|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
#include <assert.h> |
|
|
|
#include <xnnpack/gemm.h> |
|
#include <xnnpack/math.h> |
|
|
|
|
|
void xnn_f32_gemm_relu_ukernel_1x4__wasm( |
|
size_t mr, |
|
size_t nc, |
|
size_t kc, |
|
const float* restrict a, |
|
size_t a_stride, |
|
const float* restrict w, |
|
float* restrict c, |
|
size_t cm_stride, |
|
size_t cn_stride, |
|
const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)]) |
|
{ |
|
assert(mr != 0); |
|
assert(mr <= 1); |
|
assert(nc != 0); |
|
assert(kc != 0); |
|
assert(kc % sizeof(float) == 0); |
|
assert(a != NULL); |
|
assert(w != NULL); |
|
assert(c != NULL); |
|
|
|
const float* a0 = a; |
|
float* c0 = c; |
|
|
|
do { |
|
float vacc00 = w[0]; |
|
float vacc01 = w[1]; |
|
float vacc02 = w[2]; |
|
float vacc03 = w[3]; |
|
w += 4; |
|
|
|
size_t k = kc; |
|
do { |
|
const float va0 = *a0++; |
|
|
|
const float vb0 = w[0]; |
|
const float vb1 = w[1]; |
|
const float vb2 = w[2]; |
|
const float vb3 = w[3]; |
|
w += 4; |
|
|
|
vacc00 = math_muladd_f32(va0, vb0, vacc00); |
|
vacc01 = math_muladd_f32(va0, vb1, vacc01); |
|
vacc02 = math_muladd_f32(va0, vb2, vacc02); |
|
vacc03 = math_muladd_f32(va0, vb3, vacc03); |
|
|
|
k -= sizeof(float); |
|
} while (k != 0); |
|
|
|
vacc00 = __builtin_wasm_max_f32(vacc00, 0.0f); |
|
vacc01 = __builtin_wasm_max_f32(vacc01, 0.0f); |
|
vacc02 = __builtin_wasm_max_f32(vacc02, 0.0f); |
|
vacc03 = __builtin_wasm_max_f32(vacc03, 0.0f); |
|
|
|
if XNN_LIKELY(nc >= 4) { |
|
c0[0] = vacc00; |
|
c0[1] = vacc01; |
|
c0[2] = vacc02; |
|
c0[3] = vacc03; |
|
c0 = (float*) ((uintptr_t) c0 + cn_stride); |
|
|
|
a0 = (const void*) ((uintptr_t) a0 - kc); |
|
|
|
nc -= 4; |
|
} else { |
|
if (nc & 2) { |
|
c0[0] = vacc00; |
|
c0[1] = vacc01; |
|
vacc00 = vacc02; |
|
c0 += 2; |
|
} |
|
if (nc & 1) { |
|
c0[0] = vacc00; |
|
} |
|
|
|
nc = 0; |
|
} |
|
} while (nc != 0); |
|
} |
|
|