Remove pair quantization

The intrinsics version of the pair quant is slower than running it
individually.

Change-Id: I7b4ea8599d4aab04be0a5a0c59b8b29a7fc283f4

1 files changed, 4 insertions, 120 deletions

diff --git a/vp8/encoder/arm/neon/fastquantizeb_neon.c b/vp8/encoder/arm/neon/fastquantizeb_neon.c
index 48764284f..caa763765 100644
--- a/vp8/encoder/arm/neon/fastquantizeb_neon.c
+++ b/vp8/encoder/arm/neon/fastquantizeb_neon.c
@@ -10,13 +10,12 @@
 
 #include <arm_neon.h>
 #include "vp8/encoder/block.h"
-#include "vpx_mem/vpx_mem.h"
 
 static const uint16_t inv_zig_zag[16] = {
-    0x0001, 0x0002, 0x0006, 0x0007,
-    0x0003, 0x0005, 0x0008, 0x000d,
-    0x0004, 0x0009, 0x000c, 0x000e,
-    0x000a, 0x000b, 0x000f, 0x0010
+    1,  2,  6,   7,
+    3,  5,  8,  13,
+    4,  9,  12, 14,
+    10, 11, 15, 16
 };
 
 void vp8_fast_quantize_b_neon(BLOCK *b, BLOCKD *d) {
@@ -88,118 +87,3 @@ void vp8_fast_quantize_b_neon(BLOCK *b, BLOCKD *d) {
 
