blob: d4873d701eee50cdd3cc66cbb8dc33f1b92f2116 [file] [log] [blame]
// Copyright 2019 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.
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>
#include <xnnpack/math.h>
#include <xnnpack/spmm.h>
void xnn_f32_spmm_minmax_ukernel_${MR}x${NR}__scalar${"_unroll" + str(UNROLL) if UNROLL > 1 else ""}(
uint32_t m,
uint32_t n,
const float*restrict a,
const float*restrict weights,
const int32_t*restrict widx_dmap,
const uint32_t*restrict nidx_nnzmap,
float*restrict c,
const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
assert(m != 0);
const float vmin = params->scalar.min;
const float vmax = params->scalar.max;
size_t i = m;
while (i >= ${MR}) {
const float*restrict w = weights;
const int32_t* dmap = widx_dmap;
const uint32_t* nnzmap = nidx_nnzmap;
size_t j = n;
while (j >= ${NR}) {
uint32_t nnz = *nnzmap++;
$for N in range(0, NR, 1):
float vacc0x${N} = *w++;
$for M in range(1, MR):
float vacc${ABC[M]}x${N} = vacc0x${N};
if XNN_LIKELY(nnz != 0) {
do {
const intptr_t diff = *dmap++;
$for M in range(MR):
const float va${ABC[M]} = a[${M}];
a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);
$for N in range(0, NR, 1):
const float vb${N} = *w++;
$for N in range(0, NR, 1):
$for M in range(MR):
vacc${ABC[M]}x${N} += va${ABC[M]} * vb${N};
} while (--nnz != 0);
}
$for N in range(NR):
$for M in range(MR):
float vout${ABC[M]}x${N} = math_min_f32(vacc${ABC[M]}x${N}, vmax);
$for N in range(NR):
$for M in range(MR):
vout${ABC[M]}x${N} = math_max_f32(vout${ABC[M]}x${N}, vmin);
$for N in range(NR):
$for M in range(MR):
c[${N} * m + ${M}] = vout${ABC[M]}x${N};
c += ${NR} * m;
j -= ${NR};
}
if XNN_UNLIKELY(j != 0) {
do {
uint32_t nnz = *nnzmap++;
float vacc0 = *w++;
$for M in range(1, MR):
float vacc${ABC[M]} = vacc0;
if XNN_LIKELY(nnz != 0) {
do {
const intptr_t diff = *dmap++;
$for M in range(MR):
const float va${ABC[M]} = a[${M}];
a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);
const float vb = *w++;
$for M in range(MR):
vacc${ABC[M]} += va${ABC[M]} * vb;
} while (--nnz != 0);
}
$for M in range(MR):
float vout${ABC[M]} = math_min_f32(vacc${ABC[M]}, vmax);
$for M in range(MR):
vout${ABC[M]} = math_max_f32(vout${ABC[M]}, vmin);
$for M in range(MR):
c[${M}] = vout${ABC[M]};
c += m;
j -= 1;
} while (j != 0);
}
c -= m * n;
c += ${MR};
a += ${MR};
i -= ${MR};
}
if XNN_UNLIKELY(i != 0) {
$for LOG2M in reversed(range((MR - 1).bit_length())):
$SUBMR = 1 << LOG2M
if (i & ${SUBMR}) {
const float*restrict w = weights;
const int32_t* dmap = widx_dmap;
const uint32_t* nnzmap = nidx_nnzmap;
size_t j = n;
while (j >= ${NR}) {
uint32_t nnz = *nnzmap++;
$for N in range(0, NR, 1):
float vacc0x${N} = *w++;
$for M in range(1, SUBMR):
float vacc${ABC[M]}x${N} = vacc0x${N};
if XNN_LIKELY(nnz != 0) {
do {
const intptr_t diff = *dmap++;
$for M in range(SUBMR):
const float va${ABC[M]} = a[${M}];
a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);
$for N in range(0, NR, 1):
const float vb${N} = *w++;
$for N in range(0, NR, 1):
$for M in range(SUBMR):
vacc${ABC[M]}x${N} += va${ABC[M]} * vb${N};
} while (--nnz != 0);
}
$for N in range(0, NR, 1):
$for M in range(SUBMR):
float vout${ABC[M]}x${N} = math_min_f32(vacc${ABC[M]}x${N}, vmax);
$for N in range(0, NR, 1):
$for M in range(SUBMR):
vout${ABC[M]}x${N} = math_max_f32(vout${ABC[M]}x${N}, vmin);
$for N in range(0, NR, 1):
$for M in range(SUBMR):
c[${N} * m + ${M}] = vout${ABC[M]}x${N};
c += ${NR} * m;
j -= ${NR};
}
if XNN_UNLIKELY(j != 0) {
do {
uint32_t nnz = *nnzmap++;
float vacc0 = *w++;
$for M in range(1, SUBMR):
float vacc${ABC[M]} = vacc0;
if XNN_LIKELY(nnz != 0) {
do {
const intptr_t diff = *dmap++;
$for M in range(SUBMR):
const float va${ABC[M]} = a[${M}];
a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff);
const float vb = *w++;
$for M in range(SUBMR):
vacc${ABC[M]} += va${ABC[M]} * vb;
} while (--nnz != 0);
}
$for M in range(SUBMR):
float vout${ABC[M]} = math_min_f32(vacc${ABC[M]}, vmax);
$for M in range(SUBMR):
vout${ABC[M]} = math_max_f32(vout${ABC[M]}, vmin);
$for M in range(SUBMR):
c[${M}] = vout${ABC[M]};
c += m;
j -= 1;
} while (j != 0);
}
c -= m * n;
c += ${SUBMR};
a += ${SUBMR};
}
}
}