[FFmpeg-devel] [PATCH] avfilter/vf_convolution: add 16-column operation for filter_column() to prepare for x86 SIMD.

chen chenm003 at 163.com
Mon Dec 2 09:23:46 EET 2019


I have a little suggest on filter_column16(..) [the function]


Firstly, the function is confused with filter16_column(..)


Secondly, the function's algoritym based on row direction, it means reduced address calculate operators and less cache performance, cost of them may more than calculate cost.


For more clear, I give my toy in here, I verify my patch with cmdline in below


 ./ffmpeg -s 1280*720 -pix_fmt yuv420p -i ~/git/sister_720x1280.yuv -vf convolution="1 2 3 4 5 6 7 8 9:1 2 3 4 5 6 7 8 9:1 2 3 4 5 6 7 8 9:1 2 3 4 5 6 7 8 9:1/45:1/45:1/45:1/45:1:2:3:4:column:column:column:column" -an -vframes 2000 -benchmark -f null /dev/null


The result:
Origin version:   utime=7.359s stime=0.138s rtime=1.664s
Song version:    utime=5.320s stime=0.133s rtime=1.250s
My version:       utime=2.930s stime=0.122s rtime=0.794s


ps: since the function processing up to 16-pixels each time, if we split path into 16 and non-16, I have got 1.934s in here.




My patch based on today head, I have also corrected Song's merge conflict.


************ Patch Start ********************
diff --git a/libavfilter/vf_convolution.c b/libavfilter/vf_convolution.c
index 5909fea..708732a 100644
--- a/libavfilter/vf_convolution.c
+++ b/libavfilter/vf_convolution.c
@@ -521,6 +521,61 @@ static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
             continue;
         }
 
+        if (mode == MATRIX_COLUMN && s->filter[plane] != filter_column){
+            for (y = slice_start; y < slice_end - 16; y+=16) {
+                const int xoff = (y - slice_start) * bpc;
+                const int yoff = radius * stride;
+                for (x = 0; x < radius; x++) {
+                    const int xoff = (y - slice_start) * bpc;
+                    const int yoff = x * stride;
+
+                    s->setup[plane](radius, c, src, stride, x, width, y, height, bpc);
+                    s->filter[plane](dst + yoff + xoff, 1, rdiv,
+                                    bias, matrix, c, 16, radius,
+                                    dstride, stride);
+                }
+                s->setup[plane](radius, c, src, stride, radius, width, y, height, bpc);
+                s->filter[plane](dst + yoff + xoff, sizew - 2 * radius,
+                                rdiv, bias, matrix, c, 16, radius,
+                                dstride, stride);
+                for (x = sizew - radius; x < sizew; x++) {
+                    const int xoff = (y - slice_start) * bpc;
+                    const int yoff = x * stride;
+
+                    s->setup[plane](radius, c, src, stride, x, width, y, height, bpc);
+                    s->filter[plane](dst + yoff + xoff, 1, rdiv,
+                                    bias, matrix, c, 16, radius,
+                                    dstride, stride);
+                }
+            }
+            if (y < slice_end){
+                const int xoff = (y - slice_start) * bpc;
+                const int yoff = radius * stride;
+                for (x = 0; x < radius; x++) {
+                    const int xoff = (y - slice_start) * bpc;
+                    const int yoff = x * stride;
+
+                    s->setup[plane](radius, c, src, stride, x, width, y, height, bpc);
+                    s->filter[plane](dst + yoff + xoff, 1, rdiv,
+                                    bias, matrix, c, slice_end - y, radius,
+                                    dstride, stride);
+                }
+                s->setup[plane](radius, c, src, stride, radius, width, y, height, bpc);
+                s->filter[plane](dst + yoff + xoff, sizew - 2 * radius,
+                                rdiv, bias, matrix, c, slice_end - y, radius,
+                                dstride, stride);
+                for (x = sizew - radius; x < sizew; x++) {
+                    const int xoff = (y - slice_start) * bpc;
+                    const int yoff = x * stride;
+
+                    s->setup[plane](radius, c, src, stride, x, width, y, height, bpc);
+                    s->filter[plane](dst + yoff + xoff, 1, rdiv,
+                                    bias, matrix, c, slice_end - y, radius,
+                                    dstride, stride);
+                }
+            }
+        }
+        else {
         for (y = slice_start; y < slice_end; y++) {
             const int xoff = mode == MATRIX_COLUMN ? (y - slice_start) * bpc : radius * bpc;
             const int yoff = mode == MATRIX_COLUMN ? radius * stride : 0;
@@ -551,6 +606,7 @@ static int filter_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
                 dst += dstride;
         }
     }
+    }
 
     return 0;
 }
diff --git a/libavfilter/x86/vf_convolution_init.c b/libavfilter/x86/vf_convolution_init.c
index 5143240..fcc9ae8 100644
--- a/libavfilter/x86/vf_convolution_init.c
+++ b/libavfilter/x86/vf_convolution_init.c
@@ -29,6 +29,56 @@ void ff_filter_3x3_sse4(uint8_t *dst, int width,
                         const uint8_t *c[], int peak, int radius,
                         int dstride, int stride);
 
