path: root/0002-1160.patch

From 541a62936532c3edd83edf6eb7ec83ab3e8bac5f Mon Sep 17 00:00:00 2001
From: "Nathan E. Egge" <unlord@xiph.org>
Date: Tue, 29 Dec 2020 06:58:33 -0500
Subject: [PATCH 1/3] Add bpc suffix to lr functions

---
 src/x86/looprestoration.asm         |  36 ++--
 src/x86/looprestoration_init_tmpl.c | 266 ++++++++++++++--------------
 src/x86/looprestoration_sse.asm     |  60 +++----
 3 files changed, 177 insertions(+), 185 deletions(-)

diff --git a/src/x86/looprestoration.asm b/src/x86/looprestoration.asm
index 44aaaf49..71e3e0d2 100644
--- a/src/x86/looprestoration.asm
+++ b/src/x86/looprestoration.asm
@@ -88,8 +88,8 @@ SECTION .text
 DECLARE_REG_TMP 4, 9, 7, 11, 12, 13, 14 ; ring buffer pointers
 
 INIT_YMM avx2
-cglobal wiener_filter7, 5, 15, 16, -384*12-16, dst, dst_stride, left, lpf, \
-                                               lpf_stride, w, edge, flt, h
+cglobal wiener_filter7_8bpc, 5, 15, 16, -384*12-16, dst, dst_stride, left, lpf, \
+                                                    lpf_stride, w, edge, flt, h
     mov           fltq, fltmp
     mov          edged, r8m
     mov             wd, wm
@@ -436,8 +436,8 @@ ALIGN function_align
     add           dstq, dst_strideq
     ret
 
-cglobal wiener_filter5, 5, 13, 16, 384*8+16, dst, dst_stride, left, lpf, \
-                                             lpf_stride, w, edge, flt, h
+cglobal wiener_filter5_8bpc, 5, 13, 16, 384*8+16, dst, dst_stride, left, lpf, \
+                                                  lpf_stride, w, edge, flt, h
     mov           fltq, fltmp
     mov          edged, r8m
     mov             wd, wm
@@ -554,7 +554,7 @@ cglobal wiener_filter5, 5, 13, 16, 384*8+16, dst, dst_stride, left, lpf, \
     jnz .h_have_right
     cmp           r10d, -33
     jl .h_have_right
-    call mangle(private_prefix %+ _wiener_filter7_avx2).extend_right
+    call mangle(private_prefix %+ _wiener_filter7_8bpc_avx2).extend_right
 .h_have_right:
     pshufb          m0, m4, m6
     pmaddubsw       m0, m12
@@ -613,7 +613,7 @@ ALIGN function_align
     jnz .hv_have_right
     cmp           r10d, -33
     jl .hv_have_right
-    call mangle(private_prefix %+ _wiener_filter7_avx2).extend_right
+    call mangle(private_prefix %+ _wiener_filter7_8bpc_avx2).extend_right
 .hv_have_right:
     pshufb          m0, m4, m6
     pmaddubsw       m0, m12
@@ -727,8 +727,8 @@ ALIGN function_align
     jl .v_loop
     ret
 
-cglobal sgr_filter_5x5, 5, 13, 16, 400*24+16, dst, dst_stride, left, lpf, \
-                                              lpf_stride, w, edge, params, h
+cglobal sgr_filter_5x5_8bpc, 5, 13, 16, 400*24+16, dst, dst_stride, left, lpf, \
+                                                   lpf_stride, w, edge, params, h
 %define base r12-sgr_x_by_x-256*4
     lea            r12, [sgr_x_by_x+256*4]
     mov        paramsq, paramsmp
@@ -1187,8 +1187,8 @@ ALIGN function_align
     add           dstq, dst_strideq
     ret
 
-cglobal sgr_filter_3x3, 5, 15, 15, -400*28-16, dst, dst_stride, left, lpf, \
-                                               lpf_stride, w, edge, params, h
+cglobal sgr_filter_3x3_8bpc, 5, 15, 15, -400*28-16, dst, dst_stride, left, lpf, \
+                                                    lpf_stride, w, edge, params, h
 %define base r14-sgr_x_by_x-256*4
     mov        paramsq, paramsmp
     mov          edged, r8m
@@ -1298,7 +1298,7 @@ cglobal sgr_filter_3x3, 5, 15, 15, -400*28-16, dst, dst_stride, left, lpf, \
     jnz .h_have_right
     cmp           r10d, -17
     jl .h_have_right
-    call mangle(private_prefix %+ _sgr_filter_5x5_avx2).extend_right
+    call mangle(private_prefix %+ _sgr_filter_5x5_8bpc_avx2).extend_right
 .h_have_right:
     pshufb          m0, m5, m8
     pmullw          m2, m0, m0
@@ -1346,7 +1346,7 @@ ALIGN function_align
     jnz .hv_have_right
     cmp           r10d, -17
     jl .hv_have_right
-    call mangle(private_prefix %+ _sgr_filter_5x5_avx2).extend_right
+    call mangle(private_prefix %+ _sgr_filter_5x5_8bpc_avx2).extend_right
 .hv_have_right:
     pshufb          m0, m5, m8
     pmullw          m3, m0, m0
@@ -1546,8 +1546,8 @@ ALIGN function_align
     add           dstq, dst_strideq
     ret
 
-cglobal sgr_filter_mix, 5, 13, 16, 400*56+8, dst, dst_stride, left, lpf, \
-                                             lpf_stride, w, edge, params, h
+cglobal sgr_filter_mix_8bpc, 5, 13, 16, 400*56+8, dst, dst_stride, left, lpf, \
+                                                  lpf_stride, w, edge, params, h
 %define base r12-sgr_x_by_x-256*4
     lea            r12, [sgr_x_by_x+256*4]
     mov        paramsq, paramsmp
@@ -1573,7 +1573,7 @@ cglobal sgr_filter_mix, 5, 13, 16, 400*56+8, dst, dst_stride, left, lpf, \
     call .h_top
     add           lpfq, lpf_strideq
     mov             t2, t1
-    call mangle(private_prefix %+ _sgr_filter_5x5_avx2).top_fixup
+    call mangle(private_prefix %+ _sgr_filter_5x5_8bpc_avx2).top_fixup
     add             t1, 400*12
     call .h_top
     lea            r10, [lpfq+lpf_strideq*4]
@@ -1681,7 +1681,7 @@ cglobal sgr_filter_mix, 5, 13, 16, 400*56+8, dst, dst_stride, left, lpf, \
     jnz .h_have_right
     cmp           r10d, -18
     jl .h_have_right
-    call mangle(private_prefix %+ _sgr_filter_5x5_avx2).extend_right
+    call mangle(private_prefix %+ _sgr_filter_5x5_8bpc_avx2).extend_right
 .h_have_right:
     pshufb          m6, m5, m9
     pshufb          m4, m5, m10
@@ -1742,7 +1742,7 @@ ALIGN function_align
     jnz .hv0_have_right
     cmp           r10d, -18
     jl .hv0_have_right
-    call mangle(private_prefix %+ _sgr_filter_5x5_avx2).extend_right
+    call mangle(private_prefix %+ _sgr_filter_5x5_8bpc_avx2).extend_right
 .hv0_have_right:
     pshufb          m6, m5, m9
     pshufb          m4, m5, m10
@@ -1853,7 +1853,7 @@ ALIGN function_align
     jnz .hv1_have_right
     cmp           r10d, -18
     jl .hv1_have_right
-    call mangle(private_prefix %+ _sgr_filter_5x5_avx2).extend_right
+    call mangle(private_prefix %+ _sgr_filter_5x5_8bpc_avx2).extend_right
 .hv1_have_right:
     pshufb          m6, m5, m9
     pshufb          m3, m5, m10
diff --git a/src/x86/looprestoration_init_tmpl.c b/src/x86/looprestoration_init_tmpl.c
index 57727787..a9aa5acf 100644
--- a/src/x86/looprestoration_init_tmpl.c
+++ b/src/x86/looprestoration_init_tmpl.c
@@ -30,179 +30,171 @@
 
 #include "common/intops.h"
 
