blob: f50ad11621bc516d16ba9efefb8dd116c3bcdc04 [file]
// Copyright 2022 The IREE Authors
//
// Licensed under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
#include "iree/builtins/ukernel/unpack_internal.h"
enum { iree_uk_unpack_tmp_buf_size = 4096 };
// Holds some information and a temporary buffer for performing padding.
typedef struct iree_uk_unpack_tmpbuf_helper_t {
// Temporary buffer to pad the source data into, to pass to the tile_func.
// Cache line alignment helps in pack_benchmark on ARM Cortex-A510/A710.
IREE_UK_ATTRIBUTE_ALIGNED(64) char tmp_buf[iree_uk_unpack_tmp_buf_size];
// How many tiles fit in `tmp_buf`.
int max_tiles_in_tmp_buf;
} iree_uk_unpack_tmpbuf_helper_t;
// Initializes a `iree_uk_unpack_padding_helper`. Asserts if the temporary
// buffer is smaller than one tile.
static void iree_uk_unpack_tmpbuf_helper_init(
iree_uk_index_t tile_size0, iree_uk_index_t tile_size1,
iree_uk_index_t elem_size, iree_uk_unpack_tmpbuf_helper_t* helper) {
helper->max_tiles_in_tmp_buf =
iree_uk_unpack_tmp_buf_size >>
iree_uk_ceil_log2_u32(tile_size0 * tile_size1 * elem_size);
IREE_UK_ASSERT(helper->max_tiles_in_tmp_buf > 0);
}
static void iree_uk_unpack_validate(const iree_uk_unpack_params_t* params) {
#ifdef IREE_UK_ENABLE_ASSERTS
const iree_uk_uint32_t allflags = IREE_UK_FLAG_UNPACK_TRANSPOSE_INNER |
IREE_UK_FLAG_UNPACK_TRANSPOSE_OUTER |
IREE_UK_FLAG_UNPACK_TYPE_MASK;
IREE_UK_ASSERT(!(params->flags & ~allflags));
iree_uk_uint32_t flags_type = params->flags & IREE_UK_FLAG_UNPACK_TYPE_MASK;
IREE_UK_ASSERT(flags_type == IREE_UK_FLAG_UNPACK_TYPE_F32F32 ||
flags_type == IREE_UK_FLAG_UNPACK_TYPE_I32I32 ||
flags_type == IREE_UK_FLAG_UNPACK_TYPE_F16F16 ||
flags_type == IREE_UK_FLAG_UNPACK_TYPE_BF16BF16);
IREE_UK_ASSERT(params->out_size0 >= 0);
IREE_UK_ASSERT(params->out_size1 >= 0);
IREE_UK_ASSERT(params->in_size0 >= 0);
IREE_UK_ASSERT(params->in_size1 >= 0);
IREE_UK_ASSERT(params->in_size2 >= 0);
IREE_UK_ASSERT(params->in_size3 >= 0);
IREE_UK_ASSERT(params->in_stride0 >= 0);
IREE_UK_ASSERT(params->in_stride1 == params->in_size2 * params->in_size3);
IREE_UK_ASSERT(params->out_stride0 >= 0);
IREE_UK_ASSERT(params->out_stride1 >= 0);
// Check that the input and output shapes match, give or take padding that
// must not exceed the inner tile size.s
iree_uk_index_t outer_size0 = params->in_size0;
iree_uk_index_t outer_size1 = params->in_size1;
iree_uk_index_t tile_size0 = params->in_size2;
iree_uk_index_t tile_size1 = params->in_size3;
if (params->flags & IREE_UK_FLAG_UNPACK_TRANSPOSE_OUTER) {
iree_uk_index_swap(&outer_size0, &outer_size1);
}
if (params->flags & IREE_UK_FLAG_UNPACK_TRANSPOSE_INNER) {
iree_uk_index_swap(&tile_size0, &tile_size1);
}
IREE_UK_ASSERT(outer_size0 * tile_size0 >= params->out_size0);
IREE_UK_ASSERT(outer_size1 * tile_size1 >= params->out_size1);
// TODO(#11632): reenable these conditions.
