ref: 8befcd008924a111c08dc58fa740e905bc0b0d5f
parent: af9d7aa9fb81e722a719efd8c2fea7a7a80382e9
author: James Zern <jzern@google.com>
date: Wed Nov 23 15:48:00 EST 2016
enable vpx_idct16x16_10_add_neon in hbd builds BUG=webm:1294 Change-Id: Ibad079f25e673d4f5181961896a8a8333a51e825
--- a/vpx_dsp/arm/idct16x16_add_neon.asm
+++ b/vpx_dsp/arm/idct16x16_add_neon.asm
@@ -18,6 +18,8 @@
AREA ||.text||, CODE, READONLY, ALIGN=2
+ INCLUDE vpx_dsp/arm/idct_neon.asm.S
+
; Transpose a 8x8 16bit data matrix. Datas are loaded in q8-q15.
MACRO
TRANSPOSE8X8
@@ -753,9 +755,10 @@
bx lr
ENDP ; |vpx_idct16x16_256_add_neon_pass2|
-;void |vpx_idct16x16_10_add_neon_pass1|(const int16_t *input, int16_t *output)
+;void |vpx_idct16x16_10_add_neon_pass1|(const tran_low_t *input,
+; int16_t *output)
;
-; r0 const int16_t *input
+; r0 const tran_low_t *input
; r1 int16_t *output
; idct16 stage1 - stage6 on all the elements loaded in q8-q15. The output
@@ -765,14 +768,14 @@
; TODO(hkuang): Find a better way to load the elements.
; load elements of 0, 2, 4, 6, 8, 10, 12, 14 into q8 - q15
- vld2.s16 {q8,q9}, [r0]!- vld2.s16 {q9,q10}, [r0]!- vld2.s16 {q10,q11}, [r0]!- vld2.s16 {q11,q12}, [r0]!- vld2.s16 {q12,q13}, [r0]!- vld2.s16 {q13,q14}, [r0]!- vld2.s16 {q14,q15}, [r0]!- vld2.s16 {q1,q2}, [r0]!+ LOAD_TRAN_LOW_TO_S16X2 d16, d17, d18, d19, r0
+ LOAD_TRAN_LOW_TO_S16X2 d18, d19, d20, d21, r0
+ LOAD_TRAN_LOW_TO_S16X2 d20, d21, d22, d23, r0
+ LOAD_TRAN_LOW_TO_S16X2 d22, d23, d24, d25, r0
+ LOAD_TRAN_LOW_TO_S16X2 d24, d25, d26, d27, r0
+ LOAD_TRAN_LOW_TO_S16X2 d26, d27, d28, d29, r0
+ LOAD_TRAN_LOW_TO_S16X2 d28, d29, d30, d31, r0
+ LOAD_TRAN_LOW_TO_S16X2 d2, d3, d4, d5, r0
vmov.s16 q15, q1
; cospi_28_64*2 = 6392
@@ -857,10 +860,10 @@
bx lr
ENDP ; |vpx_idct16x16_10_add_neon_pass1|
-;void vpx_idct16x16_10_add_neon_pass2(const int16_t *src, int16_t *output,
+;void vpx_idct16x16_10_add_neon_pass2(const tran_low_t *src, int16_t *output,
; int16_t *pass1_output)
;
-; r0 const int16_t *src
+; r0 const tran_low_t *src
; r1 int16_t *output
; r2 int16_t *pass1_output
@@ -872,14 +875,14 @@
; TODO(hkuang): Find a better way to load the elements.
