scale by 1 for neon implemented

scale by 1 for neon implemented


void HalfFloat1Row_NEON(const uint16* src, uint16* dst, float, int width) {
  asm volatile (
  "1:                                          \n"
    MEMACCESS(0)
    "ld1        {v1.16b}, [%0], #16            \n"  // load 8 shorts
    "subs       %w2, %w2, #8                   \n"  // 8 pixels per loop
    "uxtl       v2.4s, v1.4h                   \n"  // 8 int's
    "uxtl2      v1.4s, v1.8h                   \n"
    "scvtf      v2.4s, v2.4s                   \n"  // 8 floats
    "scvtf      v1.4s, v1.4s                   \n"
    "fcvtn      v4.4h, v2.4s                   \n"  // 8 floatsgit
    "fcvtn2     v4.8h, v1.4s                   \n"
   MEMACCESS(1)
    "st1        {v4.16b}, [%1], #16            \n"  // store 8 shorts
    "b.gt       1b                             \n"
  : "+r"(src),    // %0
    "+r"(dst),    // %1
    "+r"(width)   // %2
  :
  : "cc", "memory", "v1", "v2", "v4"
  );
}

void HalfFloatRow_NEON(const uint16* src, uint16* dst, float scale, int width) {
  asm volatile (
  "1:                                          \n"
    MEMACCESS(0)
    "ld1        {v1.16b}, [%0], #16            \n"  // load 8 shorts
    "subs       %w2, %w2, #8                   \n"  // 8 pixels per loop
    "uxtl       v2.4s, v1.4h                   \n"  // 8 int's
    "uxtl2      v1.4s, v1.8h                   \n"
    "scvtf      v2.4s, v2.4s                   \n"  // 8 floats
    "scvtf      v1.4s, v1.4s                   \n"
    "fmul       v2.4s, v2.4s, %3.s[0]          \n"  // adjust exponent
    "fmul       v1.4s, v1.4s, %3.s[0]          \n"
    "uqshrn     v4.4h, v2.4s, #13              \n"  // isolate halffloat
    "uqshrn2    v4.8h, v1.4s, #13              \n"
   MEMACCESS(1)
    "st1        {v4.16b}, [%1], #16            \n"  // store 8 shorts
    "b.gt       1b                             \n"
  : "+r"(src),    // %0
    "+r"(dst),    // %1
    "+r"(width)   // %2
  : "w"(scale * 1.9259299444e-34f)    // %3
  : "cc", "memory", "v1", "v2", "v4"
  );
}

TEST=LibYUVPlanarTest.TestHalfFloatPlane_One
BUG=libyuv:560
R=hubbe@chromium.org

Committed: https://chromium.googlesource.com/libyuv/libyuv/+/451af5e922e026c266d25abc92e7519acfc9a4c5

[fbarchard1, 2016-10-20 23:38:42]

fbarchard@google.com changed reviewers:
+ hubbe@chromium.org, wangcheng@google.com

[fbarchard1, 2016-10-20 23:38:42]

util/android/test_runner.py gtest -s libyuv_unittest -t 7200 --verbose --release
--gtest_filter=*$* -a "--libyuv_width=640 --libyuv_height=360
--libyuv_repeat=39999 --libyuv_flags=-1  --libyuv_cpu_info=-1"

TestHalfFloatPlane_denormal (32108 ms)
TestHalfFloatPlane_Opt (24340 ms)
TestHalfFloatPlane_One (13821 ms)
TestHalfFloatPlane_Offby1 (27022 ms)

[fbarchard1, 2016-10-21 19:29:37]

Test performance is not super consistent.  In this CL I changed the 3
allocations to 1 in the test and added more tests.

TestHalfFloatPlane_16bit_denormal (869 ms)
TestHalfFloatPlane_10bit_Opt (914 ms)
TestHalfFloatPlane_9bit_Opt (900 ms)
TestHalfFloatPlane_Opt (883 ms)
TestHalfFloatPlane_One (803 ms)
TestHalfFloatPlane_Offby1 (922 ms)

[hubbe, 2016-10-21 19:32:52]

lgtm

[fbarchard1, 2016-10-21 21:30:06]

