Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | |
| 9 | #include <assert.h> |
| 10 | #include <stdint.h> |
| 11 | #include <stddef.h> |
| 12 | |
| 13 | #include <fp16/bitcasts.h> |
| 14 | |
Marat Dukhan | f86ee8b | 2021-05-24 23:41:31 -0700 | [diff] [blame] | 15 | #include <xnnpack/math.h> |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 16 | #include <xnnpack/requantization-stubs.h> |
| 17 | |
| 18 | |
Marat Dukhan | 0671624 | 2021-05-26 15:56:39 -0700 | [diff] [blame] | 19 | void xnn_qu8_requantize_rndna__scalar_unsigned32( |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 20 | size_t n, |
| 21 | const int32_t* input, |
| 22 | float scale, |
| 23 | uint8_t zero_point, |
| 24 | uint8_t qmin, |
| 25 | uint8_t qmax, |
| 26 | uint8_t* output) |
| 27 | { |
| 28 | assert(n % 4 == 0); |
| 29 | assert(scale < 1.0f); |
| 30 | assert(scale >= 0x1.0p-32f); |
| 31 | |
| 32 | const uint32_t scale_bits = fp32_to_bits(scale); |
| 33 | const uint32_t multiplier = (scale_bits << 8) | UINT32_C(0x80000000); |
| 34 | const uint32_t shift = 127 + 31 - (scale_bits >> 23); |
| 35 | assert(shift >= 32); |
| 36 | assert(shift < 64); |
| 37 | |
| 38 | const uint64_t rounding = UINT64_C(1) << (shift - 1); |
Marat Dukhan | 062bee3 | 2021-05-27 20:31:07 -0700 | [diff] [blame] | 39 | const uint32_t rounding_hi = (uint32_t) (rounding >> 32); |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 40 | const uint32_t rounding_lo = (uint32_t) rounding; |
| 41 | const uint32_t shift_minus_32 = shift - 32; |
Marat Dukhan | 062bee3 | 2021-05-27 20:31:07 -0700 | [diff] [blame] | 42 | const int32_t smin = (int32_t) (uint32_t) qmin - (int32_t) (uint32_t) zero_point; |
| 43 | const int32_t smax = (int32_t) (uint32_t) qmax - (int32_t) (uint32_t) zero_point; |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 44 | for (; n != 0; n -= 4) { |
| 45 | const int32_t x = input[0]; |
| 46 | const int32_t y = input[1]; |
| 47 | const int32_t z = input[2]; |
| 48 | const int32_t w = input[3]; |
| 49 | input += 4; |
| 50 | |
| 51 | // Compute absolute value of input as unsigned 32-bit int. |
| 52 | // All further computations will work with unsigned values to avoid undefined behaviour on signed operations. |
| 53 | const uint32_t x_abs = (x >= 0) ? (uint32_t) x : -(uint32_t) x; |
| 54 | const uint32_t y_abs = (y >= 0) ? (uint32_t) y : -(uint32_t) y; |
| 55 | const uint32_t z_abs = (z >= 0) ? (uint32_t) z : -(uint32_t) z; |
| 56 | const uint32_t w_abs = (w >= 0) ? (uint32_t) w : -(uint32_t) w; |
| 57 | |
| 58 | // Compute full 64-bit product of 32-bit factors. |
| 59 | const uint64_t x_product = (uint64_t) x_abs * (uint64_t) multiplier; |
| 60 | const uint64_t y_product = (uint64_t) y_abs * (uint64_t) multiplier; |
| 61 | const uint64_t z_product = (uint64_t) z_abs * (uint64_t) multiplier; |
| 62 | const uint64_t w_product = (uint64_t) w_abs * (uint64_t) multiplier; |
| 63 | |
| 64 | // Shift the full 64-bit product right with rounding. |
| 65 | // Rounding is performed towards closest integer, with midpoints rounded up (same as away from zero). |
| 66 | // |
| 67 | // Generally, this operation requires both 64-bit addition and 64-bit shift, but we use two tricks to replace |
| 68 | // 64-bit operations with 32-bit operations. |
| 69 | // |
| 70 | // To avoid full 64-bit addition we make use of three facts: |
| 71 | // - 64-bit rounding value added before the shift is a power of 2, and thus has only one bit set. |
| 72 | // - When 0x1.0p-32f <= scale < 0x1.0p-31f, then the non-zero bit in rounding is in the low 32 bits, and |
| 73 | // rounding is exactly 0x80000000 (2**31), because rounding is 2**(scale-1) and scale >= 32. In this case, |
| 74 | // addition of rounding can affect high 32 bits of the product only through overflow, which happens if |
| 75 | // low 32-bit part of the product equals or exceeds 0x80000000. We can reformulate the latter condition |
| 76 | // as low 32-bit part of the product has the bit 31 set, and then overflow happens if both the low 32-bit part |
| 77 | // of the product and the low 32-bit part of the rounding value have bit 31 set. Since 32-bit numbers with the |
| 78 | // bit 31 set are negative when interpreted as signed integers, we can check the overflow condition as |
