsubq_s16(step2[2], step2[13]);
+ out[14] = vqsubq_s16(step2[1], step2[14]);
+ out[15] = vqsubq_s16(step2[0], step2[15]);
+#else
+ out[0] = vaddq_s16(step2[0], step2[15]);
+ out[1] = vaddq_s16(step2[1], step2[14]);
+ out[2] = vaddq_s16(step2[2], step2[13]);
+ out[3] = vaddq_s16(step2[3], step2[12]);
+ out[4] = vaddq_s16(step2[4], step2[11]);
+ out[5] = vaddq_s16(step2[5], step2[10]);
+ out[6] = vaddq_s16(step2[6], step2[9]);
+ out[7] = vaddq_s16(step2[7], step2[8]);
+ out[8] = vsubq_s16(step2[7], step2[8]);
+ out[9] = vsubq_s16(step2[6], step2[9]);
+ out[10] = vsubq_s16(step2[5], step2[10]);
+ out[11] = vsubq_s16(step2[4], step2[11]);
+ out[12] = vsubq_s16(step2[3], step2[12]);
+ out[13] = vsubq_s16(step2[2], step2[13]);
+ out[14] = vsubq_s16(step2[1], step2[14]);
+ out[15] = vsubq_s16(step2[0], step2[15]);
+#endif
+}
+
+static INLINE void idct16x16_store_pass1(const int16x8_t *const out,
+ int16_t *output) {+ // Save the result into output
+ vst1q_s16(output, out[0]);
+ output += 16;
+ vst1q_s16(output, out[1]);
+ output += 16;
+ vst1q_s16(output, out[2]);
+ output += 16;
+ vst1q_s16(output, out[3]);
+ output += 16;
+ vst1q_s16(output, out[4]);
+ output += 16;
+ vst1q_s16(output, out[5]);
+ output += 16;
+ vst1q_s16(output, out[6]);
+ output += 16;
+ vst1q_s16(output, out[7]);
+ output += 16;
+ vst1q_s16(output, out[8]);
+ output += 16;
+ vst1q_s16(output, out[9]);
+ output += 16;
+ vst1q_s16(output, out[10]);
+ output += 16;
+ vst1q_s16(output, out[11]);
+ output += 16;
+ vst1q_s16(output, out[12]);
+ output += 16;
+ vst1q_s16(output, out[13]);
+ output += 16;
+ vst1q_s16(output, out[14]);
+ output += 16;
+ vst1q_s16(output, out[15]);
+}
+
+static INLINE void idct16x16_add_store(const int16x8_t *const out,
+ uint8_t *dest, const int stride) {+ // Add the result to dest
+ idct16x16_add8x1(out[0], &dest, stride);
+ idct16x16_add8x1(out[1], &dest, stride);
+ idct16x16_add8x1(out[2], &dest, stride);
+ idct16x16_add8x1(out[3], &dest, stride);
+ idct16x16_add8x1(out[4], &dest, stride);
+ idct16x16_add8x1(out[5], &dest, stride);
+ idct16x16_add8x1(out[6], &dest, stride);
+ idct16x16_add8x1(out[7], &dest, stride);
+ idct16x16_add8x1(out[8], &dest, stride);
+ idct16x16_add8x1(out[9], &dest, stride);
+ idct16x16_add8x1(out[10], &dest, stride);
+ idct16x16_add8x1(out[11], &dest, stride);
+ idct16x16_add8x1(out[12], &dest, stride);
+ idct16x16_add8x1(out[13], &dest, stride);
+ idct16x16_add8x1(out[14], &dest, stride);
+ idct16x16_add8x1(out[15], &dest, stride);
+}
+
static void idct16x16_256_add_half1d(const void *const input, int16_t *output,
- uint8_t *dest, int stride) {+ uint8_t *dest, const int stride) {const int16x8_t cospis0 = vld1q_s16(kCospi);
const int16x8_t cospis1 = vld1q_s16(kCospi + 8);
const int16x4_t cospi_0_8_16_24 = vget_low_s16(cospis0);
@@ -346,77 +443,171 @@
step2[15] = step1[15];
// stage 7
- out[0] = vaddq_s16(step2[0], step2[15]);
- out[1] = vaddq_s16(step2[1], step2[14]);
- out[2] = vaddq_s16(step2[2], step2[13]);
