1// Copyright 2021 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$assert BATCH_TILE % 8 == 0 7$assert BATCH_TILE >= 8 8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <wasm_simd128.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/intrinsics-polyfill.h> 15#include <xnnpack/vcvt.h> 16 17 18$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE] 19$WASM_X16X8_LOAD8X8 = {"QS8": "wasm_i16x8_load8x8", "QU8": "wasm_u16x8_load8x8"}[DATATYPE] 20void xnn_${DATATYPE.lower()}_f32_vcvt_ukernel__wasmsimd_x${BATCH_TILE}( 21 size_t n, 22 const ${XINT8_T}* x, 23 float* y, 24 const union xnn_${DATATYPE.lower()}_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 25{ 26 assert(n != 0); 27 assert(n % sizeof(${XINT8_T}) == 0); 28 assert(x != NULL); 29 assert(y != NULL); 30 31 const v128_t vminus_zero_point = wasm_v128_load64_splat(params->wasmsimd.minus_zero_point); 32 const v128_t vscale = wasm_v128_load64_splat(params->wasmsimd.scale); 33 $if BATCH_TILE > 8: 34 for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) { 35 v128_t vx${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(x); 36 $for N in range(8, BATCH_TILE, 8): 37 v128_t vx${ABC[N:N+8]} = ${WASM_X16X8_LOAD8X8}(x + ${N}); 38 x += ${BATCH_TILE}; 39 40 $for N in range(0, BATCH_TILE, 8): 41 vx${ABC[N:N+8]} = wasm_i16x8_add(vx${ABC[N:N+8]}, vminus_zero_point); 42 43 $for N in range(0, BATCH_TILE, 8): 44 v128_t vy${ABC[N:N+4]} = wasm_i32x4_extend_low_i16x8(vx${ABC[N:N+8]}); 45 v128_t vy${ABC[N+4:N+8]} = wasm_i32x4_extend_high_i16x8(vx${ABC[N:N+8]}); 46 47 $for N in range(0, BATCH_TILE, 4): 48 vy${ABC[N:N+4]} = wasm_f32x4_convert_i32x4(vy${ABC[N:N+4]}); 49 50 $for N in range(0, BATCH_TILE, 4): 51 vy${ABC[N:N+4]} = wasm_f32x4_mul(vy${ABC[N:N+4]}, vscale); 52 53 wasm_v128_store(y, vy${ABC[0:4]}); 54 $for N in range(4, BATCH_TILE, 4): 55 wasm_v128_store(y + ${N}, vy${ABC[N:N+4]}); 56 y += ${BATCH_TILE}; 57 } 58 for (; n >= 8 * sizeof(${XINT8_T}); n -= 8 * sizeof(${XINT8_T})) { 59 v128_t vx = ${WASM_X16X8_LOAD8X8}(x); 60 vx = wasm_i16x8_add(vx, vminus_zero_point); 61 x += 8; 62 63 v128_t vy_lo = wasm_i32x4_extend_low_i16x8(vx); 64 v128_t vy_hi = wasm_i32x4_extend_high_i16x8(vx); 65 66 vy_lo = wasm_f32x4_convert_i32x4(vy_lo); 67 vy_hi = wasm_f32x4_convert_i32x4(vy_hi); 68 69 vy_lo = wasm_f32x4_mul(vy_lo, vscale); 70 vy_hi = wasm_f32x4_mul(vy_hi, vscale); 71 72 wasm_v128_store(y, vy_lo); 73 wasm_v128_store(y + 4, vy_hi); 74 y += 8; 75 } 76 if XNN_UNLIKELY(n != 0) { 77 assert(n >= 1 * sizeof(${XINT8_T})); 78 assert(n <= 7 * sizeof(${XINT8_T})); 79 80 v128_t vx = ${WASM_X16X8_LOAD8X8}(x); 81 vx = wasm_i16x8_add(vx, vminus_zero_point); 82 x += 8; 83 84 v128_t vy = wasm_i32x4_extend_low_i16x8(vx); 85 vy = wasm_f32x4_convert_i32x4(vy); 86 vy = wasm_f32x4_mul(vy, vscale); 87 88 if (n & (4 * sizeof(${XINT8_T}))) { 89 wasm_v128_store(y, vy); y += 4; 90 vy = wasm_i32x4_extend_high_i16x8(vx); 91 vy = wasm_f32x4_convert_i32x4(vy); 92 vy = wasm_f32x4_mul(vy, vscale); 93 } 94 if (n & (2 * sizeof(${XINT8_T}))) { 95 *((double*) y) = wasm_f64x2_extract_lane(vy, 0); 96 vy = wasm_v64x2_shuffle(vy, vy, 1, 1); 97 y += 2; 98 } 99 if (n & (1 * sizeof(${XINT8_T}))) { 100 *y = wasm_f32x4_extract_lane(vy, 0); 101 } 102 } 103} 104