-#define WIENER_FILTER(ext) \
-void dav1d_wiener_filter7_##ext(pixel *dst, ptrdiff_t dst_stride, \
-                                const pixel (*left)[4], const pixel *lpf, \
-                                ptrdiff_t lpf_stride, int w, int h, \
-                                const LooprestorationParams *params, \
-                                enum LrEdgeFlags edges); \
-void dav1d_wiener_filter5_##ext(pixel *dst, ptrdiff_t dst_stride, \
-                                const pixel (*left)[4], const pixel *lpf, \
-                                ptrdiff_t lpf_stride, int w, int h, \
-                                const LooprestorationParams *params, \
-                                enum LrEdgeFlags edges);
+#define decl_wiener_filter_fns(ext) \
+decl_lr_filter_fn(BF(dav1d_wiener_filter7, ext)); \
+decl_lr_filter_fn(BF(dav1d_wiener_filter5, ext))
 
-#define SGR_FILTER(ext) \
-void dav1d_sgr_filter_5x5_##ext(pixel *dst, ptrdiff_t dst_stride, \
-                                const pixel (*left)[4], const pixel *lpf, \
-                                ptrdiff_t lpf_stride, int w, int h, \
-                                const LooprestorationParams *params, \
-                                enum LrEdgeFlags edges); \
-void dav1d_sgr_filter_3x3_##ext(pixel *dst, ptrdiff_t dst_stride, \
-                                const pixel (*left)[4], const pixel *lpf, \
-                                ptrdiff_t lpf_stride, int w, int h, \
-                                const LooprestorationParams *params, \
-                                enum LrEdgeFlags edges); \
-void dav1d_sgr_filter_mix_##ext(pixel *dst, ptrdiff_t dst_stride, \
-                                const pixel (*left)[4], const pixel *lpf, \
-                                ptrdiff_t lpf_stride, int w, int h, \
-                                const LooprestorationParams *params, \
-                                enum LrEdgeFlags edges);
+#define decl_sgr_filter_fns(ext) \
+void BF(dav1d_sgr_filter_5x5, ext)(pixel *dst, ptrdiff_t dst_stride, \
+                                   const pixel (*left)[4], const pixel *lpf, \
+                                   ptrdiff_t lpf_stride, int w, int h, \
+                                   const LooprestorationParams *params, \
+                                   enum LrEdgeFlags edges); \
+void BF(dav1d_sgr_filter_3x3, ext)(pixel *dst, ptrdiff_t dst_stride, \
+                                   const pixel (*left)[4], const pixel *lpf, \
+                                   ptrdiff_t lpf_stride, int w, int h, \
+                                   const LooprestorationParams *params, \
+                                   enum LrEdgeFlags edges); \
+void BF(dav1d_sgr_filter_mix, ext)(pixel *dst, ptrdiff_t dst_stride, \
+                                   const pixel (*left)[4], const pixel *lpf, \
+                                   ptrdiff_t lpf_stride, int w, int h, \
+                                   const LooprestorationParams *params, \
+                                   enum LrEdgeFlags edges);
 
