avfilter/unsharp: OpenCL unsharpen filter optimization: substitute N^2 filter computation with 2N+C

i7-4770K luma 21% faster, chroma 18% faster A10-7850K luma 42% faster, chroma 37% faster on 1920x1080 res
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

libavfilter/unsharp.h

@@ -41,6 +41,10 @@ typedef struct {
     cl_kernel kernel_chroma;
     cl_mem cl_luma_mask;
     cl_mem cl_chroma_mask;
+    cl_mem cl_luma_mask_x;
+    cl_mem cl_chroma_mask_x;
+    cl_mem cl_luma_mask_y;
+    cl_mem cl_chroma_mask_y;
     int in_plane_size[8];
     int out_plane_size[8];
     int plane_num;

libavfilter/unsharp_opencl.c

@@ -87,11 +87,12 @@ end:
     return ret;
 }
 
-static int compute_mask_matrix(cl_mem cl_mask_matrix, int step_x, int step_y)
+static int copy_separable_masks(cl_mem cl_mask_x, cl_mem cl_mask_y, int step_x, int step_y)
 {
-    int i, j, ret = 0;
-    uint32_t *mask_matrix, *mask_x, *mask_y;
-    size_t size_matrix = sizeof(uint32_t) * (2 * step_x + 1) * (2 * step_y + 1);
+    int ret = 0;
+    uint32_t *mask_x, *mask_y;
+    size_t size_mask_x = sizeof(uint32_t) * (2 * step_x + 1);
+    size_t size_mask_y = sizeof(uint32_t) * (2 * step_y + 1);
     mask_x = av_mallocz_array(2 * step_x + 1, sizeof(uint32_t));
     if (!mask_x) {
         ret = AVERROR(ENOMEM);
@@ -102,37 +103,36 @@ static int compute_mask_matrix(cl_mem cl_mask_matrix, int step_x, int step_y)
         ret = AVERROR(ENOMEM);
         goto end;
     }
-    mask_matrix = av_mallocz(size_matrix);
-    if (!mask_matrix) {
-        ret = AVERROR(ENOMEM);
-        goto end;
-    }
+
     ret = compute_mask(step_x, mask_x);
     if (ret < 0)
         goto end;
     ret = compute_mask(step_y, mask_y);
     if (ret < 0)
         goto end;
-    for (j = 0; j < 2 * step_y + 1; j++) {
-        for (i = 0; i < 2 * step_x + 1; i++) {
-            mask_matrix[i + j * (2 * step_x + 1)] = mask_y[j] * mask_x[i];
-        }
-    }
-    ret = av_opencl_buffer_write(cl_mask_matrix, (uint8_t *)mask_matrix, size_matrix);
+
+    ret = av_opencl_buffer_write(cl_mask_x, (uint8_t *)mask_x, size_mask_x);
+    ret = av_opencl_buffer_write(cl_mask_y, (uint8_t *)mask_y, size_mask_y);
 end:
     av_freep(&mask_x);
     av_freep(&mask_y);
-    av_freep(&mask_matrix);
+
     return ret;
 }
 
 static int generate_mask(AVFilterContext *ctx)
 {
-    UnsharpContext *unsharp = ctx->priv;
-    int i, ret = 0, step_x[2], step_y[2];
+    cl_mem masks[4];
     cl_mem mask_matrix[2];
+    int i, ret = 0, step_x[2], step_y[2];
+
+    UnsharpContext *unsharp = ctx->priv;
     mask_matrix[0] = unsharp->opencl_ctx.cl_luma_mask;
     mask_matrix[1] = unsharp->opencl_ctx.cl_chroma_mask;
+    masks[0] = unsharp->opencl_ctx.cl_luma_mask_x;
+    masks[1] = unsharp->opencl_ctx.cl_luma_mask_y;
+    masks[2] = unsharp->opencl_ctx.cl_chroma_mask_x;
+    masks[3] = unsharp->opencl_ctx.cl_chroma_mask_y;
     step_x[0] = unsharp->luma.steps_x;
     step_x[1] = unsharp->chroma.steps_x;
     step_y[0] = unsharp->luma.steps_y;
@@ -144,12 +144,16 @@ static int generate_mask(AVFilterContext *ctx)
     else
         unsharp->opencl_ctx.use_fast_kernels = 1;
 
