mirrors/xmrig
1
0
Fork 1
mirror of https://github.com/xmrig/xmrig.git synced 2025-12-26 14:02:53 -05:00
Code Issues Projects Releases Wiki Activity

Cleanup MoneroOcean patchset

Browse Source
This commit is contained in:
Tony Butler
2019-12-22 09:54:34 -07:00
parent aaaaf5c1ed
commit c4ff8c4064
49 changed files with 2094 additions and 2237 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
src/crypto/cn/gpu/cn_gpu_avx.cpp
View File

@@ -45,11 +45,11 @@ inline void prep_dv_avx(__m256i* idx, __m256i& v, __m256& n01)
n01 = _mm256_cvtepi32_ps(v);
}
inline __m256 fma_break(const __m256& x)
{
// Break the dependency chain by setitng the exp to ?????01
__m256 xx = _mm256_and_ps(_mm256_castsi256_ps(_mm256_set1_epi32(0xFEFFFFFF)), x);
return _mm256_or_ps(_mm256_castsi256_ps(_mm256_set1_epi32(0x00800000)), xx);
inline __m256 fma_break(const __m256& x)
{
// Break the dependency chain by setting the exp to ?????01
__m256 xx = _mm256_and_ps(_mm256_castsi256_ps(_mm256_set1_epi32(0xFEFFFFFF)), x);
return _mm256_or_ps(_mm256_castsi256_ps(_mm256_set1_epi32(0x00800000)), xx);
}
// 14
@@ -151,13 +151,13 @@ void cn_gpu_inner_avx(const uint8_t* spad, uint8_t* lpad)
__m256 n01, n23;
prep_dv_avx(idx0, v01, n01);
prep_dv_avx(idx2, v23, n23);
__m256i out, out2;
__m256 n10, n22, n33;
n10 = _mm256_permute2f128_ps(n01, n01, 0x01);
n22 = _mm256_permute2f128_ps(n23, n23, 0x00);
n33 = _mm256_permute2f128_ps(n23, n23, 0x11);
out = _mm256_setzero_si256();
double_compute_wrap<0>(n01, n10, n22, n33, 1.3437500f, 1.4296875f, rc, suma, out);
double_compute_wrap<1>(n01, n22, n33, n10, 1.2812500f, 1.3984375f, rc, suma, out);
@@ -166,7 +166,7 @@ void cn_gpu_inner_avx(const uint8_t* spad, uint8_t* lpad)
_mm256_store_si256(idx0, _mm256_xor_si256(v01, out));
sum0 = _mm256_add_ps(suma, sumb);
out2 = out;
__m256 n11, n02, n30;
n11 = _mm256_permute2f128_ps(n01, n01, 0x11);
n02 = _mm256_permute2f128_ps(n01, n23, 0x20);
@@ -191,7 +191,7 @@ void cn_gpu_inner_avx(const uint8_t* spad, uint8_t* lpad)
__m128 sum = _mm256_castps256_ps128(sum0);
sum = _mm_and_ps(_mm_castsi128_ps(_mm_set1_epi32(0x7fffffff)), sum); // take abs(va) by masking the float sign bit
// vs range 0 - 64
// vs range 0 - 64
__m128i v0 = _mm_cvttps_epi32(_mm_mul_ps(sum, _mm_set1_ps(16777216.0f)));
v0 = _mm_xor_si128(v0, _mm256_castsi256_si128(out2));
__m128i v1 = _mm_shuffle_epi32(v0, _MM_SHUFFLE(0, 1, 2, 3));

18
src/crypto/cn/gpu/cn_gpu_ssse3.cpp
View File

@@ -39,11 +39,11 @@ inline void prep_dv(__m128i* idx, __m128i& v, __m128& n)
n = _mm_cvtepi32_ps(v);
}
inline __m128 fma_break(__m128 x)
{
// Break the dependency chain by setitng the exp to ?????01
x = _mm_and_ps(_mm_castsi128_ps(_mm_set1_epi32(0xFEFFFFFF)), x);
return _mm_or_ps(_mm_castsi128_ps(_mm_set1_epi32(0x00800000)), x);
inline __m128 fma_break(__m128 x)
{
// Break the dependency chain by setting the exp to ?????01
x = _mm_and_ps(_mm_castsi128_ps(_mm_set1_epi32(0xFEFFFFFF)), x);
return _mm_or_ps(_mm_castsi128_ps(_mm_set1_epi32(0x00800000)), x);
}
// 14
@@ -136,13 +136,13 @@ void cn_gpu_inner_ssse3(const uint8_t* spad, uint8_t* lpad)
__m128i* idx2 = scratchpad_ptr<MASK>(lpad, s, 2);
__m128i* idx3 = scratchpad_ptr<MASK>(lpad, s, 3);
__m128 sum0 = _mm_setzero_ps();
for(size_t i = 0; i < ITER; i++)
{
__m128 n0, n1, n2, n3;
__m128i v0, v1, v2, v3;
__m128 suma, sumb, sum1, sum2, sum3;
prep_dv(idx0, v0, n0);
prep_dv(idx1, v1, n1);
prep_dv(idx2, v2, n2);
@@ -158,7 +158,7 @@ void cn_gpu_inner_ssse3(const uint8_t* spad, uint8_t* lpad)
sum0 = _mm_add_ps(suma, sumb);
_mm_store_si128(idx0, _mm_xor_si128(v0, out));
out2 = out;
out = _mm_setzero_si128();
single_compute_wrap<0>(n1, n0, n2, n3, 1.4296875f, rc, suma, out);
single_compute_wrap<1>(n1, n2, n3, n0, 1.3984375f, rc, suma, out);
@@ -190,7 +190,7 @@ void cn_gpu_inner_ssse3(const uint8_t* spad, uint8_t* lpad)
sum0 = _mm_add_ps(sum0, sum2);
sum0 = _mm_and_ps(_mm_castsi128_ps(_mm_set1_epi32(0x7fffffff)), sum0); // take abs(va) by masking the float sign bit
// vs range 0 - 64
// vs range 0 - 64
n0 = _mm_mul_ps(sum0, _mm_set1_ps(16777216.0f));
v0 = _mm_cvttps_epi32(n0);
v0 = _mm_xor_si128(v0, out2);