Neon shuffle GEMM and IGEMM kernels.
M1 is 7.1% faster on mobilenet_v2
M2 is 6.5% faster on mobilenet_v2
PiperOrigin-RevId: 281623279
diff --git a/src/f32-gemm/6x8s4-neonfma.c b/src/f32-gemm/6x8s4-neonfma.c
new file mode 100644
index 0000000..9a1e7c5
--- /dev/null
+++ b/src/f32-gemm/6x8s4-neonfma.c
@@ -0,0 +1,318 @@
+// Auto-generated file. Do not edit!
+// Template: src/f32-gemm/neon-shuffle.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 <arm_neon.h>
+
+#include <xnnpack/gemm.h>
+
+
+void xnn_f32_gemm_ukernel_6x8s4__neonfma(
+ 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_output_params params[restrict static 1])
+{
+ assert(mr != 0);
+ assert(mr <= 6);
+ 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;
+ const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
+ float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
+ if XNN_UNPREDICTABLE(mr < 2) {
+ a1 = a0;
+ c1 = c0;
+ }
+ const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
+ float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
+ if XNN_UNPREDICTABLE(mr <= 2) {
+ a2 = a1;
+ c2 = c1;
+ }
+ const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
+ float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
+ if XNN_UNPREDICTABLE(mr < 4) {
+ a3 = a2;
+ c3 = c2;
+ }
+ const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
+ float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
+ if XNN_UNPREDICTABLE(mr <= 4) {
+ a4 = a3;
+ c4 = c3;
+ }
+ const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
+ float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
+ if XNN_UNPREDICTABLE(mr != 6) {
+ a5 = a4;
+ c5 = c4;
+ }
+
+ do {
+ float32x4_t vacc0x0123 = vld1q_f32(w); w += 4;
+ float32x4_t vacc0x4567 = vld1q_f32(w); w += 4;
+ float32x4_t vacc1x0123 = vacc0x0123;
+ float32x4_t vacc1x4567 = vacc0x4567;
+ float32x4_t vacc2x0123 = vacc0x0123;
+ float32x4_t vacc2x4567 = vacc0x4567;
+ float32x4_t vacc3x0123 = vacc0x0123;
+ float32x4_t vacc3x4567 = vacc0x4567;
+ float32x4_t vacc4x0123 = vacc0x0123;
+ float32x4_t vacc4x4567 = vacc0x4567;
+ float32x4_t vacc5x0123 = vacc0x0123;
+ float32x4_t vacc5x4567 = vacc0x4567;
+
+ size_t k = kc;
+ while (k >= 4 * sizeof(float)) {
+ float32x4_t va0 = vld1q_f32(a0); a0 += 4;
+ float32x4_t va1 = vld1q_f32(a1); a1 += 4;
+ float32x4_t va2 = vld1q_f32(a2); a2 += 4;
+ float32x4_t va3 = vld1q_f32(a3); a3 += 4;
+ float32x4_t va4 = vld1q_f32(a4); a4 += 4;
+ float32x4_t va5 = vld1q_f32(a5); a5 += 4;
+
+
+ const float32x4_t vb0123c0 = vld1q_f32(w + 0);
+ const float32x4_t vb4567c0 = vld1q_f32(w + 4);
+
+ vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c0);
+ vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123c0);
+ vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123c0);
+ vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123c0);
+ vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123c0);
+ vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123c0);
+ vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c0);
+ vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567c0);
+ vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567c0);
+ vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567c0);
+ vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567c0);
+ vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567c0);
+
+ va0 = vextq_f32(va0, va0, 1);
+ va1 = vextq_f32(va1, va1, 1);
+ va2 = vextq_f32(va2, va2, 1);
+ va3 = vextq_f32(va3, va3, 1);
+ va4 = vextq_f32(va4, va4, 1);
+ va5 = vextq_f32(va5, va5, 1);
+
+ const float32x4_t vb0123c1 = vld1q_f32(w + 8);
+ const float32x4_t vb4567c1 = vld1q_f32(w + 12);
+
+ vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c1);
+ vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123c1);
+ vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123c1);
+ vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123c1);
+ vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123c1);
+ vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123c1);
+ vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c1);
+ vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567c1);
+ vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567c1);
+ vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567c1);
+ vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567c1);
+ vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567c1);
+
+ va0 = vextq_f32(va0, va0, 1);
+ va1 = vextq_f32(va1, va1, 1);
+ va2 = vextq_f32(va2, va2, 1);
+ va3 = vextq_f32(va3, va3, 1);
+ va4 = vextq_f32(va4, va4, 1);
+ va5 = vextq_f32(va5, va5, 1);
+
+ const float32x4_t vb0123c2 = vld1q_f32(w + 16);
+ const float32x4_t vb4567c2 = vld1q_f32(w + 20);
+
+ vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c2);
+ vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123c2);
+ vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123c2);
+ vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123c2);
+ vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123c2);
+ vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123c2);
+ vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c2);
+ vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567c2);
+ vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567c2);
+ vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567c2);
+ vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567c2);
+ vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567c2);
+
+ va0 = vextq_f32(va0, va0, 1);
+ va1 = vextq_f32(va1, va1, 1);
+ va2 = vextq_f32(va2, va2, 1);
+ va3 = vextq_f32(va3, va3, 1);
+ va4 = vextq_f32(va4, va4, 1);
+ va5 = vextq_f32(va5, va5, 1);
+
+ const float32x4_t vb0123c3 = vld1q_f32(w + 24);
+ const float32x4_t vb4567c3 = vld1q_f32(w + 28);
+
+ vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123c3);
+ vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123c3);
+ vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123c3);
+ vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123c3);
+ vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123c3);
+ vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123c3);
+ vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567c3);
+ vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567c3);
+ vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567c3);
+ vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567c3);
+ vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567c3);
+ vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567c3);
+
+
+ w += 32;
+ k -= 4 * sizeof(float);
+ }
+ if XNN_UNLIKELY(k != 0) {
+ do {
+ const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1;
+ const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1;
+ const float32x4_t va2 = vld1q_dup_f32(a2); a2 += 1;
+ const float32x4_t va3 = vld1q_dup_f32(a3); a3 += 1;
+ const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1;
+ const float32x4_t va5 = vld1q_dup_f32(a5); a5 += 1;
+
+ const float32x4_t vb0123 = vld1q_f32(w); w += 4;
+ const float32x4_t vb4567 = vld1q_f32(w); w += 4;
+
+ vacc0x0123 = vfmaq_f32(vacc0x0123, va0, vb0123);
+ vacc1x0123 = vfmaq_f32(vacc1x0123, va1, vb0123);
+ vacc2x0123 = vfmaq_f32(vacc2x0123, va2, vb0123);
+ vacc3x0123 = vfmaq_f32(vacc3x0123, va3, vb0123);
+ vacc4x0123 = vfmaq_f32(vacc4x0123, va4, vb0123);
+ vacc5x0123 = vfmaq_f32(vacc5x0123, va5, vb0123);
+ vacc0x4567 = vfmaq_f32(vacc0x4567, va0, vb4567);
+ vacc1x4567 = vfmaq_f32(vacc1x4567, va1, vb4567);
+ vacc2x4567 = vfmaq_f32(vacc2x4567, va2, vb4567);
+ vacc3x4567 = vfmaq_f32(vacc3x4567, va3, vb4567);
+ vacc4x4567 = vfmaq_f32(vacc4x4567, va4, vb4567);
+ vacc5x4567 = vfmaq_f32(vacc5x4567, va5, vb4567);
+
