summaryrefslogtreecommitdiff
path: root/volk/spu_lib/spu_16s_vector_subtract_unaligned.c
diff options
context:
space:
mode:
authorTom Rondeau2012-04-13 18:36:53 -0400
committerTom Rondeau2012-04-13 18:36:53 -0400
commitf919f9dcbb54a08e6e26d6c229ce92fb784fa1b2 (patch)
tree7e846386b9eb1676f9a93fc4a1e55916b9accc97 /volk/spu_lib/spu_16s_vector_subtract_unaligned.c
parent6a1e9783fec6ed827f49db27c171591d30f32933 (diff)
downloadgnuradio-f919f9dcbb54a08e6e26d6c229ce92fb784fa1b2.tar.gz
gnuradio-f919f9dcbb54a08e6e26d6c229ce92fb784fa1b2.tar.bz2
gnuradio-f919f9dcbb54a08e6e26d6c229ce92fb784fa1b2.zip
Removed whitespace and added dtools/bin/remove-whitespace as a tool to do this in the future.
The sed script was provided by Moritz Fischer.
Diffstat (limited to 'volk/spu_lib/spu_16s_vector_subtract_unaligned.c')
-rw-r--r--volk/spu_lib/spu_16s_vector_subtract_unaligned.c68
1 files changed, 34 insertions, 34 deletions
diff --git a/volk/spu_lib/spu_16s_vector_subtract_unaligned.c b/volk/spu_lib/spu_16s_vector_subtract_unaligned.c
index a3ce6c2fe..ea110c8d2 100644
--- a/volk/spu_lib/spu_16s_vector_subtract_unaligned.c
+++ b/volk/spu_lib/spu_16s_vector_subtract_unaligned.c
@@ -4,15 +4,15 @@ void* libvector_16s_vector_subtract_unaligned(void* target, void* src0, void* s
//loop iterator i
int i = 0;
void* retval = target;
-
+
//put the target and source addresses into qwords
vector unsigned int address_counter_tgt = {(unsigned int)target, 0, 0, 0};
vector unsigned int address_counter_src0 = {(unsigned int)src0, 0, 0 ,0};
vector unsigned int address_counter_src1 = {(unsigned int)src1, 0, 0, 0};
-
+
//create shuffle masks
-
+
//shuffle mask building blocks:
//all from the first vector
vector unsigned char oneup = {0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
@@ -20,9 +20,9 @@ void* libvector_16s_vector_subtract_unaligned(void* target, void* src0, void* s
//all from the second vector
vector unsigned char second_oneup = {0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17,
0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f};
-
-
+
+
//gamma: second half of the second, first half of the first, break at (unsigned int)src0%16
vector unsigned char src_cmp = spu_splats((unsigned char)((unsigned int)src0%16));
vector unsigned char gt_res = spu_cmpgt(oneup, src_cmp);
@@ -30,7 +30,7 @@ void* libvector_16s_vector_subtract_unaligned(void* target, void* src0, void* s
vector unsigned char cmp_res = spu_or(gt_res, eq_res);
vector unsigned char sixteen_uchar = spu_splats((unsigned char)16);
vector unsigned char phase_change = spu_and(sixteen_uchar, cmp_res);
- vector unsigned int shuffle_mask_gamma = spu_add((vector unsigned int)phase_change,
+ vector unsigned int shuffle_mask_gamma = spu_add((vector unsigned int)phase_change,
(vector unsigned int)oneup);
shuffle_mask_gamma = spu_rlqwbyte(shuffle_mask_gamma, (unsigned int)src0%16);
@@ -41,17 +41,17 @@ void* libvector_16s_vector_subtract_unaligned(void* target, void* src0, void* s
cmp_res = spu_or(gt_res, eq_res);
sixteen_uchar = spu_splats((unsigned char)16);
phase_change = spu_and(sixteen_uchar, cmp_res);
- vector unsigned int shuffle_mask_eta = spu_add((vector unsigned int)phase_change,
+ vector unsigned int shuffle_mask_eta = spu_add((vector unsigned int)phase_change,
(vector unsigned int)oneup);
shuffle_mask_eta = spu_rlqwbyte(shuffle_mask_eta, (unsigned int)src1%16);
-
-
-
+
+
+
vector unsigned char tgt_second = spu_rlqwbyte(second_oneup, -((unsigned int)target%16));
vector unsigned char tgt_first = spu_rlqwbyte(oneup, -((unsigned int)target%16));
-
+
//alpha: first half of first, second half of second, break at (unsigned int)target%16
src_cmp = spu_splats((unsigned char)((unsigned int)target%16));
