ref: ab27b68693acd40d784146b3580fc4cd8ed9e5d5
parent: 419ce36294db6597801dd9a9c21654a187a5d060
parent: 2c56bb97f2363a21fc26c1e20f86e01657a87bb5
author: Johann Koenig <johannkoenig@google.com>
date: Tue Aug 22 15:19:14 EDT 2017
Merge changes Icfb70687,I9a963e99,Ie8ac00ef,I1272917c * changes: quantize: ignore skip_block in arm quantize: ignore skip_block in x86 quantize fp: ignore skip_block in arm quantize fp: ignore skip_block in x86
--- a/vp9/encoder/arm/neon/vp9_quantize_neon.c
+++ b/vp9/encoder/arm/neon/vp9_quantize_neon.c
@@ -9,7 +9,7 @@
*/
#include <arm_neon.h>
-
+#include <assert.h>
#include <math.h>
#include "vpx_mem/vpx_mem.h"
@@ -31,86 +31,83 @@
tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
uint16_t *eob_ptr, const int16_t *scan,
const int16_t *iscan) {+ // Quantization pass: All coefficients with index >= zero_flag are
+ // skippable. Note: zero_flag can be zero.
+ int i;
+ const int16x8_t v_zero = vdupq_n_s16(0);
+ const int16x8_t v_one = vdupq_n_s16(1);
+ int16x8_t v_eobmax_76543210 = vdupq_n_s16(-1);
+ int16x8_t v_round = vmovq_n_s16(round_ptr[1]);
+ int16x8_t v_quant = vmovq_n_s16(quant_ptr[1]);
+ int16x8_t v_dequant = vmovq_n_s16(dequant_ptr[1]);
+
(void)scan;
+ (void)skip_block;
+ assert(!skip_block);
- if (!skip_block) {- // Quantization pass: All coefficients with index >= zero_flag are
- // skippable. Note: zero_flag can be zero.
- int i;
- const int16x8_t v_zero = vdupq_n_s16(0);
- const int16x8_t v_one = vdupq_n_s16(1);
- int16x8_t v_eobmax_76543210 = vdupq_n_s16(-1);
- int16x8_t v_round = vmovq_n_s16(round_ptr[1]);
- int16x8_t v_quant = vmovq_n_s16(quant_ptr[1]);
- int16x8_t v_dequant = vmovq_n_s16(dequant_ptr[1]);
- // adjust for dc
- v_round = vsetq_lane_s16(round_ptr[0], v_round, 0);
- v_quant = vsetq_lane_s16(quant_ptr[0], v_quant, 0);
- v_dequant = vsetq_lane_s16(dequant_ptr[0], v_dequant, 0);
- // process dc and the first seven ac coeffs
- {- const int16x8_t v_iscan = vld1q_s16(&iscan[0]);
- const int16x8_t v_coeff = load_tran_low_to_s16q(coeff_ptr);
- const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
- const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
- const int32x4_t v_tmp_lo =
- vmull_s16(vget_low_s16(v_tmp), vget_low_s16(v_quant));
- const int32x4_t v_tmp_hi =
- vmull_s16(vget_high_s16(v_tmp), vget_high_s16(v_quant));
- const int16x8_t v_tmp2 =
- vcombine_s16(vshrn_n_s32(v_tmp_lo, 16), vshrn_n_s32(v_tmp_hi, 16));
- const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
- const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
- const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
- const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
- const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
- const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
- v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
- store_s16q_to_tran_low(qcoeff_ptr, v_qcoeff);
- store_s16q_to_tran_low(dqcoeff_ptr, v_dqcoeff);
- v_round = vmovq_n_s16(round_ptr[1]);
- v_quant = vmovq_n_s16(quant_ptr[1]);
- v_dequant = vmovq_n_s16(dequant_ptr[1]);
- }
- // now process the rest of the ac coeffs
- for (i = 8; i < count; i += 8) {- const int16x8_t v_iscan = vld1q_s16(&iscan[i]);
- const int16x8_t v_coeff = load_tran_low_to_s16q(coeff_ptr + i);
- const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
- const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
- const int32x4_t v_tmp_lo =
- vmull_s16(vget_low_s16(v_tmp), vget_low_s16(v_quant));
- const int32x4_t v_tmp_hi =
- vmull_s16(vget_high_s16(v_tmp), vget_high_s16(v_quant));
- const int16x8_t v_tmp2 =
- vcombine_s16(vshrn_n_s32(v_tmp_lo, 16), vshrn_n_s32(v_tmp_hi, 16));
- const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
- const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
- const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
- const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
- const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
- const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
- v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
- store_s16q_to_tran_low(qcoeff_ptr + i, v_qcoeff);
- store_s16q_to_tran_low(dqcoeff_ptr + i, v_dqcoeff);
- }
- {- const int16x4_t v_eobmax_3210 = vmax_s16(
- vget_low_s16(v_eobmax_76543210), vget_high_s16(v_eobmax_76543210));
- const int64x1_t v_eobmax_xx32 =
- vshr_n_s64(vreinterpret_s64_s16(v_eobmax_3210), 32);
- const int16x4_t v_eobmax_tmp =
- vmax_s16(v_eobmax_3210, vreinterpret_s16_s64(v_eobmax_xx32));
- const int64x1_t v_eobmax_xxx3 =
- vshr_n_s64(vreinterpret_s64_s16(v_eobmax_tmp), 16);
- const int16x4_t v_eobmax_final =
- vmax_s16(v_eobmax_tmp, vreinterpret_s16_s64(v_eobmax_xxx3));
+ // adjust for dc
+ v_round = vsetq_lane_s16(round_ptr[0], v_round, 0);
+ v_quant = vsetq_lane_s16(quant_ptr[0], v_quant, 0);
+ v_dequant = vsetq_lane_s16(dequant_ptr[0], v_dequant, 0);
+ // process dc and the first seven ac coeffs
+ {+ const int16x8_t v_iscan = vld1q_s16(&iscan[0]);
+ const int16x8_t v_coeff = load_tran_low_to_s16q(coeff_ptr);
+ const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
+ const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
+ const int32x4_t v_tmp_lo =
+ vmull_s16(vget_low_s16(v_tmp), vget_low_s16(v_quant));
+ const int32x4_t v_tmp_hi =
+ vmull_s16(vget_high_s16(v_tmp), vget_high_s16(v_quant));
+ const int16x8_t v_tmp2 =
