Marat Dukhan | d3f3d87 | 2020-06-24 13:08:25 -0700 | [diff] [blame^] | 1 | // Copyright 2020 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #include <assert.h> |
| 7 | #include <stddef.h> |
| 8 | #include <stdint.h> |
| 9 | |
| 10 | #include <wasm_simd128.h> |
| 11 | |
| 12 | #include <xnnpack/math-stubs.h> |
| 13 | |
| 14 | |
| 15 | void xnn_math_f32_roundne__wasmsimd_addsub( |
| 16 | size_t n, |
| 17 | const float* input, |
| 18 | float* output) |
| 19 | { |
| 20 | assert(n % (4 * sizeof(float)) == 0); |
| 21 | |
| 22 | // Mask for the sign bit of a floating-point number. |
| 23 | const v128_t vsign_mask = wasm_i32x4_splat(INT32_C(0x80000000)); |
| 24 | // Addition of this number to a floating-point number x cause rounding of the result to an integer. Then this magic |
| 25 | // number is subtracted back from the result to get original x rounded to integer. This trick works only for |
| 26 | // 0 <= x < 2**24, but all numbers in 2**23 <= x < 2**24 range are integers, so we can further restrict it to |
| 27 | // 0 <= x < 2**23. Then the upper bound of the validity interval is conveniently the same as the magic number. |
| 28 | const v128_t vmagic_number = wasm_f32x4_splat(0x1.000000p+23f); |
| 29 | |
| 30 | for (; n != 0; n -= 4 * sizeof(float)) { |
| 31 | const v128_t vx = wasm_v128_load(input); |
| 32 | input += 4; |
| 33 | |
| 34 | // The rounding trick works only for x >= 0, so we compute absolute value of x, round it, and restore the sign in |
| 35 | // the end. This method works for round-to-nearest-even because it is an odd function. |
| 36 | const v128_t vabsx = wasm_v128_andnot(vx, vsign_mask); |
| 37 | |
| 38 | // Compute bitmask for the bits we want to copy from x. Other bits will be copied from the rounded abs(x). |
| 39 | // If abs(x) < 2**23 or x is NaN, we want the sign bit from x and the rest from the rounded abs(x). |
| 40 | // Otherwise (abs(x) >= 2**23), we want all bits from x. |
| 41 | const v128_t vrndmask = wasm_v128_or(vsign_mask, wasm_f32x4_gt(vabsx, vmagic_number)); |
| 42 | // Addition-subtraction trick with the magic number to cause rounding to integer for abs(x). |
| 43 | // Note: the result is valid only for 0 <= abs(x) < 2**23. |
| 44 | // Note: addition-subtraction implicitly converts SNaN inputs to QNaNs. |
| 45 | const v128_t vrndabsx = wasm_f32x4_sub(wasm_f32x4_add(vabsx, vmagic_number), vmagic_number); |
| 46 | |
| 47 | // Combine abs(x) rounded via addition-subtraction trick and the input x value. |
| 48 | // For abs(x) < 2**23, the result is abs(x) rounded via addition-subtraction trick with the sign of x. |
| 49 | // For NaN inputs, the result is x converted to QNaN as a side-effect of addition-subtraction. |
| 50 | // For abs(x) >= 2**23, the result is x itself. |
| 51 | const v128_t vy = wasm_v128_bitselect(vx, vrndabsx, vrndmask); |
| 52 | |
| 53 | wasm_v128_store(output, vy); |
| 54 | output += 4; |
| 55 | } |
| 56 | } |