Skip to content

Commit 1dbddfc

Browse files
arnaud de mattiaadematti
arnaud de mattia
authored and
adematti
committed
renaming sse int4, etc. to avoid conflict with cuda
1 parent 4a1543a commit 1dbddfc

13 files changed

Lines changed: 90 additions & 86 deletions

‎mocks/DDrppi_mocks/countpairs_rp_pi_mocks_kernels.c.src‎

Lines changed: 9 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -215,9 +215,9 @@ static inline int countpairs_rp_pi_mocks_avx512_intrinsics_DOUBLE(const int64_t
215215
m_yhat1 = AVX512_LOAD_FLOATS_UNALIGNED(yhat1 + j);
216216
m_zhat1 = AVX512_LOAD_FLOATS_UNALIGNED(zhat1 + j);
217217
}
218-
union int16 union_finalbin;
219-
union float16 union_mDperp;
220-
union float16_weights union_mweight;
218+
union avx512int union_finalbin;
219+
union avx512float union_mDperp;
220+
union avx512float_weights union_mweight;
221221

222222
union_finalbin.m_ibin = AVX512_SETZERO_INT();
223223
union_mDperp.m_Dperp = AVX512_SETZERO_FLOAT();
@@ -587,9 +587,9 @@ static inline int countpairs_rp_pi_mocks_avx_intrinsics_DOUBLE(const int64_t N0,
587587
m_yhat1 = AVX_LOAD_FLOATS_UNALIGNED(yhat1 + j);
588588
m_zhat1 = AVX_LOAD_FLOATS_UNALIGNED(zhat1 + j);
589589
}
590-
union int8 union_finalbin;
591-
union float8 union_mDperp;
592-
union float8_weights union_mweight;
590+
union avxint union_finalbin;
591+
union avxfloat union_mDperp;
592+
union avxfloat_weights union_mweight;
593593

594594
localx1 += AVX_NVEC;
595595
localy1 += AVX_NVEC;
@@ -1084,9 +1084,9 @@ static inline int countpairs_rp_pi_mocks_sse_intrinsics_DOUBLE(const int64_t N0,
10841084
m_zhat1 = SSE_LOAD_FLOATS_UNALIGNED(zhat1 + j);
10851085
}
10861086

1087-
union int4 union_finalbin;
1088-
union float4 union_mDperp;
1089-
union float4_weights union_mweight;
1087+
union sseint union_finalbin;
1088+
union ssefloat union_mDperp;
1089+
union ssefloat_weights union_mweight;
10901090

10911091
localx1 += SSE_NVEC;
10921092
localy1 += SSE_NVEC;

‎mocks/DDsmu_mocks/countpairs_s_mu_mocks_impl.c.src‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -29,6 +29,7 @@
2929
#include "countpairs_s_mu_mocks_kernels_DOUBLE.c"
3030

3131
#ifdef GPU
32+
#include <cuda_runtime.h>
3233
#include "countpairs_s_mu_mocks_gpu.h"
3334
#endif
3435

‎mocks/DDsmu_mocks/countpairs_s_mu_mocks_kernels.c.src‎

Lines changed: 10 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -21,6 +21,7 @@
2121
#include "weight_functions_DOUBLE.h"
2222

2323
#ifdef GPU
24+
#include <cuda_runtime.h>
2425
#include "countpairs_s_mu_mocks_gpu.h"
2526
#endif
2627

@@ -227,9 +228,9 @@ static inline int countpairs_s_mu_mocks_avx512_intrinsics_DOUBLE(const int64_t N
227228
local_w1.weights[w] += AVX512_NVEC;
228229
}
229230

230-
union int16 union_finalbin;
231-
union float16 union_msep;
232-
union float16_weights union_mweight;
231+
union avx512int union_finalbin;
232+
union avx512float union_msep;
233+
union avx512float_weights union_mweight;
233234

