Skip to content

Commit

Permalink
f16 & f32 generic mean operator & subgraph
Browse files Browse the repository at this point in the history 
PiperOrigin-RevId: 631024901
  • Loading branch information
@alankelly @xnnpack-bot
alankelly authored and xnnpack-bot committed May 13, 2024
1 parent b489d33 commit 5084f9b
Show file tree
Hide file tree
Showing 22 changed files with 2,324 additions and 426 deletions.
6 changes: 0 additions & 6 deletions include/xnnpack.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -4862,13 +4862,10 @@ enum xnn_status xnn_reshape_mean_nd_f16(
const size_t* reduction_axes,
size_t num_input_dims,
const size_t* input_shape,
size_t* workspace_size,
size_t* workspace_alignment,
pthreadpool_t threadpool);

enum xnn_status xnn_setup_mean_nd_f16(
xnn_operator_t mean_op,
void* workspace,
const void* input,
void* output);

Expand All @@ -4882,13 +4879,10 @@ enum xnn_status xnn_reshape_mean_nd_f32(
const size_t* reduction_axes,
size_t num_input_dims,
const size_t* input_shape,
size_t* workspace_size,
size_t* workspace_alignment,
pthreadpool_t threadpool);

enum xnn_status xnn_setup_mean_nd_f32(
xnn_operator_t mean_op,
void* workspace,
const float* input,
float* output);

Expand Down
243 changes: 243 additions & 0 deletions src/amalgam/gen/avx.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3642,6 +3642,249 @@ void xnn_f32_qu8_vcvt_ukernel__avx_u32(
}
}

void xnn_f32_rdsum_ukernel_7p7x__avx_c32(
size_t rows,
size_t channels,
const float* input,
size_t input_stride,
const float* zero,
float* output,
const union xnn_f32_scale_params params[restrict XNN_MIN_ELEMENTS(1)])
{
assert(rows != 0);
assert(channels != 0);
assert(input != NULL);
assert(output != NULL);

const __m256 vscale = _mm256_set1_ps(params->avx.scale);

size_t input_increment = 7 * input_stride;
for (; channels >= 32; channels -= 32) {
const float* i0 = input;
const float* i1 = (const float*) ((uintptr_t) input + 1 * input_stride);
const float* i2 = (const float*) ((uintptr_t) input + 2 * input_stride);
const float* i3 = (const float*) ((uintptr_t) input + 3 * input_stride);
const float* i4 = (const float*) ((uintptr_t) input + 4 * input_stride);
const float* i5 = (const float*) ((uintptr_t) input + 5 * input_stride);
const float* i6 = (const float*) ((uintptr_t) input + 6 * input_stride);

__m256 vacc0 = _mm256_setzero_ps();
__m256 vacc1 = _mm256_setzero_ps();
__m256 vacc2 = _mm256_setzero_ps();
__m256 vacc3 = _mm256_setzero_ps();