     vst1_lane_s8((int8_t *)d->eob, vreinterpret_s8_u32(eob_d32), 0);
 }
-
-void vp8_fast_quantize_b_pair_neon(BLOCK *b0, BLOCK *b1,
-                                   BLOCKD *d0, BLOCKD *d1) {
-    const int16x8_t one_q = vdupq_n_s16(0xff),
-                    b0_z0 = vld1q_s16(b0->coeff),
-                    b0_z1 = vld1q_s16(b0->coeff + 8),
-                    b0_round0 = vld1q_s16(b0->round),
-                    b0_round1 = vld1q_s16(b0->round + 8),
-                    b0_quant0 = vld1q_s16(b0->quant_fast),
-                    b0_quant1 = vld1q_s16(b0->quant_fast + 8),
-                    d0_dequant0 = vld1q_s16(d0->dequant),
-                    d0_dequant1 = vld1q_s16(d0->dequant + 8),
-                    b1_z0 = vld1q_s16(b1->coeff),
-                    b1_z1 = vld1q_s16(b1->coeff + 8),
-                    b1_round0 = vld1q_s16(b1->round),
-                    b1_round1 = vld1q_s16(b1->round + 8),
-                    b1_quant0 = vld1q_s16(b1->quant_fast),
-                    b1_quant1 = vld1q_s16(b1->quant_fast + 8),
-                    d1_dequant0 = vld1q_s16(d1->dequant),
-                    d1_dequant1 = vld1q_s16(d1->dequant + 8);
-    const uint16x8_t zig_zag0 = vld1q_u16(inv_zig_zag),
-                     zig_zag1 = vld1q_u16(inv_zig_zag + 8);
-    int16x8_t b0_x0, b0_x1, b0_sz0, b0_sz1, b0_y0, b0_y1,
-              b1_x0, b1_x1, b1_sz0, b1_sz1, b1_y0, b1_y1;
-    uint16x8_t b0_eob0, b0_eob1,
-               b1_eob0, b1_eob1;
-    uint16x4_t b0_eob_d16, b1_eob_d16;
-    uint32x2_t b0_eob_d32, b1_eob_d32;
-    uint32x4_t b0_eob_q32, b1_eob_q32;
-
-    /* sign of z: z >> 15 */
-    b0_sz0 = vshrq_n_s16(b0_z0, 15);
-    b0_sz1 = vshrq_n_s16(b0_z1, 15);
-    b1_sz0 = vshrq_n_s16(b1_z0, 15);
-    b1_sz1 = vshrq_n_s16(b1_z1, 15);
-
-    /* x = abs(z) */
-    b0_x0 = vabsq_s16(b0_z0);
-    b0_x1 = vabsq_s16(b0_z1);
-    b1_x0 = vabsq_s16(b1_z0);
-    b1_x1 = vabsq_s16(b1_z1);
-
-    /* x += round */
-    b0_x0 = vaddq_s16(b0_x0, b0_round0);
-    b0_x1 = vaddq_s16(b0_x1, b0_round1);
-    b1_x0 = vaddq_s16(b1_x0, b1_round0);
-    b1_x1 = vaddq_s16(b1_x1, b1_round1);
-
-    /* y = 2 * (x * quant) >> 16 */
-    b0_y0 = vqdmulhq_s16(b0_x0, b0_quant0);
-    b0_y1 = vqdmulhq_s16(b0_x1, b0_quant1);
-    b1_y0 = vqdmulhq_s16(b1_x0, b1_quant0);
-    b1_y1 = vqdmulhq_s16(b1_x1, b1_quant1);
-
-    /* Compensate for doubling in vqdmulhq */
-    b0_y0 = vshrq_n_s16(b0_y0, 1);
-    b0_y1 = vshrq_n_s16(b0_y1, 1);
-    b1_y0 = vshrq_n_s16(b1_y0, 1);
-    b1_y1 = vshrq_n_s16(b1_y1, 1);
-
-    /* Restore sign bit */
-    b0_y0 = veorq_s16(b0_y0, b0_sz0);
-    b0_y1 = veorq_s16(b0_y1, b0_sz1);
-    b0_x0 = vsubq_s16(b0_y0, b0_sz0);
-    b0_x1 = vsubq_s16(b0_y1, b0_sz1);
-    b1_y0 = veorq_s16(b1_y0, b1_sz0);
-    b1_y1 = veorq_s16(b1_y1, b1_sz1);
-    b1_x0 = vsubq_s16(b1_y0, b1_sz0);
-    b1_x1 = vsubq_s16(b1_y1, b1_sz1);
-
-    /* find non-zero elements */
-    b0_eob0 = vtstq_s16(b0_x0, one_q);
-    b0_eob1 = vtstq_s16(b0_x1, one_q);
-    b1_eob0 = vtstq_s16(b1_x0, one_q);
-    b1_eob1 = vtstq_s16(b1_x1, one_q);
-
-    /* mask zig zag */
-    b0_eob0 = vandq_u16(b0_eob0, zig_zag0);
-    b0_eob1 = vandq_u16(b0_eob1, zig_zag1);
-    b1_eob0 = vandq_u16(b1_eob0, zig_zag0);
-    b1_eob1 = vandq_u16(b1_eob1, zig_zag1);
-
-    /* select the largest value */
-    b0_eob0 = vmaxq_u16(b0_eob0, b0_eob1);
-    b0_eob_d16 = vmax_u16(vget_low_u16(b0_eob0),
-                          vget_high_u16(b0_eob0));
-    b0_eob_q32 = vmovl_u16(b0_eob_d16);
-    b0_eob_d32 = vmax_u32(vget_low_u32(b0_eob_q32),
-                          vget_high_u32(b0_eob_q32));
-    b0_eob_d32 = vpmax_u32(b0_eob_d32, b0_eob_d32);
-
-    b1_eob0 = vmaxq_u16(b1_eob0, b1_eob1);
-    b1_eob_d16 = vmax_u16(vget_low_u16(b1_eob0),
-                          vget_high_u16(b1_eob0));
-    b1_eob_q32 = vmovl_u16(b1_eob_d16);
-    b1_eob_d32 = vmax_u32(vget_low_u32(b1_eob_q32),
-                          vget_high_u32(b1_eob_q32));
-    b1_eob_d32 = vpmax_u32(b1_eob_d32, b1_eob_d32);
-
-    /* qcoeff = x */
-    vst1q_s16(d0->qcoeff, b0_x0);
-    vst1q_s16(d0->qcoeff + 8, b0_x1);
-    vst1q_s16(d1->qcoeff, b1_x0);
-    vst1q_s16(d1->qcoeff + 8, b1_x1);
-
-    /* dqcoeff = x * dequant */
-    vst1q_s16(d0->dqcoeff, vmulq_s16(d0_dequant0, b0_x0));
-    vst1q_s16(d0->dqcoeff + 8, vmulq_s16(d0_dequant1, b0_x1));
-    vst1q_s16(d1->dqcoeff, vmulq_s16(d1_dequant0, b1_x0));
-    vst1q_s16(d1->dqcoeff + 8, vmulq_s16(d1_dequant1, b1_x1));
-
-    vst1_lane_s8((int8_t *)d0->eob, vreinterpret_s8_u32(b0_eob_d32), 0);
-    vst1_lane_s8((int8_t *)d1->eob, vreinterpret_s8_u32(b1_eob_d32), 0);
-    return;
-}