 /* FIXME: Replace with a port of the AVX2 code */
 #define SGR_FILTER_OLD(ext) \
-void dav1d_sgr_box3_h_##ext(int32_t *sumsq, int16_t *sum, \
-                            const pixel (*left)[4], \
-                            const pixel *src, const ptrdiff_t stride, \
-                            const int w, const int h, \
-                            const enum LrEdgeFlags edges); \
-void dav1d_sgr_box3_v_##ext(int32_t *sumsq, int16_t *sum, \
-                            const int w, const int h, \
-                            const enum LrEdgeFlags edges); \
-void dav1d_sgr_calc_ab1_##ext(int32_t *a, int16_t *b, \
-                              const int w, const int h, const unsigned s); \
-void dav1d_sgr_finish_filter1_##ext(coef *tmp, \
-                                    const pixel *src, const ptrdiff_t stride, \
-                                    const int32_t *a, const int16_t *b, \
-                                    const int w, const int h); \
+void BF(dav1d_sgr_box3_h, ext)(int32_t *sumsq, int16_t *sum, \
+                               const pixel (*left)[4], \
+                               const pixel *src, const ptrdiff_t stride, \
+                               const int w, const int h, \
+                               const enum LrEdgeFlags edges); \
+void BF(dav1d_sgr_box3_v, ext)(int32_t *sumsq, int16_t *sum, \
+                               const int w, const int h, \
+                               const enum LrEdgeFlags edges); \
+void BF(dav1d_sgr_calc_ab1, ext)(int32_t *a, int16_t *b, \
+                                 const int w, const int h, const unsigned s); \
+void BF(dav1d_sgr_finish_filter1, ext)(coef *tmp, \
+                                       const pixel *src, const ptrdiff_t stride, \
+                                       const int32_t *a, const int16_t *b, \
+                                       const int w, const int h); \
 \
 /* filter with a 3x3 box (radius=1) */ \
-static void dav1d_sgr_filter1_##ext(coef *tmp, \
-                                    const pixel *src, const ptrdiff_t stride, \
-                                    const pixel (*left)[4], \
-                                    const pixel *lpf, const ptrdiff_t lpf_stride, \
-                                    const int w, const int h, const int strength, \
-                                    const enum LrEdgeFlags edges) \
+static void BF(dav1d_sgr_filter1, ext)(coef *tmp, \
+                                       const pixel *src, const ptrdiff_t stride, \
+                                       const pixel (*left)[4], \
+                                       const pixel *lpf, const ptrdiff_t lpf_stride, \
+                                       const int w, const int h, const int strength, \
+                                       const enum LrEdgeFlags edges) \
 { \
     ALIGN_STK_32(int32_t, sumsq_mem, (384 + 16) * 68 + 8,); \
     int32_t *const sumsq = &sumsq_mem[(384 + 16) * 2 + 8], *const a = sumsq; \
     ALIGN_STK_32(int16_t, sum_mem, (384 + 16) * 68 + 16,); \
     int16_t *const sum = &sum_mem[(384 + 16) * 2 + 16], *const b = sum; \
 \
-    dav1d_sgr_box3_h_##ext(sumsq, sum, left, src, stride, w, h, edges); \
+    BF(dav1d_sgr_box3_h, ext)(sumsq, sum, left, src, stride, w, h, edges); \
     if (edges & LR_HAVE_TOP) \
-        dav1d_sgr_box3_h_##ext(&sumsq[-2 * (384 + 16)], &sum[-2 * (384 + 16)], \
-                              NULL, lpf, lpf_stride, w, 2, edges); \
+        BF(dav1d_sgr_box3_h, ext)(&sumsq[-2 * (384 + 16)], &sum[-2 * (384 + 16)], \
+                                  NULL, lpf, lpf_stride, w, 2, edges); \
 \
     if (edges & LR_HAVE_BOTTOM) \
-        dav1d_sgr_box3_h_##ext(&sumsq[h * (384 + 16)], &sum[h * (384 + 16)], \
-                              NULL, lpf + 6 * PXSTRIDE(lpf_stride), \
-                              lpf_stride, w, 2, edges); \
+        BF(dav1d_sgr_box3_h, ext)(&sumsq[h * (384 + 16)], &sum[h * (384 + 16)], \
+                                  NULL, lpf + 6 * PXSTRIDE(lpf_stride), \
+                                  lpf_stride, w, 2, edges); \
 \
-    dav1d_sgr_box3_v_##ext(sumsq, sum, w, h, edges); \
-    dav1d_sgr_calc_ab1_##ext(a, b, w, h, strength); \
-    dav1d_sgr_finish_filter1_##ext(tmp, src, stride, a, b, w, h); \
+    BF(dav1d_sgr_box3_v, ext)(sumsq, sum, w, h, edges); \
+    BF(dav1d_sgr_calc_ab1, ext)(a, b, w, h, strength); \
+    BF(dav1d_sgr_finish_filter1, ext)(tmp, src, stride, a, b, w, h); \
 } \
 \
-void dav1d_sgr_box5_h_##ext(int32_t *sumsq, int16_t *sum, \
-                            const pixel (*left)[4], \
-                            const pixel *src, const ptrdiff_t stride, \
-                            const int w, const int h, \
-                            const enum LrEdgeFlags edges); \
-void dav1d_sgr_box5_v_##ext(int32_t *sumsq, int16_t *sum, \
-                            const int w, const int h, \
-                            const enum LrEdgeFlags edges); \
-void dav1d_sgr_calc_ab2_##ext(int32_t *a, int16_t *b, \
-                              const int w, const int h, const int strength); \
-void dav1d_sgr_finish_filter2_##ext(coef *tmp, \
-                                    const pixel *src, const ptrdiff_t stride, \
-                                    const int32_t *a, const int16_t *b, \
-                                    const int w, const int h); \
+void BF(dav1d_sgr_box5_h, ext)(int32_t *sumsq, int16_t *sum, \
+                               const pixel (*left)[4], \
+                               const pixel *src, const ptrdiff_t stride, \
+                               const int w, const int h, \
+                               const enum LrEdgeFlags edges); \
+void BF(dav1d_sgr_box5_v, ext)(int32_t *sumsq, int16_t *sum, \
+                               const int w, const int h, \
+                               const enum LrEdgeFlags edges); \
+void BF(dav1d_sgr_calc_ab2, ext)(int32_t *a, int16_t *b, \
+                                 const int w, const int h, const int strength); \
+void BF(dav1d_sgr_finish_filter2, ext)(coef *tmp, \
+                                       const pixel *src, const ptrdiff_t stride, \
+                                       const int32_t *a, const int16_t *b, \
+                                       const int w, const int h); \
 \
 /* filter with a 5x5 box (radius=2) */ \
-static void dav1d_sgr_filter2_##ext(coef *tmp, \
-                                    const pixel *src, const ptrdiff_t stride, \
-                                    const pixel (*left)[4], \
-                                    const pixel *lpf, const ptrdiff_t lpf_stride, \
-                                    const int w, const int h, const int strength, \
-                                    const enum LrEdgeFlags edges) \
+static void BF(dav1d_sgr_filter2, ext)(coef *tmp, \
+                                       const pixel *src, const ptrdiff_t stride, \
+                                       const pixel (*left)[4], \
+                                       const pixel *lpf, const ptrdiff_t lpf_stride, \
+                                       const int w, const int h, const int strength, \
+                                       const enum LrEdgeFlags edges) \
 { \
     ALIGN_STK_32(int32_t, sumsq_mem, (384 + 16) * 68 + 8,); \
     int32_t *const sumsq = &sumsq_mem[(384 + 16) * 2 + 8], *const a = sumsq; \
     ALIGN_STK_32(int16_t, sum_mem, (384 + 16) * 68 + 16,); \
     int16_t *const sum = &sum_mem[(384 + 16) * 2 + 16], *const b = sum; \
 \
-    dav1d_sgr_box5_h_##ext(sumsq, sum, left, src, stride, w, h, edges); \
+    BF(dav1d_sgr_box5_h, ext)(sumsq, sum, left, src, stride, w, h, edges); \
     if (edges & LR_HAVE_TOP) \
-        dav1d_sgr_box5_h_##ext(&sumsq[-2 * (384 + 16)], &sum[-2 * (384 + 16)], \
-                              NULL, lpf, lpf_stride, w, 2, edges); \
+        BF(dav1d_sgr_box5_h, ext)(&sumsq[-2 * (384 + 16)], &sum[-2 * (384 + 16)], \
+                                  NULL, lpf, lpf_stride, w, 2, edges); \
 \
     if (edges & LR_HAVE_BOTTOM) \
-        dav1d_sgr_box5_h_##ext(&sumsq[h * (384 + 16)], &sum[h * (384 + 16)], \
-                              NULL, lpf + 6 * PXSTRIDE(lpf_stride), \
-                              lpf_stride, w, 2, edges); \
+        BF(dav1d_sgr_box5_h, ext)(&sumsq[h * (384 + 16)], &sum[h * (384 + 16)], \
+                                  NULL, lpf + 6 * PXSTRIDE(lpf_stride), \
+                                  lpf_stride, w, 2, edges); \
 \
-    dav1d_sgr_box5_v_##ext(sumsq, sum, w, h, edges); \
-    dav1d_sgr_calc_ab2_##ext(a, b, w, h, strength); \
-    dav1d_sgr_finish_filter2_##ext(tmp, src, stride, a, b, w, h); \
+    BF(dav1d_sgr_box5_v, ext)(sumsq, sum, w, h, edges); \
+    BF(dav1d_sgr_calc_ab2, ext)(a, b, w, h, strength); \
+    BF(dav1d_sgr_finish_filter2, ext)(tmp, src, stride, a, b, w, h); \
 } \
 \
-void dav1d_sgr_weighted1_##ext(pixel *dst, const ptrdiff_t stride, \
-                               const coef *t1, const int w, const int h, \
-                               const int wt); \
-void dav1d_sgr_weighted2_##ext(pixel *dst, const ptrdiff_t stride, \
-                               const coef *t1, const coef *t2, \
-                               const int w, const int h, \
-                               const uint32_t wt); \
+void BF(dav1d_sgr_weighted1, ext)(pixel *dst, const ptrdiff_t stride, \
+                                  const coef *t1, const int w, const int h, \
+                                  const int wt); \
+void BF(dav1d_sgr_weighted2, ext)(pixel *dst, const ptrdiff_t stride, \
+                                  const coef *t1, const coef *t2, \
+                                  const int w, const int h, \
+                                  const uint32_t wt); \
 \
-static void sgr_filter_5x5_##ext(pixel *const dst, const ptrdiff_t dst_stride, \
-                                 const pixel (*const left)[4], \
-                                 const pixel *lpf, const ptrdiff_t lpf_stride, \
-                                 const int w, const int h, \
-                                 const LooprestorationParams *const params, \
-                                 const enum LrEdgeFlags edges) \
+static void BF(sgr_filter_5x5, ext)(pixel *const dst, const ptrdiff_t dst_stride, \
+                                    const pixel (*const left)[4], \
+                                    const pixel *lpf, const ptrdiff_t lpf_stride, \
+                                    const int w, const int h, \
+                                    const LooprestorationParams *const params, \
+                                    const enum LrEdgeFlags edges) \
 { \
     ALIGN_STK_32(coef, tmp, 64 * 384,); \
-    dav1d_sgr_filter2_##ext(tmp, dst, dst_stride, left, lpf, lpf_stride, \
-                           w, h, params->sgr.s0, edges); \
-    dav1d_sgr_weighted1_##ext(dst, dst_stride, tmp, w, h, params->sgr.w0); \
+    BF(dav1d_sgr_filter2, ext)(tmp, dst, dst_stride, left, lpf, lpf_stride, \
+                               w, h, params->sgr.s0, edges); \
+    BF(dav1d_sgr_weighted1, ext)(dst, dst_stride, tmp, w, h, params->sgr.w0); \
 } \
-static void sgr_filter_3x3_##ext(pixel *const dst, const ptrdiff_t dst_stride, \
-                                 const pixel (*const left)[4], \
-                                 const pixel *lpf, const ptrdiff_t lpf_stride, \
-                                 const int w, const int h, \
-                                 const LooprestorationParams *const params, \
-                                 const enum LrEdgeFlags edges) \
+static void BF(sgr_filter_3x3, ext)(pixel *const dst, const ptrdiff_t dst_stride, \
+                                    const pixel (*const left)[4], \
+                                    const pixel *lpf, const ptrdiff_t lpf_stride, \
+                                    const int w, const int h, \
+                                    const LooprestorationParams *const params, \
+                                    const enum LrEdgeFlags edges) \
 { \
     ALIGN_STK_32(coef, tmp, 64 * 384,); \
-    dav1d_sgr_filter1_##ext(tmp, dst, dst_stride, left, lpf, lpf_stride, \
-                           w, h, params->sgr.s1, edges); \
-    dav1d_sgr_weighted1_##ext(dst, dst_stride, tmp, w, h, params->sgr.w1); \
+    BF(dav1d_sgr_filter1, ext)(tmp, dst, dst_stride, left, lpf, lpf_stride, \
+                               w, h, params->sgr.s1, edges); \
+    BF(dav1d_sgr_weighted1, ext)(dst, dst_stride, tmp, w, h, params->sgr.w1); \
 } \
-static void sgr_filter_mix_##ext(pixel *const dst, const ptrdiff_t dst_stride, \
-                                 const pixel (*const left)[4], \
-                                 const pixel *lpf, const ptrdiff_t lpf_stride, \
-                                 const int w, const int h, \
-                                 const LooprestorationParams *const params, \
-                                 const enum LrEdgeFlags edges) \
+static void BF(sgr_filter_mix, ext)(pixel *const dst, const ptrdiff_t dst_stride, \
+                                    const pixel (*const left)[4], \
+                                    const pixel *lpf, const ptrdiff_t lpf_stride, \
+                                    const int w, const int h, \
+                                    const LooprestorationParams *const params, \
+                                    const enum LrEdgeFlags edges) \
 { \
     ALIGN_STK_32(coef, tmp1, 64 * 384,); \
     ALIGN_STK_32(coef, tmp2, 64 * 384,); \
-    dav1d_sgr_filter2_##ext(tmp1, dst, dst_stride, left, lpf, lpf_stride, \
-                           w, h, params->sgr.s0, edges); \
-    dav1d_sgr_filter1_##ext(tmp2, dst, dst_stride, left, lpf, lpf_stride, \
-                           w, h, params->sgr.s1, edges); \
+    BF(dav1d_sgr_filter2, ext)(tmp1, dst, dst_stride, left, lpf, lpf_stride, \
+                               w, h, params->sgr.s0, edges); \
+    BF(dav1d_sgr_filter1, ext)(tmp2, dst, dst_stride, left, lpf, lpf_stride, \
+                               w, h, params->sgr.s1, edges); \
     const uint32_t wt = (params->sgr.w1 << 16) | (uint16_t) params->sgr.w0; \
-    dav1d_sgr_weighted2_##ext(dst, dst_stride, tmp1, tmp2, w, h, wt); \
+    BF(dav1d_sgr_weighted2, ext)(dst, dst_stride, tmp1, tmp2, w, h, wt); \
 }
 