- out[3] = vaddq_s16(step2[3], step2[12]);
- out[4] = vaddq_s16(step2[4], step2[11]);
- out[5] = vaddq_s16(step2[5], step2[10]);
- out[6] = vaddq_s16(step2[6], step2[9]);
- out[7] = vaddq_s16(step2[7], step2[8]);
- out[8] = vsubq_s16(step2[7], step2[8]);
- out[9] = vsubq_s16(step2[6], step2[9]);
- out[10] = vsubq_s16(step2[5], step2[10]);
- out[11] = vsubq_s16(step2[4], step2[11]);
- out[12] = vsubq_s16(step2[3], step2[12]);
- out[13] = vsubq_s16(step2[2], step2[13]);
- out[14] = vsubq_s16(step2[1], step2[14]);
- out[15] = vsubq_s16(step2[0], step2[15]);
+ idct16x16_add_stage7(step2, out);
if (output) {- // pass 1: save the result into output
- vst1q_s16(output, out[0]);
- output += 16;
- vst1q_s16(output, out[1]);
- output += 16;
- vst1q_s16(output, out[2]);
- output += 16;
- vst1q_s16(output, out[3]);
- output += 16;
- vst1q_s16(output, out[4]);
- output += 16;
- vst1q_s16(output, out[5]);
- output += 16;
- vst1q_s16(output, out[6]);
- output += 16;
- vst1q_s16(output, out[7]);
- output += 16;
- vst1q_s16(output, out[8]);
- output += 16;
- vst1q_s16(output, out[9]);
- output += 16;
- vst1q_s16(output, out[10]);
- output += 16;
- vst1q_s16(output, out[11]);
- output += 16;
- vst1q_s16(output, out[12]);
- output += 16;
- vst1q_s16(output, out[13]);
- output += 16;
- vst1q_s16(output, out[14]);
- output += 16;
- vst1q_s16(output, out[15]);
+ idct16x16_store_pass1(out, output);
} else {- // pass 2: add the result to dest.
- idct16x16_add8x1(out[0], &dest, stride);
- idct16x16_add8x1(out[1], &dest, stride);
- idct16x16_add8x1(out[2], &dest, stride);
- idct16x16_add8x1(out[3], &dest, stride);
- idct16x16_add8x1(out[4], &dest, stride);
- idct16x16_add8x1(out[5], &dest, stride);
- idct16x16_add8x1(out[6], &dest, stride);
- idct16x16_add8x1(out[7], &dest, stride);
- idct16x16_add8x1(out[8], &dest, stride);
- idct16x16_add8x1(out[9], &dest, stride);
- idct16x16_add8x1(out[10], &dest, stride);
- idct16x16_add8x1(out[11], &dest, stride);
- idct16x16_add8x1(out[12], &dest, stride);
- idct16x16_add8x1(out[13], &dest, stride);
- idct16x16_add8x1(out[14], &dest, stride);
- idct16x16_add8x1(out[15], &dest, stride);
+ idct16x16_add_store(out, dest, stride);
}
}
+static void idct16x16_38_add_half1d(const void *const input, int16_t *output,
+ uint8_t *dest, const int stride) {+ const int16x8_t cospis0 = vld1q_s16(kCospi);
+ const int16x8_t cospis1 = vld1q_s16(kCospi + 8);
+ const int16x8_t cospisd0 = vaddq_s16(cospis0, cospis0);
+ const int16x8_t cospisd1 = vaddq_s16(cospis1, cospis1);
+ const int16x4_t cospi_0_8_16_24 = vget_low_s16(cospis0);
+ const int16x4_t cospid_0_8_16_24 = vget_low_s16(cospisd0);
+ const int16x4_t cospid_4_12_20N_28 = vget_high_s16(cospisd0);
+ const int16x4_t cospid_2_30_10_22 = vget_low_s16(cospisd1);
+ const int16x4_t cospid_6_26_14_18N = vget_high_s16(cospisd1);
+ int16x8_t in[8], step1[16], step2[16], out[16];
+
+ // Load input (8x8)
+ if (output) {+ const tran_low_t *inputT = (const tran_low_t *)input;
+ in[0] = load_tran_low_to_s16q(inputT);
