Refactor DWCONV micro-kernels
- Fix bugs in generation of micro-kernels with large channel tiles
- Add missing unit tests
- Generate, test, and benchmark a microkernels with 2 accumulators, with 2X
channel tile, and their combinations
PiperOrigin-RevId: 279137161
diff --git a/src/f32-dwconv/up4x4-sse-acc2.c b/src/f32-dwconv/up4x4-sse-acc2.c
new file mode 100644
index 0000000..38c3905
--- /dev/null
+++ b/src/f32-dwconv/up4x4-sse-acc2.c
@@ -0,0 +1,118 @@
+// Auto-generated file. Do not edit!
+// Template: src/f32-dwconv/up-sse.c.in
+// Generator: tools/xngen
+//
+// 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.
+
+#include <assert.h>
+
+#include <xmmintrin.h>
+
+#include <xnnpack/dwconv.h>
+
+
+void xnn_f32_dwconv_ukernel_up4x4__sse_acc2(
+ size_t channels,
+ size_t output_width,
+ const float** input,
+ const float* weights,
+ float* output,
+ size_t input_stride,
+ size_t output_increment,
+ const union xnn_f32_output_params params[restrict static 1])
+{
+ assert(channels != 0);
+ assert(output_width != 0);
+
+ const __m128 vmax = _mm_load_ps(params->sse.max);
+ const __m128 vmin = _mm_load_ps(params->sse.min);
+ do {
+ const float* i0 = input[0];
+ const float* i1 = input[1];
+ const float* i2 = input[2];
+ const float* i3 = input[3];
+ input = (const float**) ((uintptr_t) input + input_stride);
+
+ size_t c = channels;
+ const float* w = weights;
+ for (; c >= 4; c -= 4) {
+ __m128 vacc0123p0 = _mm_load_ps(w);
+
+
+ const __m128 vi0x0123 = _mm_loadu_ps(i0);
+ i0 += 4;
+
+ const __m128 vk0x0123 = _mm_load_ps(w + 4);
+ vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
+
+ const __m128 vi1x0123 = _mm_loadu_ps(i1);
+ i1 += 4;
+
+ const __m128 vk1x0123 = _mm_load_ps(w + 8);
+ __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
+
+ const __m128 vi2x0123 = _mm_loadu_ps(i2);
+ i2 += 4;
+
+ const __m128 vk2x0123 = _mm_load_ps(w + 12);
+ vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
+
+ const __m128 vi3x0123 = _mm_loadu_ps(i3);
+ i3 += 4;
+
+ const __m128 vk3x0123 = _mm_load_ps(w + 16);
+ vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
+
+ w += 20;
+
+ // Add up all accumulators to vacc0123p0
+ vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
+
+ __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
+ vacc0123 = _mm_min_ps(vacc0123, vmax);
+
+ _mm_storeu_ps(output, vacc0123);
+ output += 4;
+ }
+ if XNN_UNLIKELY(c != 0) {
+ __m128 vacc0123p0 = _mm_load_ps(w);
+
+ const __m128 vi0x0123 = _mm_loadu_ps(i0);
+ const __m128 vk0x0123 = _mm_load_ps(w + 4);
+ vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));
+
+ const __m128 vi1x0123 = _mm_loadu_ps(i1);
+ const __m128 vk1x0123 = _mm_load_ps(w + 8);
+ __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123);
+
+ const __m128 vi2x0123 = _mm_loadu_ps(i2);
+ const __m128 vk2x0123 = _mm_load_ps(w + 12);
+ vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));
+
+ const __m128 vi3x0123 = _mm_loadu_ps(i3);
+ const __m128 vk3x0123 = _mm_load_ps(w + 16);
+ vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123));
+
+ // Add up all accumulators to vacc0123p0
+ vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1);
+
+ __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
+ vacc0123 = _mm_min_ps(vacc0123, vmax);
+
+ if (c & 2) {
+ _mm_storel_pi((__m64*) output, vacc0123);
+ vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
+ output += 2;
+ }
+ if (c & 1) {
+ _mm_store_ss(output, vacc0123);
+ output += 1;
+ }
+ }
+
+ output = (float*) ((uintptr_t) output + output_increment);
+ } while (--output_width != 0);
+}