 #if BITDEPTH == 8
-WIENER_FILTER(sse2)
-WIENER_FILTER(ssse3)
+decl_wiener_filter_fns(sse2);
+decl_wiener_filter_fns(ssse3);
 SGR_FILTER_OLD(ssse3)
 # if ARCH_X86_64
-WIENER_FILTER(avx2)
-SGR_FILTER(avx2)
+decl_wiener_filter_fns(avx2);
+decl_sgr_filter_fns(avx2)
 # endif
 #endif
 
@@ -211,25 +203,25 @@ COLD void bitfn(dav1d_loop_restoration_dsp_init_x86)(Dav1dLoopRestorationDSPCont
 
     if (!(flags & DAV1D_X86_CPU_FLAG_SSE2)) return;
 #if BITDEPTH == 8
-    c->wiener[0] = dav1d_wiener_filter7_sse2;
-    c->wiener[1] = dav1d_wiener_filter5_sse2;
+    c->wiener[0] = BF(dav1d_wiener_filter7, sse2);
+    c->wiener[1] = BF(dav1d_wiener_filter5, sse2);
 #endif
 
     if (!(flags & DAV1D_X86_CPU_FLAG_SSSE3)) return;
 #if BITDEPTH == 8
-    c->wiener[0] = dav1d_wiener_filter7_ssse3;
-    c->wiener[1] = dav1d_wiener_filter5_ssse3;
-    c->sgr[0] = sgr_filter_5x5_ssse3;
-    c->sgr[1] = sgr_filter_3x3_ssse3;
-    c->sgr[2] = sgr_filter_mix_ssse3;
+    c->wiener[0] = BF(dav1d_wiener_filter7, ssse3);
+    c->wiener[1] = BF(dav1d_wiener_filter5, ssse3);
+    c->sgr[0] = BF(sgr_filter_5x5, ssse3);
+    c->sgr[1] = BF(sgr_filter_3x3, ssse3);
+    c->sgr[2] = BF(sgr_filter_mix, ssse3);
 #endif
 
     if (!(flags & DAV1D_X86_CPU_FLAG_AVX2)) return;
 #if BITDEPTH == 8 && ARCH_X86_64
-    c->wiener[0] = dav1d_wiener_filter7_avx2;
-    c->wiener[1] = dav1d_wiener_filter5_avx2;
-    c->sgr[0] = dav1d_sgr_filter_5x5_avx2;
-    c->sgr[1] = dav1d_sgr_filter_3x3_avx2;
-    c->sgr[2] = dav1d_sgr_filter_mix_avx2;
+    c->wiener[0] = BF(dav1d_wiener_filter7, avx2);
+    c->wiener[1] = BF(dav1d_wiener_filter5, avx2);
+    c->sgr[0] = BF(dav1d_sgr_filter_5x5, avx2);
+    c->sgr[1] = BF(dav1d_sgr_filter_3x3, avx2);
+    c->sgr[2] = BF(dav1d_sgr_filter_mix, avx2);
 #endif
 }
diff --git a/src/x86/looprestoration_sse.asm b/src/x86/looprestoration_sse.asm
index 5d3ca492..4b77138d 100644
--- a/src/x86/looprestoration_sse.asm
+++ b/src/x86/looprestoration_sse.asm
@@ -97,8 +97,8 @@ SECTION .text
 %macro WIENER 0
 %if ARCH_X86_64
 DECLARE_REG_TMP 4, 10, 7, 11, 12, 13, 14 ; ring buffer pointers
-cglobal wiener_filter7, 5, 15, 16, -384*12-16, dst, dst_stride, left, lpf, \
-                                               lpf_stride, w, edge, flt, h, x
+cglobal wiener_filter7_8bpc, 5, 15, 16, -384*12-16, dst, dst_stride, left, lpf, \
+                                                    lpf_stride, w, edge, flt, h, x
     %define base 0
     mov           fltq, fltmp
     mov          edged, r8m
@@ -139,7 +139,7 @@ DECLARE_REG_TMP 4, 0, _, 5
     %define m11         [stk+96]
     %define stk_off     112
 %endif
-cglobal wiener_filter7, 0, 7, 8, -384*12-stk_off, _, x, left, lpf, lpf_stride
+cglobal wiener_filter7_8bpc, 0, 7, 8, -384*12-stk_off, _, x, left, lpf, lpf_stride
     %define base        r6-pb_right_ext_mask-21
     %define stk         esp
     %define dstq        leftq
@@ -245,7 +245,7 @@ cglobal wiener_filter7, 0, 7, 8, -384*12-stk_off, _, x, left, lpf, lpf_stride
     add           lpfq, [rsp+gprsize*1]
     call .hv_bottom
 .v1:
-    call mangle(private_prefix %+ _wiener_filter7_ssse3).v
+    call mangle(private_prefix %+ _wiener_filter7_8bpc_ssse3).v
     RET
 .no_top:
     lea             t3, [lpfq+lpf_strideq*4]
@@ -281,9 +281,9 @@ cglobal wiener_filter7, 0, 7, 8, -384*12-stk_off, _, x, left, lpf, lpf_stride
     dec             hd
     jnz .main
 .v3:
-    call mangle(private_prefix %+ _wiener_filter7_ssse3).v
+    call mangle(private_prefix %+ _wiener_filter7_8bpc_ssse3).v
 .v2:
-    call mangle(private_prefix %+ _wiener_filter7_ssse3).v
+    call mangle(private_prefix %+ _wiener_filter7_8bpc_ssse3).v
     jmp .v1
 .extend_right:
     movd            m2, [lpfq-4]
@@ -685,8 +685,8 @@ ALIGN function_align
 %endif
 
 %if ARCH_X86_64
-cglobal wiener_filter5, 5, 13, 16, 384*8+16, dst, dst_stride, left, lpf, \
-                                             lpf_stride, w, edge, flt, h, x
+cglobal wiener_filter5_8bpc, 5, 13, 16, 384*8+16, dst, dst_stride, left, lpf, \
+                                                  lpf_stride, w, edge, flt, h, x
     mov           fltq, fltmp
     mov          edged, r8m
     mov             wd, wm
@@ -720,7 +720,7 @@ cglobal wiener_filter5, 5, 13, 16, 384*8+16, dst, dst_stride, left, lpf, \
     %define m11         [stk+80]
     %define stk_off     96
 %endif
-cglobal wiener_filter5, 0, 7, 8, -384*8-stk_off, _, x, left, lpf, lpf_stride
+cglobal wiener_filter5_8bpc, 0, 7, 8, -384*8-stk_off, _, x, left, lpf, lpf_stride
     %define stk         esp
     %define leftmp      [stk+28]
     %define m8          [base+pw_m16380]
@@ -827,14 +827,14 @@ cglobal wiener_filter5, 0, 7, 8, -384*8-stk_off, _, x, left, lpf, lpf_stride
     dec             hd
     jnz .main
 .v2:
-    call mangle(private_prefix %+ _wiener_filter5_ssse3).v
+    call mangle(private_prefix %+ _wiener_filter5_8bpc_ssse3).v
     add           dstq, dst_strideq
     mov             t4, t3
     mov             t3, t2
     mov             t2, t1
     movifnidn    dstmp, dstq
 .v1:
-    call mangle(private_prefix %+ _wiener_filter5_ssse3).v
+    call mangle(private_prefix %+ _wiener_filter5_8bpc_ssse3).v
     jmp .end
 .h:
     %define stk esp+4
@@ -873,7 +873,7 @@ cglobal wiener_filter5, 0, 7, 8, -384*8-stk_off, _, x, left, lpf, lpf_stride
     jnz .h_have_right
     cmp             xd, -17
     jl .h_have_right
-    call mangle(private_prefix %+ _wiener_filter7 %+ SUFFIX).extend_right
+    call mangle(private_prefix %+ _wiener_filter7_8bpc %+ SUFFIX).extend_right
 .h_have_right:
 %macro %%h5 0
 %if cpuflag(ssse3)
@@ -991,7 +991,7 @@ ALIGN function_align
     jnz .hv_have_right
     cmp             xd, -17
     jl .hv_have_right
-    call mangle(private_prefix %+ _wiener_filter7 %+ SUFFIX).extend_right
+    call mangle(private_prefix %+ _wiener_filter7_8bpc %+ SUFFIX).extend_right
 .hv_have_right:
     %%h5
     mova            m2, [t3+xq*2]
@@ -1161,7 +1161,7 @@ WIENER
 %endmacro
 