+ vcombine_s16(vshrn_n_s32(v_tmp_lo, 16), vshrn_n_s32(v_tmp_hi, 16));
+ const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
+ const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
+ const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
+ const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
+ const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
+ const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+ v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
+ store_s16q_to_tran_low(qcoeff_ptr, v_qcoeff);
+ store_s16q_to_tran_low(dqcoeff_ptr, v_dqcoeff);
+ v_round = vmovq_n_s16(round_ptr[1]);
+ v_quant = vmovq_n_s16(quant_ptr[1]);
+ v_dequant = vmovq_n_s16(dequant_ptr[1]);
+ }
+ // now process the rest of the ac coeffs
+ for (i = 8; i < count; i += 8) {+ const int16x8_t v_iscan = vld1q_s16(&iscan[i]);
+ const int16x8_t v_coeff = load_tran_low_to_s16q(coeff_ptr + i);
+ const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
+ const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
+ const int32x4_t v_tmp_lo =
+ vmull_s16(vget_low_s16(v_tmp), vget_low_s16(v_quant));
+ const int32x4_t v_tmp_hi =
+ vmull_s16(vget_high_s16(v_tmp), vget_high_s16(v_quant));
+ const int16x8_t v_tmp2 =
+ vcombine_s16(vshrn_n_s32(v_tmp_lo, 16), vshrn_n_s32(v_tmp_hi, 16));
+ const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
+ const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
+ const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
+ const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
+ const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
+ const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+ v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
+ store_s16q_to_tran_low(qcoeff_ptr + i, v_qcoeff);
+ store_s16q_to_tran_low(dqcoeff_ptr + i, v_dqcoeff);
+ }
+ {+ const int16x4_t v_eobmax_3210 = vmax_s16(vget_low_s16(v_eobmax_76543210),
+ vget_high_s16(v_eobmax_76543210));
+ const int64x1_t v_eobmax_xx32 =
+ vshr_n_s64(vreinterpret_s64_s16(v_eobmax_3210), 32);
+ const int16x4_t v_eobmax_tmp =
+ vmax_s16(v_eobmax_3210, vreinterpret_s16_s64(v_eobmax_xx32));
+ const int64x1_t v_eobmax_xxx3 =
+ vshr_n_s64(vreinterpret_s64_s16(v_eobmax_tmp), 16);
+ const int16x4_t v_eobmax_final =
+ vmax_s16(v_eobmax_tmp, vreinterpret_s16_s64(v_eobmax_xxx3));
- *eob_ptr = (uint16_t)vget_lane_s16(v_eobmax_final, 0);
- }
- } else {- memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
- memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
- *eob_ptr = 0;
+ *eob_ptr = (uint16_t)vget_lane_s16(v_eobmax_final, 0);
}
}
--- a/vp9/encoder/x86/vp9_quantize_sse2.c
+++ b/vp9/encoder/x86/vp9_quantize_sse2.c
@@ -8,6 +8,7 @@
* be found in the AUTHORS file in the root of the source tree.
*/
+#include <assert.h>
#include <emmintrin.h>
#include <xmmintrin.h>
@@ -25,8 +26,12 @@
__m128i zero;
__m128i thr;
int16_t nzflag;
+ __m128i eob;
+ __m128i round, quant, dequant;
(void)scan_ptr;
+ (void)skip_block;
+ assert(!skip_block);
coeff_ptr += n_coeffs;
iscan_ptr += n_coeffs;
@@ -35,40 +40,106 @@
n_coeffs = -n_coeffs;
zero = _mm_setzero_si128();
- if (!skip_block) {- __m128i eob;
- __m128i round, quant, dequant;
+ {+ __m128i coeff0, coeff1;
+
+ // Setup global values
{- __m128i coeff0, coeff1;
+ round = _mm_load_si128((const __m128i *)round_ptr);
+ quant = _mm_load_si128((const __m128i *)quant_ptr);
+ dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+ }
- // Setup global values
- {- round = _mm_load_si128((const __m128i *)round_ptr);
- quant = _mm_load_si128((const __m128i *)quant_ptr);
- dequant = _mm_load_si128((const __m128i *)dequant_ptr);
- }
+ {+ __m128i coeff0_sign, coeff1_sign;
+ __m128i qcoeff0, qcoeff1;
+ __m128i qtmp0, qtmp1;
+ // Do DC and first 15 AC
+ coeff0 = load_tran_low(coeff_ptr + n_coeffs);
+ coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
- {- __m128i coeff0_sign, coeff1_sign;
- __m128i qcoeff0, qcoeff1;
- __m128i qtmp0, qtmp1;
- // Do DC and first 15 AC
- coeff0 = load_tran_low(coeff_ptr + n_coeffs);
- coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
+ // Poor man's sign extract
+ coeff0_sign = _mm_srai_epi16(coeff0, 15);
+ coeff1_sign = _mm_srai_epi16(coeff1, 15);
+ qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+ qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+ qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+ qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
- // Poor man's sign extract
- coeff0_sign = _mm_srai_epi16(coeff0, 15);
- coeff1_sign = _mm_srai_epi16(coeff1, 15);
- qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
- qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
- qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
- qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+ qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+ round = _mm_unpackhi_epi64(round, round);
+ qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+ qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+ quant = _mm_unpackhi_epi64(quant, quant);
+ qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+ // Reinsert signs
+ qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+ qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+ qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+ qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+ store_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
