summaryrefslogtreecommitdiff
path: root/libavutil
diff options
context:
space:
mode:
Diffstat (limited to 'libavutil')
-rw-r--r--libavutil/arm/float_dsp_init_neon.c4
-rw-r--r--libavutil/arm/float_dsp_neon.S47
-rw-r--r--libavutil/float_dsp.c20
-rw-r--r--libavutil/float_dsp.h19
-rw-r--r--libavutil/ppc/float_dsp_altivec.c33
-rw-r--r--libavutil/ppc/float_dsp_altivec.h4
-rw-r--r--libavutil/ppc/float_dsp_init.c4
-rw-r--r--libavutil/x86/float_dsp_init.c75
8 files changed, 206 insertions, 0 deletions
diff --git a/libavutil/arm/float_dsp_init_neon.c b/libavutil/arm/float_dsp_init_neon.c
index 88eb4b3d2a..16ea47154a 100644
--- a/libavutil/arm/float_dsp_init_neon.c
+++ b/libavutil/arm/float_dsp_init_neon.c
@@ -32,9 +32,13 @@ void ff_vector_fmac_scalar_neon(float *dst, const float *src, float mul,
void ff_vector_fmul_scalar_neon(float *dst, const float *src, float mul,
int len);
+void ff_vector_fmul_window_neon(float *dst, const float *src0,
+ const float *src1, const float *win, int len);
+
void ff_float_dsp_init_neon(AVFloatDSPContext *fdsp)
{
fdsp->vector_fmul = ff_vector_fmul_neon;
fdsp->vector_fmac_scalar = ff_vector_fmac_scalar_neon;
fdsp->vector_fmul_scalar = ff_vector_fmul_scalar_neon;
+ fdsp->vector_fmul_window = ff_vector_fmul_window_neon;
}
diff --git a/libavutil/arm/float_dsp_neon.S b/libavutil/arm/float_dsp_neon.S
index 6d7bd5236e..540cfc6907 100644
--- a/libavutil/arm/float_dsp_neon.S
+++ b/libavutil/arm/float_dsp_neon.S
@@ -146,3 +146,50 @@ NOVFP vdup.32 q8, r2
bx lr
.unreq len
endfunc
+
+function ff_vector_fmul_window_neon, export=1
+ push {r4,r5,lr}
+ ldr lr, [sp, #12]
+ sub r2, r2, #8
+ sub r5, lr, #2
+ add r2, r2, r5, lsl #2
+ add r4, r3, r5, lsl #3
+ add ip, r0, r5, lsl #3
+ mov r5, #-16
+ vld1.32 {d0,d1}, [r1,:128]!
+ vld1.32 {d2,d3}, [r2,:128], r5
+ vld1.32 {d4,d5}, [r3,:128]!
+ vld1.32 {d6,d7}, [r4,:128], r5
+1: subs lr, lr, #4
+ vmul.f32 d22, d0, d4
+ vrev64.32 q3, q3
+ vmul.f32 d23, d1, d5
+ vrev64.32 q1, q1
+ vmul.f32 d20, d0, d7
+ vmul.f32 d21, d1, d6
+ beq 2f
+ vmla.f32 d22, d3, d7
+ vld1.32 {d0,d1}, [r1,:128]!
+ vmla.f32 d23, d2, d6
+ vld1.32 {d18,d19},[r2,:128], r5
+ vmls.f32 d20, d3, d4
+ vld1.32 {d24,d25},[r3,:128]!
+ vmls.f32 d21, d2, d5
+ vld1.32 {d6,d7}, [r4,:128], r5
+ vmov q1, q9
+ vrev64.32 q11, q11
+ vmov q2, q12
+ vswp d22, d23
+ vst1.32 {d20,d21},[r0,:128]!
+ vst1.32 {d22,d23},[ip,:128], r5
+ b 1b
+2: vmla.f32 d22, d3, d7
+ vmla.f32 d23, d2, d6
+ vmls.f32 d20, d3, d4
+ vmls.f32 d21, d2, d5
+ vrev64.32 q11, q11
+ vswp d22, d23
+ vst1.32 {d20,d21},[r0,:128]!
