ref: dd4cc5b596846f2cad83244477d65b694c96d315
parent: 0cc23242b035cc1c102dbeecc524592dc93cfb7d
parent: 07a0bf038f9cb57308c58ec7a1f78b8f14ed98a7
author: Kyle Siefring <kylesiefring@gmail.com>
date: Mon Nov 20 17:37:57 EST 2017
Merge "Optimize AVX2 get16x16var and get32x16var functions"
--- a/test/variance_test.cc
+++ b/test/variance_test.cc
@@ -492,7 +492,7 @@
vpx_usec_timer timer;
vpx_usec_timer_start(&timer);
- for (int i = 0; i < 100000000 / block_size(); ++i) {+ for (int i = 0; i < (1 << 30) / block_size(); ++i) {const uint32_t variance = params_.func(src_, width(), ref_, width(), &sse);
// Ignore return value.
(void)variance;
--- a/vpx_dsp/x86/variance_avx2.c
+++ b/vpx_dsp/x86/variance_avx2.c
@@ -31,6 +31,11 @@
2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14,
2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14,
};
+
+DECLARE_ALIGNED(32, static const int8_t, adjacent_sub_avx2[32]) = {+ 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1,
+ 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1
+};
/* clang-format on */
void vpx_get16x16var_avx2(const unsigned char *src_ptr, int source_stride,
@@ -37,66 +42,54 @@
const unsigned char *ref_ptr, int recon_stride,
unsigned int *sse, int *sum) {unsigned int i, src_2strides, ref_2strides;
- __m256i sum_ref_src = _mm256_setzero_si256();
- __m256i madd_ref_src = _mm256_setzero_si256();
+ __m256i sum_reg = _mm256_setzero_si256();
+ __m256i sse_reg = _mm256_setzero_si256();
// process two 16 byte locations in a 256 bit register
src_2strides = source_stride << 1;
ref_2strides = recon_stride << 1;
for (i = 0; i < 8; ++i) {- const __m256i zero_reg = _mm256_setzero_si256();
+ // convert up values in 128 bit registers across lanes
const __m256i src0 =
- _mm256_castsi128_si256(_mm_loadu_si128((__m128i const *)(src_ptr)));
- const __m256i src = _mm256_inserti128_si256(
- src0, _mm_loadu_si128((__m128i const *)(src_ptr + source_stride)), 1);
+ _mm256_cvtepu8_epi16(_mm_loadu_si128((__m128i const *)(src_ptr)));
+ const __m256i src1 = _mm256_cvtepu8_epi16(
+ _mm_loadu_si128((__m128i const *)(src_ptr + source_stride)));
const __m256i ref0 =
- _mm256_castsi128_si256(_mm_loadu_si128((__m128i const *)(ref_ptr)));
- const __m256i ref = _mm256_inserti128_si256(
- ref0, _mm_loadu_si128((__m128i const *)(ref_ptr + recon_stride)), 1);
- const __m256i src_lo = _mm256_unpacklo_epi8(src, zero_reg);
- const __m256i src_hi = _mm256_unpackhi_epi8(src, zero_reg);
- const __m256i ref_lo = _mm256_unpacklo_epi8(ref, zero_reg);
- const __m256i ref_hi = _mm256_unpackhi_epi8(ref, zero_reg);
- const __m256i diff_lo = _mm256_sub_epi16(src_lo, ref_lo);
- const __m256i diff_hi = _mm256_sub_epi16(src_hi, ref_hi);
- const __m256i madd_lo = _mm256_madd_epi16(diff_lo, diff_lo);
- const __m256i madd_hi = _mm256_madd_epi16(diff_hi, diff_hi);
- const __m256i src_ref_diff_sum = _mm256_add_epi16(diff_lo, diff_hi);