; load elements of 1, 3, 5, 7, 9, 11, 13, 15 into q8 - q15
- vld2.s16 {q8,q9}, [r0]!- vld2.s16 {q9,q10}, [r0]!- vld2.s16 {q10,q11}, [r0]!- vld2.s16 {q11,q12}, [r0]!- vld2.s16 {q12,q13}, [r0]!- vld2.s16 {q13,q14}, [r0]!- vld2.s16 {q14,q15}, [r0]!- vld2.s16 {q0,q1}, [r0]!+ LOAD_TRAN_LOW_TO_S16X2 d16, d17, d18, d19, r0
+ LOAD_TRAN_LOW_TO_S16X2 d18, d19, d20, d21, r0
+ LOAD_TRAN_LOW_TO_S16X2 d20, d21, d22, d23, r0
+ LOAD_TRAN_LOW_TO_S16X2 d22, d23, d24, d25, r0
+ LOAD_TRAN_LOW_TO_S16X2 d24, d25, d26, d27, r0
+ LOAD_TRAN_LOW_TO_S16X2 d26, d27, d28, d29, r0
+ LOAD_TRAN_LOW_TO_S16X2 d28, d29, d30, d31, r0
+ LOAD_TRAN_LOW_TO_S16X2 d0, d1, d2, d3, r0
vmov.s16 q15, q0;
; 2*cospi_30_64 = 3212
--- a/vpx_dsp/arm/idct16x16_add_neon.c
+++ b/vpx_dsp/arm/idct16x16_add_neon.c
@@ -10,8 +10,7 @@
#include <arm_neon.h>
-#include "./vpx_config.h"
-#include "vpx_dsp/arm/transpose_neon.h"
+#include "vpx_dsp/arm/idct_neon.h"
#include "vpx_dsp/txfm_common.h"
void vpx_idct16x16_256_add_neon_pass1(const int16_t *in, int16_t *out) {@@ -761,7 +760,7 @@
}
}
-void vpx_idct16x16_10_add_neon_pass1(const int16_t *in, int16_t *out) {+void vpx_idct16x16_10_add_neon_pass1(const tran_low_t *in, int16_t *out) {int16x4_t d4s16;
int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
int16x8_t q0s16, q1s16, q2s16, q4s16, q5s16, q6s16, q7s16;
@@ -770,28 +769,28 @@
int32x4_t q10s32, q11s32, q12s32, q15s32;
int16x8x2_t q0x2s16;
- q0x2s16 = vld2q_s16(in);
+ q0x2s16 = load_tran_low_to_s16x2q(in);
q8s16 = q0x2s16.val[0];
in += 16;
- q0x2s16 = vld2q_s16(in);
+ q0x2s16 = load_tran_low_to_s16x2q(in);
q9s16 = q0x2s16.val[0];
in += 16;
- q0x2s16 = vld2q_s16(in);
+ q0x2s16 = load_tran_low_to_s16x2q(in);
q10s16 = q0x2s16.val[0];
in += 16;
- q0x2s16 = vld2q_s16(in);
+ q0x2s16 = load_tran_low_to_s16x2q(in);
q11s16 = q0x2s16.val[0];
in += 16;
- q0x2s16 = vld2q_s16(in);
+ q0x2s16 = load_tran_low_to_s16x2q(in);
q12s16 = q0x2s16.val[0];
in += 16;
- q0x2s16 = vld2q_s16(in);
+ q0x2s16 = load_tran_low_to_s16x2q(in);
q13s16 = q0x2s16.val[0];
in += 16;
- q0x2s16 = vld2q_s16(in);
+ q0x2s16 = load_tran_low_to_s16x2q(in);
q14s16 = q0x2s16.val[0];
in += 16;
- q0x2s16 = vld2q_s16(in);
+ q0x2s16 = load_tran_low_to_s16x2q(in);
q15s16 = q0x2s16.val[0];
transpose_s16_8x8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
@@ -859,7 +858,7 @@
vst1q_s16(out, q15s16);
}
-void vpx_idct16x16_10_add_neon_pass2(const int16_t *src, int16_t *out,
+void vpx_idct16x16_10_add_neon_pass2(const tran_low_t *src, int16_t *out,
int16_t *pass1_output) {int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
@@ -874,28 +873,28 @@
int32x4_t q10s32, q11s32, q12s32, q13s32;
int16x8x2_t q0x2s16;
- q0x2s16 = vld2q_s16(src);
+ q0x2s16 = load_tran_low_to_s16x2q(src);