234235
const AVX512_FLOATS m_perpx = AVX512_MASKZ_SUBTRACT_FLOATS(m_mask_left, m_x2, m_xpos);
235236
const AVX512_FLOATS m_perpy = AVX512_MASKZ_SUBTRACT_FLOATS(m_mask_left, m_y2, m_ypos);
@@ -595,9 +596,9 @@ static inline int countpairs_s_mu_mocks_avx_intrinsics_DOUBLE(const int64_t N0,
595596
local_w1.weights[w] += AVX_NVEC;
596597
}
597598

598-
union int8 union_finalbin;
599-
union float8 union_msep;
600-
union float8_weights union_mweight;
599+
union avxint union_finalbin;
600+
union avxfloat union_msep;
601+
union avxfloat_weights union_mweight;
601602

602603
const AVX_FLOATS m_perpx = AVX_SUBTRACT_FLOATS(m_x2, m_xpos);
603604
const AVX_FLOATS m_perpy = AVX_SUBTRACT_FLOATS(m_y2, m_ypos);
@@ -1076,9 +1077,9 @@ static inline int countpairs_s_mu_mocks_sse_intrinsics_DOUBLE(const int64_t N0,
10761077
local_w1.weights[w] += SSE_NVEC;
10771078
}
10781079

1079-
union int4 union_finalbin;
1080-
union float4 union_msep;
1081-
union float4_weights union_mweight;
1080+
union sseint union_finalbin;
1081+
union ssefloat union_msep;
1082+
union ssefloat_weights union_mweight;
10821083

10831084
const SSE_FLOATS m_perpx = SSE_SUBTRACT_FLOATS(m_x2, m_xpos);
10841085
const SSE_FLOATS m_perpy = SSE_SUBTRACT_FLOATS(m_y2, m_ypos);

‎mocks/DDtheta_mocks/countpairs_theta_mocks_kernels.c.src‎

Lines changed: 9 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -409,9 +409,9 @@ static inline int countpairs_theta_mocks_sse_intrinsics_DOUBLE(const int64_t N0,
409409
const SSE_FLOATS m_y1 = SSE_SET_FLOAT(ypos);
410410
const SSE_FLOATS m_z1 = SSE_SET_FLOAT(zpos);
411411
for(;j<=(N1-SSE_NVEC);j+=SSE_NVEC){
412-
union int4 union_rpbin;
413-
union float4 union_mDperp;
414-
union float4_weights union_mweight;
412+
union sseint union_rpbin;
413+
union ssefloat union_mDperp;
414+
union ssefloat_weights union_mweight;
415415

416416
const SSE_FLOATS m_costhetamax = SSE_SET_FLOAT(costhetamax);
417417
const SSE_FLOATS m_costhetamin = SSE_SET_FLOAT(costhetamin);
@@ -751,9 +751,9 @@ static inline int countpairs_theta_mocks_avx_intrinsics_DOUBLE(const int64_t N0,
751751
const AVX_FLOATS m_y1 = AVX_SET_FLOAT(ypos);
752752
const AVX_FLOATS m_z1 = AVX_SET_FLOAT(zpos);
753753
for(;j<=(N1-AVX_NVEC);j+=AVX_NVEC){
754-
union int8 union_rpbin;
755-
union float8 union_mDperp;
756-
union float8_weights union_mweight;
754+
union avxint union_rpbin;
755+
union avxfloat union_mDperp;
756+
union avxfloat_weights union_mweight;
757757

758758
const AVX_FLOATS m_costhetamax = AVX_SET_FLOAT(costhetamax);
759759
const AVX_FLOATS m_costhetamin = AVX_SET_FLOAT(costhetamin);
@@ -1096,9 +1096,9 @@ static inline int countpairs_theta_mocks_avx512_intrinsics_DOUBLE(const int64_t
10961096
const AVX512_FLOATS m_y1 = AVX512_MASKZ_LOAD_FLOATS_UNALIGNED(m_mask_left, localy1);
10971097
const AVX512_FLOATS m_z1 = AVX512_MASKZ_LOAD_FLOATS_UNALIGNED(m_mask_left, localz1);
10981098