+ _mm256_cvtepu8_epi16(_mm_loadu_si128((__m128i const *)(ref_ptr)));
+ const __m256i ref1 = _mm256_cvtepu8_epi16(
+ _mm_loadu_si128((__m128i const *)(ref_ptr + recon_stride)));
+ const __m256i diff0 = _mm256_sub_epi16(src0, ref0);
+ const __m256i diff1 = _mm256_sub_epi16(src1, ref1);
+ const __m256i madd0 = _mm256_madd_epi16(diff0, diff0);
+ const __m256i madd1 = _mm256_madd_epi16(diff1, diff1);
- sum_ref_src = _mm256_add_epi16(sum_ref_src, src_ref_diff_sum);
- // add high to low
- madd_ref_src =
- _mm256_add_epi32(madd_ref_src, _mm256_add_epi32(madd_lo, madd_hi));
+ // add to the running totals
+ sum_reg = _mm256_add_epi16(sum_reg, _mm256_add_epi16(diff0, diff1));
+ sse_reg = _mm256_add_epi32(sse_reg, _mm256_add_epi32(madd0, madd1));
src_ptr += src_2strides;
ref_ptr += ref_2strides;
}
-
{- const __m128i zero_reg = _mm_setzero_si128();
// extract the low lane and add it to the high lane
- const __m128i sum_ref_src_128 =
- _mm_add_epi16(_mm256_castsi256_si128(sum_ref_src),
- _mm256_extractf128_si256(sum_ref_src, 1));
- const __m128i madd_ref_src_128 =
- _mm_add_epi32(_mm256_castsi256_si128(madd_ref_src),
- _mm256_extractf128_si256(madd_ref_src, 1));
- // 16 -> 32 sign extended
- const __m128i sum_lo =
- _mm_srai_epi32(_mm_unpacklo_epi16(zero_reg, sum_ref_src_128), 16);
- // 16 -> 32 sign extended
- const __m128i sum_hi =
- _mm_srai_epi32(_mm_unpackhi_epi16(zero_reg, sum_ref_src_128), 16);
- const __m128i sum_hl = _mm_add_epi32(sum_lo, sum_hi);
- const __m128i madd_lo = _mm_unpacklo_epi32(madd_ref_src_128, zero_reg);
- const __m128i madd_hi = _mm_unpackhi_epi32(madd_ref_src_128, zero_reg);
- const __m128i madd = _mm_add_epi32(madd_lo, madd_hi);
- const __m128i ex_sum_lo = _mm_unpacklo_epi32(sum_hl, zero_reg);
- const __m128i ex_sum_hi = _mm_unpackhi_epi32(sum_hl, zero_reg);
- const __m128i ex_sum = _mm_add_epi32(ex_sum_lo, ex_sum_hi);
- *((int *)sse) =
- _mm_cvtsi128_si32(_mm_add_epi32(madd, _mm_srli_si128(madd, 8)));
- *((int *)sum) =
- _mm_cvtsi128_si32(_mm_add_epi32(ex_sum, _mm_srli_si128(ex_sum, 8)));
+ const __m128i sum_reg_128 = _mm_add_epi16(
+ _mm256_castsi256_si128(sum_reg), _mm256_extractf128_si256(sum_reg, 1));
+ const __m128i sse_reg_128 = _mm_add_epi32(
+ _mm256_castsi256_si128(sse_reg), _mm256_extractf128_si256(sse_reg, 1));
+
+ // sum upper and lower 64 bits together and convert up to 32 bit values
+ const __m128i sum_reg_64 =
+ _mm_add_epi16(sum_reg_128, _mm_srli_si128(sum_reg_128, 8));
+ const __m128i sum_int32 = _mm_cvtepi16_epi32(sum_reg_64);
+
+ // unpack sse and sum registers and add
+ const __m128i sse_sum_lo = _mm_unpacklo_epi32(sse_reg_128, sum_int32);
+ const __m128i sse_sum_hi = _mm_unpackhi_epi32(sse_reg_128, sum_int32);
+ const __m128i sse_sum = _mm_add_epi32(sse_sum_lo, sse_sum_hi);
+
+ // perform the final summation and extract the results