 %if ARCH_X86_64
-cglobal sgr_box3_h, 5, 11, 8, sumsq, sum, left, src, stride, x, h, edge, w, xlim
+cglobal sgr_box3_h_8bpc, 5, 11, 8, sumsq, sum, left, src, stride, x, h, edge, w, xlim
     mov        xlimd, edgem
     movifnidn     xd, xm
     mov           hd, hm
@@ -1170,7 +1170,7 @@ cglobal sgr_box3_h, 5, 11, 8, sumsq, sum, left, src, stride, x, h, edge, w, xlim
     add           xd, xlimd
     xor        xlimd, 2                             ; 2*!have_right
 %else
-cglobal sgr_box3_h, 6, 7, 8, sumsq, sum, left, src, stride, x, h, edge, w, xlim
+cglobal sgr_box3_h_8bpc, 6, 7, 8, sumsq, sum, left, src, stride, x, h, edge, w, xlim
  %define wq     r0m
  %define xlimd  r1m
  %define hd     hmp
@@ -1287,10 +1287,10 @@ cglobal sgr_box3_h, 6, 7, 8, sumsq, sum, left, src, stride, x, h, edge, w, xlim
     RET
 
 %if ARCH_X86_64
-cglobal sgr_box3_v, 4, 10, 9, sumsq, sum, w, h, edge, x, y, sumsq_base, sum_base, ylim
+cglobal sgr_box3_v_8bpc, 4, 10, 9, sumsq, sum, w, h, edge, x, y, sumsq_base, sum_base, ylim
     movifnidn  edged, edgem
 %else
-cglobal sgr_box3_v, 3, 7, 8, -28, sumsq, sum, w, edge, h, x, y
+cglobal sgr_box3_v_8bpc, 3, 7, 8, -28, sumsq, sum, w, edge, h, x, y
  %define sumsq_baseq dword [esp+0]
  %define sum_baseq   dword [esp+4]
  %define ylimd       dword [esp+8]
@@ -1383,7 +1383,7 @@ cglobal sgr_box3_v, 3, 7, 8, -28, sumsq, sum, w, edge, h, x, y
     jl .loop_x
     RET
 
-cglobal sgr_calc_ab1, 4, 7, 12, a, b, w, h, s
+cglobal sgr_calc_ab1_8bpc, 4, 7, 12, a, b, w, h, s
     movifnidn     sd, sm
     sub           aq, (384+16-1)*4
     sub           bq, (384+16-1)*2
@@ -1463,8 +1463,8 @@ cglobal sgr_calc_ab1, 4, 7, 12, a, b, w, h, s
     RET
 
 %if ARCH_X86_64
-cglobal sgr_finish_filter1, 5, 13, 16, t, src, stride, a, b, w, h, \
-                                       tmp_base, src_base, a_base, b_base, x, y
+cglobal sgr_finish_filter1_8bpc, 5, 13, 16, t, src, stride, a, b, w, h, \
+                                            tmp_base, src_base, a_base, b_base, x, y
     movifnidn     wd, wm
     mov           hd, hm
     mova         m15, [pw_16]
@@ -1474,7 +1474,7 @@ cglobal sgr_finish_filter1, 5, 13, 16, t, src, stride, a, b, w, h, \
     mov      b_baseq, bq
     xor           xd, xd
 %else
-cglobal sgr_finish_filter1, 7, 7, 8, -144, t, src, stride, a, b, x, y
+cglobal sgr_finish_filter1_8bpc, 7, 7, 8, -144, t, src, stride, a, b, x, y
  %define tmp_baseq  [esp+8]
  %define src_baseq  [esp+12]
  %define a_baseq    [esp+16]
@@ -1688,7 +1688,7 @@ cglobal sgr_finish_filter1, 7, 7, 8, -144, t, src, stride, a, b, x, y
     jl .loop_x
     RET
 
-cglobal sgr_weighted1, 4, 7, 8, dst, stride, t, w, h, wt
+cglobal sgr_weighted1_8bpc, 4, 7, 8, dst, stride, t, w, h, wt
     movifnidn     hd, hm
 %if ARCH_X86_32
     SETUP_PIC r6, 0
@@ -1726,14 +1726,14 @@ cglobal sgr_weighted1, 4, 7, 8, dst, stride, t, w, h, wt
     RET
 
 %if ARCH_X86_64
-cglobal sgr_box5_h, 5, 11, 12, sumsq, sum, left, src, stride, w, h, edge, x, xlim
+cglobal sgr_box5_h_8bpc, 5, 11, 12, sumsq, sum, left, src, stride, w, h, edge, x, xlim
     mov        edged, edgem
     movifnidn     wd, wm
     mov           hd, hm
     mova         m10, [pb_0]
     mova         m11, [pb_0_1]
 %else
-cglobal sgr_box5_h, 7, 7, 8, sumsq, sum, left, src, xlim, x, h, edge
+cglobal sgr_box5_h_8bpc, 7, 7, 8, sumsq, sum, left, src, xlim, x, h, edge
  %define edgeb      byte edgem
  %define wd         xd
  %define wq         wd
@@ -1909,11 +1909,11 @@ cglobal sgr_box5_h, 7, 7, 8, sumsq, sum, left, src, xlim, x, h, edge
     RET
 
 %if ARCH_X86_64
-cglobal sgr_box5_v, 4, 10, 15, sumsq, sum, w, h, edge, x, y, sumsq_ptr, sum_ptr, ylim
+cglobal sgr_box5_v_8bpc, 4, 10, 15, sumsq, sum, w, h, edge, x, y, sumsq_ptr, sum_ptr, ylim
     movifnidn  edged, edgem
     mov        ylimd, edged
 %else
-cglobal sgr_box5_v, 5, 7, 8, -44, sumsq, sum, x, y, ylim, sumsq_ptr, sum_ptr
+cglobal sgr_box5_v_8bpc, 5, 7, 8, -44, sumsq, sum, x, y, ylim, sumsq_ptr, sum_ptr
  %define wm     [esp+0]
  %define hm     [esp+4]
  %define edgem  [esp+8]
@@ -2127,7 +2127,7 @@ cglobal sgr_box5_v, 5, 7, 8, -44, sumsq, sum, x, y, ylim, sumsq_ptr, sum_ptr
     jmp .sum_loop_y_noload
 %endif
 
-cglobal sgr_calc_ab2, 4, 7, 11, a, b, w, h, s
+cglobal sgr_calc_ab2_8bpc, 4, 7, 11, a, b, w, h, s
     movifnidn     sd, sm
     sub           aq, (384+16-1)*4
     sub           bq, (384+16-1)*2
@@ -2205,7 +2205,7 @@ cglobal sgr_calc_ab2, 4, 7, 11, a, b, w, h, s
     RET
 
 %if ARCH_X86_64
-cglobal sgr_finish_filter2, 5, 13, 14, t, src, stride, a, b, w, h, \
+cglobal sgr_finish_filter2_8bpc, 5, 13, 14, t, src, stride, a, b, w, h, \
                                        tmp_base, src_base, a_base, b_base, x, y
     movifnidn     wd, wm
     mov           hd, hm
@@ -2219,7 +2219,7 @@ cglobal sgr_finish_filter2, 5, 13, 14, t, src, stride, a, b, w, h, \
     psrlw        m11, m12, 1                    ; pw_128
     pxor         m13, m13
 %else
-cglobal sgr_finish_filter2, 6, 7, 8, t, src, stride, a, b, x, y
+cglobal sgr_finish_filter2_8bpc, 6, 7, 8, t, src, stride, a, b, x, y
  %define tmp_baseq  r0m
  %define src_baseq  r1m
  %define a_baseq    r3m
@@ -2378,7 +2378,7 @@ cglobal sgr_finish_filter2, 6, 7, 8, t, src, stride, a, b, x, y
     RET
 