1099-
union int16 union_rpbin;
1100-
union float16 union_mDperp;
1101-
union float16_weights union_mweight;
1099+
union avx512int union_rpbin;
1100+
union avx512float union_mDperp;
1101+
union avx512float_weights union_mweight;
11021102

11031103
const AVX512_FLOATS m_costhetamax = AVX512_SET_FLOAT(costhetamax);
11041104
const AVX512_FLOATS m_costhetamin = AVX512_SET_FLOAT(costhetamin);

‎theory/DD/countpairs_kernels.c.src‎

Lines changed: 9 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -160,9 +160,9 @@ static inline int countpairs_avx512_intrinsics_DOUBLE(const int64_t N0, DOUBLE *
160160
const AVX512_FLOATS m_y1 = AVX512_MASKZ_LOAD_FLOATS_UNALIGNED(m_mask_left, localy1);
161161
const AVX512_FLOATS m_z1 = AVX512_MASKZ_LOAD_FLOATS_UNALIGNED(m_mask_left, localz1);
162162

163-
union int16 union_rpbin;
164-
union float16 union_mDperp;
165-
union float16_weights union_mweight;
163+
union avx512int union_rpbin;
164+
union avx512float union_mDperp;
165+
union avx512float_weights union_mweight;
166166

167167
union_rpbin.m_ibin = AVX512_SETZERO_INT();
168168
union_mDperp.m_Dperp = AVX512_SETZERO_FLOAT();
@@ -427,9 +427,9 @@ static inline int countpairs_avx_intrinsics_DOUBLE(const int64_t N0, DOUBLE *x0,
427427
const AVX_FLOATS m_zpos = AVX_SET_FLOAT(zpos);
428428
const AVX_FLOATS m_norm0 = AVX_SET_FLOAT(norm0);
429429

430-
union int8 union_rpbin;
431-
union float8 union_mDperp;
432-
union float8_weights union_mweight;
430+
union avxint union_rpbin;
431+
union avxfloat union_mDperp;
432+
union avxfloat_weights union_mweight;
433433

434434
const AVX_FLOATS m_x1 = AVX_LOAD_FLOATS_UNALIGNED(localx1);
435435
const AVX_FLOATS m_y1 = AVX_LOAD_FLOATS_UNALIGNED(localy1);
@@ -765,9 +765,9 @@ static inline int countpairs_sse_intrinsics_DOUBLE(const int64_t N0, DOUBLE *x0,
765765
}
766766

767767
for(;j<=(N1 - SSE_NVEC);j+=SSE_NVEC){
768-
union int4 union_rpbin;
769-
union float4 union_mDperp;
770-
union float4_weights union_mweight;
768+
union sseint union_rpbin;
769+
union ssefloat union_mDperp;
770+
union ssefloat_weights union_mweight;
771771

772772
const SSE_FLOATS m_xpos = SSE_SET_FLOAT(xpos);
773773
const SSE_FLOATS m_ypos = SSE_SET_FLOAT(ypos);

‎theory/DDrppi/countpairs_rp_pi_kernels.c.src‎

Lines changed: 10 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -185,9 +185,9 @@ static inline int countpairs_rp_pi_avx512_intrinsics_DOUBLE(const int64_t N0, DO
185185
localy1 += AVX512_NVEC;
186186
localz1 += AVX512_NVEC;
187187

188-
union int16 union_finalbin;
189-
union float16 union_mDperp;
190-
union float16_weights union_mweight;
188+
union avx512int union_finalbin;
189+
union avx512float union_mDperp;
190+
union avx512float_weights union_mweight;
191191

