Mirrors/overte
3
0
Fork 0
mirror of https://github.com/overte-org/overte.git synced 2025-04-11 14:53:59 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

Full SIMD implementation of packBlendshapeOffsets() using AVX2.

Browse source 
6x speedup over the existing (partial SIMD) version.
60x speedup over the original (pure GLM) version.
This commit is contained in:
Ken Cooke 2019-06-23 10:32:57 -07:00
parent cceff21cd0
commit 8653118b6c
2 changed files with 294 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
libraries/render-utils/src/Model.cpp
View file

@ -1708,7 +1708,7 @@ static void packBlendshapeOffsetTo_Pos_F32_3xSN10_Nor_3xSN10_Tan_3xSN10(glm::uve
);
}
static void packBlendshapeOffsets(BlendshapeOffsetUnpacked* unpacked, BlendshapeOffsetPacked* packed, int size) {
static void packBlendshapeOffsets_ref(BlendshapeOffsetUnpacked* unpacked, BlendshapeOffsetPacked* packed, int size) {
for (int i = 0; i < size; ++i) {
packBlendshapeOffsetTo_Pos_F32_3xSN10_Nor_3xSN10_Tan_3xSN10((*packed).packedPosNorTan, (*unpacked));
++unpacked;
@ -1716,6 +1716,14 @@ static void packBlendshapeOffsets(BlendshapeOffsetUnpacked* unpacked, Blendshape
}
}
void packBlendshapeOffsets_AVX2(float (*unpacked)[9], uint32_t (*packed)[4], int size);
static void packBlendshapeOffsets(BlendshapeOffsetUnpacked* unpacked, BlendshapeOffsetPacked* packed, int size) {
static_assert(sizeof(BlendshapeOffsetUnpacked) == 9 * sizeof(float), "struct BlendshapeOffsetUnpacked size doesn't match.");
static_assert(sizeof(BlendshapeOffsetPacked) == 4 * sizeof(uint32_t), "struct BlendshapeOffsetPacked size doesn't match.");
packBlendshapeOffsets_AVX2((float(*)[9])unpacked, (uint32_t(*)[4])packed, size);
}
class Blender : public QRunnable {
public:

285
libraries/render-utils/src/avx2/BlendshapePacking_avx2.cpp Normal file
View file

@ -0,0 +1,285 @@
//
// BlendshapePacking_avx2.cpp
//
// Created by Ken Cooke on 6/22/19.
// Copyright 2019 High Fidelity, Inc.
//
// Distributed under the Apache License, Version 2.0.
// See the accompanying file LICENSE or http://www.apache.org/licenses/LICENSE-2.0.html
//
#ifdef __AVX2__
#include <stdint.h>
#include <immintrin.h>
void packBlendshapeOffsets_AVX2(float (*unpacked)[9], uint32_t (*packed)[4], int size) {
int i = 0;
for (; i < size - 7; i += 8) { // blocks of 8
//
// deinterleave (8x9 to 9x8 matrix transpose)
//
__m256 s0 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&unpacked[i+0][0])), _mm_load_ps(&unpacked[i+4][0]), 1);
__m256 s1 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&unpacked[i+1][0])), _mm_load_ps(&unpacked[i+5][0]), 1);
__m256 s2 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&unpacked[i+2][0])), _mm_load_ps(&unpacked[i+6][0]), 1);
__m256 s3 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&unpacked[i+3][0])), _mm_load_ps(&unpacked[i+7][0]), 1);
__m256 s4 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&unpacked[i+0][4])), _mm_load_ps(&unpacked[i+4][4]), 