 %undef t2
-cglobal sgr_weighted2, 4, 7, 12, dst, stride, t1, t2, w, h, wt
+cglobal sgr_weighted2_8bpc, 4, 7, 12, dst, stride, t1, t2, w, h, wt
     movifnidn     wd, wm
     movd          m0, wtm
 %if ARCH_X86_64
-- 
GitLab


From 1a4489861e55f0e4d70df60ecf15559dfda70aee Mon Sep 17 00:00:00 2001
From: "Nathan E. Egge" <unlord@xiph.org>
Date: Sun, 10 Jan 2021 14:12:10 -0500
Subject: [PATCH 2/3] x86: lr: Add AVX2 implementation of wiener filter for 16
 bpc

Relative speed-ups over C code (compared with gcc-9.3.0):

                                           C      AVX2
wiener_5tap_10bpc:                  194892.0   14831.9    13.14x
wiener_5tap_12bpc:                  194295.4   14828.9    13.10x
wiener_7tap_10bpc:                  194391.7   19461.4     9.99x
wiener_7tap_12bpc:                  194136.1   19418.7    10.00x
---
 src/x86/looprestoration16_avx2.asm | 480 +++++++++++++++++++++++++++++
 1 file changed, 480 insertions(+)
 create mode 100644 src/x86/looprestoration16_avx2.asm