// IREE_UK_ASSERT((outer_size0 - 1) * tile_size0 < params->out_size0);
// IREE_UK_ASSERT((outer_size1 - 1) * tile_size1 < params->out_size1);
// Initialize a padding helper, just to get the assertion that the tile size
// does not exceed the internal temporary buffer size, without having to
// duplicate this arithmetic. Generally, we want to hit all failure modes
// in the validation function so that the subsequent ukernel code can be
// treated as infallible.
iree_uk_unpack_tmpbuf_helper_t helper;
iree_uk_unpack_type_t unpack_type = iree_uk_unpack_type(params->flags);
iree_uk_type_t elem_type = iree_uk_unpack_in_type(unpack_type);
iree_uk_index_t elem_size = iree_uk_type_size(elem_type);
iree_uk_unpack_tmpbuf_helper_init(tile_size0, tile_size1, elem_size, &helper);
#endif // IREE_UK_ENABLE_ASSERTS
}
// Early-return implementation for this ukernel. Returns true if already done.
static bool iree_uk_unpack_early(const iree_uk_unpack_params_t* params) {
return (params->out_size0 == 0 || params->out_size1 == 0);
}
static void iree_uk_copy_1d_unstrided_to_strided_x8(
iree_uk_uint8_t* IREE_UK_RESTRICT dst,
const iree_uk_uint8_t* IREE_UK_RESTRICT src, iree_uk_index_t num_elems,
iree_uk_index_t stride) {
for (iree_uk_index_t i = 0; i < num_elems; ++i) {
dst[i * stride] = src[i];
}
}
static void iree_uk_copy_1d_unstrided_to_strided_x16(
iree_uk_uint16_t* IREE_UK_RESTRICT dst,
const iree_uk_uint16_t* IREE_UK_RESTRICT src, iree_uk_index_t num_elems,
iree_uk_index_t stride) {
for (iree_uk_index_t i = 0; i < num_elems; ++i) {
dst[i * stride] = src[i];
}
}
static void iree_uk_copy_1d_unstrided_to_strided_x32(
iree_uk_uint32_t* IREE_UK_RESTRICT dst,
const iree_uk_uint32_t* IREE_UK_RESTRICT src, iree_uk_index_t num_elems,
iree_uk_index_t stride) {
for (iree_uk_index_t i = 0; i < num_elems; ++i) {
dst[i * stride] = src[i];
}
}
static void iree_uk_copy_1d_unstrided_to_strided(
void* IREE_UK_RESTRICT dst, const void* IREE_UK_RESTRICT src,
iree_uk_index_t num_elems, iree_uk_index_t elem_size,
iree_uk_index_t stride) {
if (stride == 1) {
iree_uk_memcpy(dst, src, num_elems * elem_size);
return;
}
switch (elem_size) {
case 1:
iree_uk_copy_1d_unstrided_to_strided_x8(dst, src, num_elems, stride);
return;
case 2:
iree_uk_copy_1d_unstrided_to_strided_x16(dst, src, num_elems, stride);
return;
case 4:
iree_uk_copy_1d_unstrided_to_strided_x32(dst, src, num_elems, stride);
return;
default:
for (iree_uk_index_t i = 0; i < num_elems; i++) {
iree_uk_memcpy(((char*)dst) + i * stride * elem_size,
((const char*)src) + i * elem_size, elem_size);
}
}
}
static void iree_uk_copy_and_extract_slice(
iree_uk_index_t src_stride0, const char* src_buf, iree_uk_index_t dst_size0,
iree_uk_index_t dst_size1, iree_uk_index_t dst_stride0,
iree_uk_index_t dst_stride1, char* dst_buf, iree_uk_index_t elem_size) {
for (iree_uk_index_t in_i0 = 0; in_i0 < dst_size0; in_i0++) {
iree_uk_copy_1d_unstrided_to_strided(dst_buf, src_buf, dst_size1, elem_size,
dst_stride1);
dst_buf += dst_stride0 * elem_size;
src_buf += src_stride0 * elem_size;
}
}
// Unpacks an entire row, going through the temporary buffer to handle
// incomplete tiles. In cases involving only complete tiles, it is faster to
// call tile_func directly.