+ const __m128i res = _mm_add_epi32(sse_sum, _mm_srli_si128(sse_sum, 8));
+ *((int *)sse) = _mm_cvtsi128_si32(res);
+ *((int *)sum) = _mm_extract_epi32(res, 1);
}
}
@@ -103,65 +96,69 @@
static void get32x16var_avx2(const unsigned char *src_ptr, int source_stride,
const unsigned char *ref_ptr, int recon_stride,
unsigned int *sse, int *sum) {- unsigned int i;
- const __m256i zero_reg = _mm256_setzero_si256();
- __m256i sum_ref_src = _mm256_setzero_si256();
- __m256i madd_ref_src = _mm256_setzero_si256();
+ unsigned int i, src_2strides, ref_2strides;
+ const __m256i adj_sub = _mm256_load_si256((__m256i const *)adjacent_sub_avx2);
+ __m256i sum_reg = _mm256_setzero_si256();
+ __m256i sse_reg = _mm256_setzero_si256();
- // processing 32 elements in parallel
- for (i = 0; i < 16; i++) {- const __m256i src = _mm256_loadu_si256((__m256i const *)(src_ptr));
- const __m256i ref = _mm256_loadu_si256((__m256i const *)(ref_ptr));
- const __m256i src_lo = _mm256_unpacklo_epi8(src, zero_reg);
- const __m256i src_hi = _mm256_unpackhi_epi8(src, zero_reg);
- const __m256i ref_lo = _mm256_unpacklo_epi8(ref, zero_reg);
- const __m256i ref_hi = _mm256_unpackhi_epi8(ref, zero_reg);
- const __m256i diff_lo = _mm256_sub_epi16(src_lo, ref_lo);
- const __m256i diff_hi = _mm256_sub_epi16(src_hi, ref_hi);
- const __m256i madd_lo = _mm256_madd_epi16(diff_lo, diff_lo);
- const __m256i madd_hi = _mm256_madd_epi16(diff_hi, diff_hi);
- // add high to low
- const __m256i diff_sum = _mm256_add_epi16(diff_lo, diff_hi);
+ // process 64 elements in an iteration
+ src_2strides = source_stride << 1;
+ ref_2strides = recon_stride << 1;
+ for (i = 0; i < 8; i++) {+ const __m256i src0 = _mm256_loadu_si256((__m256i const *)(src_ptr));
+ const __m256i src1 =
+ _mm256_loadu_si256((__m256i const *)(src_ptr + source_stride));
+ const __m256i ref0 = _mm256_loadu_si256((__m256i const *)(ref_ptr));
+ const __m256i ref1 =
+ _mm256_loadu_si256((__m256i const *)(ref_ptr + recon_stride));
- sum_ref_src = _mm256_add_epi16(sum_ref_src, diff_sum);
+ // unpack into pairs of source and reference values
+ const __m256i src_ref0 = _mm256_unpacklo_epi8(src0, ref0);
+ const __m256i src_ref1 = _mm256_unpackhi_epi8(src0, ref0);
+ const __m256i src_ref2 = _mm256_unpacklo_epi8(src1, ref1);
+ const __m256i src_ref3 = _mm256_unpackhi_epi8(src1, ref1);
- // add high to low
- madd_ref_src =
- _mm256_add_epi32(madd_ref_src, _mm256_add_epi32(madd_lo, madd_hi));
+ // subtract adjacent elements using src*1 + ref*-1
+ const __m256i diff0 = _mm256_maddubs_epi16(src_ref0, adj_sub);
+ const __m256i diff1 = _mm256_maddubs_epi16(src_ref1, adj_sub);
+ const __m256i diff2 = _mm256_maddubs_epi16(src_ref2, adj_sub);
+ const __m256i diff3 = _mm256_maddubs_epi16(src_ref3, adj_sub);
+ const __m256i madd0 = _mm256_madd_epi16(diff0, diff0);
+ const __m256i madd1 = _mm256_madd_epi16(diff1, diff1);