+ inputT += 16;
+ in[1] = load_tran_low_to_s16q(inputT);
+ inputT += 16;
+ in[2] = load_tran_low_to_s16q(inputT);
+ inputT += 16;
+ in[3] = load_tran_low_to_s16q(inputT);
+ inputT += 16;
+ in[4] = load_tran_low_to_s16q(inputT);
+ inputT += 16;
+ in[5] = load_tran_low_to_s16q(inputT);
+ inputT += 16;
+ in[6] = load_tran_low_to_s16q(inputT);
+ inputT += 16;
+ in[7] = load_tran_low_to_s16q(inputT);
+ } else {+ const int16_t *inputT = (const int16_t *)input;
+ in[0] = vld1q_s16(inputT);
+ inputT += 16;
+ in[1] = vld1q_s16(inputT);
+ inputT += 16;
+ in[2] = vld1q_s16(inputT);
+ inputT += 16;
+ in[3] = vld1q_s16(inputT);
+ inputT += 16;
+ in[4] = vld1q_s16(inputT);
+ inputT += 16;
+ in[5] = vld1q_s16(inputT);
+ inputT += 16;
+ in[6] = vld1q_s16(inputT);
+ inputT += 16;
+ in[7] = vld1q_s16(inputT);
+ }
+
+ // Transpose
+ transpose_s16_8x8(&in[0], &in[1], &in[2], &in[3], &in[4], &in[5], &in[6],
+ &in[7]);
+
+ // stage 1
+ step1[0] = in[0 / 2];
+ step1[2] = in[8 / 2];
+ step1[4] = in[4 / 2];
+ step1[6] = in[12 / 2];
+ step1[8] = in[2 / 2];
+ step1[10] = in[10 / 2];
+ step1[12] = in[6 / 2];
+ step1[14] = in[14 / 2]; // 0 in pass 1
+
+ // stage 2
+ step2[0] = step1[0];
+ step2[2] = step1[2];
+ step2[4] = step1[4];
+ step2[6] = step1[6];
+ step2[8] = vqrdmulhq_lane_s16(step1[8], cospid_2_30_10_22, 1);
+ step2[9] = vqrdmulhq_lane_s16(step1[14], cospid_6_26_14_18N, 3);
+ step2[10] = vqrdmulhq_lane_s16(step1[10], cospid_2_30_10_22, 3);
+ step2[11] = vqrdmulhq_lane_s16(step1[12], cospid_6_26_14_18N, 1);
+ step2[12] = vqrdmulhq_lane_s16(step1[12], cospid_6_26_14_18N, 0);
+ step2[13] = vqrdmulhq_lane_s16(step1[10], cospid_2_30_10_22, 2);
+ step2[14] = vqrdmulhq_lane_s16(step1[14], cospid_6_26_14_18N, 2);
+ step2[15] = vqrdmulhq_lane_s16(step1[8], cospid_2_30_10_22, 0);
+
+ // stage 3
+ step1[0] = step2[0];
+ step1[2] = step2[2];
+ step1[4] = vqrdmulhq_lane_s16(step2[4], cospid_4_12_20N_28, 3);
+ step1[5] = vqrdmulhq_lane_s16(step2[6], cospid_4_12_20N_28, 2);
+ step1[6] = vqrdmulhq_lane_s16(step2[6], cospid_4_12_20N_28, 1);
+ step1[7] = vqrdmulhq_lane_s16(step2[4], cospid_4_12_20N_28, 0);
+ step1[8] = vaddq_s16(step2[8], step2[9]);
+ step1[9] = vsubq_s16(step2[8], step2[9]);
+ step1[10] = vsubq_s16(step2[11], step2[10]);
+ step1[11] = vaddq_s16(step2[11], step2[10]);
+ step1[12] = vaddq_s16(step2[12], step2[13]);
+ step1[13] = vsubq_s16(step2[12], step2[13]);
+ step1[14] = vsubq_s16(step2[15], step2[14]);
+ step1[15] = vaddq_s16(step2[15], step2[14]);
+
+ // stage 4
+ step2[0] = step2[1] = vqrdmulhq_lane_s16(step1[0], cospid_0_8_16_24, 2);
+ step2[2] = vqrdmulhq_lane_s16(step1[2], cospid_0_8_16_24, 3);
+ step2[3] = vqrdmulhq_lane_s16(step1[2], cospid_0_8_16_24, 1);
+ step2[4] = vaddq_s16(step1[4], step1[5]);
+ step2[5] = vsubq_s16(step1[4], step1[5]);
+ step2[6] = vsubq_s16(step1[7], step1[6]);
+ step2[7] = vaddq_s16(step1[7], step1[6]);
+ step2[8] = step1[8];
+ idct_cospi_8_24_q(step1[14], step1[9], cospi_0_8_16_24, &step2[9],
+ &step2[14]);
