hook up av1_quantize_fp_neon
2-7% speed up on QVGA and 5-7% on VGA.
Change-Id: Ic1d3e78750626c2b4d73c9231ebc3c396645a24f
diff --git a/av1/common/arm/mem_neon.h b/av1/common/arm/mem_neon.h
index beae4ed..10e8a16 100644
--- a/av1/common/arm/mem_neon.h
+++ b/av1/common/arm/mem_neon.h
@@ -13,6 +13,7 @@
#include <arm_neon.h>
#include <string.h>
+#include "aom_dsp/aom_dsp_common.h"
static INLINE void store_row2_u8_8x8(uint8_t *s, int p, const uint8x8_t s0,
const uint8x8_t s1) {
@@ -500,4 +501,19 @@
vst1q_u32(s, s4);
}
+static INLINE int16x8_t load_tran_low_to_s16q(const tran_low_t *buf) {
+ const int32x4_t v0 = vld1q_s32(buf);
+ const int32x4_t v1 = vld1q_s32(buf + 4);
+ const int16x4_t s0 = vmovn_s32(v0);
+ const int16x4_t s1 = vmovn_s32(v1);
+ return vcombine_s16(s0, s1);
+}
+
+static INLINE void store_s16q_to_tran_low(tran_low_t *buf, const int16x8_t a) {
+ const int32x4_t v0 = vmovl_s16(vget_low_s16(a));
+ const int32x4_t v1 = vmovl_s16(vget_high_s16(a));
+ vst1q_s32(buf, v0);
+ vst1q_s32(buf + 4, v1);
+}
+
#endif // AOM_AV1_COMMON_ARM_MEM_NEON_H_
diff --git a/av1/common/av1_rtcd_defs.pl b/av1/common/av1_rtcd_defs.pl
index 3b87f60..08f7dde 100644
--- a/av1/common/av1_rtcd_defs.pl
+++ b/av1/common/av1_rtcd_defs.pl
@@ -210,7 +210,7 @@
specialize qw/av1_block_error avx2/;
add_proto qw/void av1_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, 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, const int16_t *iscan";
- specialize qw/av1_quantize_fp sse2 avx2/;
+ specialize qw/av1_quantize_fp sse2 avx2 neon/;
add_proto qw/void av1_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, 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, const int16_t *iscan";
specialize qw/av1_quantize_fp_32x32 avx2/;
diff --git a/av1/encoder/arm/neon/quantize_neon.c b/av1/encoder/arm/neon/quantize_neon.c
index 36e7d33..4c53544 100644
--- a/av1/encoder/arm/neon/quantize_neon.c
+++ b/av1/encoder/arm/neon/quantize_neon.c
@@ -17,102 +17,97 @@
#include "av1/common/quant_common.h"
#include "av1/common/seg_common.h"
+#include "av1/common/arm/mem_neon.h"
#include "av1/encoder/av1_quantize.h"
#include "av1/encoder/encoder.h"
#include "av1/encoder/rd.h"
-void av1_quantize_fp_neon(const int16_t *coeff_ptr, intptr_t count,
- int skip_block, const int16_t *zbin_ptr,
- const int16_t *round_ptr, const int16_t *quant_ptr,
- const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
- int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
- uint16_t *eob_ptr, const int16_t *scan,
- const int16_t *iscan) {
+void av1_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t count,
+ const int16_t *zbin_ptr, const int16_t *round_ptr,
+ const int16_t *quant_ptr,
+ 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, const int16_t *iscan) {
// TODO(jingning) Decide the need of these arguments after the
// quantization process is completed.
(void)zbin_ptr;
(void)quant_shift_ptr;
(void)scan;
- 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 = vld1q_s16(&coeff_ptr[0]);
- 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);
- vst1q_s16(&qcoeff_ptr[0], v_qcoeff);
- vst1q_s16(&dqcoeff_ptr[0], 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 = vld1q_s16(&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);
- vst1q_s16(&qcoeff_ptr[i], v_qcoeff);
- vst1q_s16(&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));
+ // 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[0]);
+ 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[0], v_qcoeff);
+ store_s16q_to_tran_low(&dqcoeff_ptr[0], 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(int16_t));
- memset(dqcoeff_ptr, 0, count * sizeof(int16_t));
- *eob_ptr = 0;
+ *eob_ptr = (uint16_t)vget_lane_s16(v_eobmax_final, 0);
}
}