static void iree_uk_unpack_row_using_tile_func(
iree_uk_unpack_tile_func_t tile_func, iree_uk_index_t dim1_tile_start,
iree_uk_index_t dim1_tile_end, iree_uk_index_t dim0_write_size,
iree_uk_index_t tile_size0, iree_uk_index_t tile_size1,
iree_uk_index_t elem_size, iree_uk_index_t out_size1,
iree_uk_index_t out_stride0, iree_uk_index_t out_stride1,
iree_uk_index_t in_stride1, bool whole_tiles,
iree_uk_unpack_tmpbuf_helper_t* helper, const char* in_buf, char* out_buf) {
if (whole_tiles && out_stride1 == 1) {
tile_func(out_buf + dim1_tile_start * tile_size1 * out_stride1 * elem_size,
in_buf + dim1_tile_start * in_stride1 * elem_size,
dim1_tile_end - dim1_tile_start, out_stride0, in_stride1,
elem_size, tile_size0, tile_size1);
return;
}
iree_uk_index_t dim1_tile = dim1_tile_start;
while (dim1_tile < dim1_tile_end) {
iree_uk_index_t dim1_chunk_tiles = iree_uk_index_clamp(
dim1_tile_end - dim1_tile, 0, helper->max_tiles_in_tmp_buf);
iree_uk_index_t dim1_chunk_src_width = dim1_chunk_tiles * tile_size1;
iree_uk_index_t dim1_chunk_src_pos = dim1_tile * tile_size1;
iree_uk_index_t dim1_write_size = iree_uk_index_clamp(
out_size1 - dim1_chunk_src_pos, 0, dim1_chunk_src_width);
tile_func(helper->tmp_buf, in_buf + dim1_tile * in_stride1 * elem_size,
dim1_chunk_tiles, dim1_chunk_src_width, in_stride1, elem_size,
tile_size0, tile_size1);
iree_uk_copy_and_extract_slice(
dim1_chunk_src_width, helper->tmp_buf, dim0_write_size, dim1_write_size,
out_stride0, out_stride1,
out_buf + dim1_tile * tile_size1 * out_stride1 * elem_size, elem_size);
dim1_tile += dim1_chunk_tiles;
}
}
static void iree_uk_unpack_using_tile_func(
const iree_uk_unpack_params_t* params,
iree_uk_unpack_tile_func_t tile_func) {
// For now, the input and output element types are always the same.
iree_uk_unpack_type_t unpack_type = iree_uk_unpack_type(params->flags);
iree_uk_type_t elem_type = iree_uk_unpack_in_type(unpack_type);
iree_uk_index_t elem_size = iree_uk_type_size(elem_type);
iree_uk_index_t outer_size0 = params->in_size0;
iree_uk_index_t outer_size1 = params->in_size1;
iree_uk_index_t tile_size0 = params->in_size2;
iree_uk_index_t tile_size1 = params->in_size3;
iree_uk_index_t in_stride0 = params->in_stride0;
iree_uk_index_t in_stride1 = params->in_size3 * params->in_size2;
if (params->flags & IREE_UK_FLAG_UNPACK_TRANSPOSE_OUTER) {
iree_uk_index_swap(&outer_size0, &outer_size1);
iree_uk_index_swap(&in_stride0, &in_stride1);
}
if (params->flags & IREE_UK_FLAG_UNPACK_TRANSPOSE_INNER) {
iree_uk_index_swap(&tile_size0, &tile_size1);
}
const char* in_buf =
(const char*)params->in_buffer + (params->in_offset * elem_size);
char* out_buf = (char*)params->out_buffer + (params->out_offset * elem_size);
// Prepare for handling incomplete tiles with a temporary buffer.