q8s16 = q0x2s16.val[0];
src += 16;
- q0x2s16 = vld2q_s16(src);
+ q0x2s16 = load_tran_low_to_s16x2q(src);
q9s16 = q0x2s16.val[0];
src += 16;
- q0x2s16 = vld2q_s16(src);
+ q0x2s16 = load_tran_low_to_s16x2q(src);
q10s16 = q0x2s16.val[0];
src += 16;
- q0x2s16 = vld2q_s16(src);
+ q0x2s16 = load_tran_low_to_s16x2q(src);
q11s16 = q0x2s16.val[0];
src += 16;
- q0x2s16 = vld2q_s16(src);
+ q0x2s16 = load_tran_low_to_s16x2q(src);
q12s16 = q0x2s16.val[0];
src += 16;
- q0x2s16 = vld2q_s16(src);
+ q0x2s16 = load_tran_low_to_s16x2q(src);
q13s16 = q0x2s16.val[0];
src += 16;
- q0x2s16 = vld2q_s16(src);
+ q0x2s16 = load_tran_low_to_s16x2q(src);
q14s16 = q0x2s16.val[0];
src += 16;
- q0x2s16 = vld2q_s16(src);
+ q0x2s16 = load_tran_low_to_s16x2q(src);
q15s16 = q0x2s16.val[0];
transpose_s16_8x8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
--- a/vpx_dsp/arm/idct16x16_neon.c
+++ b/vpx_dsp/arm/idct16x16_neon.c
@@ -16,8 +16,8 @@
int16_t *pass1_output,
int16_t skip_adding, uint8_t *dest,
int dest_stride);
-void vpx_idct16x16_10_add_neon_pass1(const int16_t *input, int16_t *output);
-void vpx_idct16x16_10_add_neon_pass2(const int16_t *src, int16_t *output,
+void vpx_idct16x16_10_add_neon_pass1(const tran_low_t *input, int16_t *output);
+void vpx_idct16x16_10_add_neon_pass2(const tran_low_t *src, int16_t *output,
int16_t *pass1_output);
#if HAVE_NEON_ASM
@@ -90,7 +90,7 @@
#endif
}
-void vpx_idct16x16_10_add_neon(const int16_t *input, uint8_t *dest,
+void vpx_idct16x16_10_add_neon(const tran_low_t *input, uint8_t *dest,
int dest_stride) {#if HAVE_NEON_ASM
int64_t store_reg[8];
--- a/vpx_dsp/arm/idct_neon.asm
+++ b/vpx_dsp/arm/idct_neon.asm
@@ -10,8 +10,9 @@
INCLUDE ./vpx_config.asm
- ; Helper function used to load tran_low_t into int16, narrowing if
+ ; Helper functions used to load tran_low_t into int16, narrowing if
; necessary.
+
; $dst0..3 are d registers with the pairs assumed to be contiguous in
; non-high-bitdepth builds. q0-q3 are used as temporaries in high-bitdepth.
MACRO
@@ -25,6 +26,21 @@
vmovn.i32 $dst3, q3
ELSE
vld1.s16 {$dst0-$dst1,$dst2-$dst3}, [$src]!+ ENDIF
+ MEND
+
+ ; $dst0..3 are d registers. q0-q3 are used as temporaries in high-bitdepth.
+ MACRO
+ LOAD_TRAN_LOW_TO_S16X2 $dst0, $dst1, $dst2, $dst3, $src
+ IF CONFIG_VP9_HIGHBITDEPTH
+ vld2.s32 {q0,q1}, [$src]!+ vld2.s32 {q2,q3}, [$src]!+ vmovn.i32 $dst0, q0
+ vmovn.i32 $dst1, q2
+ vmovn.i32 $dst2, q1
+ vmovn.i32 $dst3, q3
+ ELSE
+ vld2.s16 {$dst0,$dst1,$dst2,$dst3}, [$src]!ENDIF
MEND
END
--- a/vpx_dsp/arm/idct_neon.h
+++ b/vpx_dsp/arm/idct_neon.h
@@ -20,6 +20,23 @@
//------------------------------------------------------------------------------
// Helper functions used to load tran_low_t into int16, narrowing if necessary.