+    if (!masks[0] || !masks[1] || !masks[2] || !masks[3]) {
+        av_log(ctx, AV_LOG_ERROR, "Luma mask and chroma mask should not be NULL\n");
+        return AVERROR(EINVAL);
+    }
     if (!mask_matrix[0] || !mask_matrix[1]) {
         av_log(ctx, AV_LOG_ERROR, "Luma mask and chroma mask should not be NULL\n");
         return AVERROR(EINVAL);
     }
     for (i = 0; i < 2; i++) {
-        ret = compute_mask_matrix(mask_matrix[i], step_x[i], step_y[i]);
+        ret = copy_separable_masks(masks[2*i], masks[2*i+1], step_x[i], step_y[i]);
         if (ret < 0)
             return ret;
     }
@@ -184,7 +188,8 @@ int ff_opencl_apply_unsharp(AVFilterContext *ctx, AVFrame *in, AVFrame *out)
         ret = avpriv_opencl_set_parameter(&kernel1,
                                       FF_OPENCL_PARAM_INFO(unsharp->opencl_ctx.cl_inbuf),
                                       FF_OPENCL_PARAM_INFO(unsharp->opencl_ctx.cl_outbuf),
-                                      FF_OPENCL_PARAM_INFO(unsharp->opencl_ctx.cl_luma_mask),
+                                      FF_OPENCL_PARAM_INFO(unsharp->opencl_ctx.cl_luma_mask_x),
+                                      FF_OPENCL_PARAM_INFO(unsharp->opencl_ctx.cl_luma_mask_y),
                                       FF_OPENCL_PARAM_INFO(unsharp->luma.amount),
                                       FF_OPENCL_PARAM_INFO(unsharp->luma.scalebits),
                                       FF_OPENCL_PARAM_INFO(unsharp->luma.halfscale),
@@ -201,7 +206,8 @@ int ff_opencl_apply_unsharp(AVFilterContext *ctx, AVFrame *in, AVFrame *out)
         ret = avpriv_opencl_set_parameter(&kernel2,
                                       FF_OPENCL_PARAM_INFO(unsharp->opencl_ctx.cl_inbuf),
                                       FF_OPENCL_PARAM_INFO(unsharp->opencl_ctx.cl_outbuf),
-                                      FF_OPENCL_PARAM_INFO(unsharp->opencl_ctx.cl_chroma_mask),
+                                      FF_OPENCL_PARAM_INFO(unsharp->opencl_ctx.cl_chroma_mask_x),
+                                      FF_OPENCL_PARAM_INFO(unsharp->opencl_ctx.cl_chroma_mask_y),
                                       FF_OPENCL_PARAM_INFO(unsharp->chroma.amount),
                                       FF_OPENCL_PARAM_INFO(unsharp->chroma.scalebits),
                                       FF_OPENCL_PARAM_INFO(unsharp->chroma.halfscale),
@@ -264,7 +270,9 @@ int ff_opencl_apply_unsharp(AVFilterContext *ctx, AVFrame *in, AVFrame *out)
             return AVERROR_EXTERNAL;
         }
     }
-    clFinish(unsharp->opencl_ctx.command_queue);
+    //blocking map is suffficient, no need for clFinish
+    //clFinish(unsharp->opencl_ctx.command_queue);
+
     return av_opencl_buffer_read_image(out->data, unsharp->opencl_ctx.out_plane_size,
                                        unsharp->opencl_ctx.plane_num, unsharp->opencl_ctx.cl_outbuf,
                                        unsharp->opencl_ctx.cl_outbuf_size);
@@ -286,6 +294,27 @@ int ff_opencl_unsharp_init(AVFilterContext *ctx)
     ret = av_opencl_buffer_create(&unsharp->opencl_ctx.cl_chroma_mask,
                                   sizeof(uint32_t) * (2 * unsharp->chroma.steps_x + 1) * (2 * unsharp->chroma.steps_y + 1),
                                   CL_MEM_READ_ONLY, NULL);
+    // separable filters
+    if (ret < 0)
+        return ret;
+    ret = av_opencl_buffer_create(&unsharp->opencl_ctx.cl_luma_mask_x,
+                                  sizeof(uint32_t) * (2 * unsharp->luma.steps_x + 1),
+                                  CL_MEM_READ_ONLY, NULL);
+    if (ret < 0)
+        return ret;
+    ret = av_opencl_buffer_create(&unsharp->opencl_ctx.cl_luma_mask_y,
+                                  sizeof(uint32_t) * (2 * unsharp->luma.steps_y + 1),
+                                  CL_MEM_READ_ONLY, NULL);
+    if (ret < 0)
+        return ret;
+    ret = av_opencl_buffer_create(&unsharp->opencl_ctx.cl_chroma_mask_x,
+                                  sizeof(uint32_t) * (2 * unsharp->chroma.steps_x + 1),
+                                  CL_MEM_READ_ONLY, NULL);
+    if (ret < 0)
+        return ret;
+    ret = av_opencl_buffer_create(&unsharp->opencl_ctx.cl_chroma_mask_y,
+                                  sizeof(uint32_t) * (2 * unsharp->chroma.steps_y + 1),
+                                  CL_MEM_READ_ONLY, NULL);
     if (ret < 0)
         return ret;
     ret = generate_mask(ctx);
@@ -339,6 +368,10 @@ void ff_opencl_unsharp_uninit(AVFilterContext *ctx)
     av_opencl_buffer_release(&unsharp->opencl_ctx.cl_outbuf);
     av_opencl_buffer_release(&unsharp->opencl_ctx.cl_luma_mask);
     av_opencl_buffer_release(&unsharp->opencl_ctx.cl_chroma_mask);
+    av_opencl_buffer_release(&unsharp->opencl_ctx.cl_luma_mask_x);
+    av_opencl_buffer_release(&unsharp->opencl_ctx.cl_chroma_mask_x);
+    av_opencl_buffer_release(&unsharp->opencl_ctx.cl_luma_mask_y);
+    av_opencl_buffer_release(&unsharp->opencl_ctx.cl_chroma_mask_y);
     clReleaseKernel(unsharp->opencl_ctx.kernel_default);
     clReleaseKernel(unsharp->opencl_ctx.kernel_luma);
     clReleaseKernel(unsharp->opencl_ctx.kernel_chroma);