192192
for(int w = 0; w < pair.num_weights; w++){
193193
pair.weights1[w].a512 = AVX512_MASKZ_LOAD_FLOATS_UNALIGNED(m_mask_left, local_w1.weights[w]);
@@ -481,9 +481,9 @@ static inline int countpairs_rp_pi_avx_intrinsics_DOUBLE(const int64_t N0, DOUBL
481481
const AVX_FLOATS m_zpos = AVX_SET_FLOAT(zpos);
482482
const AVX_FLOATS m_norm0 = AVX_SET_FLOAT(norm0);
483483

484-
union int8 union_finalbin;
485-
union float8 union_mDperp;
486-
union float8_weights union_mweight;
484+
union avxint union_finalbin;
485+
union avxfloat union_mDperp;
486+
union avxfloat_weights union_mweight;
487487

488488
const AVX_FLOATS m_x1 = AVX_LOAD_FLOATS_UNALIGNED(localx1);
489489
const AVX_FLOATS m_y1 = AVX_LOAD_FLOATS_UNALIGNED(localy1);
@@ -868,8 +868,8 @@ static inline int countpairs_rp_pi_sse_intrinsics_DOUBLE(const int64_t N0, DOUBL
868868

869869
for(;j<=(N1 - SSE_NVEC);j+=SSE_NVEC){
870870

871-
union int4 union_finalbin;
872-
union float4 union_mDperp;
871+
union sseint union_finalbin;
872+
union ssefloat union_mDperp;
873873

874874
const SSE_FLOATS m_xpos = SSE_SET_FLOAT(xpos);
875875
const SSE_FLOATS m_ypos = SSE_SET_FLOAT(ypos);
@@ -895,11 +895,11 @@ static inline int countpairs_rp_pi_sse_intrinsics_DOUBLE(const int64_t N0, DOUBL
895895
local_w1.weights[w] += SSE_NVEC;
896896
}
897897

898-
union float4_weights{
898+
union ssefloat_weights{
899899
SSE_FLOATS m_weights;
900900
DOUBLE weights[SSE_NVEC];
901901
};
902-
union float4_weights union_mweight;
902+
union ssefloat_weights union_mweight;
903903

904904
const SSE_FLOATS m_pimax = SSE_SET_FLOAT((DOUBLE) pimax);
905905
const SSE_FLOATS m_sqr_rpmax = m_rupp_sqr[nbin-1];

‎theory/DDsmu/countpairs_s_mu_impl.c.src‎

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -26,6 +26,7 @@
2626
#include "gridlink_utils_DOUBLE.h" //for associated helper routines
2727

2828
#ifdef GPU
29+
#include <cuda_runtime.h>
2930
#include "countpairs_s_mu_gpu.h"
3031
#endif
3132

@@ -576,7 +577,7 @@ int countpairs_s_mu_DOUBLE(const int64_t ND1, DOUBLE *X1, DOUBLE *Y1, DOUBLE *Z1
576577
gpu_start_idx1[icell] = pcount2;
577578
pcount1 += (&lattice1[icell])->nelements;
578579
pcount2 += (&lattice2[icell])->nelements;
579-
580+
580581
this_cell_threads = (&lattice1[icell])->nelements * (&lattice2[icell])->nelements;
581582
nblocks = (this_cell_threads+THREADS_PER_BLOCK-1)/THREADS_PER_BLOCK;
582583
if (nblocks > max_blocks) {

‎theory/DDsmu/countpairs_s_mu_kernels.c.src‎

Lines changed: 10 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -19,6 +19,7 @@
1919
#include "weight_functions_DOUBLE.h"
2020

2121
#ifdef GPU
22+
#include <cuda_runtime.h>
2223
#include "countpairs_s_mu_gpu.h"
2324
#endif
2425

@@ -210,9 +211,9 @@ static inline int countpairs_s_mu_avx512_intrinsics_DOUBLE(const int64_t N0, DOU
210211
local_w1.weights[w] += AVX512_NVEC;
211212
}
212213

213-
union int16 union_finalbin;
214-
union float16 union_mDperp;
215-
union float16_weights union_mweight;
214+
union avx512int union_finalbin;
215+
union avx512float union_mDperp;
216+
union avx512float_weights union_mweight;
216217

