From: Luke Kenneth Casson Leighton <lkcl@lkcl.net>
Date: Thu, 29 Jul 2021 21:58:02 +0000 (+0100)
Subject: use coefficient table in iDCT
X-Git-Tag: xlen-bcd~183
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=47070117f47ca66f8792bd382e48f02dfbd8085e;p=openpower-isa.git

use coefficient table in iDCT
---

diff --git a/src/openpower/decoder/isa/fastdct-test.py b/src/openpower/decoder/isa/fastdct-test.py
index c3aacd6b..72414d5c 100644
--- a/src/openpower/decoder/isa/fastdct-test.py
+++ b/src/openpower/decoder/isa/fastdct-test.py
@@ -28,7 +28,7 @@ import fastdctlee, naivedct, remap_dct_yield
 class FastDctTest(unittest.TestCase):
 
     def test_fast_dct_lee_vs_naive(self):
-        for i in range(3, 10):
+        for i in range(3, 4):
             n = 2**i
             vector = FastDctTest.nonrandom_vector(n)
             expect = naivedct.transform(vector)
diff --git a/src/openpower/decoder/isa/fastdctlee.py b/src/openpower/decoder/isa/fastdctlee.py
index 2ab421fb..16de9cb7 100644
--- a/src/openpower/decoder/isa/fastdctlee.py
+++ b/src/openpower/decoder/isa/fastdctlee.py
@@ -347,13 +347,13 @@ def inverse_transform_iter(vec):
     # this table could be cached and used multiple times rather than
     # computed every time.
     ctable = []
-    size = n
-    while size >= 2:
+    size = 2
+    while size <= n:
         halfsize = size // 2
         for i in range(n//size):
             for ci in range(halfsize):
                 ctable.append((math.cos((ci + 0.5) * math.pi / size) * 2.0))
-        size //= 2
+        size *= 2
 
     # first divide element 0 by 2
     vec[0] /= 2.0
@@ -408,8 +408,8 @@ def inverse_transform_iter(vec):
             for ci, (jl, jh) in enumerate(zip(j, jr)):
                 #t1, t2 = vec[ri[ji[jl]]], vec[ri[ji[jh]]]
                 t1, t2 = vec[ji[jl]], vec[ji[jl+halfsize]]
-                coeff = (math.cos((ci + 0.5) * math.pi / size) * 2.0)
-                #coeff = ctable[k]
+                #coeff = (math.cos((ci + 0.5) * math.pi / size) * 2.0)
+                coeff = ctable[k]
                 k += 1
                 # normally DCT would use jl+halfsize not jh, here.
                 # to be able to work in-place, the idea is to perform a