summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorPaul Davis <paul@linuxaudiosystems.com>2015-05-12 21:27:33 -0400
committerPaul Davis <paul@linuxaudiosystems.com>2015-06-29 14:18:13 -0400
commit0d78359d297a1cffb498f2d104e6fb0ff249a226 (patch)
treeb09b89203491645309ee3829ee5c9d7a835dbebb
parent8bb54f41c66ff7db27dff8a06cc28e04bcd0c263 (diff)
meter computation must use absolute value of all samples (ignoring sign bit)
-rw-r--r--libs/ardour/sse_avx_functions_64bit_win.s13
1 files changed, 13 insertions, 0 deletions
diff --git a/libs/ardour/sse_avx_functions_64bit_win.s b/libs/ardour/sse_avx_functions_64bit_win.s
index 22829db98a..2e82ab8728 100644
--- a/libs/ardour/sse_avx_functions_64bit_win.s
+++ b/libs/ardour/sse_avx_functions_64bit_win.s
@@ -494,6 +494,12 @@ x86_sse_avx_compute_peak:
cmp $0, %rdx
je .CP_END
+ #; create the "abs" mask in %xmm3
+ #; if will be used to discard sign bit
+ pushq $2147483647
+ movss (%rsp), %xmm3
+ addq $8, %rsp
+
#; Check for alignment
movq %rcx, %r8 #; buf => %rdx
andq $28, %r8 #; mask bits 1 & 2
@@ -507,6 +513,7 @@ x86_sse_avx_compute_peak:
#; Load next value from the buffer
movss (%rcx), %xmm1
+ andps %xmm3, %xmm1 #; mask out sign bit
maxss %xmm1, %xmm0
#; increment buffer, decrement counter
@@ -535,9 +542,14 @@ x86_sse_avx_compute_peak:
vshufps $0x00, %ymm0, %ymm0, %ymm0 #; spread single float value to the all 128 bits of xmm0 register
vperm2f128 $0x00, %ymm0, %ymm0, %ymm0 #; extend the first 128 bits of ymm0 register to higher 128 bits
+ #; broadcast sign mask to the whole ymm3 register
+ vshufps $0x00, %ymm3, %ymm3, %ymm3 #; spread single float value to the all 128 bits of xmm3 register
+ vperm2f128 $0x00, %ymm3, %ymm3, %ymm3 #; extend the first 128 bits of ymm3 register to higher 128 bits
+
.LP_AVX:
vmovaps (%rcx), %ymm1
+ vandps %ymm3, %ymm1, %ymm1 #; mask out sign bit
vmaxps %ymm1, %ymm0, %ymm0
addq $32, %rcx #; buf+=8
@@ -569,6 +581,7 @@ x86_sse_avx_compute_peak:
.POST_START:
movss (%rcx), %xmm1
+ andps %xmm3, %xmm1 #; mask out sign bit
maxss %xmm1, %xmm0
addq $4, %rcx #; buf++;