+ vst1.32 {d22,d23},[ip,:128], r5
+ pop {r4,r5,pc}
+endfunc
diff --git a/libavutil/float_dsp.c b/libavutil/float_dsp.c
index 22139defe4..cf33df303c 100644
--- a/libavutil/float_dsp.c
+++ b/libavutil/float_dsp.c
@@ -52,12 +52,32 @@ static void vector_dmul_scalar_c(double *dst, const double *src, double mul,
dst[i] = src[i] * mul;
}
+static void vector_fmul_window_c(float *dst, const float *src0,
+ const float *src1, const float *win, int len)
+{
+ int i, j;
+
+ dst += len;
+ win += len;
+ src0 += len;
+
+ for (i = -len, j = len - 1; i < 0; i++, j--) {
+ float s0 = src0[i];
+ float s1 = src1[j];
+ float wi = win[i];
+ float wj = win[j];
+ dst[i] = s0 * wj - s1 * wi;
+ dst[j] = s0 * wi + s1 * wj;
+ }
+}
+
void avpriv_float_dsp_init(AVFloatDSPContext *fdsp, int bit_exact)
{
fdsp->vector_fmul = vector_fmul_c;
fdsp->vector_fmac_scalar = vector_fmac_scalar_c;
fdsp->vector_fmul_scalar = vector_fmul_scalar_c;
fdsp->vector_dmul_scalar = vector_dmul_scalar_c;
+ fdsp->vector_fmul_window = vector_fmul_window_c;
#if ARCH_ARM
ff_float_dsp_init_arm(fdsp);
diff --git a/libavutil/float_dsp.h b/libavutil/float_dsp.h
index 41b73c5b26..d0ceaaf5e1 100644
--- a/libavutil/float_dsp.h
+++ b/libavutil/float_dsp.h
@@ -81,6 +81,25 @@ typedef struct AVFloatDSPContext {
*/
void (*vector_dmul_scalar)(double *dst, const double *src, double mul,
int len);
+
+ /**
+ * Overlap/add with window function.
+ * Used primarily by MDCT-based audio codecs.
+ * Source and destination vectors must overlap exactly or not at all.
+ *
+ * @param dst result vector
+ * constraints: 16-byte aligned
+ * @param src0 first source vector
+ * constraints: 16-byte aligned
+ * @param src1 second source vector
+ * constraints: 16-byte aligned
+ * @param win half-window vector
+ * constraints: 16-byte aligned
+ * @param len length of vector
+ * constraints: multiple of 4
+ */
+ void (*vector_fmul_window)(float *dst, const float *src0,
+ const float *src1, const float *win, int len);
} AVFloatDSPContext;
/**
diff --git a/libavutil/ppc/float_dsp_altivec.c b/libavutil/ppc/float_dsp_altivec.c
index 55e3fbe91f..e5fd9aba33 100644
--- a/libavutil/ppc/float_dsp_altivec.c
+++ b/libavutil/ppc/float_dsp_altivec.c
@@ -36,3 +36,36 @@ void ff_vector_fmul_altivec(float *dst, const float *src0, const float *src1,
vec_st(d1, 16, dst + i);
}
}
+
+void ff_vector_fmul_window_altivec(float *dst, const float *src0,
+ const float *src1, const float *win, int len)
+{
+ vector float zero, t0, t1, s0, s1, wi, wj;
+ const vector unsigned char reverse = vcprm(3, 2, 1, 0);
+ int i, j;
+
+ dst += len;
+ win += len;
+ src0 += len;
+
+ zero = (vector float)vec_splat_u32(0);
+
+ for (i = -len * 4, j = len * 4 - 16; i < 0; i += 16, j -= 16) {
+ s0 = vec_ld(i, src0);
+ s1 = vec_ld(j, src1);
+ wi = vec_ld(i, win);
+ wj = vec_ld(j, win);
+
+ s1 = vec_perm(s1, s1, reverse);
+ wj = vec_perm(wj, wj, reverse);
+
+ t0 = vec_madd(s0, wj, zero);
+ t0 = vec_nmsub(s1, wi, t0);
+ t1 = vec_madd(s0, wi, zero);
+ t1 = vec_madd(s1, wj, t1);
+ t1 = vec_perm(t1, t1, reverse);
+
+ vec_st(t0, i, dst);
+ vec_st(t1, j, dst);
+ }
+}
diff --git a/libavutil/ppc/float_dsp_altivec.h b/libavutil/ppc/float_dsp_altivec.h
index 0b9425bef4..4d46edf61a 100644
--- a/libavutil/ppc/float_dsp_altivec.h
+++ b/libavutil/ppc/float_dsp_altivec.h
@@ -24,4 +24,8 @@
extern void ff_vector_fmul_altivec(float *dst, const float *src0,
const float *src1, int len);
+extern void ff_vector_fmul_window_altivec(float *dst, const float *src0,