+ store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
+
+ coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+ dequant = _mm_unpackhi_epi64(dequant, dequant);
+ coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+ store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
+ store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
+ }
+
+ {+ // Scan for eob
+ __m128i zero_coeff0, zero_coeff1;
+ __m128i nzero_coeff0, nzero_coeff1;
+ __m128i iscan0, iscan1;
+ __m128i eob1;
+ zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+ zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+ nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+ nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+ iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
+ iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
+ // Add one to convert from indices to counts
+ iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+ iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+ eob = _mm_and_si128(iscan0, nzero_coeff0);
+ eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+ eob = _mm_max_epi16(eob, eob1);
+ }
+ n_coeffs += 8 * 2;
+ }
+
+ thr = _mm_srai_epi16(dequant, 1);
+
+ // AC only loop
+ while (n_coeffs < 0) {+ __m128i coeff0, coeff1;
+ {+ __m128i coeff0_sign, coeff1_sign;
+ __m128i qcoeff0, qcoeff1;
+ __m128i qtmp0, qtmp1;
+
+ coeff0 = load_tran_low(coeff_ptr + n_coeffs);
+ coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
+
+ // Poor man's sign extract
+ coeff0_sign = _mm_srai_epi16(coeff0, 15);
+ coeff1_sign = _mm_srai_epi16(coeff1, 15);
+ qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+ qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+ qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+ qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+ nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
+ _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
+
+ if (nzflag) {qcoeff0 = _mm_adds_epi16(qcoeff0, round);
- round = _mm_unpackhi_epi64(round, round);
qcoeff1 = _mm_adds_epi16(qcoeff1, round);
qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
- quant = _mm_unpackhi_epi64(quant, quant);
qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
// Reinsert signs
@@ -81,131 +152,51 @@
store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
- dequant = _mm_unpackhi_epi64(dequant, dequant);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
- }
+ } else {+ store_zero_tran_low(qcoeff_ptr + n_coeffs);
+ store_zero_tran_low(qcoeff_ptr + n_coeffs + 8);
- {- // Scan for eob
- __m128i zero_coeff0, zero_coeff1;
- __m128i nzero_coeff0, nzero_coeff1;
- __m128i iscan0, iscan1;
- __m128i eob1;
- zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
- zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
- nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
- nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
- iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
- iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
- // Add one to convert from indices to counts
- iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
- iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
- eob = _mm_and_si128(iscan0, nzero_coeff0);
- eob1 = _mm_and_si128(iscan1, nzero_coeff1);
- eob = _mm_max_epi16(eob, eob1);
+ store_zero_tran_low(dqcoeff_ptr + n_coeffs);
+ store_zero_tran_low(dqcoeff_ptr + n_coeffs + 8);
}
- n_coeffs += 8 * 2;
}
- thr = _mm_srai_epi16(dequant, 1);
-
- // AC only loop
- while (n_coeffs < 0) {- __m128i coeff0, coeff1;
- {- __m128i coeff0_sign, coeff1_sign;
- __m128i qcoeff0, qcoeff1;
- __m128i qtmp0, qtmp1;
-
- coeff0 = load_tran_low(coeff_ptr + n_coeffs);
- coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
-
- // Poor man's sign extract
- coeff0_sign = _mm_srai_epi16(coeff0, 15);
- coeff1_sign = _mm_srai_epi16(coeff1, 15);
- qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
- qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
- qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
- qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
-
- nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
- _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
-
- if (nzflag) {- qcoeff0 = _mm_adds_epi16(qcoeff0, round);
- qcoeff1 = _mm_adds_epi16(qcoeff1, round);
- qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
- qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
-
- // Reinsert signs
- qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
- qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
- qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
- qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
-
- store_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
- store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
-
- coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
- coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
-
- store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
- store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
- } else {- store_zero_tran_low(qcoeff_ptr + n_coeffs);
- store_zero_tran_low(qcoeff_ptr + n_coeffs + 8);
-
- store_zero_tran_low(dqcoeff_ptr + n_coeffs);
- store_zero_tran_low(dqcoeff_ptr + n_coeffs + 8);
- }
- }
-
- if (nzflag) {- // Scan for eob
- __m128i zero_coeff0, zero_coeff1;