1);
__m256 s5 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&unpacked[i+1][4])), _mm_load_ps(&unpacked[i+5][4]), 1);
__m256 s6 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&unpacked[i+2][4])), _mm_load_ps(&unpacked[i+6][4]), 1);
__m256 s7 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(&unpacked[i+3][4])), _mm_load_ps(&unpacked[i+7][4]), 1);
__m256 t0 = _mm256_unpacklo_ps(s0, s1);
__m256 t1 = _mm256_unpackhi_ps(s0, s1);
__m256 t2 = _mm256_unpacklo_ps(s2, s3);
__m256 t3 = _mm256_unpackhi_ps(s2, s3);
__m256 t4 = _mm256_unpacklo_ps(s4, s5);
__m256 t5 = _mm256_unpackhi_ps(s4, s5);
__m256 t6 = _mm256_unpacklo_ps(s6, s7);
__m256 t7 = _mm256_unpackhi_ps(s6, s7);
__m256 px = _mm256_shuffle_ps(t0, t2, _MM_SHUFFLE(1,0,1,0));
__m256 py = _mm256_shuffle_ps(t0, t2, _MM_SHUFFLE(3,2,3,2));
__m256 pz = _mm256_shuffle_ps(t1, t3, _MM_SHUFFLE(1,0,1,0));
__m256 nx = _mm256_shuffle_ps(t1, t3, _MM_SHUFFLE(3,2,3,2));
__m256 ny = _mm256_shuffle_ps(t4, t6, _MM_SHUFFLE(1,0,1,0));
__m256 nz = _mm256_shuffle_ps(t4, t6, _MM_SHUFFLE(3,2,3,2));
__m256 tx = _mm256_shuffle_ps(t5, t7, _MM_SHUFFLE(1,0,1,0));
__m256 ty = _mm256_shuffle_ps(t5, t7, _MM_SHUFFLE(3,2,3,2));
__m256 tz = _mm256_i32gather_ps(unpacked[i+0], _mm256_setr_epi32(8,17,26,35,44,53,62,71), sizeof(float));
// abs(pos)
__m256 apx = _mm256_andnot_ps(_mm256_set1_ps(-0.0f), px);
__m256 apy = _mm256_andnot_ps(_mm256_set1_ps(-0.0f), py);
__m256 apz = _mm256_andnot_ps(_mm256_set1_ps(-0.0f), pz);
// len = compMax(abs(pos))
__m256 len = _mm256_max_ps(_mm256_max_ps(apx, apy), apz);
// detect zeros
__m256 mask = _mm256_cmp_ps(len, _mm256_setzero_ps(), _CMP_EQ_OQ);
// rcp = 1.0f / len
__m256 rcp = _mm256_div_ps(_mm256_set1_ps(1.0f), len);
// replace +inf with 1.0f
rcp = _mm256_blendv_ps(rcp, _mm256_set1_ps(1.0f), mask);
len = _mm256_blendv_ps(len, _mm256_set1_ps(1.0f), mask);
// pos *= 1.0f / len
px = _mm256_mul_ps(px, rcp);
py = _mm256_mul_ps(py, rcp);
pz = _mm256_mul_ps(pz, rcp);
// clamp(vec, -1.0f, 1.0f)
px = _mm256_min_ps(_mm256_max_ps(px, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
py = _mm256_min_ps(_mm256_max_ps(py, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
pz = _mm256_min_ps(_mm256_max_ps(pz, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
nx = _mm256_min_ps(_mm256_max_ps(nx, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
ny = _mm256_min_ps(_mm256_max_ps(ny, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
nz = _mm256_min_ps(_mm256_max_ps(nz, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
tx = _mm256_min_ps(_mm256_max_ps(tx, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
ty = _mm256_min_ps(_mm256_max_ps(ty, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
tz = _mm256_min_ps(_mm256_max_ps(tz, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
// vec *= 511.0f
px = _mm256_mul_ps(px, _mm256_set1_ps(511.0f));
