lavfi/deshake_opencl: optimze transform filter

Reviewed-by: Wei Gao <highgod0401@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

libavfilter/deshake.h

@@ -1,5 +1,6 @@
 /*
  * Copyright (C) 2013 Wei Gao <weigao@multicorewareinc.com>
+ * Copyright (C) 2013 Lenny Wang
  *
  * This file is part of FFmpeg.
  *
@@ -57,12 +58,8 @@ typedef struct {
 typedef struct {
     cl_command_queue command_queue;
     cl_program program;
-    cl_kernel kernel;
-    size_t matrix_size;
-    float matrix_y[9];
-    float matrix_uv[9];
-    cl_mem cl_matrix_y;
-    cl_mem cl_matrix_uv;
+    cl_kernel kernel_luma;
+    cl_kernel kernel_chroma;
     int in_plane_size[8];
     int out_plane_size[8];
     int plane_num;

libavfilter/deshake_opencl.c

@@ -1,5 +1,6 @@
 /*
  * Copyright (C) 2013 Wei Gao <weigao@multicorewareinc.com>
+ * Copyright (C) 2013 Lenny Wang
  *
  * This file is part of FFmpeg.
  *
@@ -29,8 +30,8 @@
 #include "deshake_opencl.h"
 #include "libavutil/opencl_internal.h"
 
-#define MATRIX_SIZE 6
 #define PLANE_NUM 3
+#define ROUND_TO_16(a) ((((a - 1)/16)+1)*16)
 
 int ff_opencl_transform(AVFilterContext *ctx,
                         int width, int height, int cw, int ch,
@@ -39,29 +40,40 @@ int ff_opencl_transform(AVFilterContext *ctx,
                         enum FillMethod fill, AVFrame *in, AVFrame *out)
 {
     int ret = 0;
-    const size_t global_work_size = width * height + 2 * ch * cw;
     cl_int status;
     DeshakeContext *deshake = ctx->priv;
-    FFOpenclParam opencl_param = {0};
-
-    opencl_param.ctx = ctx;
-    opencl_param.kernel = deshake->opencl_ctx.kernel;
-    ret = av_opencl_buffer_write(deshake->opencl_ctx.cl_matrix_y, (uint8_t *)matrix_y, deshake->opencl_ctx.matrix_size * sizeof(cl_float));
-    if (ret < 0)
-        return ret;
-    ret = av_opencl_buffer_write(deshake->opencl_ctx.cl_matrix_uv, (uint8_t *)matrix_uv, deshake->opencl_ctx.matrix_size * sizeof(cl_float));
-    if (ret < 0)
-        return ret;
+    float4 packed_matrix_lu = {matrix_y[0], matrix_y[1], matrix_y[2], matrix_y[5]};
+    float4 packed_matrix_ch = {matrix_uv[0], matrix_uv[1], matrix_uv[2], matrix_uv[5]};
+    size_t global_worksize_lu[2] = {(size_t)ROUND_TO_16(width), (size_t)ROUND_TO_16(height)};
+    size_t global_worksize_ch[2] = {(size_t)ROUND_TO_16(cw), (size_t)(2*ROUND_TO_16(ch))};
+    size_t local_worksize[2] = {16, 16};
+    FFOpenclParam param_lu = {0};
+    FFOpenclParam param_ch = {0};
+    param_lu.ctx = param_ch.ctx = ctx;
+    param_lu.kernel = deshake->opencl_ctx.kernel_luma;
+    param_ch.kernel = deshake->opencl_ctx.kernel_chroma;
 
