Separate WAsm micro-kernels and scalar micro-kernels
- WAsm-specific microkernels explicitly use f32.min/f32.max WAsm instructions
- About 2% end-to-end improvement on x86, no change on ARM64
PiperOrigin-RevId: 283845483
diff --git a/src/f32-dwconv/gen/up2x4-wasm.c b/src/f32-dwconv/gen/up2x4-wasm.c
new file mode 100644
index 0000000..b1fa49a
--- /dev/null
+++ b/src/f32-dwconv/gen/up2x4-wasm.c
@@ -0,0 +1,118 @@
+// Auto-generated file. Do not edit!
+// Template: src/f32-dwconv/up-scalar.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 <xnnpack/dwconv.h>
+#include <xnnpack/math.h>
+
+
+void xnn_f32_dwconv_ukernel_up2x4__wasm(
+ 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 float vmin = params->scalar.min;
+ const float vmax = params->scalar.max;
+ 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 >= 2; c -= 2) {
+ float vacc0p0 = w[0];
+ float vacc1p0 = w[1];
+
+
+ const float vi0x0 = i0[0];
+ const float vi0x1 = i0[1];
+ i0 += 2;
+
+ const float vk0x0 = w[2];
+ vacc0p0 += vi0x0 * vk0x0;
+ const float vk0x1 = w[3];
+ vacc1p0 += vi0x1 * vk0x1;
+
+ const float vi1x0 = i1[0];
+ const float vi1x1 = i1[1];
+ i1 += 2;
+
+ const float vk1x0 = w[4];
+ vacc0p0 += vi1x0 * vk1x0;
+ const float vk1x1 = w[5];
+ vacc1p0 += vi1x1 * vk1x1;
+
+ const float vi2x0 = i2[0];
+ const float vi2x1 = i2[1];
+ i2 += 2;
+
+ const float vk2x0 = w[6];
+ vacc0p0 += vi2x0 * vk2x0;
+ const float vk2x1 = w[7];
+ vacc1p0 += vi2x1 * vk2x1;
+
+ const float vi3x0 = i3[0];
+ const float vi3x1 = i3[1];
+ i3 += 2;
+
+ const float vk3x0 = w[8];
+ vacc0p0 += vi3x0 * vk3x0;
+ const float vk3x1 = w[9];
+ vacc1p0 += vi3x1 * vk3x1;
+
+ w += 10;
+
+
+ float vacc0 = __builtin_wasm_max_f32(vacc0p0, vmin);
+ float vacc1 = __builtin_wasm_max_f32(vacc1p0, vmin);
+
+ vacc0 = __builtin_wasm_min_f32(vacc0, vmax);
+ vacc1 = __builtin_wasm_min_f32(vacc1, vmax);
+
+ output[0] = vacc0;
+ output[1] = vacc1;
+ output += 2;
+ }
+ for (; c >= 1; c -= 1) {
+ float vacc0p0 = *w++;
+
+ const float vi0 = *i0++;
+ const float vk0 = w[1];
+ vacc0p0 += vi0 * vk0;
+ const float vi1 = *i1++;
+ const float vk1 = w[3];
+ vacc0p0 += vi1 * vk1;
+ const float vi2 = *i2++;
+ const float vk2 = w[5];
+ vacc0p0 += vi2 * vk2;
+ const float vi3 = *i3++;
+ const float vk3 = w[7];
+ vacc0p0 += vi3 * vk3;
+
+
+ float vacc0 = __builtin_wasm_max_f32(vacc0p0, vmin);
+ vacc0 = __builtin_wasm_min_f32(vacc0, vmax);
+ *output++ = vacc0;
+ }
+
+ output = (float*) ((uintptr_t) output + output_increment);
+ } while (--output_width != 0);
+}