diff --git a/src/x86/looprestoration16_avx2.asm b/src/x86/looprestoration16_avx2.asm
new file mode 100644
index 00000000..4eb1b805
--- /dev/null
+++ b/src/x86/looprestoration16_avx2.asm
@@ -0,0 +1,480 @@
+; Copyright (c) 2017-2021, The rav1e contributors
+; Copyright (c) 2021, Nathan Egge
+; All rights reserved.
+;
+; Redistribution and use in source and binary forms, with or without
+; modification, are permitted provided that the following conditions are met:
+;
+; 1. Redistributions of source code must retain the above copyright notice, this
+;    list of conditions and the following disclaimer.
+;
+; 2. Redistributions in binary form must reproduce the above copyright notice,
+;    this list of conditions and the following disclaimer in the documentation
+;    and/or other materials provided with the distribution.
+;
+; THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+; ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+; WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+; DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+; ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+; (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+; ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+; (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+; SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+%include "config.asm"
+%include "ext/x86/x86inc.asm"
+
+%if ARCH_X86_64
+
+SECTION_RODATA 32
+
+wiener5_shufB:  db  0,  1,  2,  3,  2,  3,  4,  5,  4,  5,  6,  7,  6,  7,  8,  9
+wiener5_shufC:  db  8,  9,  6,  7, 10, 11,  8,  9, 12, 13, 10, 11, 14, 15, 12, 13
+wiener5_shufD:  db  4,  5, -1, -1,  6,  7, -1, -1,  8,  9, -1, -1, 10, 11, -1, -1
+wiener5_l_shuf: db  4,  5,  4,  5,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15
+pb_0to31:       db  0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15
+                db 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
+
+wiener7_shufC:  db  4,  5,  2,  3,  6,  7,  4,  5,  8,  9,  6,  7, 10, 11,  8,  9
+wiener7_shufD:  db  4,  5,  6,  7,  6,  7,  8,  9,  8,  9, 10, 11, 10, 11, 12, 13
+wiener7_shufE:  db  8,  9, -1, -1, 10, 11, -1, -1, 12, 13, -1, -1, 14, 15, -1, -1
+rev_w:          db 14, 15, 12, 13, 10, 11,  8,  9,  6,  7,  4,  5,  2,  3,  0,  1
+rev_d:          db 12, 13, 14, 15,  8,  9, 10, 11,  4,  5,  6,  7,  0,  1,  2,  3
+wiener7_l_shuf: db  6,  7,  6,  7,  6,  7,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15
+                db 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
+
+pq_3:      dq (6 - 4) + 1
+pq_5:      dq (6 - 2) + 1
+pd_65540:  dd (1 << (8 + (6 - 4) + 6)) + (1 << (6 - 4))
+pd_262160: dd (1 << (8 + (6 - 2) + 6)) + (1 << (6 - 2))
+
+pq_11:      dq 12 - (6 - 4) + 1
+pq_9:       dq 12 - (6 - 2) + 1
+nd_1047552: dd (1 << (12 - (6 - 4))) - (1 << (12 + 8))
+nd_1048320: dd (1 << (12 - (6 - 2))) - (1 << (12 + 8))
+
+pb_wiener5_l: times 2 db  2,  3
+pb_wiener5_r: times 2 db -6, -5
+
+pb_wiener7_l: times 2 db  4,  5
+pb_wiener7_m: times 2 db -4, -3
+pb_wiener7_r: times 2 db -8, -7
+
+SECTION .text
+
+INIT_YMM avx2
+cglobal wiener_filter5_h_16bpc, 6, 9, 14, dst, left, src, ss, f, w, h, edge, bdmax
+  movifnidn      wd, wm
+  movifnidn      hd, hm
+  movifnidn   edgeb, edgem
+  vbroadcasti128 m6, [wiener5_shufB]
+  vpbroadcastd  m12, [fq + 2]
+  vbroadcasti128 m7, [wiener5_shufC]
+  vpbroadcastw  m13, [fq + 6]
+  vbroadcasti128 m8, [wiener5_shufD]
+  popcnt     bdmaxd, bdmaxm
+  vpbroadcastd   m9, [pd_65540]
+  movq         xm10, [pq_3]
+  cmp        bdmaxd, 10
+  je .bits10
+  vpbroadcastd   m9, [pd_262160]
+  movq         xm10, [pq_5]
+.bits10:
+  pxor          m11, m11
+  add            wq, wq
+  add          srcq, wq
+  add          dstq, wq
+  neg            wq
+  DEFINE_ARGS dst, left, src, ss, f, w, h, edge, x
+.v_loop:
+  mov            xq, wq
+  test        edgeb, 1 ; LR_HAVE_LEFT
+  jz .h_extend_left
+  test        leftq, leftq
+  jz .h_loop
+  movd          xm4, [leftq + 4]
+  vpblendd       m4, [srcq + xq - 4], 0xfe
+  add         leftq, 8
+  jmp .h_main
+.h_extend_left:
+  vbroadcasti128 m5, [srcq + xq]
+  mova           m4, [srcq + xq]
+  palignr        m4, m5, 12
+  pshufb         m4, [wiener5_l_shuf]
+  jmp .h_main
+.h_loop:
+  movu           m4, [srcq + xq - 4]
+.h_main:
+  movu           m5, [srcq + xq + 4]
+  test        edgeb, 2 ; LR_HAVE_RIGHT
+  jnz .h_have_right
+  cmp            xd, -36
+  jl .h_have_right
+  movd          xm2, xd
+  vpbroadcastd   m0, [pb_wiener5_l]
+  vpbroadcastd   m1, [pb_wiener5_r]
+  vpbroadcastb   m2, xm2
+  movu           m3, [pb_0to31]
+  psubb          m0, m2
+  psubb          m1, m2
+  pminub         m0, m3
+  pminub         m1, m3
+  pshufb         m4, m0
+  pshufb         m5, m1
+.h_have_right:
+  pshufb         m0, m4, m6
+  pshufb         m2, m4, m7
+  paddw          m0, m2
+  pmaddwd        m0, m12
+  pshufb         m1, m5, m6
+  pshufb         m3, m5, m7
+  paddw          m1, m3
+  pmaddwd        m1, m12
+  pshufb         m4, m8
+  pmaddwd        m4, m13
+  pshufb         m5, m8
+  pmaddwd        m5, m13
+  paddd          m0, m4
+  paddd          m1, m5
+  paddd          m0, m9
+  paddd          m1, m9
+  psrad          m0, xm10
+  psrad          m1, xm10
+  packssdw       m0, m1
+  pmaxsw         m0, m11
+  mova  [dstq + xq], m0
+  add            xq, 32
+  jl .h_loop
+  add          srcq, ssq
+  add          dstq, 384*2
+  dec            hd
+  jg .v_loop
+  RET
+
+DECLARE_REG_TMP 8, 9, 10, 11, 12, 13, 14
+
+INIT_YMM avx2
+cglobal wiener_filter5_v_16bpc, 6, 13, 12, dst, ds, mid, f, w, h, edge, bdmax
+  movifnidn    wd, wm
+  movifnidn    hd, hm
+  movifnidn edgeb, edgem
+  pxor         m6, m6
+  vpbroadcastd m7, [fq + 2]
+  vpbroadcastd m8, [fq + 6]
+  popcnt   bdmaxd, bdmaxm
+  vpbroadcastd m9, [nd_1047552]
+  movq       xm10, [pq_11]
+  cmp      bdmaxd, 10
+  je .bits10
+  vpbroadcastd m9, [nd_1048320]
+  movq       xm10, [pq_9]
+.bits10:
+  vpbroadcastw m11, bdmaxm
+  add          wq, wq
+  add        midq, wq
+  add        dstq, wq
+  neg          wq
+  DEFINE_ARGS dst, ds, mid, ms, w, h, edge, x
+  mov         msq, 2*384
+  mov          t0, midq
+  lea          t1, [t0 + msq]
+  lea          t2, [t1 + msq]
+  lea          t3, [t2 + msq]
+  lea          t4, [t3 + msq]
+  test      edgeb, 4 ; LR_HAVE_TOP
+  jnz .have_top
+  mov          t0, t2
+  mov          t1, t2
+.have_top:
+  test      edgeb, 8 ; LR_HAVE_BOTTOM
+  jnz .v_loop
+  cmp          hd, 2
+  jg .v_loop
+  cmp          hd, 1
+  jne .limit_v
+  mov          t3, t2
+.limit_v:
+  mov          t4, t3
+.v_loop:
+  mov          xq, wq
+.h_loop:
+  mova         m1, [t0 + xq]
+  mova         m2, [t1 + xq]
+  mova         m3, [t2 + xq]
+  mova         m4, [t3 + xq]
+  mova         m5, [t4 + xq]
+  punpcklwd    m0, m1, m2
+  pmaddwd      m0, m7
+  punpckhwd    m1, m2
+  pmaddwd      m1, m7
+  punpcklwd    m2, m5, m4
+  pmaddwd      m2, m7
+  punpckhwd    m5, m4
+  pmaddwd      m5, m7
+  paddd        m0, m2
+  paddd        m1, m5
+  punpcklwd    m2, m3, m6
+  pmaddwd      m2, m8
+  punpckhwd    m3, m6
+  pmaddwd      m3, m8
+  paddd        m0, m2
+  paddd        m1, m3
+  paddd        m0, m9
+  paddd        m1, m9
+  psrad        m0, xm10
+  psrad        m1, xm10
+  packusdw     m0, m1
+  pminuw       m0, m11
+  mova [dstq + xq], m0
+  add          xq, 32
+  jl .h_loop
+  add        dstq, dsq
+  mov          t0, t1
+  mov          t1, t2
+  mov          t2, t3
+  mov          t3, t4
+  add          t4, msq
+  test      edgeb, 8 ; LR_HAVE_BOTTOM
+  jnz .have_bottom
+  cmp          hd, 3
+  jg .have_bottom
+  mov          t4, t3
+.have_bottom:
+  dec          hd
+  jg .v_loop
+  RET
+
+INIT_YMM avx2
+cglobal wiener_filter7_h_16bpc, 6, 10, 16, dst, left, src, ss, f, w, h, edge, bdmax, rh
+  movifnidn       wd, wm
+  movifnidn       hd, hm
+  movifnidn    edgeb, edgem
+  vpbroadcastd    m7, [fq]
+  vpbroadcastd    m8, [fq + 4]
+  vbroadcasti128 m10, [rev_w]
+  vbroadcasti128 m11, [wiener5_shufB]
+  vbroadcasti128 m12, [wiener7_shufC]
+  vbroadcasti128 m13, [wiener7_shufD]
+  vbroadcasti128 m14, [wiener7_shufE]
+  vbroadcasti128 m15, [rev_d]
+  popcnt      bdmaxd, bdmaxm
+  vpbroadcastd    m9, [pd_65540]
+  mov            rhq, [pq_3]
+  cmp         bdmaxd, 10
+  je .bits10
+  vpbroadcastd    m9, [pd_262160]
+  mov            rhq, [pq_5]
+.bits10:
+  add             wq, wq
+  add           srcq, wq
+  add           dstq, wq
+  neg             wq
+  DEFINE_ARGS dst, left, src, ss, f, w, h, edge, x, rh
+.v_loop:
+  mov             xq, wq
+  test         edgeb, 1 ; LR_HAVE_LEFT
+  jz .h_extend_left
+  test         leftq, leftq
+  jz .h_loop
+  movq           xm4, [leftq + 2]
+  vpblendw       xm4, [srcq + xq - 6], 0xf8
+  vinserti128     m4, [srcq + xq + 10], 1
+  add          leftq, 8
+  jmp .h_main
+.h_extend_left:
+  vbroadcasti128  m5, [srcq + xq]
+  mova            m4, [srcq + xq]
+  palignr         m4, m5, 10
+  pshufb          m4, [wiener7_l_shuf]
+  jmp .h_main
+.h_loop:
+  movu            m4, [srcq + xq - 6]
+.h_main:
+  movu            m5, [srcq + xq + 2]
+  movu            m6, [srcq + xq + 6]
+  test         edgeb, 2 ; LR_HAVE_RIGHT
+  jnz .h_have_right
+  cmp             xd, -38
+  jl .h_have_right
+  movd           xm3, xd
+  vpbroadcastd    m0, [pb_wiener7_l]
+  vpbroadcastd    m1, [pb_wiener7_m]
+  vpbroadcastd    m2, [pb_wiener7_r]
+  vpbroadcastb    m3, xm3
+  psubb           m0, m3
+  psubb           m1, m3
+  psubb           m2, m3
+  movu            m3, [pb_0to31]
+  pminub          m0, m3
+  pminub          m1, m3
+  pminub          m2, m3
+  pshufb          m4, m0
+  pshufb          m5, m1
+  pshufb          m6, m2
+  cmp             xd, -9*2
+  jne .hack
+  vpbroadcastw   xm3, [srcq + xq + 16]
+  vinserti128     m5, xm3, 1
+  jmp .h_have_right
+.hack:
+  cmp             xd, -1*2
+  jne .h_have_right
+  vpbroadcastw   xm5, [srcq + xq]
+.h_have_right:
+  pshufb          m6, m10
+  pshufb          m0, m4, m11
+  pshufb          m2, m5, m12
+  paddw           m0, m2
+  pmaddwd         m0, m7
+  pshufb          m2, m4, m13
+  pshufb          m4, m14
+  paddw           m2, m4
+  pmaddwd         m2, m8
+  pshufb          m1, m6, m11
+  pshufb          m5, m11
+  pmaddwd         m1, m7
+  pmaddwd         m5, m7
+  pshufb          m3, m6, m13
+  pshufb          m6, m14
+  paddw           m3, m6
+  pmaddwd         m3, m8
+  paddd           m0, m2
+  paddd           m1, m3
+  pshufb          m1, m15
+  paddd           m1, m5
+  movq           xm4, rhq
+  pxor            m5, m5
+  paddd           m0, m9
+  paddd           m1, m9
+  psrad           m0, xm4
+  psrad           m1, xm4
+  packssdw        m0, m1
+  pmaxsw          m0, m5
+  mova   [dstq + xq], m0
+  add             xq, 32
+  jl .h_loop
+  add           srcq, ssq
+  add           dstq, 384*2
+  dec             hd
+  jg .v_loop
+  RET
+
+INIT_YMM avx2
+cglobal wiener_filter7_v_16bpc, 6, 15, 13, dst, ds, mid, f, w, h, edge, bdmax
+  movifnidn     wd, wm
+  movifnidn     hd, hm
+  movifnidn  edgeb, edgem
+  pxor          m6, m6
+  vpbroadcastd  m7, [fq]
+  vpbroadcastw  m8, [fq + 4]
+  vpbroadcastd  m9, [fq + 6]
+  popcnt    bdmaxd, bdmaxm
+  vpbroadcastd m10, [nd_1047552]
+  movq        xm11, [pq_11]
+  cmp       bdmaxd, 10
+  je .bits10
+  vpbroadcastd m10, [nd_1048320]
+  movq        xm11, [pq_9]
+.bits10:
+  vpbroadcastw m12, bdmaxm
+  add           wq, wq
+  add         midq, wq
+  add         dstq, wq
+  neg           wq
+  DEFINE_ARGS dst, ds, mid, ms, w, h, edge, x
+  mov          msq, 2*384
+  mov           t0, midq
+  mov           t1, t0
+  lea           t2, [t1 + msq]
+  lea           t3, [t2 + msq]
+  lea           t4, [t3 + msq]
+  lea           t5, [t4 + msq]
+  lea           t6, [t5 + msq]
+  test       edgeb, 4 ; LR_HAVE_TOP
+  jnz .have_top
+  mov           t0, t3
+  mov           t1, t3
+  mov           t2, t3
+.have_top:
+  cmp           hd, 3
+  jg .v_loop
+  test       edgeb, 8 ; LR_HAVE_BOTTOM
+  jz .no_bottom0
+  cmp           hd, 1
+  jg .v_loop
+  jmp .h3
+.no_bottom0:
+  cmp           hd, 2
+  je .h2
+  jns .h3
+.h1:
+  mov           t4, t3
+.h2:
+  mov           t5, t4
+.h3:
+  mov           t6, t5
+.v_loop:
+  mov           xq, wq
+.h_loop:
+  mova          m1, [t0 + xq]
+  mova          m2, [t1 + xq]
+  mova          m3, [t5 + xq]
+  mova          m4, [t6 + xq]
+  punpcklwd     m0, m1, m2
+  pmaddwd       m0, m7
+  punpckhwd     m1, m2
+  pmaddwd       m1, m7
+  punpcklwd     m2, m4, m3
+  pmaddwd       m2, m7
+  punpckhwd     m4, m3
+  pmaddwd       m4, m7
+  paddd         m0, m2
+  paddd         m1, m4
+  mova          m3, [t2 + xq]
+  mova          m4, [t4 + xq]
+  punpcklwd     m2, m3, m4
+  pmaddwd       m2, m8
+  punpckhwd     m3, m4
+  pmaddwd       m3, m8
+  paddd         m0, m2
+  paddd         m1, m3
+  mova          m3, [t3 + xq]
+  punpcklwd     m2, m3, m6
+  pmaddwd       m2, m9
+  punpckhwd     m3, m6
+  pmaddwd       m3, m9
+  paddd         m0, m2
+  paddd         m1, m3
+  paddd         m0, m10
+  paddd         m1, m10
+  psrad         m0, xm11
+  psrad         m1, xm11
+  packusdw      m0, m1
+  pminuw        m0, m12
+  mova [dstq + xq], m0
+  add           xq, 32
+  jl .h_loop
+  add         dstq, dsq
+  mov           t0, t1
+  mov           t1, t2
+  mov           t2, t3
+  mov           t3, t4
+  mov           t4, t5
+  mov           t5, t6
+  add           t6, msq
+  cmp           hd, 4
+  jg .next_row
+  test       edgeb, 8 ; LR_HAVE_BOTTOM
+  jz .no_bottom
+  cmp           hd, 2
+  jg .next_row
+.no_bottom:
+  mov           t6, t5
+.next_row:
+  dec           hd
+  jg .v_loop
+  RET
+
+%endif ; ARCH_X86_64
-- 
GitLab