Description was changed from

==========
scale by 1 for neon implemented


void HalfFloat1Row_NEON(const uint16* src, uint16* dst, float, int width) {
  asm volatile (
  "1:                                          \n"
    MEMACCESS(0)
    "ld1        {v1.16b}, [%0], #16            \n"  // load 8 shorts
    "subs       %w2, %w2, #8                   \n"  // 8 pixels per loop
    "uxtl       v2.4s, v1.4h                   \n"  // 8 int's
    "uxtl2      v1.4s, v1.8h                   \n"
    "scvtf      v2.4s, v2.4s                   \n"  // 8 floats
    "scvtf      v1.4s, v1.4s                   \n"
    "fcvtn      v4.4h, v2.4s                   \n"  // 8 floatsgit
    "fcvtn2     v4.8h, v1.4s                   \n"
   MEMACCESS(1)
    "st1        {v4.16b}, [%1], #16            \n"  // store 8 shorts
    "b.gt       1b                             \n"
  : "+r"(src),    // %0
    "+r"(dst),    // %1
    "+r"(width)   // %2
  :
  : "cc", "memory", "v1", "v2", "v4"
  );
}

void HalfFloatRow_NEON(const uint16* src, uint16* dst, float scale, int width) {
  asm volatile (
  "1:                                          \n"
    MEMACCESS(0)
    "ld1        {v1.16b}, [%0], #16            \n"  // load 8 shorts
    "subs       %w2, %w2, #8                   \n"  // 8 pixels per loop
    "uxtl       v2.4s, v1.4h                   \n"  // 8 int's
    "uxtl2      v1.4s, v1.8h                   \n"
    "scvtf      v2.4s, v2.4s                   \n"  // 8 floats
    "scvtf      v1.4s, v1.4s                   \n"
    "fmul       v2.4s, v2.4s, %3.s[0]          \n"  // adjust exponent
    "fmul       v1.4s, v1.4s, %3.s[0]          \n"
    "uqshrn     v4.4h, v2.4s, #13              \n"  // isolate halffloat
    "uqshrn2    v4.8h, v1.4s, #13              \n"
   MEMACCESS(1)
    "st1        {v4.16b}, [%1], #16            \n"  // store 8 shorts
    "b.gt       1b                             \n"
  : "+r"(src),    // %0
    "+r"(dst),    // %1
    "+r"(width)   // %2
  : "w"(scale * 1.9259299444e-34f)    // %3
  : "cc", "memory", "v1", "v2", "v4"
  );
}

TEST=LibYUVPlanarTest.TestHalfFloatPlane_One
BUG=libyuv:560
==========

to

==========
scale by 1 for neon implemented


void HalfFloat1Row_NEON(const uint16* src, uint16* dst, float, int width) {
  asm volatile (
  "1:                                          \n"
    MEMACCESS(0)
    "ld1        {v1.16b}, [%0], #16            \n"  // load 8 shorts
    "subs       %w2, %w2, #8                   \n"  // 8 pixels per loop
    "uxtl       v2.4s, v1.4h                   \n"  // 8 int's
    "uxtl2      v1.4s, v1.8h                   \n"
    "scvtf      v2.4s, v2.4s                   \n"  // 8 floats
    "scvtf      v1.4s, v1.4s                   \n"
    "fcvtn      v4.4h, v2.4s                   \n"  // 8 floatsgit
    "fcvtn2     v4.8h, v1.4s                   \n"
   MEMACCESS(1)
    "st1        {v4.16b}, [%1], #16            \n"  // store 8 shorts
    "b.gt       1b                             \n"
  : "+r"(src),    // %0
    "+r"(dst),    // %1
    "+r"(width)   // %2
  :
  : "cc", "memory", "v1", "v2", "v4"
  );
}

void HalfFloatRow_NEON(const uint16* src, uint16* dst, float scale, int width) {
  asm volatile (
  "1:                                          \n"
    MEMACCESS(0)
    "ld1        {v1.16b}, [%0], #16            \n"  // load 8 shorts
    "subs       %w2, %w2, #8                   \n"  // 8 pixels per loop
    "uxtl       v2.4s, v1.4h                   \n"  // 8 int's
    "uxtl2      v1.4s, v1.8h                   \n"
    "scvtf      v2.4s, v2.4s                   \n"  // 8 floats
    "scvtf      v1.4s, v1.4s                   \n"
    "fmul       v2.4s, v2.4s, %3.s[0]          \n"  // adjust exponent
    "fmul       v1.4s, v1.4s, %3.s[0]          \n"
    "uqshrn     v4.4h, v2.4s, #13              \n"  // isolate halffloat
    "uqshrn2    v4.8h, v1.4s, #13              \n"
   MEMACCESS(1)
    "st1        {v4.16b}, [%1], #16            \n"  // store 8 shorts
    "b.gt       1b                             \n"
  : "+r"(src),    // %0
    "+r"(dst),    // %1
    "+r"(width)   // %2
  : "w"(scale * 1.9259299444e-34f)    // %3
  : "cc", "memory", "v1", "v2", "v4"
  );
}

TEST=LibYUVPlanarTest.TestHalfFloatPlane_One
BUG=libyuv:560
R=hubbe@chromium.org

Committed:
https://chromium.googlesource.com/libyuv/libyuv/+/451af5e922e026c266d25abc92e...
==========

[fbarchard1, 2016-10-21 21:30:07]

Committed patchset #6 (id:100001) manually as
451af5e922e026c266d25abc92e7519acfc9a4c5 (presubmit successful).