Initial open-source release

PiperOrigin-RevId: 271685289
diff --git a/src/f32-gemm/neon-ld64.c.in b/src/f32-gemm/neon-ld64.c.in
new file mode 100644
index 0000000..5a89619
--- /dev/null
+++ b/src/f32-gemm/neon-ld64.c.in
@@ -0,0 +1,164 @@
+// 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.
+
+$assert NR % 4 == 0
+$ABC = "0123456789ABCDEFGHIJKLMN"
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/gemm.h>
+
+
+void xnn_f32_gemm${"inc" if INC else ""}_ukernel_${MR}x${NR}__${"neonfma" if FMA else "neon"}_ld64(
+    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,
+    $if INC:
+      const float*restrict acc,
+    const union xnn_f32_output_params params[restrict static 1])
+{
+  assert(mr != 0);
+  assert(mr <= ${MR});
+  assert(nc != 0);
+  assert(kc != 0);
+  assert(kc % sizeof(float) == 0);
+  assert(a != NULL);
+  assert(w != NULL);
+  assert(c != NULL);
+  $if INC:
+    assert(acc != NULL);
+
+  const float* a0 = a;
+  float* c0 = c;
+  $for M in range(1, MR):
+    const float* a${M} = (const float*) ((uintptr_t) a${M-1} + a_stride);
+    float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride);
+    $if M % 2 == 0:
+      if XNN_UNPREDICTABLE(mr <= ${M}) {
+        a${M} = a${M-1};
+        c${M} = c${M-1};
+      }
+    $elif M + 1 == MR:
+      if XNN_UNPREDICTABLE(mr != ${M+1}) {
+        a${M} = a${M-1};
+        c${M} = c${M-1};
+      }
+    $else:
+      if XNN_UNPREDICTABLE(mr < ${M+1}) {
+        a${M} = a${M-1};
+        c${M} = c${M-1};
+      }
+
+  do {
+    $if INC:
+      $for M in range(MR):
+        $for N in range(0, NR, 4):
+          float32x4_t vacc${M}x${ABC[N:N+4]} = vld1q_f32(acc); acc += 4;
+    $else:
+      $for N in range(0, NR, 4):
+        float32x4_t vacc0x${ABC[N:N+4]} = vld1q_f32(w); w += 4;
+      $for M in range(1, MR):
+        $for N in range(0, NR, 4):
+          float32x4_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};
+
+    size_t k = kc;
+    for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) {
+      $for M in range(MR):
+        const float32x2_t va${M} = vld1_f32(a${M}); a${M} += 2;
+
+      $for L in range(2):
+        $for N in range(0, NR, 4):
+          const float32x4_t vb${ABC[N:N+4]}c${L} = vld1q_f32(w); w += 4;
+
+        $if FMA:
+          #if defined(__aarch64__)
+            $for N in range(0, NR, 4):
+              $for M in range(MR):
+                vacc${M}x${ABC[N:N+4]} = vfmaq_lane_f32(vacc${M}x${ABC[N:N+4]},   vb${ABC[N:N+4]}c${L}, va${M}, ${L});
+          #else
+            $for M in range(MR):
+              const float32x4_t va${M}c${L} = vdupq_lane_f32(va${M}, ${L});
+            $for N in range(0, NR, 4):
+              $for M in range(MR):
+                vacc${M}x${ABC[N:N+4]} = vfmaq_f32(vacc${M}x${ABC[N:N+4]},   va${M}c${L}, vb${ABC[N:N+4]}c${L});
+          #endif
+        $else:
+          $for N in range(0, NR, 4):
+            $for M in range(MR):
+              vacc${M}x${ABC[N:N+4]} = vmlaq_lane_f32(vacc${M}x${ABC[N:N+4]},   vb${ABC[N:N+4]}c${L}, va${M}, ${L});
+    }
+    if XNN_UNLIKELY(k != 0) {
+      $for M in range(MR):
+        const float32x4_t va${M} = vld1q_dup_f32(a${M}); a${M} += 1;
+
+      $for N in range(0, NR, 4):
+        const float32x4_t vb${ABC[N:N+4]} = vld1q_f32(w); w += 4;
+
+      $for N in range(0, NR, 4):
+        $for M in range(MR):
+          $if FMA:
+            vacc${M}x${ABC[N:N+4]} = vfmaq_f32(vacc${M}x${ABC[N:N+4]}, va${M},   vb${ABC[N:N+4]});
+          $else:
+            vacc${M}x${ABC[N:N+4]} = vmlaq_f32(vacc${M}x${ABC[N:N+4]}, va${M},   vb${ABC[N:N+4]});
+    }
+    const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
+    $for N in range(0, NR, 4):
+      $for M in range(MR):
+        vacc${M}x${ABC[N:N+4]} = vminq_f32(vacc${M}x${ABC[N:N+4]}, vmax);
+
+    const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
+    $for N in range(0, NR, 4):
+      $for M in range(MR):
+        vacc${M}x${ABC[N:N+4]} = vmaxq_f32(vacc${M}x${ABC[N:N+4]}, vmin);
+
+    if XNN_LIKELY(nc >= ${NR}) {
+      $for M in reversed(range(MR)):
+        vst1q_f32(c${M}, vacc${M}x${ABC[0:4]});
+        $for N in range(4, NR, 4):
+          vst1q_f32(c${M} + ${N}, vacc${M}x${ABC[N:N+4]});
+        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);
+
+      $for M in reversed(range(MR)):
+        a${M} = (const float*) ((uintptr_t) a${M} - kc);
+
+      nc -= ${NR};
+
+    } else {
+      $for LOG2N in reversed(range(NR.bit_length())):
+        $if NR != 1 << LOG2N:
+          if (nc & ${1 << LOG2N}) {
+            $if LOG2N >= 2:
+              $for N in range(0, 1 << LOG2N, 4):
+                $for M in reversed(range(MR)):
+                  vst1q_f32(c${M}, vacc${M}x${ABC[N:N+4]}); c${M} += 4;
+
+              $for M in reversed(range(MR)):
+                $for N in range(0, 1 << (LOG2N - 1), 4):
+                  vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 <<   LOG2N)+4]};
+            $elif LOG2N == 1:
+              $for M in reversed(range(MR)):
+                vst1_f32(c${M}, vacc${M}x${ABC[0:2]}); c${M} += 2;
+
+              $for M in reversed(range(MR)):
+                vacc${M}x${ABC[0:2]} = vget_high_f32(vacc${M}x${ABC[0:4]});
+            $elif LOG2N == 0:
+              $for M in reversed(range(MR)):
+                vst1_lane_f32(c${M}, vacc${M}x${ABC[0:2]}, 0);
+          }
+          $if LOG2N == 2:
+            $for M in reversed(range(MR)):
+              float32x2_t vacc${M}x${ABC[0:2]} = vget_low_f32(vacc${M}x${ABC[0:4]});
+
+      nc = 0;
+    }
+  } while (nc != 0);
+}