Changeset 8964

Timestamp:

07/21/08 16:03:03

Author:

eb

Message:

work-in-progress on altivec

Files:

• gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_fir_fff_vmx.cc (modified) (6 diffs)
• gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/qa_gr_fir_fff.cc (modified) (2 diffs)

gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_fir_fff_vmx.cc

@@ -27 +27 @@
-#include <stdlib.h>
-#include <stdexcept>
-
-
-
+
+
+
+
+
-
-union v_float_u {
@@ -95 +97 @@
-
-float
-unsigned
+size_t
-{
-  float sum = 0;
-unsigned in
+size_
-    sum += a[i] * b[i];
-  }
@@ -104 +106 @@
-}
-
-
-
+
-/*
- *  preconditions:
@@ -114 +115 @@
- */
-float
-unsigned
-
-unsigned
-
-unsigned
+size_t
+
+size_t
+
+size_t
-  vector float acc0 = {0, 0, 0, 0};
-
@@ -137 +138 @@
-  b += FLOATS_PER_VEC;
-  
-unsigned
+size_t
-    lsq_a0 = vec_ld(0, a);
-    lsq_b0 = vec_ld(0, b);
@@ -154 +155 @@
-  return horizontal_add_f(acc0);
-}
+
+#else
+/*
+ *  preconditions:
+ *
+ *    n > 0 and a multiple of 4
+ *    a   4-byte aligned
+ *    b  16-byte aligned
+ */
+float
+dotprod_fff_vmx(const float *_a, const float *_b, size_t n)
+{
+  const vector float *a = (const vector float *) _a;
+  const vector float *b = (const vector float *) _b;
+
+  static const size_t UNROLL_CNT = 4;
+
+  size_t loop_cnt = n / (UNROLL_CNT * FLOATS_PER_VEC);
+  size_t nleft = n % (UNROLL_CNT * FLOATS_PER_VEC);
+
+  // printf("n = %zd, loop_cnt = %zd, nleft = %zd\n", n, loop_cnt, nleft);
+
+  // Used with vperm to build a* from p*
+  vector unsigned char lvsl_a = vec_lvsl(0, _a);
+
+  vector float p0, p1, p2, p3;
+  vector float a0, a1, a2, a3;
+  vector float b0, b1, b2, b3;
+  vector float acc0 = {0, 0, 0, 0};
+  vector float acc1 = {0, 0, 0, 0};
+  vector float acc2 = {0, 0, 0, 0};
+  vector float acc3 = {0, 0, 0, 0};
+
+  // wind in
+
+  p0 = vec_ld(0*VS, a);
+  p1 = vec_ld(1*VS, a);
+  p2 = vec_ld(2*VS, a);
+  p3 = vec_ld(3*VS, a);
+
+  a0 = vec_perm(p0, p1, lvsl_a);
+  b0 = vec_ld(0*VS, b);
+  p0 = vec_ld((UNROLL_CNT + 0)*VS, a);
+
+  for (size_t i = 0; i < loop_cnt; i++){
+
+    a1 = vec_perm(p1, p2, lvsl_a);
+    b1 = vec_ld(1*VS, b);
+    p1 = vec_ld((UNROLL_CNT + 1)*VS, a);
+    acc0 = vec_madd(a0, b0, acc0);
+
+    a2 = vec_perm(p2, p3, lvsl_a);
+    b2 = vec_ld(2*VS, b);
+    p2 = vec_ld((UNROLL_CNT + 2)*VS, a);
+    acc1 = vec_madd(a1, b1, acc1);
+
+    a3 = vec_perm(p3, p0, lvsl_a);
+    b3 = vec_ld(3*VS, b);
+    p3 = vec_ld((UNROLL_CNT + 3)*VS, a);
+    acc2 = vec_madd(a2, b2, acc2);
+
+    a += UNROLL_CNT;
+    b += UNROLL_CNT;
+
+    a0 = vec_perm(p0, p1, lvsl_a);
+    b0 = vec_ld(0*VS, b);
+    p0 = vec_ld((UNROLL_CNT + 0)*VS, a);
+    acc3 = vec_madd(a3, b3, acc3);
+  }
+
+  assert((nleft % 4) == 0);
+
+  switch (nleft/4){
+  case 0:
+    break;
+    
+  case 1:
+    acc0 = vec_madd(a0, b0, acc0);
+    break;
+
+  case 2:
+    a1 = vec_perm(p1, p2, lvsl_a);
+    b1 = vec_ld(1*VS, b);
+    acc0 = vec_madd(a0, b0, acc0);
+    acc1 = vec_madd(a1, b1, acc1);
+    break;
+
+  case 3:
+    a1 = vec_perm(p1, p2, lvsl_a);
+    b1 = vec_ld(1*VS, b);
+    acc0 = vec_madd(a0, b0, acc0);
+    a2 = vec_perm(p2, p3, lvsl_a);
+    b2 = vec_ld(2*VS, b);
+    acc1 = vec_madd(a1, b1, acc1);
+    acc2 = vec_madd(a2, b2, acc2);
+    break;
+
+  default:
+    assert(0);
+    break;
+  }
+            
+  acc0 = acc0 + acc1;
+  acc2 = acc2 + acc3;
+  acc0 = acc0 + acc2;
+
+  return horizontal_add_f(acc0);
+}
+
-#endif
-

gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/qa_gr_fir_fff.cc

@@ -144 +144 @@
-test_random_io (fir_maker_t maker)  
-{
-9
+32
-  const int     OUTPUT_LEN      = 17;
-  const int     INPUT_LEN       = MAX_TAPS + OUTPUT_LEN;
@@ -188 +188 @@
-      for (int o = 0; o < ol; o++){
-        CPPUNIT_ASSERT_DOUBLES_EQUAL (expected_output[o], actual_output[o],
-    fabs (expected_output[o]) * 1e-4
+             fabs (expected_output[o]) * 5e-3
-      }
-