+static INLINE int16x8x2_t load_tran_low_to_s16x2q(const tran_low_t *buf) {+#if CONFIG_VP9_HIGHBITDEPTH
+ const int32x4x2_t v0 = vld2q_s32(buf);
+ const int32x4x2_t v1 = vld2q_s32(buf + 8);
+ const int16x4_t s0 = vmovn_s32(v0.val[0]);
+ const int16x4_t s1 = vmovn_s32(v0.val[1]);
+ const int16x4_t s2 = vmovn_s32(v1.val[0]);
+ const int16x4_t s3 = vmovn_s32(v1.val[1]);
+ int16x8x2_t res;
+ res.val[0] = vcombine_s16(s0, s2);
+ res.val[1] = vcombine_s16(s1, s3);
+ return res;
+#else
+ return vld2q_s16(buf);
+#endif
+}
+
static INLINE int16x8_t load_tran_low_to_s16q(const tran_low_t *buf) {#if CONFIG_VP9_HIGHBITDEPTH
const int32x4_t v0 = vld1q_s32(buf);
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -204,14 +204,6 @@
DSP_SRCS-$(HAVE_NEON_ASM) += arm/save_reg_neon$(ASM)
ifneq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
-ifeq ($(HAVE_NEON_ASM),yes)
-DSP_SRCS-yes += arm/idct16x16_add_neon$(ASM)
-else
-ifeq ($(HAVE_NEON),yes)
-DSP_SRCS-yes += arm/idct16x16_add_neon.c
-endif # HAVE_NEON
-endif # HAVE_NEON_ASM
-DSP_SRCS-$(HAVE_NEON) += arm/idct16x16_neon.c
DSP_SRCS-$(HAVE_NEON) += arm/idct32x32_add_neon.c
DSP_SRCS-$(HAVE_MSA) += mips/inv_txfm_msa.h
@@ -235,6 +227,7 @@
DSP_SRCS-yes += arm/idct8x8_1_add_neon$(ASM)
DSP_SRCS-yes += arm/idct8x8_add_neon$(ASM)
DSP_SRCS-yes += arm/idct16x16_1_add_neon$(ASM)
+DSP_SRCS-yes += arm/idct16x16_add_neon$(ASM)
else
DSP_SRCS-$(HAVE_NEON) += arm/idct4x4_1_add_neon.c
DSP_SRCS-$(HAVE_NEON) += arm/idct4x4_add_neon.c
@@ -241,8 +234,10 @@
DSP_SRCS-$(HAVE_NEON) += arm/idct8x8_1_add_neon.c
DSP_SRCS-$(HAVE_NEON) += arm/idct8x8_add_neon.c
DSP_SRCS-$(HAVE_NEON) += arm/idct16x16_1_add_neon.c
+DSP_SRCS-$(HAVE_NEON) += arm/idct16x16_add_neon.c
endif # HAVE_NEON_ASM
DSP_SRCS-$(HAVE_NEON) += arm/idct_neon.h
+DSP_SRCS-$(HAVE_NEON) += arm/idct16x16_neon.c
DSP_SRCS-$(HAVE_NEON) += arm/idct32x32_1_add_neon.c
DSP_SRCS-$(HAVE_NEON) += arm/idct32x32_34_add_neon.c
DSP_SRCS-$(HAVE_NEON) += arm/idct32x32_135_add_neon.c
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -689,7 +689,7 @@
specialize qw/vpx_idct16x16_256_add sse2/;
add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
- specialize qw/vpx_idct16x16_10_add sse2/;
+ specialize qw/vpx_idct16x16_10_add neon sse2/;
add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
specialize qw/vpx_idct16x16_1_add neon sse2/;