for (int r = rows; r > 0; r -= 7) {
if XNN_UNPREDICTABLE(r < 2) {
i1 = zero;
}
if XNN_UNPREDICTABLE(r <= 2) {
i2 = zero;
}
if XNN_UNPREDICTABLE(r < 4) {
i3 = zero;
}
if XNN_UNPREDICTABLE(r <= 4) {
i4 = zero;
}
if XNN_UNPREDICTABLE(r < 6) {
i5 = zero;
}
if XNN_UNPREDICTABLE(r <= 6) {
i6 = zero;
}
__m256 vin0;
__m256 vin1;
__m256 vin2;
__m256 vin3;
vin0 = _mm256_loadu_ps(&i0[0]);
vin1 = _mm256_loadu_ps(&i0[8]);
vin2 = _mm256_loadu_ps(&i0[16]);
vin3 = _mm256_loadu_ps(&i0[24]);
vacc0 = _mm256_add_ps(vin0, vacc0);
vacc1 = _mm256_add_ps(vin1, vacc1);
vacc2 = _mm256_add_ps(vin2, vacc2);
vacc3 = _mm256_add_ps(vin3, vacc3);
vin0 = _mm256_loadu_ps(&i1[0]);
vin1 = _mm256_loadu_ps(&i1[8]);
vin2 = _mm256_loadu_ps(&i1[16]);
vin3 = _mm256_loadu_ps(&i1[24]);
vacc0 = _mm256_add_ps(vin0, vacc0);
vacc1 = _mm256_add_ps(vin1, vacc1);
vacc2 = _mm256_add_ps(vin2, vacc2);
vacc3 = _mm256_add_ps(vin3, vacc3);
vin0 = _mm256_loadu_ps(&i2[0]);
vin1 = _mm256_loadu_ps(&i2[8]);
vin2 = _mm256_loadu_ps(&i2[16]);
vin3 = _mm256_loadu_ps(&i2[24]);
vacc0 = _mm256_add_ps(vin0, vacc0);
vacc1 = _mm256_add_ps(vin1, vacc1);
vacc2 = _mm256_add_ps(vin2, vacc2);
vacc3 = _mm256_add_ps(vin3, vacc3);
vin0 = _mm256_loadu_ps(&i3[0]);
vin1 = _mm256_loadu_ps(&i3[8]);
vin2 = _mm256_loadu_ps(&i3[16]);
vin3 = _mm256_loadu_ps(&i3[24]);
vacc0 = _mm256_add_ps(vin0, vacc0);
vacc1 = _mm256_add_ps(vin1, vacc1);
vacc2 = _mm256_add_ps(vin2, vacc2);
vacc3 = _mm256_add_ps(vin3, vacc3);
vin0 = _mm256_loadu_ps(&i4[0]);
vin1 = _mm256_loadu_ps(&i4[8]);
vin2 = _mm256_loadu_ps(&i4[16]);
vin3 = _mm256_loadu_ps(&i4[24]);
vacc0 = _mm256_add_ps(vin0, vacc0);
vacc1 = _mm256_add_ps(vin1, vacc1);
vacc2 = _mm256_add_ps(vin2, vacc2);
vacc3 = _mm256_add_ps(vin3, vacc3);
vin0 = _mm256_loadu_ps(&i5[0]);
vin1 = _mm256_loadu_ps(&i5[8]);
vin2 = _mm256_loadu_ps(&i5[16]);
vin3 = _mm256_loadu_ps(&i5[24]);
vacc0 = _mm256_add_ps(vin0, vacc0);
vacc1 = _mm256_add_ps(vin1, vacc1);
vacc2 = _mm256_add_ps(vin2, vacc2);
vacc3 = _mm256_add_ps(vin3, vacc3);
vin0 = _mm256_loadu_ps(&i6[0]);
vin1 = _mm256_loadu_ps(&i6[8]);
vin2 = _mm256_loadu_ps(&i6[16]);
vin3 = _mm256_loadu_ps(&i6[24]);
vacc0 = _mm256_add_ps(vin0, vacc0);
vacc1 = _mm256_add_ps(vin1, vacc1);
vacc2 = _mm256_add_ps(vin2, vacc2);
vacc3 = _mm256_add_ps(vin3, vacc3);
i0 = (const float*) ((uintptr_t) i0 + input_increment);
i1 = (const float*) ((uintptr_t) i1 + input_increment);
i2 = (const float*) ((uintptr_t) i2 + input_increment);
i3 = (const float*) ((uintptr_t) i3 + input_increment);
i4 = (const float*) ((uintptr_t) i4 + input_increment);
i5 = (const float*) ((uintptr_t) i5 + input_increment);
i6 = (const float*) ((uintptr_t) i6 + input_increment);
}
vacc0 = _mm256_mul_ps(vacc0, vscale);
vacc1 = _mm256_mul_ps(vacc1, vscale);
vacc2 = _mm256_mul_ps(vacc2, vscale);
vacc3 = _mm256_mul_ps(vacc3, vscale);

const float* o = output;
__m256 vo0 = _mm256_loadu_ps(o); o += 8;
__m256 vo1 = _mm256_loadu_ps(o); o += 8;
__m256 vo2 = _mm256_loadu_ps(o); o += 8;
__m256 vo3 = _mm256_loadu_ps(o); o += 8;
vacc0 = _mm256_add_ps(vo0, vacc0);
vacc1 = _mm256_add_ps(vo1, vacc1);
vacc2 = _mm256_add_ps(vo2, vacc2);
vacc3 = _mm256_add_ps(vo3, vacc3);
_mm256_storeu_ps(output, vacc0); output += 8;
_mm256_storeu_ps(output, vacc1); output += 8;
_mm256_storeu_ps(output, vacc2); output += 8;
_mm256_storeu_ps(output, vacc3); output += 8;

input = (const float*) ((uintptr_t) input + 32 * sizeof(float));
}
__m256i vmask;
if (channels != 0) {
input_increment = 7 * input_stride;
const float* i0 = input;
const float* i1 = (const float*) ((uintptr_t) input + 1 * input_stride);
const float* i2 = (const float*) ((uintptr_t) input + 2 * input_stride);
const float* i3 = (const float*) ((uintptr_t) input + 3 * input_stride);
const float* i4 = (const float*) ((uintptr_t) input + 4 * input_stride);
const float* i5 = (const float*) ((uintptr_t) input + 5 * input_stride);
const float* i6 = (const float*) ((uintptr_t) input + 6 * input_stride);
__m256 vacc[4];
vacc[0] = _mm256_setzero_ps();
vacc[1] = _mm256_setzero_ps();
vacc[2] = _mm256_setzero_ps();
vacc[3] = _mm256_setzero_ps();

