File size: 10,224 Bytes
c699c4c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 | // SPDX-FileCopyrightText: © 2026 Tenstorrent USA, Inc.
// SPDX-License-Identifier: Apache-2.0
//
// Bit-exact Pillow bilinear resize of one 8-bit plane (H x W, DRAM interleaved, one page per source
// row) to the OUT_H x OUT_W uint8 network input (L1 height-sharded, 2 * OUT_ROWS rows per core),
// written straight into the core's input shard. Integer arithmetic of Pillow's Resample.c (8bpc):
// out = clamp(((1 << 21) + sum(px * k)) >> 22, 0, 255), horizontal pass then vertical pass, with the
// 22-bit fixed-point coefficients computed on the host (models/tt/resize_r8.py).
//
// Coefficient tensor (uint32, 256-byte pages):
// HT at word 0: per output column xx: [xmin, k_0 .. k_{kh-1}] (stride kh + 1)
// VT at word vt_off: per output row yy: [ymin, k_0 .. k_{kv-1}] (stride kv + 1)
// Taps past Pillow's count carry k = 0: they read padding / spare rows and add nothing, so every
// output uses exactly kh (kv) taps and the loops are unrolled per (odd) tap count.
//
// Both data-movement RISCs of every core run this kernel (PROC 0 / 1):
// 1. each reads half of the HT and the source rows of its half of the core's source-row span
// [r0, r1) (the rows the core's 2 * OUT_ROWS output rows read); sync
// 2. horizontal pass of its rows (coefficients of a column held in registers across the rows)
// into the shared [rows, OUT_W] buffer; sync
// 3. vertical pass of its OUT_ROWS output rows (4 pixels per 32-bit load) into the shard.
// Semaphores 0 / 1 are the per-RISC progress counters (reset by PROC 0 at the end).
//
// RT args: src_addr, coef_addr, W, kh, kv, vt_off_words, Y0 (first output row of the core), r0, nrows
// CT args: cb, PROC, OUT_W, OUT_ROWS, KHMAX, KVMAX, NRMAX, RBMAX, then TensorAccessorArgs(src), (coef)
#include <stdint.h>
#include "api/dataflow/dataflow_api.h"
static inline uint32_t clip8(int32_t ss) {
if (ss >= (1 << 30)) {
return 255;
}
if (ss <= 0) {
return 0;
}
return (uint32_t)ss >> 22;
}
template <uint32_t KH, uint32_t OUT_W>
static inline void hpass(const int32_t* ht, const uint8_t* rows, uint32_t rb, uint32_t nrows, uint8_t* tmp) {
for (uint32_t xx = 0; xx < OUT_W; ++xx, ht += KH + 1) {
int32_t k[KH];
#pragma GCC unroll 16
for (uint32_t j = 0; j < KH; ++j) {
k[j] = ht[1 + j];
}
const uint8_t* s = rows + ht[0];
uint8_t* o = tmp + xx;
for (uint32_t r = 0; r < nrows; ++r, s += rb, o += OUT_W) {
int32_t ss = 1 << 21;
#pragma GCC unroll 16
for (uint32_t j = 0; j < KH; ++j) {
ss += (int32_t)s[j] * k[j];
}
*o = (uint8_t)clip8(ss);
}
}
}
template <uint32_t KV, uint32_t OUT_W>
static inline void vpass(const int32_t* v, const uint8_t* t0, uint32_t* orow) {
int32_t k[KV];
#pragma GCC unroll 24
for (uint32_t j = 0; j < KV; ++j) {
k[j] = v[j];
}
for (uint32_t xx = 0; xx < OUT_W; xx += 4) {
int32_t a0 = 1 << 21, a1 = 1 << 21, a2 = 1 << 21, a3 = 1 << 21;
const uint32_t* t = reinterpret_cast<const uint32_t*>(t0 + xx);
#pragma GCC unroll 24
for (uint32_t j = 0; j < KV; ++j) {
const uint32_t w = t[j * (OUT_W / 4)];
a0 += (int32_t)(w & 0xFF) * k[j];
a1 += (int32_t)((w >> 8) & 0xFF) * k[j];
a2 += (int32_t)((w >> 16) & 0xFF) * k[j];
a3 += (int32_t)(w >> 24) * k[j];
}
orow[xx / 4] = clip8(a0) | (clip8(a1) << 8) | (clip8(a2) << 16) | (clip8(a3) << 24);
}
}
static inline void wait_ge(volatile tt_l1_ptr uint32_t* s, uint32_t v) {
do {