+static void filter_column16(uint8_t *dst, int height,
+                          float rdiv, float bias, const int *const matrix,
+                          const uint8_t *c[], int length, int radius,
+                          int dstride, int stride)
+{
+    int y, off16;
+
+#if 1
+    #define __assume(cond) do { if (!(cond)) __builtin_unreachable(); } while (0)
+    assert(length <= 16);
+    __assume(length <= 16);
+    // NOTE: alignment to 64-bytes, so 16 of int can be fill into full of a cache line
+    int __attribute__ ((aligned(64))) sum[16];
+    for (y = 0; y < height; y++) {
+        int i;
+        memset(sum, 0, sizeof(sum));
+
+        for (i = 0; i < 2 * radius + 1; i++) {
+            for (off16 = 0; off16 < length; off16++){
+                sum[off16] += c[i][0 + y * stride + off16] * matrix[i];
+            }
+        }
+
+        for (off16 = 0; off16 < length; off16++){
+            sum[off16] = (int)(sum[off16] * rdiv + bias + 0.5f);
+            dst[off16] = av_clip_uint8(sum[off16]);
+        }
+        dst += dstride;
+
+    }
+    #undef __assume
+
+#else
+
+    assert(length <= 16);
+    for (y = 0; y < height; y++) {
+        for (off16 = 0; off16 < length; off16++){
+            int i, sum = 0;
+
+            for (i = 0; i < 2 * radius + 1; i++)
+                sum += c[i][0 + y * stride + off16] * matrix[i];
+
+            sum = (int)(sum * rdiv + bias + 0.5f);
+            dst[off16] = av_clip_uint8(sum);
+        }
+        dst += dstride;
+    }
+#endif
+}
+
 av_cold void ff_convolution_init_x86(ConvolutionContext *s)
 {
 #if ARCH_X86_64
@@ -41,6 +91,8 @@ av_cold void ff_convolution_init_x86(ConvolutionContext *s)
                     s->filter[i] = ff_filter_3x3_sse4;
             }
         }
+        if (s->mode[i] == MATRIX_COLUMN)
+            s->filter[i] = filter_column16;
     }
 #endif
 }
************ End ********************




