frnet-p150 / code /tt_frnet /tt /kernels /epi_writer.cpp
changh95's picture
tt-model push frnet-p150 (container)
7a24154 verified
Raw History Blame Contribute Delete
2.05 kB
// SPDX-License-Identifier: Apache-2.0
// EPI writer (frnet-p150 OPT round 2, tt/epi.py): this core's tiles (work split as the reader) to the fp32 map
// (CB 16, OUT32) and / or its bf16 copy (CB 17, OUT16), flushed per tile, one barrier at the end.
// CT args: [0] T, [1] NC, [2] GX, [3] OUT32, [4] OUT16, then the TensorAccessorArgs of the fp32 and the bf16 output
// (repeats of the present one when one is absent). Common RT args: [out32_addr, out16_addr].
#include <cstdint>
#include "api/dataflow/dataflow_api.h"
constexpr uint32_t T = get_compile_time_arg_val(0);
constexpr uint32_t NC = get_compile_time_arg_val(1);
constexpr uint32_t GX = get_compile_time_arg_val(2);
constexpr uint32_t OUT32 = get_compile_time_arg_val(3);
constexpr uint32_t OUT16 = get_compile_time_arg_val(4);
constexpr auto o32_args = TensorAccessorArgs<5>();
constexpr auto o16_args = TensorAccessorArgs<o32_args.next_compile_time_args_offset()>();
constexpr uint32_t cb32 = 16, cb16 = 17, TB32 = 4096, TB16 = 2048;
void kernel_main() {
const uint32_t core = get_absolute_logical_y() * GX + get_absolute_logical_x();
if (core >= NC) {
return;
}
constexpr uint32_t B = T / NC, E = T % NC;
const uint32_t t0 = core * B + (core < E ? core : E);
const uint32_t n = B + (core < E ? 1 : 0);
const auto o32 = TensorAccessor(o32_args, get_common_arg_val<uint32_t>(0), TB32);
const auto o16 = TensorAccessor(o16_args, get_common_arg_val<uint32_t>(1), TB16);
for (uint32_t t = t0; t < t0 + n; ++t) {
if constexpr (OUT32) {
cb_wait_front(cb32, 1);
noc_async_write(get_read_ptr(cb32), o32.get_noc_addr(t), TB32);
}
if constexpr (OUT16) {
cb_wait_front(cb16, 1);
noc_async_write(get_read_ptr(cb16), o16.get_noc_addr(t), TB16);
}
noc_async_writes_flushed();
if constexpr (OUT32) {
cb_pop_front(cb32, 1);
}
if constexpr (OUT16) {
cb_pop_front(cb16, 1);
}
}
noc_async_write_barrier();
}