     if ((unsigned int)interpolate > INTERPOLATE_BIQUADRATIC) {
         av_log(ctx, AV_LOG_ERROR, "Selected interpolate method is invalid\n");
         return AVERROR(EINVAL);
     }
-    ret = ff_opencl_set_parameter(&opencl_param,
+    ret = ff_opencl_set_parameter(&param_lu,
                                   FF_OPENCL_PARAM_INFO(deshake->opencl_ctx.cl_inbuf),
                                   FF_OPENCL_PARAM_INFO(deshake->opencl_ctx.cl_outbuf),
-                                  FF_OPENCL_PARAM_INFO(deshake->opencl_ctx.cl_matrix_y),
-                                  FF_OPENCL_PARAM_INFO(deshake->opencl_ctx.cl_matrix_uv),
+                                  FF_OPENCL_PARAM_INFO(packed_matrix_lu),
+                                  FF_OPENCL_PARAM_INFO(interpolate),
+                                  FF_OPENCL_PARAM_INFO(fill),
+                                  FF_OPENCL_PARAM_INFO(in->linesize[0]),
+                                  FF_OPENCL_PARAM_INFO(out->linesize[0]),
+                                  FF_OPENCL_PARAM_INFO(height),
+                                  FF_OPENCL_PARAM_INFO(width),
+                                  NULL);
+    if (ret < 0)
+        return ret;
+    ret = ff_opencl_set_parameter(&param_ch,
+                                  FF_OPENCL_PARAM_INFO(deshake->opencl_ctx.cl_inbuf),
+                                  FF_OPENCL_PARAM_INFO(deshake->opencl_ctx.cl_outbuf),
+                                  FF_OPENCL_PARAM_INFO(packed_matrix_ch),
                                   FF_OPENCL_PARAM_INFO(interpolate),
                                   FF_OPENCL_PARAM_INFO(fill),
                                   FF_OPENCL_PARAM_INFO(in->linesize[0]),
@@ -76,13 +88,15 @@ int ff_opencl_transform(AVFilterContext *ctx,
     if (ret < 0)
         return ret;
     status = clEnqueueNDRangeKernel(deshake->opencl_ctx.command_queue,
-                                    deshake->opencl_ctx.kernel, 1, NULL,
-                                    &global_work_size, NULL, 0, NULL, NULL);
+                                    deshake->opencl_ctx.kernel_luma, 2, NULL,
+                                    global_worksize_lu, local_worksize, 0, NULL, NULL);
+    status |= clEnqueueNDRangeKernel(deshake->opencl_ctx.command_queue,
+                                    deshake->opencl_ctx.kernel_chroma, 2, NULL,
+                                    global_worksize_ch, local_worksize, 0, NULL, NULL);
     if (status != CL_SUCCESS) {
         av_log(ctx, AV_LOG_ERROR, "OpenCL run kernel error occurred: %s\n", av_opencl_errstr(status));
         return AVERROR_EXTERNAL;
     }
-    clFinish(deshake->opencl_ctx.command_queue);
     ret = av_opencl_buffer_read_image(out->data, deshake->opencl_ctx.out_plane_size,
                                       deshake->opencl_ctx.plane_num, deshake->opencl_ctx.cl_outbuf,
                                       deshake->opencl_ctx.cl_outbuf_size);
@@ -98,16 +112,7 @@ int ff_opencl_deshake_init(AVFilterContext *ctx)
     ret = av_opencl_init(NULL);
     if (ret < 0)
         return ret;
-    deshake->opencl_ctx.matrix_size = MATRIX_SIZE;
-    deshake->opencl_ctx.plane_num   = PLANE_NUM;
-    ret = av_opencl_buffer_create(&deshake->opencl_ctx.cl_matrix_y,
-        deshake->opencl_ctx.matrix_size*sizeof(cl_float), CL_MEM_READ_ONLY, NULL);
-    if (ret < 0)
-        return ret;
-    ret = av_opencl_buffer_create(&deshake->opencl_ctx.cl_matrix_uv,
-        deshake->opencl_ctx.matrix_size*sizeof(cl_float), CL_MEM_READ_ONLY, NULL);
-    if (ret < 0)
-        return ret;
+    deshake->opencl_ctx.plane_num = PLANE_NUM;
     deshake->opencl_ctx.command_queue = av_opencl_get_command_queue();
     if (!deshake->opencl_ctx.command_queue) {
         av_log(ctx, AV_LOG_ERROR, "Unable to get OpenCL command queue in filter 'deshake'\n");
@@ -118,10 +123,19 @@ int ff_opencl_deshake_init(AVFilterContext *ctx)
         av_log(ctx, AV_LOG_ERROR, "OpenCL failed to compile program 'avfilter_transform'\n");
         return AVERROR(EINVAL);
     }
-    if (!deshake->opencl_ctx.kernel) {
-        deshake->opencl_ctx.kernel = clCreateKernel(deshake->opencl_ctx.program, "avfilter_transform", &ret);
+    if (!deshake->opencl_ctx.kernel_luma) {
+        deshake->opencl_ctx.kernel_luma = clCreateKernel(deshake->opencl_ctx.program,
+                                                         "avfilter_transform_luma", &ret);
         if (ret != CL_SUCCESS) {
-            av_log(ctx, AV_LOG_ERROR, "OpenCL failed to create kernel 'avfilter_transform'\n");
+            av_log(ctx, AV_LOG_ERROR, "OpenCL failed to create kernel 'avfilter_transform_luma'\n");
+            return AVERROR(EINVAL);
+        }
+    }
+    if (!deshake->opencl_ctx.kernel_chroma) {
+        deshake->opencl_ctx.kernel_chroma = clCreateKernel(deshake->opencl_ctx.program,
+                                                           "avfilter_transform_chroma", &ret);
+        if (ret != CL_SUCCESS) {
+            av_log(ctx, AV_LOG_ERROR, "OpenCL failed to create kernel 'avfilter_transform_chroma'\n");
             return AVERROR(EINVAL);
         }
     }
@@ -133,9 +147,8 @@ void ff_opencl_deshake_uninit(AVFilterContext *ctx)
     DeshakeContext *deshake = ctx->priv;
     av_opencl_buffer_release(&deshake->opencl_ctx.cl_inbuf);
     av_opencl_buffer_release(&deshake->opencl_ctx.cl_outbuf);
-    av_opencl_buffer_release(&deshake->opencl_ctx.cl_matrix_y);
-    av_opencl_buffer_release(&deshake->opencl_ctx.cl_matrix_uv);
-    clReleaseKernel(deshake->opencl_ctx.kernel);
+    clReleaseKernel(deshake->opencl_ctx.kernel_luma);
+    clReleaseKernel(deshake->opencl_ctx.kernel_chroma);
     clReleaseProgram(deshake->opencl_ctx.program);
     deshake->opencl_ctx.command_queue = NULL;
     av_opencl_uninit();