invalidate_l1_cache();
} while (*s < v);
}
void kernel_main() {
const uint32_t src_addr = get_arg_val<uint32_t>(0);
const uint32_t coef_addr = get_arg_val<uint32_t>(1);
const uint32_t W = get_arg_val<uint32_t>(2);
const uint32_t kh = get_arg_val<uint32_t>(3);
const uint32_t kv = get_arg_val<uint32_t>(4);
const uint32_t vt_off = get_arg_val<uint32_t>(5);
const uint32_t Y0 = get_arg_val<uint32_t>(6);
const uint32_t r0 = get_arg_val<uint32_t>(7); // core's source-row span [r0, r0 + nrows)
const uint32_t nrows = get_arg_val<uint32_t>(8);
constexpr uint32_t cb = get_compile_time_arg_val(0);
constexpr uint32_t PROC = get_compile_time_arg_val(1);
constexpr uint32_t OUT_W = get_compile_time_arg_val(2);
constexpr uint32_t OUT_ROWS = get_compile_time_arg_val(3);
constexpr uint32_t KHMAX = get_compile_time_arg_val(4);
constexpr uint32_t KVMAX = get_compile_time_arg_val(5);
constexpr uint32_t NRMAX = get_compile_time_arg_val(6);
constexpr uint32_t RBMAX = get_compile_time_arg_val(7); // source-row buffer bytes
constexpr uint32_t PAGE = 256;
constexpr uint32_t CORE_ROWS = 2 * OUT_ROWS;
constexpr uint32_t HT_BYTES = ((OUT_W * (KHMAX + 1) * 4 + PAGE - 1) / PAGE) * PAGE;
constexpr uint32_t VT_BYTES = ((CORE_ROWS * (KVMAX + 1) * 4 + 2 * PAGE - 1) / PAGE + 1) * PAGE;
static_assert(OUT_W % 4 == 0, "vertical pass works on 32-bit words");
constexpr auto s_args = TensorAccessorArgs<8>();
constexpr auto c_args = TensorAccessorArgs<s_args.next_compile_time_args_offset()>();
const uint32_t rb = ((W + 63) & ~63u) + 64; // row stride: aligned row + room for padded taps
const auto sacc = TensorAccessor(s_args, src_addr, W);
const auto cacc = TensorAccessor(c_args, coef_addr, PAGE);
const uint32_t base = get_write_ptr(cb);
const uint32_t ht_l1 = base;
const uint32_t vt_l1 = ht_l1 + HT_BYTES + PROC * VT_BYTES;
const uint32_t rows_l1 = ht_l1 + HT_BYTES + 2 * VT_BYTES;
uint8_t* tmp = reinterpret_cast<uint8_t*>(rows_l1 + RBMAX);
// vertical entries of the core's rows: words [vt_off + Y0 * (kv + 1), + CORE_ROWS * (kv + 1))
const uint32_t vw0 = vt_off + Y0 * (kv + 1);
const uint32_t vw1 = vw0 + CORE_ROWS * (kv + 1);
const uint32_t vp0 = vw0 / 64, vp1 = (vw1 + 63) / 64;
for (uint32_t p = vp0; p < vp1; ++p) {
noc_async_read(cacc.get_noc_addr(p), vt_l1 + (p - vp0) * PAGE, PAGE);
}
const uint32_t hp = (OUT_W * (kh + 1) + 63) / 64;
const uint32_t hsplit = hp / 2;
for (uint32_t p = PROC == 0 ? 0 : hsplit; p < (PROC == 0 ? hsplit : hp); ++p) {
noc_async_read(cacc.get_noc_addr(p), ht_l1 + p * PAGE, PAGE);
}
noc_async_read_barrier();
const int32_t* vt = reinterpret_cast<const int32_t*>(vt_l1) + (vw0 - vp0 * 64);
const uint32_t half = (nrows + 1) / 2;
const uint32_t ra = PROC == 0 ? 0 : half, rz = PROC == 0 ? half : nrows;
const uint32_t row_bytes = (W + 63) & ~63u;
for (uint32_t r = ra; r < rz; ++r) {
noc_async_read(sacc.get_noc_addr(r0 + r), rows_l1 + r * rb, row_bytes);
}
noc_async_read_barrier();
volatile tt_l1_ptr uint32_t* s_me = reinterpret_cast<volatile tt_l1_ptr uint32_t*>(get_semaphore(PROC));
volatile tt_l1_ptr uint32_t* s_ot = reinterpret_cast<volatile tt_l1_ptr uint32_t*>(get_semaphore(1 - PROC));
#ifdef RSZ5
// SP_RSZ5: 5 RISCs share the work by output columns (resize_r8_fn.inc rsz5_work). PROC 0 stores the output shard
// address for the TRISCs BEFORE the DM sync below, then (after it, all rows in L1) pushes the scratch CB that starts them.