py = _mm256_mul_ps(py, _mm256_set1_ps(511.0f));
pz = _mm256_mul_ps(pz, _mm256_set1_ps(511.0f));
nx = _mm256_mul_ps(nx, _mm256_set1_ps(511.0f));
ny = _mm256_mul_ps(ny, _mm256_set1_ps(511.0f));
nz = _mm256_mul_ps(nz, _mm256_set1_ps(511.0f));
tx = _mm256_mul_ps(tx, _mm256_set1_ps(511.0f));
ty = _mm256_mul_ps(ty, _mm256_set1_ps(511.0f));
tz = _mm256_mul_ps(tz, _mm256_set1_ps(511.0f));
// veci = lrint(vec) & 03ff
__m256i pxi = _mm256_and_si256(_mm256_cvtps_epi32(px), _mm256_set1_epi32(0x3ff));
__m256i pyi = _mm256_and_si256(_mm256_cvtps_epi32(py), _mm256_set1_epi32(0x3ff));
__m256i pzi = _mm256_and_si256(_mm256_cvtps_epi32(pz), _mm256_set1_epi32(0x3ff));
__m256i nxi = _mm256_and_si256(_mm256_cvtps_epi32(nx), _mm256_set1_epi32(0x3ff));
__m256i nyi = _mm256_and_si256(_mm256_cvtps_epi32(ny), _mm256_set1_epi32(0x3ff));
__m256i nzi = _mm256_and_si256(_mm256_cvtps_epi32(nz), _mm256_set1_epi32(0x3ff));
__m256i txi = _mm256_and_si256(_mm256_cvtps_epi32(tx), _mm256_set1_epi32(0x3ff));
__m256i tyi = _mm256_and_si256(_mm256_cvtps_epi32(ty), _mm256_set1_epi32(0x3ff));
__m256i tzi = _mm256_and_si256(_mm256_cvtps_epi32(tz), _mm256_set1_epi32(0x3ff));
// pack = (xi << 0) | (yi << 10) | (zi << 20);
__m256i li = _mm256_castps_si256(len); // length
__m256i pi = _mm256_or_si256(_mm256_or_si256(pxi, _mm256_slli_epi32(pyi, 10)), _mm256_slli_epi32(pzi, 20)); // position
__m256i ni = _mm256_or_si256(_mm256_or_si256(nxi, _mm256_slli_epi32(nyi, 10)), _mm256_slli_epi32(nzi, 20)); // normal
__m256i ti = _mm256_or_si256(_mm256_or_si256(txi, _mm256_slli_epi32(tyi, 10)), _mm256_slli_epi32(tzi, 20)); // tangent
//
// interleave (4x4 matrix transpose)
//
__m256i u0 = _mm256_unpacklo_epi32(li, pi);
__m256i u1 = _mm256_unpackhi_epi32(li, pi);
__m256i u2 = _mm256_unpacklo_epi32(ni, ti);
__m256i u3 = _mm256_unpackhi_epi32(ni, ti);
__m256i v0 = _mm256_unpacklo_epi64(u0, u2);
__m256i v1 = _mm256_unpackhi_epi64(u0, u2);
__m256i v2 = _mm256_unpacklo_epi64(u1, u3);
__m256i v3 = _mm256_unpackhi_epi64(u1, u3);
__m256i w0 = _mm256_permute2f128_si256(v0, v1, 0x20);
__m256i w1 = _mm256_permute2f128_si256(v2, v3, 0x20);
__m256i w2 = _mm256_permute2f128_si256(v0, v1, 0x31);
__m256i w3 = _mm256_permute2f128_si256(v2, v3, 0x31);
// store pack x 8
_mm256_storeu_si256((__m256i*)packed[i+0], w0);
_mm256_storeu_si256((__m256i*)packed[i+2], w1);
_mm256_storeu_si256((__m256i*)packed[i+4], w2);
_mm256_storeu_si256((__m256i*)packed[i+6], w3);
}
if (i < size) { // remainder
int rem = size - i;
//
// deinterleave (8x9 to 9x8 matrix transpose)
//
__m256 s0 = _mm256_setzero_ps();
__m256 s1 = _mm256_setzero_ps();
__m256 s2 = _mm256_setzero_ps();
__m256 s3 = _mm256_setzero_ps();
__m256 s4 = _mm256_setzero_ps();
__m256 s5 = _mm256_setzero_ps();
__m256 s6 = _mm256_setzero_ps();
__m256 s7 = _mm256_setzero_ps();
switch (rem) {
case 7: s6 = _mm256_loadu_ps(unpacked[i+6]);
case 6: s5 = _mm256_loadu_ps(unpacked[i+5]);