+ const float *src1, const float *win,
+ int len);
+
#endif /* AVUTIL_PPC_FLOAT_DSP_ALTIVEC_H */
diff --git a/libavutil/ppc/float_dsp_init.c b/libavutil/ppc/float_dsp_init.c
index 20527642cc..1134b56926 100644
--- a/libavutil/ppc/float_dsp_init.c
+++ b/libavutil/ppc/float_dsp_init.c
@@ -32,5 +32,9 @@ void ff_float_dsp_init_ppc(AVFloatDSPContext *fdsp, int bit_exact)
return;
fdsp->vector_fmul = ff_vector_fmul_altivec;
+
+ if (!bit_exact) {
+ fdsp->vector_fmul_window = ff_vector_fmul_window_altivec;
+ }
#endif
}
diff --git a/libavutil/x86/float_dsp_init.c b/libavutil/x86/float_dsp_init.c
index b3b7ff4c3a..56fb32b0d5 100644
--- a/libavutil/x86/float_dsp_init.c
+++ b/libavutil/x86/float_dsp_init.c
@@ -21,6 +21,7 @@
#include "libavutil/cpu.h"
#include "libavutil/float_dsp.h"
#include "cpu.h"
+#include "asm.h"
extern void ff_vector_fmul_sse(float *dst, const float *src0, const float *src1,
int len);
@@ -40,10 +41,84 @@ extern void ff_vector_dmul_scalar_sse2(double *dst, const double *src,
extern void ff_vector_dmul_scalar_avx(double *dst, const double *src,
double mul, int len);
+#if HAVE_6REGS
+static void vector_fmul_window_3dnowext(float *dst, const float *src0,
+ const float *src1, const float *win,
+ int len)
+{
+ x86_reg i = -len * 4;
+ x86_reg j = len * 4 - 8;
+ __asm__ volatile (
+ "1: \n"
+ "pswapd (%5, %1), %%mm1 \n"
+ "movq (%5, %0), %%mm0 \n"
+ "pswapd (%4, %1), %%mm5 \n"
+ "movq (%3, %0), %%mm4 \n"
+ "movq %%mm0, %%mm2 \n"
+ "movq %%mm1, %%mm3 \n"
+ "pfmul %%mm4, %%mm2 \n" // src0[len + i] * win[len + i]
+ "pfmul %%mm5, %%mm3 \n" // src1[j] * win[len + j]
+ "pfmul %%mm4, %%mm1 \n" // src0[len + i] * win[len + j]
+ "pfmul %%mm5, %%mm0 \n" // src1[j] * win[len + i]
+ "pfadd %%mm3, %%mm2 \n"
+ "pfsub %%mm0, %%mm1 \n"
+ "pswapd %%mm2, %%mm2 \n"
+ "movq %%mm1, (%2, %0) \n"
+ "movq %%mm2, (%2, %1) \n"
+ "sub $8, %1 \n"
+ "add $8, %0 \n"
+ "jl 1b \n"
+ "femms \n"
+ : "+r"(i), "+r"(j)
+ : "r"(dst + len), "r"(src0 + len), "r"(src1), "r"(win + len)
+ );
+}
+
+static void vector_fmul_window_sse(float *dst, const float *src0,
+ const float *src1, const float *win, int len)
+{
+ x86_reg i = -len * 4;
+ x86_reg j = len * 4 - 16;
+ __asm__ volatile (
+ "1: \n"
+ "movaps (%5, %1), %%xmm1 \n"
+ "movaps (%5, %0), %%xmm0 \n"
+ "movaps (%4, %1), %%xmm5 \n"
+ "movaps (%3, %0), %%xmm4 \n"
+ "shufps $0x1b, %%xmm1, %%xmm1 \n"
+ "shufps $0x1b, %%xmm5, %%xmm5 \n"
+ "movaps %%xmm0, %%xmm2 \n"
+ "movaps %%xmm1, %%xmm3 \n"
+ "mulps %%xmm4, %%xmm2 \n" // src0[len + i] * win[len + i]
+ "mulps %%xmm5, %%xmm3 \n" // src1[j] * win[len + j]
+ "mulps %%xmm4, %%xmm1 \n" // src0[len + i] * win[len + j]
+ "mulps %%xmm5, %%xmm0 \n" // src1[j] * win[len + i]
+ "addps %%xmm3, %%xmm2 \n"
+ "subps %%xmm0, %%xmm1 \n"
+ "shufps $0x1b, %%xmm2, %%xmm2 \n"
+ "movaps %%xmm1, (%2, %0) \n"
+ "movaps %%xmm2, (%2, %1) \n"
+ "sub $16, %1 \n"
+ "add $16, %0 \n"
+ "jl 1b \n"
+ : "+r"(i), "+r"(j)
+ : "r"(dst + len), "r"(src0 + len), "r"(src1), "r"(win + len)
+ );
+}
+#endif /* HAVE_6REGS */
+
void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp)
{
int mm_flags = av_get_cpu_flags();
+#if HAVE_6REGS
+ if (INLINE_AMD3DNOWEXT(mm_flags)) {
+ fdsp->vector_fmul_window = vector_fmul_window_3dnowext;
+ }
+ if (INLINE_SSE(mm_flags)) {
+ fdsp->vector_fmul_window = vector_fmul_window_sse;
+ }
+#endif
if (EXTERNAL_SSE(mm_flags)) {
fdsp->vector_fmul = ff_vector_fmul_sse;
fdsp->vector_fmac_scalar = ff_vector_fmac_scalar_sse;
generated by cgit v1.2.3 (git 2.39.1) at 2024-06-16 02:04:37 +0000