iree_uk_unpack_tmpbuf_helper_t helper;
iree_uk_unpack_tmpbuf_helper_init(tile_size0, tile_size1, elem_size, &helper);
// Compute number of tiles along both dimensions that fit entirely within the
// source buffer's boundaries.
int dim1_full_tiles = params->out_size1 >> iree_uk_ceil_log2_u32(tile_size1);
iree_uk_index_t i0 = 0;
for (; i0 <= params->out_size0 - tile_size0; i0 += tile_size0) {
// Pack whole tiles that do not require padding (entirely within the source
// buffer's boundaries).
iree_uk_unpack_row_using_tile_func(
tile_func, 0, dim1_full_tiles, tile_size0, tile_size0, tile_size1,
elem_size, params->out_size1, params->out_stride0, params->out_stride1,
in_stride1, /*whole_tiles=*/true, &helper, in_buf, out_buf);
// Right-padding.
iree_uk_unpack_row_using_tile_func(
tile_func, dim1_full_tiles, outer_size1, tile_size0, tile_size0,
tile_size1, elem_size, params->out_size1, params->out_stride0,
params->out_stride1, in_stride1, /*whole_tiles=*/false, &helper, in_buf,
out_buf);
out_buf += tile_size0 * params->out_stride0 * elem_size;
in_buf += in_stride0 * elem_size;
}
// Bottom-padding.
for (; i0 < outer_size0 * tile_size0; i0 += tile_size0) {
iree_uk_index_t dim0_write_size =
iree_uk_index_clamp(params->out_size0 - i0, 0, tile_size0);
iree_uk_unpack_row_using_tile_func(
tile_func, 0, outer_size1, dim0_write_size, tile_size0, tile_size1,
elem_size, params->out_size1, params->out_stride0, params->out_stride1,
in_stride1, /*whole_tiles=*/false, &helper, in_buf, out_buf);
out_buf += tile_size0 * params->out_stride0 * elem_size;
in_buf += in_stride0 * elem_size;
}
}
void iree_uk_unpack_p(const iree_uk_unpack_params_t* params) {
iree_uk_unpack_validate(params);
if (iree_uk_unpack_early(params)) return;
// Select a target-specific tile_func and use that with generic outer loops.
iree_uk_unpack_tile_func_t func = iree_uk_unpack_select_tile_func(params);
iree_uk_unpack_using_tile_func(params, func);
}
IREE_UK_EXPORT void iree_uk_unpack(
const void* in_buffer, iree_uk_index_t in_offset,
iree_uk_index_t in_stride0, iree_uk_index_t in_stride1, void* out_buffer,
iree_uk_index_t out_offset, iree_uk_index_t out_stride0,
iree_uk_index_t out_stride1, iree_uk_index_t in_size0,
iree_uk_index_t in_size1, iree_uk_index_t in_size2,
iree_uk_index_t in_size3, iree_uk_index_t out_size0,
iree_uk_index_t out_size1, iree_uk_uint32_t flags,
const iree_uk_uint64_t* cpu_data) {
iree_uk_unpack_params_t params = {.in_buffer = in_buffer,
.in_offset = in_offset,
.in_stride0 = in_stride0,
.in_stride1 = in_stride1,
.out_buffer = out_buffer,
.out_offset = out_offset,
.out_stride0 = out_stride0,
.out_stride1 = out_stride1,
.in_size0 = in_size0,
.in_size1 = in_size1,
.in_size2 = in_size2,
.in_size3 = in_size3,
.out_size0 = out_size0,
.out_size1 = out_size1,
.flags = flags,
.cpu_data = cpu_data};
iree_uk_unpack_p(&params);
}