217218
const AVX512_FLOATS m_max_dz = AVX512_SET_FLOAT(max_dz);
218219
const AVX512_FLOATS m_sqr_smax = m_supp_sqr[nsbin-1];
@@ -552,9 +553,9 @@ static inline int countpairs_s_mu_avx_intrinsics_DOUBLE(const int64_t N0, DOUBLE
552553
pair.weights1[w].a = AVX_LOAD_FLOATS_UNALIGNED(local_w1.weights[w]);
553554
local_w1.weights[w] += AVX_NVEC;
554555
}
555-
union int8 union_finalbin;
556-
union float8 union_mDperp;
557-
union float8_weights union_mweight;
556+
union avxint union_finalbin;
557+
union avxfloat union_mDperp;
558+
union avxfloat_weights union_mweight;
558559

559560
const AVX_FLOATS m_max_dz = AVX_SET_FLOAT(max_dz);
560561
const AVX_FLOATS m_sqr_smax = m_supp_sqr[nsbin-1];
@@ -990,9 +991,9 @@ static inline int countpairs_s_mu_sse_intrinsics_DOUBLE(const int64_t N0, DOUBLE
990991
local_w1.weights[w] += SSE_NVEC;
991992
}
992993

993-
union int4 union_finalbin;
994-
union float4 union_mDperp;
995-
union float4_weights union_mweight;
994+
union sseint union_finalbin;
995+
union ssefloat union_mDperp;
996+
union ssefloat_weights union_mweight;
996997

997998
const SSE_FLOATS m_max_dz = SSE_SET_FLOAT(max_dz);
998999
const SSE_FLOATS m_sqr_smax = m_supp_sqr[nsbin-1];

‎theory/wp/wp_kernels.c.src‎

Lines changed: 12 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -171,9 +171,9 @@ static inline int wp_avx512_intrinsics_DOUBLE(DOUBLE *x0, DOUBLE *y0, DOUBLE *z0
171171
const AVX512_FLOATS m_y1 = AVX512_MASKZ_LOAD_FLOATS_UNALIGNED(m_mask_left, localy1);
172172
const AVX512_FLOATS m_z1 = AVX512_MASKZ_LOAD_FLOATS_UNALIGNED(m_mask_left, localz1);
173173

174-
union int16 union_rpbin;
175-
union float16 union_mDperp;
176-
union float16_weights union_mweight;
174+
union avx512int union_rpbin;
175+
union avx512float union_mDperp;
176+
union avx512float_weights union_mweight;
177177

178178
union_rpbin.m_ibin = AVX512_SETZERO_INT();
179179
union_mDperp.m_Dperp = AVX512_SETZERO_FLOAT();
@@ -481,9 +481,9 @@ static inline int wp_avx2_intrinsics_DOUBLE(DOUBLE *x0, DOUBLE *y0, DOUBLE *z0,
481481
const AVX2_FLOATS m_ypos = AVX2_SET_FLOAT(ypos);
482482
const AVX2_FLOATS m_zpos = AVX2_SET_FLOAT(zpos);
483483

484-
union int8 union_rpbin;
485-
union float8 union_mDperp;
486-
union float8_weights union_mweight;
484+
union avxint union_rpbin;
485+
union avxfloat union_mDperp;
486+
union avxfloat_weights union_mweight;
487487

488488
const AVX2_FLOATS m_x1 = AVX2_LOAD_FLOATS_UNALIGNED(localx1);
489489
const AVX2_FLOATS m_y1 = AVX2_LOAD_FLOATS_UNALIGNED(localy1);
@@ -855,9 +855,9 @@ static inline int wp_avx_intrinsics_DOUBLE(DOUBLE *x0, DOUBLE *y0, DOUBLE *z0, c
855855
const AVX_FLOATS m_ypos = AVX_SET_FLOAT(ypos);
856856
const AVX_FLOATS m_zpos = AVX_SET_FLOAT(zpos);
857857