- __m128i nzero_coeff0, nzero_coeff1;
- __m128i iscan0, iscan1;
- __m128i eob0, eob1;
- zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
- zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
- nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
- nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
- iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
- iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
- // Add one to convert from indices to counts
- iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
- iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
- eob0 = _mm_and_si128(iscan0, nzero_coeff0);
- eob1 = _mm_and_si128(iscan1, nzero_coeff1);
- eob0 = _mm_max_epi16(eob0, eob1);
- eob = _mm_max_epi16(eob, eob0);
- }
- n_coeffs += 8 * 2;
+ if (nzflag) {+ // Scan for eob
+ __m128i zero_coeff0, zero_coeff1;
+ __m128i nzero_coeff0, nzero_coeff1;
+ __m128i iscan0, iscan1;
+ __m128i eob0, eob1;
+ zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+ zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+ nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+ nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+ iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
+ iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
+ // Add one to convert from indices to counts
+ iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+ iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+ eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+ eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+ eob0 = _mm_max_epi16(eob0, eob1);
+ eob = _mm_max_epi16(eob, eob0);
}
+ n_coeffs += 8 * 2;
+ }
- // Accumulate EOB
- {- __m128i eob_shuffled;
- eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
- eob = _mm_max_epi16(eob, eob_shuffled);
- eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
- eob = _mm_max_epi16(eob, eob_shuffled);
- eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
- eob = _mm_max_epi16(eob, eob_shuffled);
- *eob_ptr = _mm_extract_epi16(eob, 1);
- }
- } else {- do {- store_zero_tran_low(qcoeff_ptr + n_coeffs);
- store_zero_tran_low(qcoeff_ptr + n_coeffs + 8);
-
- store_zero_tran_low(dqcoeff_ptr + n_coeffs);
- store_zero_tran_low(dqcoeff_ptr + n_coeffs + 8);
- n_coeffs += 8 * 2;
- } while (n_coeffs < 0);
- *eob_ptr = 0;
+ // Accumulate EOB
+ {+ __m128i eob_shuffled;
+ eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+ eob = _mm_max_epi16(eob, eob_shuffled);
+ eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+ eob = _mm_max_epi16(eob, eob_shuffled);
+ eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+ eob = _mm_max_epi16(eob, eob_shuffled);
+ *eob_ptr = _mm_extract_epi16(eob, 1);
}
}
--- a/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
+++ b/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
@@ -22,8 +22,6 @@
cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, round, quant, \
qcoeff, dqcoeff, dequant, \
eob, scan, iscan
- cmp dword skipm, 0
- jne .blank
; actual quantize loop - setup pointers, rounders, etc.
movifnidn coeffq, coeffmp
@@ -172,27 +170,6 @@
pmaxsw m8, m7
pextrw r6, m8, 0
mov [r2], r6
- RET
-
- ; skip-block, i.e. just write all zeroes
-.blank:
- mov r0, dqcoeffmp
- movifnidn ncoeffq, ncoeffmp
- mov r2, qcoeffmp
- mov r3, eobmp
-
- lea r0q, [r0q+ncoeffq*2]
- lea r2q, [r2q+ncoeffq*2]
- neg ncoeffq
- pxor m7, m7
-.blank_loop:
- STORE_ZERO_TRAN_LOW 7, r0q, ncoeffq
- STORE_ZERO_TRAN_LOW 7, r0q, ncoeffq + 8
- STORE_ZERO_TRAN_LOW 7, r2q, ncoeffq
- STORE_ZERO_TRAN_LOW 7, r2q, ncoeffq + 8
- add ncoeffq, mmsize
- jl .blank_loop
- mov word [r3q], 0
RET
%endmacro
--- a/vpx_dsp/arm/quantize_neon.c
+++ b/vpx_dsp/arm/quantize_neon.c
@@ -9,6 +9,7 @@
*/
#include <arm_neon.h>
+#include <assert.h>
#include "./vpx_dsp_rtcd.h"
#include "vpx_dsp/arm/mem_neon.h"
@@ -20,25 +21,13 @@
tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
uint16_t *eob_ptr, const int16_t *scan_ptr,
const int16_t *iscan_ptr) {- const int16x8_t zero = vdupq_n_s16(0);
const int16x8_t one = vdupq_n_s16(1);
const int16x8_t neg_one = vdupq_n_s16(-1);
uint16x8_t eob_max;
(void)scan_ptr;
+ (void)skip_block;
+ assert(!skip_block);
- if (skip_block) {- do {- store_s16q_to_tran_low(qcoeff_ptr, zero);
- store_s16q_to_tran_low(dqcoeff_ptr, zero);
- qcoeff_ptr += 8;
- dqcoeff_ptr += 8;
- n_coeffs -= 8;
- } while (n_coeffs > 0);
-
- *eob_ptr = 0;
- return;
- }
-
// Process first 8 values which include a dc component.
{// Only the first element of each vector is DC.
@@ -162,7 +151,6 @@
const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
const int16_t *scan_ptr, const int16_t *iscan_ptr) {- const int16x8_t zero = vdupq_n_s16(0);
const int16x8_t one = vdupq_n_s16(1);
const int16x8_t neg_one = vdupq_n_s16(-1);
uint16x8_t eob_max;
@@ -169,17 +157,8 @@
int i;
(void)scan_ptr;
(void)n_coeffs; // Because we will always calculate 32*32.
-
- if (skip_block) {- for (i = 0; i < 32 * 32 / 8; ++i) {- store_s16q_to_tran_low(qcoeff_ptr, zero);
- store_s16q_to_tran_low(dqcoeff_ptr, zero);
- qcoeff_ptr += 8;
- dqcoeff_ptr += 8;
- }
- *eob_ptr = 0;
- return;
- }
+ (void)skip_block;
+ assert(!skip_block);
// Process first 8 values which include a dc component.
{--- a/vpx_dsp/x86/highbd_quantize_intrin_sse2.c
+++ b/vpx_dsp/x86/highbd_quantize_intrin_sse2.c
@@ -8,6 +8,7 @@
* be found in the AUTHORS file in the root of the source tree.