case 5: s4 = _mm256_loadu_ps(unpacked[i+4]);
case 4: s3 = _mm256_loadu_ps(unpacked[i+3]);
case 3: s2 = _mm256_loadu_ps(unpacked[i+2]);
case 2: s1 = _mm256_loadu_ps(unpacked[i+1]);
case 1: s0 = _mm256_loadu_ps(unpacked[i+0]);
}
__m256 t0 = _mm256_unpacklo_ps(s0, s1);
__m256 t1 = _mm256_unpackhi_ps(s0, s1);
__m256 t2 = _mm256_unpacklo_ps(s2, s3);
__m256 t3 = _mm256_unpackhi_ps(s2, s3);
__m256 t4 = _mm256_unpacklo_ps(s4, s5);
__m256 t5 = _mm256_unpackhi_ps(s4, s5);
__m256 t6 = _mm256_unpacklo_ps(s6, s7);
__m256 t7 = _mm256_unpackhi_ps(s6, s7);
s0 = _mm256_shuffle_ps(t0, t2, _MM_SHUFFLE(1,0,1,0));
s1 = _mm256_shuffle_ps(t0, t2, _MM_SHUFFLE(3,2,3,2));
s2 = _mm256_shuffle_ps(t1, t3, _MM_SHUFFLE(1,0,1,0));
s3 = _mm256_shuffle_ps(t1, t3, _MM_SHUFFLE(3,2,3,2));
s4 = _mm256_shuffle_ps(t4, t6, _MM_SHUFFLE(1,0,1,0));
s5 = _mm256_shuffle_ps(t4, t6, _MM_SHUFFLE(3,2,3,2));
s6 = _mm256_shuffle_ps(t5, t7, _MM_SHUFFLE(1,0,1,0));
s7 = _mm256_shuffle_ps(t5, t7, _MM_SHUFFLE(3,2,3,2));
__m256 px = _mm256_permute2f128_ps(s0, s4, 0x20);
__m256 py = _mm256_permute2f128_ps(s1, s5, 0x20);
__m256 pz = _mm256_permute2f128_ps(s2, s6, 0x20);
__m256 nx = _mm256_permute2f128_ps(s3, s7, 0x20);
__m256 ny = _mm256_permute2f128_ps(s0, s4, 0x31);
__m256 nz = _mm256_permute2f128_ps(s1, s5, 0x31);
__m256 tx = _mm256_permute2f128_ps(s2, s6, 0x31);
__m256 ty = _mm256_permute2f128_ps(s3, s7, 0x31);
__m256i loadmask = _mm256_cvtepi8_epi32(_mm_cvtsi64_si128(0xffffffffffffffffULL >> (64 - 8 * rem)));
__m256 tz = _mm256_mask_i32gather_ps(_mm256_setzero_ps(), unpacked[i+0], _mm256_setr_epi32(8,17,26,35,44,53,62,71),
_mm256_castsi256_ps(loadmask), sizeof(float));
// abs(pos)
__m256 apx = _mm256_andnot_ps(_mm256_set1_ps(-0.0f), px);
__m256 apy = _mm256_andnot_ps(_mm256_set1_ps(-0.0f), py);
__m256 apz = _mm256_andnot_ps(_mm256_set1_ps(-0.0f), pz);
// len = compMax(abs(pos))
__m256 len = _mm256_max_ps(_mm256_max_ps(apx, apy), apz);
// detect zeros
__m256 mask = _mm256_cmp_ps(len, _mm256_setzero_ps(), _CMP_EQ_OQ);
// rcp = 1.0f / len
__m256 rcp = _mm256_div_ps(_mm256_set1_ps(1.0f), len);
// replace +inf with 1.0f
rcp = _mm256_blendv_ps(rcp, _mm256_set1_ps(1.0f), mask);
len = _mm256_blendv_ps(len, _mm256_set1_ps(1.0f), mask);
// pos *= 1.0f / len
px = _mm256_mul_ps(px, rcp);
py = _mm256_mul_ps(py, rcp);
pz = _mm256_mul_ps(pz, rcp);
// clamp(vec, -1.0f, 1.0f)
px = _mm256_min_ps(_mm256_max_ps(px, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
py = _mm256_min_ps(_mm256_max_ps(py, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
pz = _mm256_min_ps(_mm256_max_ps(pz, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
nx = _mm256_min_ps(_mm256_max_ps(nx, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
ny = _mm256_min_ps(_mm256_max_ps(ny, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
nz = _mm256_min_ps(_mm256_max_ps(nz, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