const size_t num_full_chunks = channels >> 3;
const size_t num_chunks = round_up_po2(channels, 8) >> 3;
const size_t remainder = channels & 0x7;
for (int r = rows; r > 0; r -= 7) {
if XNN_UNPREDICTABLE(r < 2) {
i1 = zero;
}
if XNN_UNPREDICTABLE(r <= 2) {
i2 = zero;
}
if XNN_UNPREDICTABLE(r < 4) {
i3 = zero;
}
if XNN_UNPREDICTABLE(r <= 4) {
i4 = zero;
}
if XNN_UNPREDICTABLE(r < 6) {
i5 = zero;
}
if XNN_UNPREDICTABLE(r <= 6) {
i6 = zero;
}
for (int i = 0; i < num_full_chunks; ++i) {
vacc[i] = _mm256_add_ps(_mm256_loadu_ps(&i0[i*8]), vacc[i]);
vacc[i] = _mm256_add_ps(_mm256_loadu_ps(&i1[i*8]), vacc[i]);
vacc[i] = _mm256_add_ps(_mm256_loadu_ps(&i2[i*8]), vacc[i]);
vacc[i] = _mm256_add_ps(_mm256_loadu_ps(&i3[i*8]), vacc[i]);
vacc[i] = _mm256_add_ps(_mm256_loadu_ps(&i4[i*8]), vacc[i]);
vacc[i] = _mm256_add_ps(_mm256_loadu_ps(&i5[i*8]), vacc[i]);
vacc[i] = _mm256_add_ps(_mm256_loadu_ps(&i6[i*8]), vacc[i]);
}

if (remainder) {
vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->avx.mask_table[7] - (channels & 0x7) * sizeof(float)));
vacc[num_full_chunks] = _mm256_add_ps(_mm256_maskload_ps(&i0[num_full_chunks*8], vmask), vacc[num_full_chunks]);
vacc[num_full_chunks] = _mm256_add_ps(_mm256_maskload_ps(&i1[num_full_chunks*8], vmask), vacc[num_full_chunks]);
vacc[num_full_chunks] = _mm256_add_ps(_mm256_maskload_ps(&i2[num_full_chunks*8], vmask), vacc[num_full_chunks]);
vacc[num_full_chunks] = _mm256_add_ps(_mm256_maskload_ps(&i3[num_full_chunks*8], vmask), vacc[num_full_chunks]);
vacc[num_full_chunks] = _mm256_add_ps(_mm256_maskload_ps(&i4[num_full_chunks*8], vmask), vacc[num_full_chunks]);
vacc[num_full_chunks] = _mm256_add_ps(_mm256_maskload_ps(&i5[num_full_chunks*8], vmask), vacc[num_full_chunks]);
vacc[num_full_chunks] = _mm256_add_ps(_mm256_maskload_ps(&i6[num_full_chunks*8], vmask), vacc[num_full_chunks]);
}
i0 = (const float*) ((uintptr_t) i0 + input_increment);
i1 = (const float*) ((uintptr_t) i1 + input_increment);
i2 = (const float*) ((uintptr_t) i2 + input_increment);
i3 = (const float*) ((uintptr_t) i3 + input_increment);
i4 = (const float*) ((uintptr_t) i4 + input_increment);
i5 = (const float*) ((uintptr_t) i5 + input_increment);
i6 = (const float*) ((uintptr_t) i6 + input_increment);
}
for (size_t i = 0; i < num_chunks; ++i) {
vacc[i] = _mm256_mul_ps(vacc[i], vscale);
}

__m256 vo[4];
const float* o = output;
for (int i = 0; i < channels >> 3; ++i) {
vo[i] = _mm256_loadu_ps(o); o += 8;
}
for (int i = 0; i < channels >> 3; ++i) {
vacc[i] = _mm256_add_ps(vo[i], vacc[i]);
}
for (int i = 0; i < channels >> 3; ++i) {
_mm256_storeu_ps(output, vacc[i]); output += 8;
}
if (remainder) {
const size_t pos = num_full_chunks;
__m256 vout = vacc[pos];
const __m256 vdata = _mm256_maskload_ps(output, vmask);
vout = _mm256_add_ps(vout, vdata);
__m128 vout_lo = _mm256_castps256_ps128(vout);
if (channels & 4) {
_mm_storeu_ps(output, vout_lo);
vout_lo = _mm256_extractf128_ps(vout, 1);
output += 4;
}
if (channels & 2) {
_mm_storel_pi((__m64*) output, vout_lo);
vout_lo = _mm_movehl_ps(vout_lo, vout_lo);
output += 2;
}
if (channels & 1) {
_mm_store_ss(output, vout_lo);
}
}
}
}

void xnn_f32_rmax_ukernel__avx_u32_acc4(
size_t batch,
const float* input,
Expand Down

0 comments on commit 5084f9b

Please sign in to comment.