*/
+#include <assert.h>
#include <emmintrin.h>
#include "vpx_dsp/vpx_dsp_common.h"
@@ -37,54 +38,54 @@
nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
(void)scan;
+ (void)skip_block;
+ assert(!skip_block);
memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
- if (!skip_block) {- // Pre-scan pass
- for (i = ((int)count / 4) - 1; i >= 0; i--) {- __m128i coeffs, cmp1, cmp2;
- int test;
- coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
- cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
- cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
- cmp1 = _mm_and_si128(cmp1, cmp2);
- test = _mm_movemask_epi8(cmp1);
- if (test == 0xffff)
- non_zero_regs--;
- else
- break;
- }
+ // Pre-scan pass
+ for (i = ((int)count / 4) - 1; i >= 0; i--) {+ __m128i coeffs, cmp1, cmp2;
+ int test;
+ coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+ cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+ cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+ cmp1 = _mm_and_si128(cmp1, cmp2);
+ test = _mm_movemask_epi8(cmp1);
+ if (test == 0xffff)
+ non_zero_regs--;
+ else
+ break;
+ }
- // Quantization pass:
- for (i = 0; i < non_zero_regs; i++) {- __m128i coeffs, coeffs_sign, tmp1, tmp2;
- int test;
- int abs_coeff[4];
- int coeff_sign[4];
+ // Quantization pass:
+ for (i = 0; i < non_zero_regs; i++) {+ __m128i coeffs, coeffs_sign, tmp1, tmp2;
+ int test;
+ int abs_coeff[4];
+ int coeff_sign[4];
- coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
- coeffs_sign = _mm_srai_epi32(coeffs, 31);
- coeffs = _mm_sub_epi32(_mm_xor_si128(coeffs, coeffs_sign), coeffs_sign);
- tmp1 = _mm_cmpgt_epi32(coeffs, zbins[i != 0]);
- tmp2 = _mm_cmpeq_epi32(coeffs, zbins[i != 0]);
- tmp1 = _mm_or_si128(tmp1, tmp2);
- test = _mm_movemask_epi8(tmp1);
- _mm_storeu_si128((__m128i *)abs_coeff, coeffs);
- _mm_storeu_si128((__m128i *)coeff_sign, coeffs_sign);
+ coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+ coeffs_sign = _mm_srai_epi32(coeffs, 31);
+ coeffs = _mm_sub_epi32(_mm_xor_si128(coeffs, coeffs_sign), coeffs_sign);
+ tmp1 = _mm_cmpgt_epi32(coeffs, zbins[i != 0]);
+ tmp2 = _mm_cmpeq_epi32(coeffs, zbins[i != 0]);
+ tmp1 = _mm_or_si128(tmp1, tmp2);
+ test = _mm_movemask_epi8(tmp1);
+ _mm_storeu_si128((__m128i *)abs_coeff, coeffs);
+ _mm_storeu_si128((__m128i *)coeff_sign, coeffs_sign);
- for (j = 0; j < 4; j++) {- if (test & (1 << (4 * j))) {- int k = 4 * i + j;
- const int64_t tmp3 = abs_coeff[j] + round_ptr[k != 0];
- const int64_t tmp4 = ((tmp3 * quant_ptr[k != 0]) >> 16) + tmp3;
- const uint32_t abs_qcoeff =
- (uint32_t)((tmp4 * quant_shift_ptr[k != 0]) >> 16);
- qcoeff_ptr[k] = (int)(abs_qcoeff ^ coeff_sign[j]) - coeff_sign[j];
- dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0];
- if (abs_qcoeff) eob_i = iscan[k] > eob_i ? iscan[k] : eob_i;
- }
+ for (j = 0; j < 4; j++) {+ if (test & (1 << (4 * j))) {+ int k = 4 * i + j;
+ const int64_t tmp3 = abs_coeff[j] + round_ptr[k != 0];
+ const int64_t tmp4 = ((tmp3 * quant_ptr[k != 0]) >> 16) + tmp3;
+ const uint32_t abs_qcoeff =
+ (uint32_t)((tmp4 * quant_shift_ptr[k != 0]) >> 16);
+ qcoeff_ptr[k] = (int)(abs_qcoeff ^ coeff_sign[j]) - coeff_sign[j];
+ dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0];
+ if (abs_qcoeff) eob_i = iscan[k] > eob_i ? iscan[k] : eob_i;
}
}
}
@@ -105,6 +106,9 @@
const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 1);
const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 1);
(void)scan;
+ (void)skip_block;
+ assert(!skip_block);
+
zbins[0] = _mm_set_epi32(zbin1_tmp, zbin1_tmp, zbin1_tmp, zbin0_tmp);
zbins[1] = _mm_set1_epi32(zbin1_tmp);
@@ -116,38 +120,35 @@
memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
- if (!skip_block) {- // Pre-scan pass
- for (i = 0; i < n_coeffs / 4; i++) {- __m128i coeffs, cmp1, cmp2;
- int test;
- coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
- cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
- cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
- cmp1 = _mm_and_si128(cmp1, cmp2);
- test = _mm_movemask_epi8(cmp1);
- if (!(test & 0xf)) idx_arr[idx++] = i * 4;
- if (!(test & 0xf0)) idx_arr[idx++] = i * 4 + 1;
- if (!(test & 0xf00)) idx_arr[idx++] = i * 4 + 2;
- if (!(test & 0xf000)) idx_arr[idx++] = i * 4 + 3;
- }
+ // Pre-scan pass
+ for (i = 0; i < n_coeffs / 4; i++) {+ __m128i coeffs, cmp1, cmp2;
+ int test;
+ coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+ cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+ cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+ cmp1 = _mm_and_si128(cmp1, cmp2);
+ test = _mm_movemask_epi8(cmp1);
+ if (!(test & 0xf)) idx_arr[idx++] = i * 4;
+ if (!(test & 0xf0)) idx_arr[idx++] = i * 4 + 1;
+ if (!(test & 0xf00)) idx_arr[idx++] = i * 4 + 2;
+ if (!(test & 0xf000)) idx_arr[idx++] = i * 4 + 3;
+ }
- // Quantization pass: only process the coefficients selected in
- // pre-scan pass. Note: idx can be zero.