858-
union int8 union_rpbin;
859-
union float8 union_mDperp;
860-
union float8_weights union_mweight;
858+
union avxint union_rpbin;
859+
union avxfloat union_mDperp;
860+
union avxfloat_weights union_mweight;
861861

862862
const AVX_FLOATS m_x1 = AVX_LOAD_FLOATS_UNALIGNED(localx1);
863863
const AVX_FLOATS m_y1 = AVX_LOAD_FLOATS_UNALIGNED(localy1);
@@ -1257,9 +1257,9 @@ static inline int wp_sse_intrinsics_DOUBLE(DOUBLE *x0, DOUBLE *y0, DOUBLE *z0, c
12571257
local_w1.weights[w] += SSE_NVEC;
12581258
}
12591259

1260-
union int4 union_rpbin;
1261-
union float4 union_mDperp;
1262-
union float4_weights union_mweight;
1260+
union sseint union_rpbin;
1261+
union ssefloat union_mDperp;
1262+
union ssefloat_weights union_mweight;
12631263

12641264
const SSE_FLOATS m_pimax = SSE_SET_FLOAT(pimax);
12651265
const SSE_FLOATS m_sqr_rpmax = SSE_SET_FLOAT(sqr_rpmax);

‎theory/xi/xi_kernels.c.src‎

Lines changed: 9 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -161,9 +161,9 @@ static inline int xi_avx512_intrinsics_DOUBLE(DOUBLE *x1, DOUBLE *y1, DOUBLE *z1
161161
const AVX512_FLOATS m_y2 = AVX512_MASKZ_LOAD_FLOATS_UNALIGNED(m_mask_left, localy2);
162162
const AVX512_FLOATS m_z2 = AVX512_MASKZ_LOAD_FLOATS_UNALIGNED(m_mask_left, localz2);
163163

164-
union int16 union_rpbin;
165-
union float16 union_mDperp;
166-
union float16_weights union_mweight;
164+
union avx512int union_rpbin;
165+
union avx512float union_mDperp;
166+
union avx512float_weights union_mweight;
167167

168168
union_rpbin.m_ibin = AVX512_SETZERO_INT();
169169
union_mDperp.m_Dperp = AVX512_SETZERO_FLOAT();
@@ -435,9 +435,9 @@ static inline int xi_avx_intrinsics_DOUBLE(DOUBLE *x1, DOUBLE *y1, DOUBLE *z1, c
435435
const AVX_FLOATS m_ypos = AVX_SET_FLOAT(ypos);
436436
const AVX_FLOATS m_zpos = AVX_SET_FLOAT(zpos);
437437

438-
union int8 union_rbin;
439-
union float8 union_mDperp;
440-
union float8_weights union_mweight;
438+
union avxint union_rbin;
439+
union avxfloat union_mDperp;
440+
union avxfloat_weights union_mweight;
441441

442442
const AVX_FLOATS m_x2 = AVX_LOAD_FLOATS_UNALIGNED(localx2);
443443
const AVX_FLOATS m_y2 = AVX_LOAD_FLOATS_UNALIGNED(localy2);
@@ -769,9 +769,9 @@ static inline int xi_sse_intrinsics_DOUBLE(DOUBLE *x0, DOUBLE *y0, DOUBLE *z0, c
769769
local_w1.weights[w] += SSE_NVEC;
770770
}
771771

772-
union int4 union_rbin;
773-
union float4 union_mDperp;
774-
union float4_weights union_mweight;
772+
union sseint union_rbin;
773+
union ssefloat union_mDperp;
774+
union ssefloat_weights union_mweight;
775775

776776
const SSE_FLOATS m_max_dz = SSE_SET_FLOAT(max_dz);
777777
const SSE_FLOATS m_sqr_rmax = SSE_SET_FLOAT(sqr_rmax);

0 commit comments

Comments
 (0)