libavfilter/unsharp_opencl_kernel.h

@@ -36,7 +36,8 @@ inline unsigned char clip_uint8(int a)
 kernel void unsharp_luma(
                     global unsigned char *src,
                     global unsigned char *dst,
-                    global int *mask,
+                    global int *mask_x,
+                    global int *mask_y,
                     int amount,
                     int scalebits,
                     int halfscale,
@@ -59,10 +60,12 @@ kernel void unsharp_luma(
         return;
     }
 
-    local uchar l[32][32];
-    local int lc[LU_RADIUS_X*LU_RADIUS_Y];
+    local unsigned int l[32][32];
+    local unsigned int lcx[LU_RADIUS_X];
+    local unsigned int lcy[LU_RADIUS_Y];
     int indexIx, indexIy, i, j;
 
+    //load up tile: actual workspace + halo of 8 points in x and y \n
     for(i = 0; i <= 1; i++) {
         indexIy = -8 + (blockIdx.y + i) * 16 + threadIdx.y;
         indexIy = indexIy < 0 ? 0 : indexIy;
@@ -76,27 +79,54 @@ kernel void unsharp_luma(
     }
 
     int indexL = threadIdx.y*16 + threadIdx.x;
-    if (indexL < LU_RADIUS_X*LU_RADIUS_Y)
-        lc[indexL] = mask[indexL];
+    if (indexL < LU_RADIUS_X)
+        lcx[indexL] = mask_x[indexL];
+    if (indexL < LU_RADIUS_Y)
+        lcy[indexL] = mask_y[indexL];
     barrier(CLK_LOCAL_MEM_FENCE);
 
+    //needed for unsharp mask application in the end \n
+    int orig_value = (int)l[threadIdx.y + 8][threadIdx.x + 8];
+
     int idx, idy, maskIndex;
-    int sum = 0;
-    int steps_x = LU_RADIUS_X/2;
-    int steps_y = LU_RADIUS_Y/2;
+    int temp[2] = {0};
+    int steps_x = (LU_RADIUS_X-1)/2;
+    int steps_y = (LU_RADIUS_Y-1)/2;
 