+ idct_cospi_8_24_neg_q(step1[13], step1[10], cospi_0_8_16_24, &step2[13],
+ &step2[10]);
+ step2[11] = step1[11];
+ step2[12] = step1[12];
+ step2[15] = step1[15];
+
+ // stage 5
+ step1[0] = vaddq_s16(step2[0], step2[3]);
+ step1[1] = vaddq_s16(step2[1], step2[2]);
+ step1[2] = vsubq_s16(step2[1], step2[2]);
+ step1[3] = vsubq_s16(step2[0], step2[3]);
+ step1[4] = step2[4];
+ idct_cospi_16_16_q(step2[5], step2[6], cospi_0_8_16_24, &step1[5], &step1[6]);
+ step1[7] = step2[7];
+ step1[8] = vaddq_s16(step2[8], step2[11]);
+ step1[9] = vaddq_s16(step2[9], step2[10]);
+ step1[10] = vsubq_s16(step2[9], step2[10]);
+ step1[11] = vsubq_s16(step2[8], step2[11]);
+ step1[12] = vsubq_s16(step2[15], step2[12]);
+ step1[13] = vsubq_s16(step2[14], step2[13]);
+ step1[14] = vaddq_s16(step2[14], step2[13]);
+ step1[15] = vaddq_s16(step2[15], step2[12]);
+
+ // stage 6
+ step2[0] = vaddq_s16(step1[0], step1[7]);
+ step2[1] = vaddq_s16(step1[1], step1[6]);
+ step2[2] = vaddq_s16(step1[2], step1[5]);
+ step2[3] = vaddq_s16(step1[3], step1[4]);
+ step2[4] = vsubq_s16(step1[3], step1[4]);
+ step2[5] = vsubq_s16(step1[2], step1[5]);
+ step2[6] = vsubq_s16(step1[1], step1[6]);
+ step2[7] = vsubq_s16(step1[0], step1[7]);
+ idct_cospi_16_16_q(step1[10], step1[13], cospi_0_8_16_24, &step2[10],
+ &step2[13]);
+ idct_cospi_16_16_q(step1[11], step1[12], cospi_0_8_16_24, &step2[11],
+ &step2[12]);
+ step2[8] = step1[8];
+ step2[9] = step1[9];
+ step2[14] = step1[14];
+ step2[15] = step1[15];
+
+ // stage 7
+ idct16x16_add_stage7(step2, out);
+
+ if (output) {+ idct16x16_store_pass1(out, output);
+ } else {+ idct16x16_add_store(out, dest, stride);
+ }
+}
+
static void idct16x16_10_add_half1d_pass1(const tran_low_t *input,
int16_t *output) {const int16x8_t cospis0 = vld1q_s16(kCospi);
@@ -572,7 +763,7 @@
}
static void idct16x16_10_add_half1d_pass2(const int16_t *input, int16_t *output,
- uint8_t *dest, int stride) {+ uint8_t *dest, const int stride) {const int16x8_t cospis0 = vld1q_s16(kCospi);
const int16x8_t cospis1 = vld1q_s16(kCospi + 8);
const int16x8_t cospisd0 = vaddq_s16(cospis0, cospis0);
@@ -702,56 +893,9 @@
out[15] = vsubq_s16(step2[0], step2[15]);
if (output) {- // pass 1: save the result into output
- vst1q_s16(output, out[0]);
- output += 16;
- vst1q_s16(output, out[1]);
- output += 16;
- vst1q_s16(output, out[2]);
- output += 16;
- vst1q_s16(output, out[3]);
- output += 16;
- vst1q_s16(output, out[4]);
- output += 16;
- vst1q_s16(output, out[5]);
- output += 16;
- vst1q_s16(output, out[6]);
- output += 16;
- vst1q_s16(output, out[7]);
- output += 16;
- vst1q_s16(output, out[8]);
- output += 16;
- vst1q_s16(output, out[9]);
- output += 16;
- vst1q_s16(output, out[10]);
- output += 16;
- vst1q_s16(output, out[11]);
- output += 16;
- vst1q_s16(output, out[12]);
- output += 16;
- vst1q_s16(output, out[13]);
- output += 16;
- vst1q_s16(output, out[14]);
- output += 16;
- vst1q_s16(output, out[15]);
+ idct16x16_store_pass1(out, output);
} else {- // pass 2: add the result to dest.