- for (i = 0; i < idx; i++) {- const int rc = idx_arr[i];
- const int coeff = coeff_ptr[rc];
- const int coeff_sign = (coeff >> 31);
- const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
- const int64_t tmp1 =
- abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
- const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
- const uint32_t abs_qcoeff =
- (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
- qcoeff_ptr[rc] = (int)(abs_qcoeff ^ coeff_sign) - coeff_sign;
- dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
- if (abs_qcoeff) eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob;
- }
+ // Quantization pass: only process the coefficients selected in
+ // pre-scan pass. Note: idx can be zero.
+ for (i = 0; i < idx; i++) {+ const int rc = idx_arr[i];
+ const int coeff = coeff_ptr[rc];
+ const int coeff_sign = (coeff >> 31);
+ const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+ const int64_t tmp1 = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+ const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
+ const uint32_t abs_qcoeff =
+ (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
+ qcoeff_ptr[rc] = (int)(abs_qcoeff ^ coeff_sign) - coeff_sign;
+ dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+ if (abs_qcoeff) eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob;
}
*eob_ptr = eob + 1;
}
--- a/vpx_dsp/x86/quantize_avx_x86_64.asm
+++ b/vpx_dsp/x86/quantize_avx_x86_64.asm
@@ -19,10 +19,6 @@
vzeroupper
- ; If we can skip this block, then just zero the output
- cmp skipmp, 0
- jne .blank
-
%ifnidn %1, b_32x32
; Special case for ncoeff == 16, as it is frequent and we can save on
@@ -491,48 +487,6 @@
pmaxsw m8, m7
movq rax, m8
mov [r2], ax
- vzeroupper
- RET
-
- ; Skip-block, i.e. just write all zeroes
-.blank:
-
-DEFINE_ARGS coeff, ncoeff, skip, zbin, round, quant, shift, \
- qcoeff, dqcoeff, dequant, eob, scan, iscan
-
- mov r0, dqcoeffmp
- movifnidn ncoeffq, ncoeffmp
- mov r2, qcoeffmp
- mov r3, eobmp
-
-DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
-
-%if CONFIG_VP9_HIGHBITDEPTH
- lea dqcoeffq, [dqcoeffq+ncoeffq*4]
- lea qcoeffq, [ qcoeffq+ncoeffq*4]
-%else
- lea dqcoeffq, [dqcoeffq+ncoeffq*2]
- lea qcoeffq, [ qcoeffq+ncoeffq*2]
-%endif
-
- neg ncoeffq
- pxor m7, m7
-
-.blank_loop:
-%if CONFIG_VP9_HIGHBITDEPTH
- mova [dqcoeffq+ncoeffq*4+ 0], ymm7
- mova [dqcoeffq+ncoeffq*4+32], ymm7
- mova [qcoeffq+ncoeffq*4+ 0], ymm7
- mova [qcoeffq+ncoeffq*4+32], ymm7
-%else
- mova [dqcoeffq+ncoeffq*2+ 0], ymm7
- mova [qcoeffq+ncoeffq*2+ 0], ymm7
-%endif
- add ncoeffq, mmsize
- jl .blank_loop
-
- mov [eobq], word 0
-
vzeroupper
RET
%endmacro
--- a/vpx_dsp/x86/quantize_sse2.c
+++ b/vpx_dsp/x86/quantize_sse2.c
@@ -8,6 +8,7 @@
* be found in the AUTHORS file in the root of the source tree.
*/
+#include <assert.h>
#include <emmintrin.h>
#include <xmmintrin.h>
@@ -23,7 +24,12 @@
uint16_t *eob_ptr, const int16_t *scan_ptr,
const int16_t *iscan_ptr) {__m128i zero;
+ __m128i eob;
+ __m128i zbin;
+ __m128i round, quant, dequant, shift;
(void)scan_ptr;
+ (void)skip_block;
+ assert(!skip_block);
coeff_ptr += n_coeffs;
iscan_ptr += n_coeffs;
@@ -31,193 +37,179 @@
dqcoeff_ptr += n_coeffs;
n_coeffs = -n_coeffs;
zero = _mm_setzero_si128();
- if (!skip_block) {- __m128i eob;
- __m128i zbin;
- __m128i round, quant, dequant, shift;
+ {+ __m128i coeff0, coeff1;
+
+ // Setup global values
{- __m128i coeff0, coeff1;
+ __m128i pw_1;
+ zbin = _mm_load_si128((const __m128i *)zbin_ptr);
+ round = _mm_load_si128((const __m128i *)round_ptr);
+ quant = _mm_load_si128((const __m128i *)quant_ptr);
+ pw_1 = _mm_set1_epi16(1);
+ zbin = _mm_sub_epi16(zbin, pw_1);
+ dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+ shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
+ }
- // Setup global values
- {- __m128i pw_1;
- zbin = _mm_load_si128((const __m128i *)zbin_ptr);
- round = _mm_load_si128((const __m128i *)round_ptr);
- quant = _mm_load_si128((const __m128i *)quant_ptr);
- pw_1 = _mm_set1_epi16(1);
- zbin = _mm_sub_epi16(zbin, pw_1);
- dequant = _mm_load_si128((const __m128i *)dequant_ptr);
- shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
- }
+ {+ __m128i coeff0_sign, coeff1_sign;
+ __m128i qcoeff0, qcoeff1;
+ __m128i qtmp0, qtmp1;
+ __m128i cmp_mask0, cmp_mask1;
+ // Do DC and first 15 AC
+ coeff0 = load_tran_low(coeff_ptr + n_coeffs);
+ coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
- {- __m128i coeff0_sign, coeff1_sign;
- __m128i qcoeff0, qcoeff1;
- __m128i qtmp0, qtmp1;
- __m128i cmp_mask0, cmp_mask1;
- // Do DC and first 15 AC
- coeff0 = load_tran_low(coeff_ptr + n_coeffs);
- coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
+ // Poor man's sign extract
+ coeff0_sign = _mm_srai_epi16(coeff0, 15);
+ coeff1_sign = _mm_srai_epi16(coeff1, 15);
+ qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+ qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+ qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+ qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
- // Poor man's sign extract
- coeff0_sign = _mm_srai_epi16(coeff0, 15);
- coeff1_sign = _mm_srai_epi16(coeff1, 15);
- qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
- qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
- qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
- qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+ cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+ zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC
+ cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+ qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+ round = _mm_unpackhi_epi64(round, round);
+ qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+ qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+ quant = _mm_unpackhi_epi64(quant, quant);
+ qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+ qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
+ qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
+ qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
+ shift = _mm_unpackhi_epi64(shift, shift);
+ qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
- cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
- zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC
- cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