-    \n#pragma unroll\n
-    for (i = -steps_y; i <= steps_y; i++) {
-        idy = 8 + i + threadIdx.y;
-        \n#pragma unroll\n
-        for (j = -steps_x; j <= steps_x; j++) {
-            idx = 8 + j + threadIdx.x;
-            maskIndex = (i + steps_y)*LU_RADIUS_X + j + steps_x;
-            sum += (int)l[idy][idx] * lc[maskIndex];
+    // compute the actual workspace + left&right halos \n
+      \n#pragma unroll\n
+    for (j = 0; j <=1; j++) {
+      //extra work to cover left and right halos \n
+      idx = 16*j + threadIdx.x;
+      \n#pragma unroll\n
+        for (i = -steps_y; i <= steps_y; i++) {
+          idy = 8 + i + threadIdx.y;
+          maskIndex = (i + steps_y);
+          temp[j] += (int)l[idy][idx] * lcy[maskIndex];
         }
     }
-    int temp = (int)l[threadIdx.y + 8][threadIdx.x + 8];
-    int res = temp + (((temp - (int)((sum + halfscale) >> scalebits)) * amount) >> 16);
+    barrier(CLK_LOCAL_MEM_FENCE);
+    //save results from the vertical filter in local memory \n
+    idy = 8 + threadIdx.y;
+      \n#pragma unroll\n
+    for (j = 0; j <=1; j++) {
+      idx = 16*j + threadIdx.x;
+      l[idy][idx] = temp[j];
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    //compute results with the horizontal filter \n
+    int sum = 0;
+    idy = 8 + threadIdx.y;
+    \n#pragma unroll\n
+      for (j = -steps_x; j <= steps_x; j++) {
+        idx = 8 + j + threadIdx.x;
+        maskIndex = j + steps_x;
+        sum += (int)l[idy][idx] * lcx[maskIndex];
+      }
+
+    int res = orig_value + (((orig_value - (int)((sum + halfscale) >> scalebits)) * amount) >> 16);
+
     if (globalIdx.x < width && globalIdx.y < height)
         dst[globalIdx.x + globalIdx.y*dst_stride] = clip_uint8(res);
 }
@@ -104,7 +134,8 @@ kernel void unsharp_luma(
 kernel void unsharp_chroma(
                     global unsigned char *src_y,
                     global unsigned char *dst_y,
-                    global int *mask,
+                    global int *mask_x,
+                    global int *mask_y,
                     int amount,
                     int scalebits,
                     int halfscale,
@@ -141,8 +172,9 @@ kernel void unsharp_chroma(
         return;
     }
 
-    local uchar l[32][32];
-    local int lc[CH_RADIUS_X*CH_RADIUS_Y];
+    local unsigned int l[32][32];
+    local unsigned int lcx[CH_RADIUS_X];
+    local unsigned int lcy[CH_RADIUS_Y];
     int indexIx, indexIy, i, j;
     for(i = 0; i <= 1; i++) {
         indexIy = -8 + (blockIdx.y + i) * 16 + threadIdx.y;
@@ -157,27 +189,51 @@ kernel void unsharp_chroma(
     }
 
     int indexL = threadIdx.y*16 + threadIdx.x;
-    if (indexL < CH_RADIUS_X*CH_RADIUS_Y)
-        lc[indexL] = mask[indexL];
+    if (indexL < CH_RADIUS_X)
+        lcx[indexL] = mask_x[indexL];
+    if (indexL < CH_RADIUS_Y)
+        lcy[indexL] = mask_y[indexL];
     barrier(CLK_LOCAL_MEM_FENCE);
 
+    int orig_value = (int)l[threadIdx.y + 8][threadIdx.x + 8];
+
     int idx, idy, maskIndex;
-    int sum = 0;
     int steps_x = CH_RADIUS_X/2;
     int steps_y = CH_RADIUS_Y/2;
+    int temp[2] = {0,0};
 
     \n#pragma unroll\n
-    for (i = -steps_y; i <= steps_y; i++) {
-        idy = 8 + i + threadIdx.y;
+      for (j = 0; j <= 1; j++) {
+        idx = 16*j + threadIdx.x;
         \n#pragma unroll\n
-        for (j = -steps_x; j <= steps_x; j++) {
-            idx = 8 + j + threadIdx.x;
-            maskIndex = (i + steps_y)*CH_RADIUS_X + j + steps_x;
-            sum += (int)l[idy][idx] * lc[maskIndex];
-        }
+          for (i = -steps_y; i <= steps_y; i++) {
+            idy = 8 + i + threadIdx.y;
+            maskIndex = i + steps_y;
+            temp[j] += (int)l[idy][idx] * lcy[maskIndex];
+          }
+      }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+    idy = 8 + threadIdx.y;
+    \n#pragma unroll\n
+    for (j = 0; j <= 1; j++) {
+      idx = 16*j + threadIdx.x;
+      l[idy][idx] = temp[j];
     }
-    int temp = (int)l[threadIdx.y + 8][threadIdx.x + 8];
-    int res = temp + (((temp - (int)((sum + halfscale) >> scalebits)) * amount) >> 16);
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    //compute results with the horizontal filter \n
+    int sum = 0;
+    idy = 8 + threadIdx.y;
+    \n#pragma unroll\n
+      for (j = -steps_x; j <= steps_x; j++) {
+        idx = 8 + j + threadIdx.x;
+        maskIndex = j + steps_x;
+        sum += (int)l[idy][idx] * lcx[maskIndex];
+      }
+
+    int res = orig_value + (((orig_value - (int)((sum + halfscale) >> scalebits)) * amount) >> 16);
+
     if (globalIdx.x < cw && globalIdx.y < ch)
         dst[globalIdx.x + globalIdx.y*dst_stride_ch] = clip_uint8(res);
 }