libavfilter/deshake_opencl.h

@@ -23,6 +23,13 @@
 
 #include "deshake.h"
 
+typedef struct {
+    float x;
+    float y;
+    float z;
+    float w;
+} float4;
+
 int ff_opencl_deshake_init(AVFilterContext *ctx);
 
 void ff_opencl_deshake_uninit(AVFilterContext *ctx);

libavfilter/deshake_opencl_kernel.h

@@ -1,5 +1,6 @@
 /*
  * Copyright (C) 2013 Wei Gao <weigao@multicorewareinc.com>
+ * Copyright (C) 2013 Lenny Wang
  *
  *
  * This file is part of FFmpeg.
@@ -25,16 +26,16 @@
 #include "libavutil/opencl.h"
 
 const char *ff_kernel_deshake_opencl = AV_OPENCL_KERNEL(
-
-inline unsigned char pixel(global const unsigned char *src, float x, float y,
+inline unsigned char pixel(global const unsigned char *src, int x, int y,
                            int w, int h,int stride, unsigned char def)
 {
-    return (x < 0 || y < 0 || x >= w || y >= h) ? def : src[(int)x + (int)y * stride];
+    return (x < 0 || y < 0 || x >= w || y >= h) ? def : src[x + y * stride];
 }
+
 unsigned char interpolate_nearest(float x, float y, global const unsigned char *src,
                                   int width, int height, int stride, unsigned char def)
 {
-    return pixel(src, (int)(x + 0.5), (int)(y + 0.5), width, height, stride, def);
+    return pixel(src, (int)(x + 0.5f), (int)(y + 0.5f), width, height, stride, def);
 }
 
 unsigned char interpolate_bilinear(float x, float y, global const unsigned char *src,
@@ -42,21 +43,18 @@ unsigned char interpolate_bilinear(float x, float y, global const unsigned char
 {
     int x_c, x_f, y_c, y_f;
     int v1, v2, v3, v4;
+    x_f = (int)x;
+    y_f = (int)y;
+    x_c = x_f + 1;
+    y_c = y_f + 1;
 