At 2019-12-02 14:38:04, "Song, Ruiling" <ruiling.song at intel.com> wrote:
>> -----Original Message-----
>> From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of
>> xujunzz at sjtu.edu.cn
>> Sent: Wednesday, November 27, 2019 10:56 PM
>> To: ffmpeg-devel at ffmpeg.org
>> Cc: xujunzz at sjtu.edu.cn
>> Subject: [FFmpeg-devel] [PATCH] avfilter/vf_convolution: add 16-column
>> operation for filter_column() to prepare for x86 SIMD.
>> 
>> From: Xu Jun <xujunzz at sjtu.edu.cn>
>> 
>> In order to add x86 SIMD for filter_column(), I write a C function which
>> processes 16 columns at a time.
>> 
>> Signed-off-by: Xu Jun <xujunzz at sjtu.edu.cn>
>> ---
>>  libavfilter/vf_convolution.c          | 56 +++++++++++++++++++++++++++
>>  libavfilter/x86/vf_convolution_init.c | 23 +++++++++++
>>  2 files changed, 79 insertions(+)
>> 
>> diff --git a/libavfilter/vf_convolution.c b/libavfilter/vf_convolution.c
>> index d022f1a04a..5291415d48 100644
>> --- a/libavfilter/vf_convolution.c
>> +++ b/libavfilter/vf_convolution.c
>> @@ -520,6 +520,61 @@ static int filter_slice(AVFilterContext *ctx, void *arg,
>> int jobnr, int nb_jobs)
>>              continue;
>>          }
>> 
>> +        if (mode == MATRIX_COLUMN && s->filter[plane] != filter_column){
>> +            for (y = slice_start; y < slice_end - 16; y+=16) {
>Please take care of the coding style there should be white-space between variables and operators.
>And also I think this piece of change make it harder to maintain, let's try to avoid code duplicate as much as we can.
>> +                const int xoff = (y - slice_start) * bpc;
>> +                const int yoff = radius * stride;
>> +                for (x = 0; x < radius; x++) {
>> +                    const int xoff = (y - slice_start) * bpc;
>> +                    const int yoff = x * stride;
>> +
>> +                    s->setup[plane](radius, c, src, stride, x, width, y, height, bpc);
>> +                    s->filter[plane](dst + yoff + xoff, 1, rdiv,
>> +                                    bias, matrix, c, 16, radius,
>> +                                    dstride, stride);
>> +                }
>> +                s->setup[plane](radius, c, src, stride, radius, width, y, height, bpc);
>> +                s->filter[plane](dst + yoff + xoff, sizew - 2 * radius,
>> +                                rdiv, bias, matrix, c, 16, radius,
>> +                                dstride, stride);
>> +                for (x = sizew - radius; x < sizew; x++) {
>> +                    const int xoff = (y - slice_start) * bpc;
>> +                    const int yoff = x * stride;
>> +
>> +                    s->setup[plane](radius, c, src, stride, x, width, y, height, bpc);
>> +                    s->filter[plane](dst + yoff + xoff, 1, rdiv,
>> +                                    bias, matrix, c, 16, radius,
>> +                                    dstride, stride);
>> +                }
>> +            }
>> +            if (y < slice_end){
>> +                const int xoff = (y - slice_start) * bpc;
>> +                const int yoff = radius * stride;
>> +                for (x = 0; x < radius; x++) {
>> +                    const int xoff = (y - slice_start) * bpc;
>> +                    const int yoff = x * stride;
>> +
>> +                    s->setup[plane](radius, c, src, stride, x, width, y, height, bpc);
>> +                    s->filter[plane](dst + yoff + xoff, 1, rdiv,
>> +                                    bias, matrix, c, slice_end - y, radius,
>> +                                    dstride, stride);
>> +                }
>> +                s->setup[plane](radius, c, src, stride, radius, width, y, height, bpc);
>> +                s->filter[plane](dst + yoff + xoff, sizew - 2 * radius,
>> +                                rdiv, bias, matrix, c, slice_end - y, radius,
>> +                                dstride, stride);
>> +                for (x = sizew - radius; x < sizew; x++) {
>> +                    const int xoff = (y - slice_start) * bpc;
>> +                    const int yoff = x * stride;
>> +
>> +                    s->setup[plane](radius, c, src, stride, x, width, y, height, bpc);
>> +                    s->filter[plane](dst + yoff + xoff, 1, rdiv,
>> +                                    bias, matrix, c, slice_end - y, radius,
>> +                                    dstride, stride);
>> +                }
>> +            }
>> +        }
>> +        else {
>>          for (y = slice_start; y < slice_end; y++) {
>>              const int xoff = mode == MATRIX_COLUMN ? (y - slice_start) * bpc :
>> radius * bpc;
>>              const int yoff = mode == MATRIX_COLUMN ? radius * stride : 0;
>> @@ -550,6 +605,7 @@ static int filter_slice(AVFilterContext *ctx, void *arg,
>> int jobnr, int nb_jobs)
>>                  dst += dstride;
>>          }
>>      }
>> +    }
>> 
>>      return 0;
>>  }
>> diff --git a/libavfilter/x86/vf_convolution_init.c
>> b/libavfilter/x86/vf_convolution_init.c
>> index d1e8c90ceb..6b1c2f0e9f 100644
>> --- a/libavfilter/x86/vf_convolution_init.c
>> +++ b/libavfilter/x86/vf_convolution_init.c
>> @@ -34,6 +34,27 @@ void ff_filter_row_sse4(uint8_t *dst, int width,
>>                          const uint8_t *c[], int peak, int radius,
>>                          int dstride, int stride);
>> 
>This C code should not be in the x86-specific file.
>
>Ruiling
>> +static void filter_column16(uint8_t *dst, int height,
>> +                          float rdiv, float bias, const int *const matrix,
>> +                          const uint8_t *c[], int length, int radius,
>> +                          int dstride, int stride)
>> +{
>> +    int y, off16;
>> +
>> +    for (y = 0; y < height; y++) {
>> +        for (off16 = 0; off16 < length; off16++){
>> +            int i, sum = 0;
>> +
>> +            for (i = 0; i < 2 * radius + 1; i++)
>> +                sum += c[i][0 + y * stride + off16] * matrix[i];
>> +
>> +            sum = (int)(sum * rdiv + bias + 0.5f);
>> +            dst[off16] = av_clip_uint8(sum);
>> +        }
>> +        dst += dstride;
>> +    }
>> +
>> +}
>> 
>>  av_cold void ff_convolution_init_x86(ConvolutionContext *s)
>>  {
>> @@ -51,6 +72,8 @@ av_cold void
>> ff_convolution_init_x86(ConvolutionContext *s)
>>                  if (EXTERNAL_SSE4(cpu_flags))
>>                      s->filter[i] = ff_filter_row_sse4;
>>          }
>> +        if (s->mode[i] == MATRIX_COLUMN)
>> +            s->filter[i] = filter_column16;
>>      }
>>  #endif
>>  }
>> --
>> 2.17.1
>> 



More information about the ffmpeg-devel mailing list