From 2ce581302a1536559aa5e56018a03ac6a3770c0f Mon Sep 17 00:00:00 2001
From: "Nathan E. Egge" <unlord@xiph.org>
Date: Wed, 13 Jan 2021 14:54:42 -0500
Subject: [PATCH 3/3] Enable AVX2 wiener filter HBD assembly

---
 src/meson.build                     |  1 +
 src/x86/looprestoration_init_tmpl.c | 40 +++++++++++++++++++++++++++--
 2 files changed, 39 insertions(+), 2 deletions(-)

diff --git a/src/meson.build b/src/meson.build
index 27946501..25729217 100644
--- a/src/meson.build
+++ b/src/meson.build
@@ -211,6 +211,7 @@ if is_asm_enabled
             libdav1d_sources_asm += files(
                 'x86/cdef16_avx2.asm',
                 'x86/cdef16_sse.asm',
+                'x86/looprestoration16_avx2.asm',
                 'x86/mc16_avx2.asm',
             )
         endif
 
diff --git a/src/x86/looprestoration_init_tmpl.c b/src/x86/looprestoration_init_tmpl.c
index a9aa5acf..52de0faf 100644
--- a/src/x86/looprestoration_init_tmpl.c
+++ b/src/x86/looprestoration_init_tmpl.c
@@ -30,9 +30,40 @@
 
 #include "common/intops.h"
 
+#if BITDEPTH != 8
+#define decl_wiener_filter_fn(name, ext) \
+void BF(name##_h, ext)(int16_t *dst, const pixel (*left)[4], const pixel *src, \
+                       ptrdiff_t stride, const int16_t fh[7], const intptr_t w, \
+                       int h, enum LrEdgeFlags edges HIGHBD_DECL_SUFFIX); \
+void BF(name##_v, ext)(pixel *dst, ptrdiff_t stride, const int16_t *mid, \
+                       const int16_t fv[7], int w, int h, \
+                       enum LrEdgeFlags edges HIGHBD_DECL_SUFFIX); \
+static void BF(name, ext)(pixel *const dst, const ptrdiff_t dst_stride, \
+                          const pixel (*const left)[4], \
+                          const pixel *lpf, const ptrdiff_t lpf_stride, \
+                          const int w, const int h, const LooprestorationParams *params, \
+                          const enum LrEdgeFlags edges HIGHBD_DECL_SUFFIX) { \
+    ALIGN_STK_64(int16_t, mid, 68 * 384,); \
+    BF(name##_h, ext)(&mid[2*384], left, dst, dst_stride, params->filter[0], w, h, \
+                      edges HIGHBD_TAIL_SUFFIX); \
+    if (edges & LR_HAVE_TOP) { \
+        BF(name##_h, ext)(mid, NULL, lpf, lpf_stride, params->filter[0], w, 2, \
+                          edges HIGHBD_TAIL_SUFFIX); \
+    } \
+    if (edges & LR_HAVE_BOTTOM) { \
+        BF(name##_h, ext)(&mid[(2 + h)*384], NULL, lpf + 6*PXSTRIDE(lpf_stride), \
+                          lpf_stride, params->filter[0], w, 2, edges HIGHBD_TAIL_SUFFIX); \
+    } \
+    BF(name##_v, ext)(dst, dst_stride, mid, params->filter[1], w, h, edges HIGHBD_TAIL_SUFFIX); \
+}
+#define decl_wiener_filter_fns(ext) \
+decl_wiener_filter_fn(dav1d_wiener_filter7, ext); \
+decl_wiener_filter_fn(dav1d_wiener_filter5, ext)
+#else
 #define decl_wiener_filter_fns(ext) \
 decl_lr_filter_fn(BF(dav1d_wiener_filter7, ext)); \
 decl_lr_filter_fn(BF(dav1d_wiener_filter5, ext))
+#endif
 
 #define decl_sgr_filter_fns(ext) \
 void BF(dav1d_sgr_filter_5x5, ext)(pixel *dst, ptrdiff_t dst_stride, \
@@ -193,11 +224,14 @@ decl_wiener_filter_fns(sse2);
 decl_wiener_filter_fns(ssse3);
 SGR_FILTER_OLD(ssse3)
 # if ARCH_X86_64
-decl_wiener_filter_fns(avx2);
 decl_sgr_filter_fns(avx2)
 # endif
 #endif
 
+#if ARCH_X86_64
+decl_wiener_filter_fns(avx2);
+#endif
+
 COLD void bitfn(dav1d_loop_restoration_dsp_init_x86)(Dav1dLoopRestorationDSPContext *const c) {
     const unsigned flags = dav1d_get_cpu_flags();
 
@@ -217,11 +251,13 @@ COLD void bitfn(dav1d_loop_restoration_dsp_init_x86)(Dav1dLoopRestorationDSPCont
 #endif
 
     if (!(flags & DAV1D_X86_CPU_FLAG_AVX2)) return;
-#if BITDEPTH == 8 && ARCH_X86_64
+#if ARCH_X86_64
     c->wiener[0] = BF(dav1d_wiener_filter7, avx2);
     c->wiener[1] = BF(dav1d_wiener_filter5, avx2);
+# if BITDEPTH == 8
     c->sgr[0] = BF(dav1d_sgr_filter_5x5, avx2);
     c->sgr[1] = BF(dav1d_sgr_filter_3x3, avx2);
     c->sgr[2] = BF(dav1d_sgr_filter_mix, avx2);
+# endif
 #endif
 }
-- 
GitLab