diff options
| author |  mtklein <mtklein@chromium.org> | 2015-07-23 08:37:02 -0700 |
|---|---|---|
| committer |  Commit bot <commit-bot@chromium.org> | 2015-07-23 08:37:02 -0700 |
| commit | b9c47f996867ec885a7ede697bd1b0859b6d9301 (patch) | |
| tree | 4ad178012a0f89830af3d20aede0c4e23411bb7c | |
| parent | 11dfc8e57a9a485476b4aff93ea2135ee1846d4f (diff) | |
Have Windows use _mm_rsqrt_ss too.
Tidy up a little while I'm in here:
1) SIMD headers are now included by SkTypes.h as appropriate.
2) _mm_cvtss_f32() is pithier and generates the same code.
Looks like this is the only code checking for SSE wrong. After this CL:
~/skia (sse) $ git grep __SSE
include/core/SkPreConfig.h: #if defined(__SSE4_2__)
include/core/SkPreConfig.h: #elif defined(__SSE4_1__)
include/core/SkPreConfig.h: #elif defined(__SSE3__)
include/core/SkPreConfig.h: #elif defined(__SSE2__)
every other check is in SkPreConfig.h where it belongs.
This is going to affect some GMs subtly on Windows.
BUG=chromium:511458
No public API changes.
TBR=reed@google.com
Review URL: https://codereview.chromium.org/1248503004
| -rw-r--r-- | include/core/SkFloatingPoint.h | 12 |
1 files changed, 2 insertions, 10 deletions
diff --git a/include/core/SkFloatingPoint.h b/include/core/SkFloatingPoint.h index ad1669c4ee..7c34706f7a 100644 --- a/include/core/SkFloatingPoint.h +++ b/include/core/SkFloatingPoint.h @@ -143,12 +143,6 @@ extern const uint32_t gIEEENegativeInfinity; #define SK_FloatInfinity (*SkTCast<const float*>(&gIEEEInfinity)) #define SK_FloatNegativeInfinity (*SkTCast<const float*>(&gIEEENegativeInfinity)) -#if defined(__SSE__) -#include <xmmintrin.h> -#elif defined(SK_ARM_HAS_NEON) -#include <arm_neon.h> -#endif - // Fast, approximate inverse square root. // Compare to name-brand "1.0f / sk_float_sqrt(x)". Should be around 10x faster on SSE, 2x on NEON. static inline float sk_float_rsqrt(const float x) { @@ -157,10 +151,8 @@ static inline float sk_float_rsqrt(const float x) { // // We do one step of Newton's method to refine the estimates in the NEON and null paths. No // refinement is faster, but very innacurate. Two steps is more accurate, but slower than 1/sqrt. -#if defined(__SSE__) - float result; - _mm_store_ss(&result, _mm_rsqrt_ss(_mm_set_ss(x))); - return result; +#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE1 + return _mm_cvtss_f32(_mm_rsqrt_ss(_mm_set_ss(x))); #elif defined(SK_ARM_HAS_NEON) // Get initial estimate. const float32x2_t xx = vdup_n_f32(x); // Clever readers will note we're doing everything 2x. |