blob: 6cb6ea9d77795664345747a443702c5f9356e1f8 [file] [log] [blame] [edit]
// Copyright 2021 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
$from itertools import chain
$import math
$assert IN_PTRS in ["MULTI", "REUSE"]
$assert OUT_PTRS in ["MULTI", "SWITCH", "MOV", "DEC"]
$assert SIZE in [8, 16, 32]
$TILE_SIZE = int(128/SIZE)
$NUM_ITERS = int(math.log2(TILE_SIZE))
$LO_PERM=str(list(chain.from_iterable((i, i+TILE_SIZE) for i in range((TILE_SIZE>>1)))))[1:-1]
$HI_PERM=str(list(chain.from_iterable(((TILE_SIZE>>1)+i, (TILE_SIZE>>1)+i+TILE_SIZE) for i in range((TILE_SIZE>>1)))))[1:-1]
#include <wasm_simd128.h>
#include <assert.h>
#include <xnnpack/common.h>
#include <xnnpack/math.h>
#include <xnnpack/transpose.h>
void xnn_x${SIZE}_transposec_ukernel__${TILE_SIZE}x${TILE_SIZE}_${IN_PTRS.lower()}_${OUT_PTRS.lower()}_wasmsimd(
const uint${SIZE}_t* input,
uint${SIZE}_t* output,
size_t input_stride,
size_t output_stride,
size_t block_width,
size_t block_height) XNN_OOB_READS
{
assert(output_stride >= block_height * sizeof(uint${SIZE}_t));
assert(input_stride >= block_width * sizeof(uint${SIZE}_t));
const size_t tile_height = ${TILE_SIZE};
const size_t tile_width = ${TILE_SIZE};
const size_t tile_hbytes = tile_height * sizeof(uint${SIZE}_t);
const size_t tile_wbytes = tile_width * sizeof(uint${SIZE}_t);
const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
$if IN_PTRS == "MULTI":
const size_t input_offset = tile_height * input_stride;
$if OUT_PTRS in ["MOV", "DEC"]:
const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint${SIZE}_t) - tile_hbytes;
$else:
const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint${SIZE}_t);
$if IN_PTRS == "MULTI":
const uint${SIZE}_t* i0 = input;
$for N in range(1, TILE_SIZE):
const uint${SIZE}_t* i${N} = (const uint${SIZE}_t*) ((uintptr_t) i${N-1} + input_stride);
$else:
const uint${SIZE}_t* i0 = input;
$if OUT_PTRS == "MULTI":
uint${SIZE}_t* o0 = (uint${SIZE}_t*) output;
$for N in range(1, TILE_SIZE):
uint${SIZE}_t* o${N} = (uint${SIZE}_t*) ((uintptr_t) o${N-1} + output_stride);
$elif OUT_PTRS == "SWITCH":
uint${SIZE}_t* o = (uint${SIZE}_t*) output;
$else:
uint${SIZE}_t* o = (uint${SIZE}_t*) ((uintptr_t) output - tile_hbytes);
$if OUT_PTRS != "MULTI":
const size_t minus_output_stride = -output_stride;
do {
$if OUT_PTRS == "MULTI":
if XNN_UNPREDICTABLE(block_width < 2) {
o1 = o0;
}
$for N in range(2, TILE_SIZE, 2):
if XNN_UNPREDICTABLE(block_width <= ${N}) {
o${N} = o0;
}
if XNN_UNPREDICTABLE(block_width < ${N+2}) {
o${N+1} = o0;
}
$elif OUT_PTRS in ["MOV", "DEC"]:
const size_t rem = min(block_width - 1, ${TILE_SIZE-1});
const size_t oN_stride = rem * output_stride;
const size_t oN_offset = oN_stride + tile_hbytes;
$else:
const size_t rem = min(block_width - 1, ${TILE_SIZE-1});
const size_t oN_stride = rem * output_stride;
size_t bh = block_height;
