neon: vpx_quantize_b_32x32

With skip block the neon is about twice as fast as C.

The neon has no shortcut for coeff < zbin so it always takes the
same amount of time. Even if the C can take the shortcut, it is over
twice as fast in neon. If it can't, that gap increases to over 10x.

BUG=webm:1426

Change-Id: I400722146c1b5a5f6289f67d85fd642463d2bfc6

diff --git a/test/vp9_quantize_test.cc b/test/vp9_quantize_test.cc
index 7ec8f58..e398535 100644 (file)
--- a/test/vp9_quantize_test.cc
+++ b/test/vp9_quantize_test.cc
@@ -358,15 +358,18 @@ INSTANTIATE_TEST_CASE_P(DISABLED_AVX, VP9QuantizeTest,
 
 // TODO(webm:1448): dqcoeff is not handled correctly in HBD builds.
 #if HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH
-INSTANTIATE_TEST_CASE_P(NEON, VP9QuantizeTest,
-                        ::testing::Values(make_tuple(&vpx_quantize_b_neon,
-                                                     &vpx_quantize_b_c,
-                                                     VPX_BITS_8, 16)));
+INSTANTIATE_TEST_CASE_P(
+    NEON, VP9QuantizeTest,
+    ::testing::Values(make_tuple(&vpx_quantize_b_neon, &vpx_quantize_b_c,
+                                 VPX_BITS_8, 16),
+                      make_tuple(&vpx_quantize_b_32x32_neon,
+                                 &vpx_quantize_b_32x32_c, VPX_BITS_8, 32)));
 #endif  // HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH
 
 // Only useful to compare "Speed" test results.
-INSTANTIATE_TEST_CASE_P(DISABLED_C, VP9QuantizeTest,
-                        ::testing::Values(make_tuple(&vpx_quantize_b_c,
-                                                     &vpx_quantize_b_c,
-                                                     VPX_BITS_8, 16)));
+INSTANTIATE_TEST_CASE_P(
+    DISABLED_C, VP9QuantizeTest,
+    ::testing::Values(
+        make_tuple(&vpx_quantize_b_c, &vpx_quantize_b_c, VPX_BITS_8, 16),
+        make_tuple(&vpx_quantize_b_c, &vpx_quantize_b_c, VPX_BITS_8, 32)));
 }  // namespace

diff --git a/vpx_dsp/arm/quantize_neon.c b/vpx_dsp/arm/quantize_neon.c
index 8bed057..e62933a 100644 (file)
--- a/vpx_dsp/arm/quantize_neon.c
+++ b/vpx_dsp/arm/quantize_neon.c
@@ -99,7 +99,7 @@ void vpx_quantize_b_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
     const int16x8_t dequant = vdupq_n_s16(dequant_ptr[1]);
 
     do {
-      // Add one because the eob is not it's index.
+      // Add one because the eob is not its index.
       const uint16x8_t iscan =
           vreinterpretq_u16_s16(vaddq_s16(vld1q_s16(iscan_ptr), one));
 