gt_res = spu_cmpgt(oneup, src_cmp);
@@ -60,13 +60,13 @@ void* libvector_16s_vector_subtract_unaligned(void* target, void* src0, void* s
phase_change = spu_and(sixteen_uchar, cmp_res);
vector unsigned int shuffle_mask_alpha = spu_add((vector unsigned int)phase_change,
(vector unsigned int)oneup);
-
+
//delta: first half of first, first half of second, break at (unsigned int)target%16
vector unsigned char shuffle_mask_delta = spu_shuffle(oneup, tgt_second, (vector unsigned char)shuffle_mask_alpha);
//epsilon: second half of second, second half of first, break at (unsigned int)target%16
vector unsigned char shuffle_mask_epsilon = spu_shuffle(tgt_second, oneup, (vector unsigned char)shuffle_mask_alpha);
//zeta: second half of second, first half of first, break at 16 - (unsigned int)target%16
- vector unsigned int shuffle_mask_zeta = spu_rlqwbyte(shuffle_mask_alpha, (unsigned int)target%16);
+ vector unsigned int shuffle_mask_zeta = spu_rlqwbyte(shuffle_mask_alpha, (unsigned int)target%16);
//beta: first half of first, second half of second, break at num_bytes%16
src_cmp = spu_splats((unsigned char)(num_bytes%16));
@@ -76,19 +76,19 @@ void* libvector_16s_vector_subtract_unaligned(void* target, void* src0, void* s
phase_change = spu_and(sixteen_uchar, cmp_res);
vector unsigned int shuffle_mask_beta = spu_add((vector unsigned int)phase_change,
(vector unsigned int)oneup);
-
-
-
-
+
+
+
+
qword src0_past;
qword src0_present;
qword src1_past;
qword src1_present;
qword tgt_past;
qword tgt_present;
-
+
qword in_temp0;
qword in_temp1;
qword out_temp0;
@@ -99,54 +99,54 @@ void* libvector_16s_vector_subtract_unaligned(void* target, void* src0, void* s
src0_past = si_lqd((qword)address_counter_src0, 0);
src1_past = si_lqd((qword)address_counter_src1, 0);
tgt_past = si_lqd((qword)address_counter_tgt, 0);
-
+
for(i = 0; i < num_bytes/16; ++i) {
-
+
src0_present = si_lqd((qword)address_counter_src0, 16);
src1_present = si_lqd((qword)address_counter_src1, 16);
tgt_present = si_lqd((qword)address_counter_tgt, 16);
-
+
in_temp0 = spu_shuffle(src0_present, src0_past, (vector unsigned char)shuffle_mask_gamma);
in_temp1 = spu_shuffle(src1_present, src1_past, (vector unsigned char)shuffle_mask_eta);
-
+
sum = spu_sub((vector signed short)in_temp0, (vector signed short)in_temp1);
-
+
out_temp0 = spu_shuffle(tgt_past, (qword)sum, shuffle_mask_delta);
out_temp1 = spu_shuffle(tgt_present, (qword)sum, shuffle_mask_epsilon);
-
+
si_stqd(out_temp0, (qword)address_counter_tgt, 0);
si_stqd(out_temp1, (qword)address_counter_tgt, 16);
-
+
tgt_past = out_temp1;
src0_past = src0_present;
src1_past = src1_present;
address_counter_src0 = spu_add(address_counter_src0, 16);
address_counter_src1 = spu_add(address_counter_src1, 16);
address_counter_tgt = spu_add(address_counter_tgt, 16);
-
-
+
+
}
-
+
src0_present = si_lqd((qword)address_counter_src0, 16);
src1_present = si_lqd((qword)address_counter_src1, 16);
tgt_present = si_lqd((qword)address_counter_tgt, 16);
-
-
+
+
in_temp0 = spu_shuffle(src0_present, src0_past, (vector unsigned char) shuffle_mask_gamma);
in_temp1 = spu_shuffle(src1_present, src1_past, (vector unsigned char) shuffle_mask_eta);
sum = spu_sub((vector signed short)in_temp0, (vector signed short)in_temp1);
qword target_temp = spu_shuffle(tgt_present, tgt_past, (vector unsigned char) shuffle_mask_zeta);
qword meld = spu_shuffle((qword)sum, target_temp, (vector unsigned char)shuffle_mask_beta);
-
-
+
+
out_temp0 = spu_shuffle(tgt_past, meld, shuffle_mask_delta);
out_temp1 = spu_shuffle(tgt_present, meld, shuffle_mask_epsilon);
-
+
si_stqd(out_temp0, (qword)address_counter_tgt, 0);
si_stqd(out_temp1, (qword)address_counter_tgt, 16);
-
+
return retval;
}