for (; bh >= ${TILE_SIZE}; bh -= ${TILE_SIZE}) {
$for N in range(TILE_SIZE):
$if IN_PTRS == "REUSE":
const v128_t v${NUM_ITERS}_${N} = wasm_v128_load(i0);
i0 = (uint${SIZE}_t*) ((uintptr_t) i0 + input_stride);
$else:
const v128_t v${NUM_ITERS}_${N} = wasm_v128_load(i${N});
i${N} = (uint${SIZE}_t*) ((uintptr_t) i${N} + input_offset);
$for M in range(NUM_ITERS):
$for N in range(TILE_SIZE >> 1):
const v128_t v${NUM_ITERS-M-1}_${2*N} = wasm_v${SIZE}x${TILE_SIZE}_shuffle(v${NUM_ITERS-M}_${N}, v${NUM_ITERS-M}_${N+int(TILE_SIZE/2)}, ${LO_PERM});
const v128_t v${NUM_ITERS-M-1}_${2*N+1} = wasm_v${SIZE}x${TILE_SIZE}_shuffle(v${NUM_ITERS-M}_${N}, v${NUM_ITERS-M}_${N+int(TILE_SIZE/2)}, ${HI_PERM});
$if OUT_PTRS == "SWITCH":
uint${SIZE}_t *oN = (uint${SIZE}_t*) ((uintptr_t) o + oN_stride);
switch (rem) {
$for N in reversed(range(2, TILE_SIZE)):
case ${N}:
wasm_v128_store(oN, v0_${N});
oN = (uint${SIZE}_t*) ((uintptr_t) oN + minus_output_stride);
case 1:
wasm_v128_store(oN, v0_1);
case 0:
wasm_v128_store(o, v0_0);
o = (uint${SIZE}_t*) ((uintptr_t) o + tile_hbytes);
break;
default:
XNN_UNREACHABLE;
}
$elif OUT_PTRS in ["MOV", "DEC"]:
o = (uint${SIZE}_t*) ((uintptr_t) o + oN_offset);
wasm_v128_store(o, v0_${TILE_SIZE-1});
$if OUT_PTRS == "MOV":
uint${SIZE}_t *oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
$for N in reversed(range(2, TILE_SIZE-1, 2)):
if XNN_UNPREDICTABLE(block_width > ${N+1}) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
wasm_v128_store(o, v0_${N});
$if OUT_PTRS == "MOV":
oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
if XNN_UNPREDICTABLE(block_width >= ${N+1}) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
wasm_v128_store(o, v0_${N-1});
$if OUT_PTRS == "MOV":
oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
if XNN_UNPREDICTABLE(block_width > 1) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
wasm_v128_store(o, v0_0);
$else:
$for N in reversed(range(TILE_SIZE)):
wasm_v128_store(o${N}, v0_${N});
o${N} = (uint${SIZE}_t*) ((uintptr_t) o${N} + tile_hbytes);
}
$if OUT_PTRS in ["MOV", "DEC"]:
o = (uint${SIZE}_t*) ((uintptr_t) o + tile_hbytes);
if (bh != 0) {
$if IN_PTRS == "REUSE":
const v128_t v${NUM_ITERS}_0 = wasm_v128_load(i0);
$for N in range(1, TILE_SIZE - 1, 2):
const uint${SIZE}_t *i${N} = (const uint${SIZE}_t*) ((uintptr_t) i${N-1} + input_stride);
if XNN_UNPREDICTABLE(bh < ${N+1}) {
i${N} = i${N-1};
}
const v128_t v${NUM_ITERS}_${N} = wasm_v128_load(i${N});
const uint${SIZE}_t *i${N+1} = (const uint${SIZE}_t*) ((uintptr_t) i${N} + input_stride);
if XNN_UNPREDICTABLE(bh <= ${N+1}) {
i${N+1} = i${N};
}
const v128_t v${NUM_ITERS}_${N+1} = wasm_v128_load(i${N+1});
$else:
const v128_t v${NUM_ITERS}_0 = wasm_v128_load(i0);
$for N in range(1, TILE_SIZE - 1, 2):
if XNN_UNPREDICTABLE(bh < ${N+1}) {
i${N} = i0;
}
const v128_t v${NUM_ITERS}_${N} = wasm_v128_load(i${N});