- qcoeff0 = _mm_adds_epi16(qcoeff0, round);
- round = _mm_unpackhi_epi64(round, round);
- qcoeff1 = _mm_adds_epi16(qcoeff1, round);
- qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
- quant = _mm_unpackhi_epi64(quant, quant);
- qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
- qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
- qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
- qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
- shift = _mm_unpackhi_epi64(shift, shift);
- qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
+ // Reinsert signs
+ qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
+ qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
+ qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+ qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
- // Reinsert signs
- qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
- qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
- qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
- qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+ // Mask out zbin threshold coeffs
+ qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+ qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
- // Mask out zbin threshold coeffs
- qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
- qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+ store_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
+ store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
- store_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
- store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
+ coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+ dequant = _mm_unpackhi_epi64(dequant, dequant);
+ coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
- coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
- dequant = _mm_unpackhi_epi64(dequant, dequant);
- coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+ store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
+ store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
+ }
- store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
- store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
- }
-
- {- // Scan for eob
- __m128i zero_coeff0, zero_coeff1;
- __m128i nzero_coeff0, nzero_coeff1;
- __m128i iscan0, iscan1;
- __m128i eob1;
- zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
- zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
- nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
- nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
- iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
- iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
- // Add one to convert from indices to counts
- iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
- iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
- eob = _mm_and_si128(iscan0, nzero_coeff0);
- eob1 = _mm_and_si128(iscan1, nzero_coeff1);
- eob = _mm_max_epi16(eob, eob1);
- }
- n_coeffs += 8 * 2;
+ {+ // Scan for eob
+ __m128i zero_coeff0, zero_coeff1;
+ __m128i nzero_coeff0, nzero_coeff1;
+ __m128i iscan0, iscan1;
+ __m128i eob1;
+ zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+ zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+ nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+ nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+ iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
+ iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
+ // Add one to convert from indices to counts
+ iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+ iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+ eob = _mm_and_si128(iscan0, nzero_coeff0);
+ eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+ eob = _mm_max_epi16(eob, eob1);
}
+ n_coeffs += 8 * 2;
+ }
- // AC only loop
- while (n_coeffs < 0) {- __m128i coeff0, coeff1;
- {- __m128i coeff0_sign, coeff1_sign;
- __m128i qcoeff0, qcoeff1;
- __m128i qtmp0, qtmp1;
- __m128i cmp_mask0, cmp_mask1;
+ // AC only loop
+ while (n_coeffs < 0) {+ __m128i coeff0, coeff1;
+ {+ __m128i coeff0_sign, coeff1_sign;
+ __m128i qcoeff0, qcoeff1;
+ __m128i qtmp0, qtmp1;
+ __m128i cmp_mask0, cmp_mask1;
- coeff0 = load_tran_low(coeff_ptr + n_coeffs);
- coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
+ coeff0 = load_tran_low(coeff_ptr + n_coeffs);
+ coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
- // Poor man's sign extract
- coeff0_sign = _mm_srai_epi16(coeff0, 15);
- coeff1_sign = _mm_srai_epi16(coeff1, 15);
- qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
- qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
- qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
- qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+ // Poor man's sign extract
+ coeff0_sign = _mm_srai_epi16(coeff0, 15);
+ coeff1_sign = _mm_srai_epi16(coeff1, 15);
+ qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+ qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+ qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+ qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
- cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
- cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
- qcoeff0 = _mm_adds_epi16(qcoeff0, round);
- qcoeff1 = _mm_adds_epi16(qcoeff1, round);
- qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
- qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
- qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
- qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
- qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
- qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
+ cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+ cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+ qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+ qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+ qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+ qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+ qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
+ qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
+ qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
+ qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
- // Reinsert signs
- qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
- qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
- qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
- qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+ // Reinsert signs
+ qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
+ qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
+ qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+ qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
- // Mask out zbin threshold coeffs
- qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
- qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+ // Mask out zbin threshold coeffs
+ qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+ qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
- store_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
- store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
+ store_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
+ store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
- coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
- coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+ coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+ coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
- store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
- store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
- }
-
- {- // Scan for eob
- __m128i zero_coeff0, zero_coeff1;
- __m128i nzero_coeff0, nzero_coeff1;
- __m128i iscan0, iscan1;
- __m128i eob0, eob1;
- zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
- zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
- nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
- nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
- iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
- iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
- // Add one to convert from indices to counts
- iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
- iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
- eob0 = _mm_and_si128(iscan0, nzero_coeff0);
- eob1 = _mm_and_si128(iscan1, nzero_coeff1);
- eob0 = _mm_max_epi16(eob0, eob1);
- eob = _mm_max_epi16(eob, eob0);
- }
- n_coeffs += 8 * 2;
+ store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
+ store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
}
- // Accumulate EOB
{- __m128i eob_shuffled;
- eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
- eob = _mm_max_epi16(eob, eob_shuffled);
- eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
- eob = _mm_max_epi16(eob, eob_shuffled);
- eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
- eob = _mm_max_epi16(eob, eob_shuffled);
- *eob_ptr = _mm_extract_epi16(eob, 1);
+ // Scan for eob
+ __m128i zero_coeff0, zero_coeff1;
+ __m128i nzero_coeff0, nzero_coeff1;
+ __m128i iscan0, iscan1;
+ __m128i eob0, eob1;
+ zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+ zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+ nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+ nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+ iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
+ iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
+ // Add one to convert from indices to counts
+ iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+ iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+ eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+ eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+ eob0 = _mm_max_epi16(eob0, eob1);
+ eob = _mm_max_epi16(eob, eob0);
}
- } else {- do {- store_tran_low(zero, dqcoeff_ptr + n_coeffs);
- store_tran_low(zero, dqcoeff_ptr + n_coeffs + 8);
- store_tran_low(zero, qcoeff_ptr + n_coeffs);
- store_tran_low(zero, qcoeff_ptr + n_coeffs + 8);
- n_coeffs += 8 * 2;
- } while (n_coeffs < 0);
- *eob_ptr = 0;
+ n_coeffs += 8 * 2;
+ }
+
+ // Accumulate EOB
+ {+ __m128i eob_shuffled;
+ eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+ eob = _mm_max_epi16(eob, eob_shuffled);
+ eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+ eob = _mm_max_epi16(eob, eob_shuffled);
+ eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+ eob = _mm_max_epi16(eob, eob_shuffled);
+ *eob_ptr = _mm_extract_epi16(eob, 1);
}
}
--- a/vpx_dsp/x86/quantize_ssse3.c
+++ b/vpx_dsp/x86/quantize_ssse3.c
@@ -8,6 +8,7 @@
* be found in the AUTHORS file in the root of the source tree.
*/
+#include <assert.h>
#include <tmmintrin.h>
#include "./vpx_dsp_rtcd.h"
@@ -28,18 +29,8 @@
__m128i round, quant, dequant, shift;
intptr_t index = 0;
(void)scan_ptr;
-
- if (skip_block) {- do {- store_tran_low(zero, dqcoeff_ptr + index);
- store_tran_low(zero, dqcoeff_ptr + index + 8);
- store_tran_low(zero, qcoeff_ptr + index);
- store_tran_low(zero, qcoeff_ptr + index + 8);
- index += 16;
- } while (index < n_coeffs);
- *eob_ptr = 0;
- return;
- }
+ (void)skip_block;
+ assert(!skip_block);
// Setup global values
{--- a/vpx_dsp/x86/quantize_ssse3_x86_64.asm
+++ b/vpx_dsp/x86/quantize_ssse3_x86_64.asm
@@ -19,9 +19,6 @@
cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
shift, qcoeff, dqcoeff, dequant, \
eob, scan, iscan
- cmp dword skipm, 0
- jne .blank
-
; actual quantize loop - setup pointers, rounders, etc.
movifnidn coeffq, coeffmp
movifnidn ncoeffq, ncoeffmp
@@ -299,46 +296,6 @@
pmaxsw m8, m7
pextrw r6, m8, 0
mov [r2], r6
- RET
-
- ; skip-block, i.e. just write all zeroes
-.blank:
-DEFINE_ARGS coeff, ncoeff, skip, zbin, round, quant, shift, \
- qcoeff, dqcoeff, dequant, eob, scan, iscan
-
- mov r0, dqcoeffmp
- movifnidn ncoeffq, ncoeffmp
- mov r2, qcoeffmp
- mov r3, eobmp
- DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
-%if CONFIG_VP9_HIGHBITDEPTH
- lea dqcoeffq, [dqcoeffq+ncoeffq*4]
- lea qcoeffq, [ qcoeffq+ncoeffq*4]
-%else
- lea dqcoeffq, [dqcoeffq+ncoeffq*2]
- lea qcoeffq, [ qcoeffq+ncoeffq*2]
-%endif
- neg ncoeffq
- pxor m7, m7
-.blank_loop:
-%if CONFIG_VP9_HIGHBITDEPTH
- mova [dqcoeffq+ncoeffq*4+ 0], m7
- mova [dqcoeffq+ncoeffq*4+16], m7
- mova [dqcoeffq+ncoeffq*4+32], m7
- mova [dqcoeffq+ncoeffq*4+48], m7
- mova [qcoeffq+ncoeffq*4+ 0], m7
- mova [qcoeffq+ncoeffq*4+16], m7
- mova [qcoeffq+ncoeffq*4+32], m7
- mova [qcoeffq+ncoeffq*4+48], m7
-%else
- mova [dqcoeffq+ncoeffq*2+ 0], m7
- mova [dqcoeffq+ncoeffq*2+16], m7
- mova [qcoeffq+ncoeffq*2+ 0], m7
- mova [qcoeffq+ncoeffq*2+16], m7
-%endif
- add ncoeffq, mmsize
- jl .blank_loop
- mov word [eobq], 0
RET
%endmacro