-    if (x < -1 || x > width || y < -1 || y > height) {
+    if (x_f < -1 || x_f > width || y_f < -1 || y_f > height) {
         return def;
     } else {
-        x_f = (int)x;
-        x_c = x_f + 1;
-
-        y_f = (int)y;
-        y_c = y_f + 1;
-
-        v1 = pixel(src, x_c, y_c, width, height, stride, def);
+        v4 = pixel(src, x_f, y_f, width, height, stride, def);
         v2 = pixel(src, x_c, y_f, width, height, stride, def);
         v3 = pixel(src, x_f, y_c, width, height, stride, def);
-        v4 = pixel(src, x_f, y_f, width, height, stride, def);
-
+        v1 = pixel(src, x_c, y_c, width, height, stride, def);
         return (v1*(x - x_f)*(y - y_f) + v2*((x - x_f)*(y_c - y)) +
                 v3*(x_c - x)*(y - y_f) + v4*((x_c - x)*(y_c - y)));
     }
@@ -68,19 +66,18 @@ unsigned char interpolate_biquadratic(float x, float y, global const unsigned ch
     int     x_c, x_f, y_c, y_f;
     unsigned char v1,  v2,  v3,  v4;
     float   f1,  f2,  f3,  f4;
+    x_f = (int)x;
+    y_f = (int)y;
+    x_c = x_f + 1;
+    y_c = y_f + 1;
 
-    if (x < - 1 || x > width || y < -1 || y > height)
+    if (x_f < - 1 || x_f > width || y_f < -1 || y_f > height)
         return def;
     else {
-        x_f = (int)x;
-        x_c = x_f + 1;
-        y_f = (int)y;
-        y_c = y_f + 1;
-
-        v1 = pixel(src, x_c, y_c, width, height, stride, def);
+        v4 = pixel(src, x_f, y_f, width, height, stride, def);
         v2 = pixel(src, x_c, y_f, width, height, stride, def);
         v3 = pixel(src, x_f, y_c, width, height, stride, def);
-        v4 = pixel(src, x_f, y_f, width, height, stride, def);
+        v1 = pixel(src, x_c, y_c, width, height, stride, def);
 
         f1 = 1 - sqrt((x_c - x) * (y_c - y));
         f2 = 1 - sqrt((x_c - x) * (y - y_f));
@@ -107,109 +104,120 @@ inline int mirror(int v, int m)
     return v;
 }
 
-kernel void avfilter_transform(global  unsigned char *src,
-                               global  unsigned char *dst,
-                               global          float *matrix,
-                               global          float *matrix2,
-                                                 int interpolate,
-                                                 int fillmethod,
-                                                 int src_stride_lu,
-                                                 int dst_stride_lu,
-                                                 int src_stride_ch,
-                                                 int dst_stride_ch,
-                                                 int height,
-                                                 int width,
-                                                 int ch,
-                                                 int cw)
+kernel void avfilter_transform_luma(global unsigned char *src,
+                                    global unsigned char *dst,
+                                    float4 matrix,
+                                    int interpolate,
+                                    int fill,
+                                    int src_stride_lu,
+                                    int dst_stride_lu,
+                                    int height,
+                                    int width)
 {
-     int global_id = get_global_id(0);
+    int x = get_global_id(0);
+    int y = get_global_id(1);
+    int idx_dst = y * dst_stride_lu + x;
+    unsigned char def = 0;
+    float x_s = x * matrix.x + y * matrix.y + matrix.z;
+    float y_s = x * (-matrix.y) + y * matrix.x + matrix.w;
 