| 79 | // (int32_t) (LOW(product) & LOW(rounding)) < 0 |
| 80 | // - When 0x1.0p-31f <= scale < 1.0f, then the non-zero bit is in the high 32 bits of rounding. We just need |
| 81 | // to do 32-bit addition of high 32 bits of rounding and high 32 bits of product. This addition never |
| 82 | // overflows because product <= 0x80000000 * 0xFFFFFF00 < 2**63 and rounding = 2**(scale-1) <= 2**62. |
| 83 | // |
| 84 | // To avoid full 64-bit shift, we leverage the fact that shift >= 32, and do it in two steps: |
| 85 | // - Shift by 32, which can be implemented by extacting the high 32-bit word on 32-bit systems. |
| 86 | // - Shift by (shift - 32), which can be implemented as a 32-bit shift of high word of addition result. |
Marat Dukhan | 062bee3 | 2021-05-27 20:31:07 -0700 | [diff] [blame] | 87 | const uint32_t x_carry_lo = (uint32_t) ((int32_t)((uint32_t) x_product & rounding_lo) < 0); |
| 88 | const uint32_t y_carry_lo = (uint32_t) ((int32_t)((uint32_t) y_product & rounding_lo) < 0); |
| 89 | const uint32_t z_carry_lo = (uint32_t) ((int32_t)((uint32_t) z_product & rounding_lo) < 0); |
| 90 | const uint32_t w_carry_lo = (uint32_t) ((int32_t)((uint32_t) w_product & rounding_lo) < 0); |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 91 | |
Marat Dukhan | 062bee3 | 2021-05-27 20:31:07 -0700 | [diff] [blame] | 92 | const uint32_t x_product_hi = (uint32_t) (x_product >> 32); |
| 93 | const uint32_t y_product_hi = (uint32_t) (y_product >> 32); |
| 94 | const uint32_t z_product_hi = (uint32_t) (z_product >> 32); |
| 95 | const uint32_t w_product_hi = (uint32_t) (w_product >> 32); |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 96 | |
Marat Dukhan | 062bee3 | 2021-05-27 20:31:07 -0700 | [diff] [blame] | 97 | const uint32_t x_abs_scaled = (uint32_t) (x_product_hi + rounding_hi + x_carry_lo) >> shift_minus_32; |
| 98 | const uint32_t y_abs_scaled = (uint32_t) (y_product_hi + rounding_hi + y_carry_lo) >> shift_minus_32; |
| 99 | const uint32_t z_abs_scaled = (uint32_t) (z_product_hi + rounding_hi + z_carry_lo) >> shift_minus_32; |
| 100 | const uint32_t w_abs_scaled = (uint32_t) (w_product_hi + rounding_hi + w_carry_lo) >> shift_minus_32; |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 101 | |
| 102 | // Copy the sign of input to scaled absolute input value. |
Marat Dukhan | 062bee3 | 2021-05-27 20:31:07 -0700 | [diff] [blame] | 103 | const int32_t x_scaled = (int32_t) (x >= 0 ? x_abs_scaled : -x_abs_scaled); |
| 104 | const int32_t y_scaled = (int32_t) (y >= 0 ? y_abs_scaled : -y_abs_scaled); |
| 105 | const int32_t z_scaled = (int32_t) (z >= 0 ? z_abs_scaled : -z_abs_scaled); |
| 106 | const int32_t w_scaled = (int32_t) (w >= 0 ? w_abs_scaled : -w_abs_scaled); |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 107 | |
| 108 | // Clamp scaled value with zero point between (qmin - zero point) and (qmax - zero point). |
Marat Dukhan | 062bee3 | 2021-05-27 20:31:07 -0700 | [diff] [blame] | 109 | const int32_t x_clamped = math_min_s32(math_max_s32(x_scaled, smin), smax); |
| 110 | const int32_t y_clamped = math_min_s32(math_max_s32(y_scaled, smin), smax); |
| 111 | const int32_t z_clamped = math_min_s32(math_max_s32(z_scaled, smin), smax); |
| 112 | const int32_t w_clamped = math_min_s32(math_max_s32(w_scaled, smin), smax); |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 113 | |
| 114 | // Add zero point to clamped value. |
| 115 | // The result is guaranteed to be in [qmin, qmax] range. |
| 116 | // |
| 117 | // This addition can not be safely done before clamping, because scaled values are in [-2147483520, 2147483519] |
| 118 | // range, so addition of zero point (which can be up to 255) can overflow signed 32-bit integer. |
| 119 | const int32_t x_biased = x_clamped + zero_point; |
| 120 | const int32_t y_biased = y_clamped + zero_point; |
| 121 | const int32_t z_biased = z_clamped + zero_point; |
| 122 | const int32_t w_biased = w_clamped + zero_point; |
| 123 | |
| 124 | output[0] = (uint8_t) x_biased; |
| 125 | output[1] = (uint8_t) y_biased; |
| 126 | output[2] = (uint8_t) z_biased; |
| 127 | output[3] = (uint8_t) w_biased; |
| 128 | output += 4; |
| 129 | } |
| 130 | } |