if XNN_UNPREDICTABLE(bh <= ${N+1}) {
i${N+1} = i0;
}
const v128_t v${NUM_ITERS}_${N+1} = wasm_v128_load(i${N+1});
const v128_t v${NUM_ITERS}_${TILE_SIZE-1} = wasm_v128_xor(v${NUM_ITERS}_0, v${NUM_ITERS}_0);
$for M in range(NUM_ITERS-1):
$for N in range(TILE_SIZE >> 1):
const v128_t v${NUM_ITERS-M-1}_${2*N} = wasm_v${SIZE}x${TILE_SIZE}_shuffle(v${NUM_ITERS-M}_${N}, v${NUM_ITERS-M}_${N+int(TILE_SIZE/2)}, ${LO_PERM});
const v128_t v${NUM_ITERS-M-1}_${2*N+1} = wasm_v${SIZE}x${TILE_SIZE}_shuffle(v${NUM_ITERS-M}_${N}, v${NUM_ITERS-M}_${N+int(TILE_SIZE/2)}, ${HI_PERM});
$for N in range(TILE_SIZE >> 1):
v128_t v0_${2*N} = wasm_v${SIZE}x${TILE_SIZE}_shuffle(v1_${N}, v1_${N+int(TILE_SIZE/2)}, ${LO_PERM});
v128_t v0_${2*N+1} = wasm_v${SIZE}x${TILE_SIZE}_shuffle(v1_${N}, v1_${N+int(TILE_SIZE/2)}, ${HI_PERM});
if (bh & ${TILE_SIZE>>1}) {
$if OUT_PTRS == "SWITCH":
uint${SIZE}_t* oN = (uint${SIZE}_t*) ((uintptr_t) o + oN_stride);
switch (rem) {
$for N in reversed(range(2, TILE_SIZE)):
case ${N}:
*((double*) oN) = wasm_f64x2_extract_lane(v0_${N}, 0);
oN = (uint${SIZE}_t*) ((uintptr_t) oN + minus_output_stride);
case 1:
*((double*) oN) = wasm_f64x2_extract_lane(v0_1, 0);
case 0:
$if NUM_ITERS > 1:
*((double*) o) = wasm_f64x2_extract_lane(v0_0, 0);
o += ${TILE_SIZE>>1};
$else:
*((double*) o) = wasm_f64x2_extract_lane(v0_0, 0);
break;
default:
XNN_UNREACHABLE;
}
$elif OUT_PTRS in ["MOV", "DEC"]:
o = (uint${SIZE}_t*) ((uintptr_t) o + oN_stride);
*((double*) o) = wasm_f64x2_extract_lane(v0_${TILE_SIZE-1}, 0);
$if OUT_PTRS == "MOV":
uint${SIZE}_t *oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
$for N in reversed(range(2, TILE_SIZE, 2)):
if XNN_UNPREDICTABLE(block_width > ${N+1}) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
*((double*) o) = wasm_f64x2_extract_lane(v0_${N}, 0);
$if OUT_PTRS == "MOV":
oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
if XNN_UNPREDICTABLE(block_width >= ${N+1}) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
*((double*) o) = wasm_f64x2_extract_lane(v0_${N-1}, 0);
$if OUT_PTRS == "MOV":
oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
if XNN_UNPREDICTABLE(block_width > 1) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
$if NUM_ITERS > 1:
*((double*) o) = wasm_f64x2_extract_lane(v0_0, 0);
o += ${TILE_SIZE>>1};
$else:
*((double*) o) = wasm_f64x2_extract_lane(v0_0, 0);
$else:
$for N in reversed(range(TILE_SIZE)):
$if NUM_ITERS>1:
*((double*) o${N}) = wasm_f64x2_extract_lane(v0_${N}, 0);
o${N} += ${TILE_SIZE>>1};
$else:
*((double*) o${N}) = wasm_f64x2_extract_lane(v0_${N}, 0);
$if NUM_ITERS > 1:
$for N in range(TILE_SIZE):
v0_${N} = wasm_v64x2_shuffle(v0_${N}, v0_${N}, 1, 1);
}
$if NUM_ITERS>1:
if (bh & ${TILE_SIZE>>2}) {
$if OUT_PTRS == "SWITCH":
uint${SIZE}_t* oN = (uint${SIZE}_t*) ((uintptr_t) o + oN_stride);
switch (rem) {