@@ -153,3 +153,176 @@ void vpx_quantize_b_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
     vst1_lane_u16(eob_ptr, eob_max_2, 0);
   }
 }
+
+// Main difference is that zbin values are halved before comparison and dqcoeff
+// values are divided by 2. zbin is rounded but dqcoeff is not.
+void vpx_quantize_b_32x32_neon(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block,
+    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_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;
+  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;
+  }
+
+  // Process first 8 values which include a dc component.
+  {
+    // Only the first element of each vector is DC.
+    const int16x8_t zbin = vrshrq_n_s16(vld1q_s16(zbin_ptr), 1);
+    const int16x8_t round = vrshrq_n_s16(vld1q_s16(round_ptr), 1);
+    const int16x8_t quant = vld1q_s16(quant_ptr);
+    const int16x8_t quant_shift = vld1q_s16(quant_shift_ptr);
+    const int16x8_t dequant = vld1q_s16(dequant_ptr);
+    // Add one because the eob does not index from 0.
+    const uint16x8_t iscan =
+        vreinterpretq_u16_s16(vaddq_s16(vld1q_s16(iscan_ptr), one));
+
+    const int16x8_t coeff = load_tran_low_to_s16q(coeff_ptr);
+    const int16x8_t coeff_sign = vshrq_n_s16(coeff, 15);
+    const int16x8_t coeff_abs = vabsq_s16(coeff);
+
+    const int16x8_t zbin_mask =
+        vreinterpretq_s16_u16(vcgeq_s16(coeff_abs, zbin));
+
+    const int16x8_t rounded = vqaddq_s16(coeff_abs, round);
+
+    // (round * quant * 2) >> 16 >> 1 == (round * quant) >> 16
+    int16x8_t qcoeff = vshrq_n_s16(vqdmulhq_s16(rounded, quant), 1);
+    int16x8_t dqcoeff;
+    int32x4_t dqcoeff_0, dqcoeff_1, dqcoeff_0_sign, dqcoeff_1_sign;
+
+    qcoeff = vaddq_s16(qcoeff, rounded);
+
+    // (qcoeff * quant_shift * 2) >> 16 == (qcoeff * quant_shift) >> 15
+    qcoeff = vqdmulhq_s16(qcoeff, quant_shift);
+
+    // Restore the sign bit.
+    qcoeff = veorq_s16(qcoeff, coeff_sign);
+    qcoeff = vsubq_s16(qcoeff, coeff_sign);
+
+    qcoeff = vandq_s16(qcoeff, zbin_mask);
+
+    // Set non-zero elements to -1 and use that to extract values for eob.
+    eob_max = vandq_u16(vtstq_s16(qcoeff, neg_one), iscan);
+
+    coeff_ptr += 8;
+    iscan_ptr += 8;
+
+    store_s16q_to_tran_low(qcoeff_ptr, qcoeff);
+    qcoeff_ptr += 8;
+
+    dqcoeff_0 = vmull_s16(vget_low_s16(qcoeff), vget_low_s16(dequant));
+    dqcoeff_1 = vmull_s16(vget_high_s16(qcoeff), vget_high_s16(dequant));
+
+    // The way the C shifts the values requires us to convert to positive before
+    // shifting or even narrowing, then put the sign back.
+    dqcoeff_0_sign = vshrq_n_s32(dqcoeff_0, 31);
+    dqcoeff_1_sign = vshrq_n_s32(dqcoeff_1, 31);
+    dqcoeff_0 = vabsq_s32(dqcoeff_0);
+    dqcoeff_1 = vabsq_s32(dqcoeff_1);
+    dqcoeff_0 = vshrq_n_s32(dqcoeff_0, 1);
+    dqcoeff_1 = vshrq_n_s32(dqcoeff_1, 1);
+    dqcoeff_0 = veorq_s32(dqcoeff_0, dqcoeff_0_sign);
+    dqcoeff_1 = veorq_s32(dqcoeff_1, dqcoeff_1_sign);
+    dqcoeff_0 = vsubq_s32(dqcoeff_0, dqcoeff_0_sign);
+    dqcoeff_1 = vsubq_s32(dqcoeff_1, dqcoeff_1_sign);
+
+    // Narrow *without saturation* because that's what the C does.
+    dqcoeff = vcombine_s16(vmovn_s32(dqcoeff_0), vmovn_s32(dqcoeff_1));
+
+    store_s16q_to_tran_low(dqcoeff_ptr, dqcoeff);
+    dqcoeff_ptr += 8;
+  }
+
+  {
+    const int16x8_t zbin = vrshrq_n_s16(vdupq_n_s16(zbin_ptr[1]), 1);
+    const int16x8_t round = vrshrq_n_s16(vdupq_n_s16(round_ptr[1]), 1);