- idct16x16_add8x1(out[0], &dest, stride);
- idct16x16_add8x1(out[1], &dest, stride);
- idct16x16_add8x1(out[2], &dest, stride);
- idct16x16_add8x1(out[3], &dest, stride);
- idct16x16_add8x1(out[4], &dest, stride);
- idct16x16_add8x1(out[5], &dest, stride);
- idct16x16_add8x1(out[6], &dest, stride);
- idct16x16_add8x1(out[7], &dest, stride);
- idct16x16_add8x1(out[8], &dest, stride);
- idct16x16_add8x1(out[9], &dest, stride);
- idct16x16_add8x1(out[10], &dest, stride);
- idct16x16_add8x1(out[11], &dest, stride);
- idct16x16_add8x1(out[12], &dest, stride);
- idct16x16_add8x1(out[13], &dest, stride);
- idct16x16_add8x1(out[14], &dest, stride);
- idct16x16_add8x1(out[15], &dest, stride);
+ idct16x16_add_store(out, dest, stride);
}
}
@@ -772,6 +916,23 @@
// Parallel idct to get the right 8 columns
idct16x16_256_add_half1d(row_idct_output + 16 * 8, NULL, dest + 8, stride);
+}
+
+void vpx_idct16x16_38_add_neon(const tran_low_t *input, uint8_t *dest,
+ int stride) {+ int16_t row_idct_output[16 * 16];
+
+ // pass 1
+ // Parallel idct on the upper 8 rows
+ idct16x16_38_add_half1d((const int16_t *)input, row_idct_output, dest,
+ stride);
+
+ // pass 2
+ // Parallel idct to get the left 8 columns
+ idct16x16_38_add_half1d(row_idct_output, NULL, dest, stride);
+
+ // Parallel idct to get the right 8 columns
+ idct16x16_38_add_half1d(row_idct_output + 16 * 8, NULL, dest + 8, stride);
}
void vpx_idct16x16_10_add_neon(const tran_low_t *input, uint8_t *dest,
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -695,7 +695,6 @@
add_proto qw/void vpx_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int stride";
specialize qw/vpx_idct16x16_38_add neon sse2/;
- $vpx_idct16x16_38_add_neon=vpx_idct16x16_256_add_neon;
$vpx_idct16x16_38_add_sse2=vpx_idct16x16_256_add_sse2;
add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int stride";
@@ -789,7 +788,6 @@
add_proto qw/void vpx_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int stride";
specialize qw/vpx_idct16x16_38_add sse2 neon dspr2 msa/;
- $vpx_idct16x16_38_add_neon=vpx_idct16x16_256_add_neon;
$vpx_idct16x16_38_add_sse2=vpx_idct16x16_256_add_sse2;
$vpx_idct16x16_38_add_dspr2=vpx_idct16x16_256_add_dspr2;
$vpx_idct16x16_38_add_msa=vpx_idct16x16_256_add_msa;