+ const __m256i madd2 = _mm256_madd_epi16(diff2, diff2);
+ const __m256i madd3 = _mm256_madd_epi16(diff3, diff3);
- src_ptr += source_stride;
- ref_ptr += recon_stride;
+ // add to the running totals
+ sum_reg = _mm256_add_epi16(sum_reg, _mm256_add_epi16(diff0, diff1));
+ sum_reg = _mm256_add_epi16(sum_reg, _mm256_add_epi16(diff2, diff3));
+ sse_reg = _mm256_add_epi32(sse_reg, _mm256_add_epi32(madd0, madd1));
+ sse_reg = _mm256_add_epi32(sse_reg, _mm256_add_epi32(madd2, madd3));
+
+ src_ptr += src_2strides;
+ ref_ptr += ref_2strides;
}
{- // 16 -> 32 sign extended
- const __m256i sum_lo =
- _mm256_srai_epi32(_mm256_unpacklo_epi16(zero_reg, sum_ref_src), 16);
- // 16 -> 32 sign extended
- const __m256i sum_hi =
- _mm256_srai_epi32(_mm256_unpackhi_epi16(zero_reg, sum_ref_src), 16);
- const __m256i sum_hl = _mm256_add_epi32(sum_lo, sum_hi);
- const __m256i madd_lo = _mm256_unpacklo_epi32(madd_ref_src, zero_reg);
- const __m256i madd_hi = _mm256_unpackhi_epi32(madd_ref_src, zero_reg);
- const __m256i madd = _mm256_add_epi32(madd_lo, madd_hi);
- const __m256i ex_sum_lo = _mm256_unpacklo_epi32(sum_hl, zero_reg);
- const __m256i ex_sum_hi = _mm256_unpackhi_epi32(sum_hl, zero_reg);
- const __m256i ex_sum = _mm256_add_epi32(ex_sum_lo, ex_sum_hi);
+ // extract the low lane and add it to the high lane
+ const __m128i sum_reg_128 = _mm_add_epi16(
+ _mm256_castsi256_si128(sum_reg), _mm256_extractf128_si256(sum_reg, 1));
+ const __m128i sse_reg_128 = _mm_add_epi32(
+ _mm256_castsi256_si128(sse_reg), _mm256_extractf128_si256(sse_reg, 1));
- // shift 8 bytes eight
- madd_ref_src = _mm256_srli_si256(madd, 8);
- sum_ref_src = _mm256_srli_si256(ex_sum, 8);
+ // sum upper and lower 64 bits together and convert up to 32 bit values
+ const __m128i sum_reg_64 =
+ _mm_add_epi16(sum_reg_128, _mm_srli_si128(sum_reg_128, 8));
+ const __m128i sum_int32 = _mm_cvtepi16_epi32(sum_reg_64);
- madd_ref_src = _mm256_add_epi32(madd_ref_src, madd);
- sum_ref_src = _mm256_add_epi32(sum_ref_src, ex_sum);
+ // unpack sse and sum registers and add
+ const __m128i sse_sum_lo = _mm_unpacklo_epi32(sse_reg_128, sum_int32);
+ const __m128i sse_sum_hi = _mm_unpackhi_epi32(sse_reg_128, sum_int32);
+ const __m128i sse_sum = _mm_add_epi32(sse_sum_lo, sse_sum_hi);
- // extract the low lane and the high lane and add the results
- *((int *)sse) =
- _mm_cvtsi128_si32(_mm256_castsi256_si128(madd_ref_src)) +
- _mm_cvtsi128_si32(_mm256_extractf128_si256(madd_ref_src, 1));
-
- *((int *)sum) = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_ref_src)) +
- _mm_cvtsi128_si32(_mm256_extractf128_si256(sum_ref_src, 1));
+ // perform the final summation and extract the results
+ const __m128i res = _mm_add_epi32(sse_sum, _mm_srli_si128(sse_sum, 8));
+ *((int *)sse) = _mm_cvtsi128_si32(res);
+ *((int *)sum) = _mm_extract_epi32(res, 1);
}
}