+    const int16x8_t quant = vdupq_n_s16(quant_ptr[1]);
+    const int16x8_t quant_shift = vdupq_n_s16(quant_shift_ptr[1]);
+    const int16x8_t dequant = vdupq_n_s16(dequant_ptr[1]);
+
+    for (i = 1; i < 32 * 32 / 8; ++i) {
+      // Add one because the eob is not its index.
+      const uint16x8_t iscan =
+          vreinterpretq_u16_s16(vaddq_s16(vld1q_s16(iscan_ptr), one));
+
+      const int16x8_t coeff = load_tran_low_to_s16q(coeff_ptr);
+      const int16x8_t coeff_sign = vshrq_n_s16(coeff, 15);
+      const int16x8_t coeff_abs = vabsq_s16(coeff);
+
+      const int16x8_t zbin_mask =
+          vreinterpretq_s16_u16(vcgeq_s16(coeff_abs, zbin));
+
+      const int16x8_t rounded = vqaddq_s16(coeff_abs, round);
+
+      // (round * quant * 2) >> 16 >> 1 == (round * quant) >> 16
+      int16x8_t qcoeff = vshrq_n_s16(vqdmulhq_s16(rounded, quant), 1);
+      int16x8_t dqcoeff;
+      int32x4_t dqcoeff_0, dqcoeff_1, dqcoeff_0_sign, dqcoeff_1_sign;
+
+      qcoeff = vaddq_s16(qcoeff, rounded);
+
+      // (qcoeff * quant_shift * 2) >> 16 == (qcoeff * quant_shift) >> 15
+      qcoeff = vqdmulhq_s16(qcoeff, quant_shift);
+
+      // Restore the sign bit.
+      qcoeff = veorq_s16(qcoeff, coeff_sign);
+      qcoeff = vsubq_s16(qcoeff, coeff_sign);
+
+      qcoeff = vandq_s16(qcoeff, zbin_mask);
+
+      // Set non-zero elements to -1 and use that to extract values for eob.
+      eob_max =
+          vmaxq_u16(eob_max, vandq_u16(vtstq_s16(qcoeff, neg_one), iscan));
+
+      coeff_ptr += 8;
+      iscan_ptr += 8;
+
+      store_s16q_to_tran_low(qcoeff_ptr, qcoeff);
+      qcoeff_ptr += 8;
+
+      dqcoeff_0 = vmull_s16(vget_low_s16(qcoeff), vget_low_s16(dequant));
+      dqcoeff_1 = vmull_s16(vget_high_s16(qcoeff), vget_high_s16(dequant));
+
+      dqcoeff_0_sign = vshrq_n_s32(dqcoeff_0, 31);
+      dqcoeff_1_sign = vshrq_n_s32(dqcoeff_1, 31);
+      dqcoeff_0 = vabsq_s32(dqcoeff_0);
+      dqcoeff_1 = vabsq_s32(dqcoeff_1);
+      dqcoeff_0 = vshrq_n_s32(dqcoeff_0, 1);
+      dqcoeff_1 = vshrq_n_s32(dqcoeff_1, 1);
+      dqcoeff_0 = veorq_s32(dqcoeff_0, dqcoeff_0_sign);
+      dqcoeff_1 = veorq_s32(dqcoeff_1, dqcoeff_1_sign);
+      dqcoeff_0 = vsubq_s32(dqcoeff_0, dqcoeff_0_sign);
+      dqcoeff_1 = vsubq_s32(dqcoeff_1, dqcoeff_1_sign);
+
+      dqcoeff = vcombine_s16(vmovn_s32(dqcoeff_0), vmovn_s32(dqcoeff_1));
+
+      store_s16q_to_tran_low(dqcoeff_ptr, dqcoeff);
+      dqcoeff_ptr += 8;
+    }
+  }
+
+  {
+    const uint16x4_t eob_max_0 =
+        vmax_u16(vget_low_u16(eob_max), vget_high_u16(eob_max));
+    const uint16x4_t eob_max_1 = vpmax_u16(eob_max_0, eob_max_0);
+    const uint16x4_t eob_max_2 = vpmax_u16(eob_max_1, eob_max_1);
+    vst1_lane_u16(eob_ptr, eob_max_2, 0);
+  }
+}

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 498e937..c542474 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -673,7 +673,7 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
   specialize qw/vpx_quantize_b neon sse2/, "$ssse3_x86_64", "$avx_x86_64";
 
   add_proto qw/void vpx_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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/vpx_quantize_b_32x32/, "$ssse3_x86_64", "$avx_x86_64";
+  specialize qw/vpx_quantize_b_32x32 neon/, "$ssse3_x86_64", "$avx_x86_64";
 
   if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
     add_proto qw/void vpx_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, 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";