constexpr uint32_t TMP_BYTES = (NRMAX + KVMAX + 1) * OUT_W;
volatile uint32_t* flags = reinterpret_cast<volatile uint32_t*>(rows_l1 + RBMAX + TMP_BYTES);
if constexpr (PROC == 0) {
flags[8] = get_write_ptr(cb + 1);
asm volatile("fence" ::: "memory");
}
#endif
*s_me = 1;
wait_ge(s_ot, 1); // the whole HT is in L1
#ifdef RSZ5
if constexpr (PROC == 0) {
cb_reserve_back(cb, 1);
cb_push_back(cb, 1);
}
rsz5_work<RSZ_KH, RSZ_KV, OUT_W, CORE_ROWS>(PROC, reinterpret_cast<const int32_t*>(ht_l1), vt, kv + 1,
reinterpret_cast<const uint8_t*>(rows_l1), rb, nrows, tmp, r0,
get_write_ptr(cb + 1));
*s_me = 3;
if constexpr (PROC == 0) {
wait_ge(s_ot, 3);
*s_me = 0;
*s_ot = 0;
}
return;
#endif
// horizontal pass of rows [ra, rz)
const int32_t* ht = reinterpret_cast<const int32_t*>(ht_l1);
const uint8_t* rows = reinterpret_cast<const uint8_t*>(rows_l1) + ra * rb;
uint8_t* tp = tmp + ra * OUT_W;
const uint32_t n = rz - ra;
switch (kh) {
case 3: hpass<3, OUT_W>(ht, rows, rb, n, tp); break;
case 5: hpass<5, OUT_W>(ht, rows, rb, n, tp); break;
case 7: hpass<7, OUT_W>(ht, rows, rb, n, tp); break;
case 9: hpass<9, OUT_W>(ht, rows, rb, n, tp); break;
case 11: hpass<11, OUT_W>(ht, rows, rb, n, tp); break;
case 13: hpass<13, OUT_W>(ht, rows, rb, n, tp); break;
case 15: hpass<15, OUT_W>(ht, rows, rb, n, tp); break;
default: break;
}
*s_me = 2;
wait_ge(s_ot, 2); // all horizontal rows done
// vertical pass of my OUT_ROWS output rows, straight into the shard
uint32_t* out = reinterpret_cast<uint32_t*>(get_write_ptr(cb + 1) + PROC * OUT_ROWS * OUT_W);
for (uint32_t i = 0; i < OUT_ROWS; ++i) {
const int32_t* v = vt + (PROC * OUT_ROWS + i) * (kv + 1);
const uint8_t* t0 = tmp + ((uint32_t)v[0] - r0) * OUT_W;
uint32_t* orow = out + i * (OUT_W / 4);
switch (kv) {
case 3: vpass<3, OUT_W>(v + 1, t0, orow); break;
case 5: vpass<5, OUT_W>(v + 1, t0, orow); break;
case 7: vpass<7, OUT_W>(v + 1, t0, orow); break;
case 9: vpass<9, OUT_W>(v + 1, t0, orow); break;
case 11: vpass<11, OUT_W>(v + 1, t0, orow); break;
case 13: vpass<13, OUT_W>(v + 1, t0, orow); break;
case 15: vpass<15, OUT_W>(v + 1, t0, orow); break;
case 17: vpass<17, OUT_W>(v + 1, t0, orow); break;
case 19: vpass<19, OUT_W>(v + 1, t0, orow); break;
case 21: vpass<21, OUT_W>(v + 1, t0, orow); break;
case 23: vpass<23, OUT_W>(v + 1, t0, orow); break;
default: break;
}
}
*s_me = 3;
if constexpr (PROC == 0) {
wait_ge(s_ot, 3);
*s_me = 0;
*s_ot = 0;
}
}
|