diff options
| author | Developer-Ecosystem-Engineering <65677710+Developer-Ecosystem-Engineering@users.noreply.github.com> | 2022-12-07 22:26:35 -0800 |
|---|---|---|
| committer | Developer-Ecosystem-Engineering <65677710+Developer-Ecosystem-Engineering@users.noreply.github.com> | 2022-12-07 22:32:52 -0800 |
| commit | 0609a34aec0d2f4b0e404c84352dfb979baea3de (patch) | |
| tree | 53eff007a185824849ea7a14245b1fb9bb10949b | |
| parent | 2de88654b6d992fe949b7e9a68793fdb74364458 (diff) | |
| download | numpy-0609a34aec0d2f4b0e404c84352dfb979baea3de.tar.gz | |
Use mask_to_true() for logical_not to save a few ops
| -rw-r--r-- | numpy/core/src/umath/loops_logical.dispatch.c.src | 24 |
1 files changed, 16 insertions, 8 deletions
diff --git a/numpy/core/src/umath/loops_logical.dispatch.c.src b/numpy/core/src/umath/loops_logical.dispatch.c.src index 94b0c4594..793a2af19 100644 --- a/numpy/core/src/umath/loops_logical.dispatch.c.src +++ b/numpy/core/src/umath/loops_logical.dispatch.c.src @@ -35,6 +35,16 @@ NPY_FINLINE npyv_u8 byte_to_true(npyv_u8 v) // tmp is filled with 0xff/0x00, negate and mask to boolean true return npyv_andc_u8(truemask, tmp); } +/* + * convert mask vector (0xff/0x00) to boolean true. similar to byte_to_true(), + * but we've already got a mask and can skip negation. + */ +NPY_FINLINE npyv_u8 mask_to_true(npyv_b8 v) +{ + const npyv_u8 truemask = npyv_setall_u8(1 == 1); + return npyv_and_u8(truemask, npyv_cvt_u8_b8(v)); +} + /**begin repeat * #kind = logical_and, logical_or# @@ -165,12 +175,11 @@ simd_@kind@_BOOL(npy_bool * op, npy_bool * ip, npy_intp len) #if UNROLL > @unroll@ npyv_u8 v@unroll@ = npyv_load_u8(ip + vstep * @unroll@); #if @not@ - npyv_u8 r@unroll@ = npyv_cvt_u8_b8(npyv_cmpeq_u8(v@unroll@, zero)); + npyv_u8 r@unroll@ = mask_to_true(npyv_cmpeq_u8(v@unroll@, zero)); #else - // abs is kind of pointless but maybe its used for byte_to_true below - npyv_u8 r@unroll@ = v@unroll@; + npyv_u8 r@unroll@ = byte_to_true(v@unroll@); #endif - npyv_store_u8(op + vstep * @unroll@, byte_to_true(r@unroll@)); + npyv_store_u8(op + vstep * @unroll@, r@unroll@); #endif /**end repeat1**/ } @@ -180,12 +189,11 @@ simd_@kind@_BOOL(npy_bool * op, npy_bool * ip, npy_intp len) for (; len >= vstep; len -= vstep, ip += vstep, op += vstep) { npyv_u8 v = npyv_load_u8(ip); #if @not@ - npyv_u8 r = npyv_cvt_u8_b8(npyv_cmpeq_u8(v, zero)); + npyv_u8 r = mask_to_true(npyv_cmpeq_u8(v, zero)); #else - // abs is kind of pointless but maybe its used for byte_to_true below - npyv_u8 r = v; + npyv_u8 r = byte_to_true(v); #endif - npyv_store_u8(op, byte_to_true(r)); + npyv_store_u8(op, r); } // Scalar loop to finish off |