+ k -= sizeof(float);
+ } while (k != 0);
+ }
+ const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
+ vacc0x0123 = vminq_f32(vacc0x0123, vmax);
+ vacc1x0123 = vminq_f32(vacc1x0123, vmax);
+ vacc2x0123 = vminq_f32(vacc2x0123, vmax);
+ vacc3x0123 = vminq_f32(vacc3x0123, vmax);
+ vacc4x0123 = vminq_f32(vacc4x0123, vmax);
+ vacc5x0123 = vminq_f32(vacc5x0123, vmax);
+ vacc0x4567 = vminq_f32(vacc0x4567, vmax);
+ vacc1x4567 = vminq_f32(vacc1x4567, vmax);
+ vacc2x4567 = vminq_f32(vacc2x4567, vmax);
+ vacc3x4567 = vminq_f32(vacc3x4567, vmax);
+ vacc4x4567 = vminq_f32(vacc4x4567, vmax);
+ vacc5x4567 = vminq_f32(vacc5x4567, vmax);
+
+ const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
+ vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
+ vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
+ vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
+ vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
+ vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
+ vacc5x0123 = vmaxq_f32(vacc5x0123, vmin);
+ vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
+ vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
+ vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
+ vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
+ vacc4x4567 = vmaxq_f32(vacc4x4567, vmin);
+ vacc5x4567 = vmaxq_f32(vacc5x4567, vmin);
+
+ if XNN_LIKELY(nc >= 8) {
+ vst1q_f32(c5, vacc5x0123);
+ vst1q_f32(c5 + 4, vacc5x4567);
+ c5 = (float*) ((uintptr_t) c5 + cn_stride);
+ vst1q_f32(c4, vacc4x0123);
+ vst1q_f32(c4 + 4, vacc4x4567);
+ c4 = (float*) ((uintptr_t) c4 + cn_stride);
+ vst1q_f32(c3, vacc3x0123);
+ vst1q_f32(c3 + 4, vacc3x4567);
+ c3 = (float*) ((uintptr_t) c3 + cn_stride);
+ vst1q_f32(c2, vacc2x0123);
+ vst1q_f32(c2 + 4, vacc2x4567);
+ c2 = (float*) ((uintptr_t) c2 + cn_stride);
+ vst1q_f32(c1, vacc1x0123);
+ vst1q_f32(c1 + 4, vacc1x4567);
+ c1 = (float*) ((uintptr_t) c1 + cn_stride);
+ vst1q_f32(c0, vacc0x0123);
+ vst1q_f32(c0 + 4, vacc0x4567);
+ c0 = (float*) ((uintptr_t) c0 + cn_stride);
+
+ a5 = (const float*) ((uintptr_t) a5 - kc);
+ a4 = (const float*) ((uintptr_t) a4 - kc);
+ a3 = (const float*) ((uintptr_t) a3 - kc);
+ a2 = (const float*) ((uintptr_t) a2 - kc);
+ a1 = (const float*) ((uintptr_t) a1 - kc);
+ a0 = (const float*) ((uintptr_t) a0 - kc);
+
+ nc -= 8;
+
+ } else {
+ if (nc & 4) {
+ vst1q_f32(c5, vacc5x0123); c5 += 4;
+ vst1q_f32(c4, vacc4x0123); c4 += 4;
+ vst1q_f32(c3, vacc3x0123); c3 += 4;
+ vst1q_f32(c2, vacc2x0123); c2 += 4;
+ vst1q_f32(c1, vacc1x0123); c1 += 4;
+ vst1q_f32(c0, vacc0x0123); c0 += 4;
+
+ vacc5x0123 = vacc5x4567;
+ vacc4x0123 = vacc4x4567;
+ vacc3x0123 = vacc3x4567;
+ vacc2x0123 = vacc2x4567;
+ vacc1x0123 = vacc1x4567;
+ vacc0x0123 = vacc0x4567;
+ }
+ float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);
+ float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);
+ float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
+ float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
+ float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
+ float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
+ if (nc & 2) {
+ vst1_f32(c5, vacc5x01); c5 += 2;
+ vst1_f32(c4, vacc4x01); c4 += 2;
+ vst1_f32(c3, vacc3x01); c3 += 2;
+ vst1_f32(c2, vacc2x01); c2 += 2;
+ vst1_f32(c1, vacc1x01); c1 += 2;
+ vst1_f32(c0, vacc0x01); c0 += 2;
+
+ vacc5x01 = vget_high_f32(vacc5x0123);
+ vacc4x01 = vget_high_f32(vacc4x0123);
+ vacc3x01 = vget_high_f32(vacc3x0123);
+ vacc2x01 = vget_high_f32(vacc2x0123);
+ vacc1x01 = vget_high_f32(vacc1x0123);
+ vacc0x01 = vget_high_f32(vacc0x0123);
+ }
+ if (nc & 1) {
+ vst1_lane_f32(c5, vacc5x01, 0);
+ vst1_lane_f32(c4, vacc4x01, 0);
+ vst1_lane_f32(c3, vacc3x01, 0);
+ vst1_lane_f32(c2, vacc2x01, 0);
+ vst1_lane_f32(c1, vacc1x01, 0);
+ vst1_lane_f32(c0, vacc0x01, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}