tx = _mm256_min_ps(_mm256_max_ps(tx, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
ty = _mm256_min_ps(_mm256_max_ps(ty, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
tz = _mm256_min_ps(_mm256_max_ps(tz, _mm256_set1_ps(-1.0f)), _mm256_set1_ps(1.0f));
// vec *= 511.0f
px = _mm256_mul_ps(px, _mm256_set1_ps(511.0f));
py = _mm256_mul_ps(py, _mm256_set1_ps(511.0f));
pz = _mm256_mul_ps(pz, _mm256_set1_ps(511.0f));
nx = _mm256_mul_ps(nx, _mm256_set1_ps(511.0f));
ny = _mm256_mul_ps(ny, _mm256_set1_ps(511.0f));
nz = _mm256_mul_ps(nz, _mm256_set1_ps(511.0f));
tx = _mm256_mul_ps(tx, _mm256_set1_ps(511.0f));
ty = _mm256_mul_ps(ty, _mm256_set1_ps(511.0f));
tz = _mm256_mul_ps(tz, _mm256_set1_ps(511.0f));
// veci = lrint(vec) & 03ff
__m256i pxi = _mm256_and_si256(_mm256_cvtps_epi32(px), _mm256_set1_epi32(0x3ff));
__m256i pyi = _mm256_and_si256(_mm256_cvtps_epi32(py), _mm256_set1_epi32(0x3ff));
__m256i pzi = _mm256_and_si256(_mm256_cvtps_epi32(pz), _mm256_set1_epi32(0x3ff));
__m256i nxi = _mm256_and_si256(_mm256_cvtps_epi32(nx), _mm256_set1_epi32(0x3ff));
__m256i nyi = _mm256_and_si256(_mm256_cvtps_epi32(ny), _mm256_set1_epi32(0x3ff));
__m256i nzi = _mm256_and_si256(_mm256_cvtps_epi32(nz), _mm256_set1_epi32(0x3ff));
__m256i txi = _mm256_and_si256(_mm256_cvtps_epi32(tx), _mm256_set1_epi32(0x3ff));
__m256i tyi = _mm256_and_si256(_mm256_cvtps_epi32(ty), _mm256_set1_epi32(0x3ff));
__m256i tzi = _mm256_and_si256(_mm256_cvtps_epi32(tz), _mm256_set1_epi32(0x3ff));
// pack = (xi << 0) | (yi << 10) | (zi << 20);
__m256i li = _mm256_castps_si256(len); // length
__m256i pi = _mm256_or_si256(_mm256_or_si256(pxi, _mm256_slli_epi32(pyi, 10)), _mm256_slli_epi32(pzi, 20)); // position
__m256i ni = _mm256_or_si256(_mm256_or_si256(nxi, _mm256_slli_epi32(nyi, 10)), _mm256_slli_epi32(nzi, 20)); // normal
__m256i ti = _mm256_or_si256(_mm256_or_si256(txi, _mm256_slli_epi32(tyi, 10)), _mm256_slli_epi32(tzi, 20)); // tangent
//
// interleave (4x4 matrix transpose)
//
__m256i u0 = _mm256_unpacklo_epi32(li, pi);
__m256i u1 = _mm256_unpackhi_epi32(li, pi);
__m256i u2 = _mm256_unpacklo_epi32(ni, ti);
__m256i u3 = _mm256_unpackhi_epi32(ni, ti);
__m256i v0 = _mm256_unpacklo_epi64(u0, u2);
__m256i v1 = _mm256_unpackhi_epi64(u0, u2);
__m256i v2 = _mm256_unpacklo_epi64(u1, u3);
__m256i v3 = _mm256_unpackhi_epi64(u1, u3);
// store pack x 8
switch (rem) {
case 7: _mm_storeu_si128((__m128i*)packed[i+6], _mm256_extractf128_si256(v2, 1));
case 6: _mm_storeu_si128((__m128i*)packed[i+5], _mm256_extractf128_si256(v1, 1));
case 5: _mm_storeu_si128((__m128i*)packed[i+4], _mm256_extractf128_si256(v0, 1));
case 4: _mm_storeu_si128((__m128i*)packed[i+3], _mm256_castsi256_si128(v3));
case 3: _mm_storeu_si128((__m128i*)packed[i+2], _mm256_castsi256_si128(v2));
case 2: _mm_storeu_si128((__m128i*)packed[i+1], _mm256_castsi256_si128(v1));
case 1: _mm_storeu_si128((__m128i*)packed[i+0], _mm256_castsi256_si128(v0));
}
}
_mm256_zeroupper();
}
#endif