$for N in reversed(range(2, TILE_SIZE)):
case ${N}:
*((float*) oN) = wasm_f32x4_extract_lane(v0_${N}, 0);
oN = (uint${SIZE}_t*) ((uintptr_t) oN + minus_output_stride);
case 1:
*((float*) oN) = wasm_f32x4_extract_lane(v0_1, 0);
case 0:
*((float*) o) = wasm_f32x4_extract_lane(v0_0, 0);
$if SIZE < 32:
o += ${TILE_SIZE>>2};
break;
default:
XNN_UNREACHABLE;
}
$elif OUT_PTRS in ["MOV", "DEC"]:
o = (uint${SIZE}_t*) ((uintptr_t) o + oN_stride);
*((float*) o) = wasm_f32x4_extract_lane(v0_${TILE_SIZE-1}, 0);
$if OUT_PTRS == "MOV":
uint${SIZE}_t *oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
$for N in reversed(range(2, TILE_SIZE, 2)):
if XNN_UNPREDICTABLE(block_width > ${N+1}) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
*((float*) o) = wasm_f32x4_extract_lane(v0_${N}, 0);
$if OUT_PTRS == "MOV":
oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
if XNN_UNPREDICTABLE(block_width >= ${N+1}) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
*((float*) o) = wasm_f32x4_extract_lane(v0_${N-1}, 0);
$if OUT_PTRS == "MOV":
oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
if XNN_UNPREDICTABLE(block_width > 1) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
*((float*) o) = wasm_f32x4_extract_lane(v0_0, 0);
$if SIZE < 32:
o += ${TILE_SIZE>>2};
$else:
$for N in reversed(range(TILE_SIZE)):
*((float*) o${N}) = wasm_f32x4_extract_lane(v0_${N}, 0);
$if SIZE < 32:
o${N} += ${TILE_SIZE>>2};
$if NUM_ITERS > 2:
$for N in range(TILE_SIZE):
v0_${N} = wasm_u64x2_shr(v0_${N}, 32);
}
$if NUM_ITERS>2:
if (bh & ${TILE_SIZE>>3}) {
$if OUT_PTRS == "SWITCH":
uint${SIZE}_t* oN = (uint${SIZE}_t*) ((uintptr_t) o + oN_stride);
switch (rem) {
$for N in reversed(range(2, TILE_SIZE)):
case ${N}:
$if SIZE == 16:
*oN = wasm_i16x8_extract_lane(v0_${N}, 0);
oN = (uint${SIZE}_t*) ((uintptr_t) oN + minus_output_stride);
$else:
*((uint16_t*) oN) = wasm_i16x8_extract_lane(v0_${N}, 0);
oN = (uint${SIZE}_t*) ((uintptr_t) oN + minus_output_stride);
case 1:
$if SIZE == 16:
*oN = wasm_i16x8_extract_lane(v0_1, 0);
$else:
*((uint16_t*) oN) = wasm_i16x8_extract_lane(v0_1, 0);
case 0:
$if SIZE == 16:
*o = wasm_i16x8_extract_lane(v0_0, 0);
$else:
*((uint16_t*) o) = wasm_i16x8_extract_lane(v0_0, 0);
o += ${TILE_SIZE>>3};
break;
default:
XNN_UNREACHABLE;
}
$elif OUT_PTRS in ["MOV", "DEC"]:
o = (uint${SIZE}_t*) ((uintptr_t) o + oN_stride);
$if SIZE == 16:
*o = wasm_i16x8_extract_lane(v0_${TILE_SIZE-1}, 0);
$else:
*((uint16_t*) o) = wasm_i16x8_extract_lane(v0_${TILE_SIZE-1}, 0);
$if OUT_PTRS == "MOV":
uint${SIZE}_t *oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
$for N in reversed(range(2, TILE_SIZE, 2)):
if XNN_UNPREDICTABLE(block_width > ${N+1}) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
$if SIZE == 16:
*o = wasm_i16x8_extract_lane(v0_${N}, 0);
$else:
*((uint16_t*) o) = wasm_i16x8_extract_lane(v0_${N}, 0);
$if OUT_PTRS == "MOV":
oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
if XNN_UNPREDICTABLE(block_width >= ${N+1}) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
$if SIZE == 16:
*o = wasm_i16x8_extract_lane(v0_${N-1}, 0);
$else:
*((uint16_t*) o) = wasm_i16x8_extract_lane(v0_${N-1}, 0);
$if OUT_PTRS == "MOV":
oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
if XNN_UNPREDICTABLE(block_width > 1) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
$if SIZE == 16:
*o = wasm_i16x8_extract_lane(v0_0, 0);
$else:
*((uint16_t*) o) = wasm_i16x8_extract_lane(v0_0, 0);
o += ${TILE_SIZE>>3};
$else:
$for N in reversed(range(TILE_SIZE)):
$if SIZE == 16:
*o${N} = wasm_i16x8_extract_lane(v0_${N}, 0);
$else:
*((uint16_t*) o${N}) = wasm_i16x8_extract_lane(v0_${N}, 0);
o${N} += ${TILE_SIZE>>3};
$if NUM_ITERS>3:
$for N in range(TILE_SIZE):
v0_${N} = wasm_u32x4_shr(v0_${N}, 16);
}
$if SIZE == 8:
if (bh & 1) {
$if OUT_PTRS == "SWITCH":
uint${SIZE}_t* oN = (uint${SIZE}_t*) ((uintptr_t) o + oN_stride);
switch (rem) {
$for N in reversed(range(2, TILE_SIZE)):
case ${N}:
*oN = wasm_i8x16_extract_lane(v0_${N}, 0);
oN = (uint${SIZE}_t*) ((uintptr_t) oN + minus_output_stride);
case 1:
*oN = wasm_i8x16_extract_lane(v0_1, 0);
case 0:
*o = wasm_i8x16_extract_lane(v0_0, 0);
break;
default:
XNN_UNREACHABLE;
}
$elif OUT_PTRS in ["MOV", "DEC"]:
o = (uint${SIZE}_t*) ((uintptr_t) o + oN_stride);
*o = wasm_i8x16_extract_lane(v0_${TILE_SIZE-1}, 0);
$if OUT_PTRS == "MOV":
uint${SIZE}_t *oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
$for N in reversed(range(2, TILE_SIZE, 2)):
if XNN_UNPREDICTABLE(block_width > ${N+1}) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
*o = wasm_i8x16_extract_lane(v0_${N}, 0);
$if OUT_PTRS == "MOV":
oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
if XNN_UNPREDICTABLE(block_width >= ${N+1}) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
*o = wasm_i8x16_extract_lane(v0_${N-1}, 0);
$if OUT_PTRS == "MOV":
oN = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
if XNN_UNPREDICTABLE(block_width > 1) {
$if OUT_PTRS == "MOV":
o = oN;
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + minus_output_stride);
}
*o = wasm_i8x16_extract_lane(v0_0, 0);
$else:
$for N in reversed(range(TILE_SIZE)):
*o${N} = wasm_i8x16_extract_lane(v0_${N}, 0);
}
}
$if IN_PTRS == "MULTI":
i0 = (const uint${SIZE}_t*) ((uintptr_t) i0 + input_reset);
$for N in range(1, TILE_SIZE):
i${N} = (const uint${SIZE}_t*) ((uintptr_t) i${N-1} + input_stride);
$else:
i0 = (const uint${SIZE}_t*) ((uintptr_t) i0 + input_reset);
$if OUT_PTRS == "MULTI":
o0 = (uint${SIZE}_t*) ((uintptr_t) o0 + output_reset);
$for N in range(1, TILE_SIZE):
o${N} = (uint${SIZE}_t*) ((uintptr_t) o${N} + output_reset);
$else:
o = (uint${SIZE}_t*) ((uintptr_t) o + output_reset);
block_width = doz(block_width, tile_width);
} while (block_width != 0);
}