-     global unsigned char *dst_y = dst;
-     global unsigned char *dst_u = dst_y + height * dst_stride_lu;
-     global unsigned char *dst_v = dst_u + ch * dst_stride_ch;
-
-     global unsigned char *src_y = src;
-     global unsigned char *src_u = src_y + height * src_stride_lu;
-     global unsigned char *src_v = src_u + ch * src_stride_ch;
-
-     global unsigned char *tempdst;
-     global unsigned char *tempsrc;
-
-     int x;
-     int y;
-     float x_s;
-     float y_s;
-     int tempsrc_stride;
-     int tempdst_stride;
-     int temp_height;
-     int temp_width;
-     int curpos;
-     unsigned char def = 0;
-     if (global_id < width*height) {
-        y = global_id/width;
-        x = global_id%width;
-        x_s = x * matrix[0] + y * matrix[1] + matrix[2];
-        y_s = x * matrix[3] + y * matrix[4] + matrix[5];
-        tempdst = dst_y;
-        tempsrc = src_y;
-        tempsrc_stride = src_stride_lu;
-        tempdst_stride = dst_stride_lu;
-        temp_height = height;
-        temp_width = width;
-     } else if ((global_id >= width*height)&&(global_id < width*height + ch*cw)) {
-        y = (global_id - width*height)/cw;
-        x = (global_id - width*height)%cw;
-        x_s = x * matrix2[0] + y * matrix2[1] + matrix2[2];
-        y_s = x * matrix2[3] + y * matrix2[4] + matrix2[5];
-        tempdst = dst_u;
-        tempsrc = src_u;
-        tempsrc_stride = src_stride_ch;
-        tempdst_stride = dst_stride_ch;
-        temp_height = ch;
-        temp_width = cw;
-     } else {
-        y = (global_id - width*height - ch*cw)/cw;
-        x = (global_id - width*height - ch*cw)%cw;
-        x_s = x * matrix2[0] + y * matrix2[1] + matrix2[2];
-        y_s = x * matrix2[3] + y * matrix2[4] + matrix2[5];
-        tempdst = dst_v;
-        tempsrc = src_v;
-        tempsrc_stride = src_stride_ch;
-        tempdst_stride = dst_stride_ch;
-        temp_height = ch;
-        temp_width = cw;
-     }
-     curpos = y * tempdst_stride + x;
-     switch (fillmethod) {
-        case 0: //FILL_BLANK
-            def = 0;
-            break;
-        case 1: //FILL_ORIGINAL
-            def = tempsrc[y*tempsrc_stride+x];
-            break;
-        case 2: //FILL_CLAMP
-            y_s = clipf(y_s, 0, temp_height - 1);
-            x_s = clipf(x_s, 0, temp_width - 1);
-            def = tempsrc[(int)y_s * tempsrc_stride + (int)x_s];
-            break;
-        case 3: //FILL_MIRROR
-            y_s = mirror(y_s,temp_height - 1);
-            x_s = mirror(x_s,temp_width - 1);
-            def = tempsrc[(int)y_s * tempsrc_stride + (int)x_s];
-            break;
+    if (x < width && y < height) {
+        switch (fill) {
+            case 0: //FILL_BLANK
+                def = 0;
+                break;
+            case 1: //FILL_ORIGINAL
+                def = src[y*src_stride_lu + x];
+                break;
+            case 2: //FILL_CLAMP
+                y_s = clipf(y_s, 0, height - 1);
+                x_s = clipf(x_s, 0, width - 1);
+                def = src[(int)y_s * src_stride_lu + (int)x_s];
+                break;
+            case 3: //FILL_MIRROR
+                y_s = mirror(y_s, height - 1);
+                x_s = mirror(x_s, width - 1);
+                def = src[(int)y_s * src_stride_lu + (int)x_s];
+                break;
+        }
+        switch (interpolate) {
+            case 0: //INTERPOLATE_NEAREST
+                dst[idx_dst] = interpolate_nearest(x_s, y_s, src, width, height, src_stride_lu, def);
+                break;
+            case 1: //INTERPOLATE_BILINEAR
+                dst[idx_dst] = interpolate_bilinear(x_s, y_s, src, width, height, src_stride_lu, def);
+                break;
+            case 2: //INTERPOLATE_BIQUADRATIC
+                dst[idx_dst] = interpolate_biquadratic(x_s, y_s, src, width, height, src_stride_lu, def);
+                break;
+            default:
+                return;
+        }
     }
-    switch (interpolate) {
-        case 0: //INTERPOLATE_NEAREST
-            tempdst[curpos] = interpolate_nearest(x_s, y_s, tempsrc, temp_width, temp_height, tempsrc_stride, def);
-            break;
-        case 1: //INTERPOLATE_BILINEAR
-            tempdst[curpos] = interpolate_bilinear(x_s, y_s, tempsrc, temp_width, temp_height, tempsrc_stride, def);
-            break;
-        case 2: //INTERPOLATE_BIQUADRATIC
-            tempdst[curpos] = interpolate_biquadratic(x_s, y_s, tempsrc, temp_width, temp_height, tempsrc_stride, def);
-            break;
-        default:
-            return;
+}
+
+kernel void avfilter_transform_chroma(global unsigned char *src,
+                                      global unsigned char *dst,
+                                      float4 matrix,
+                                      int interpolate,
+                                      int fill,
+                                      int src_stride_lu,
+                                      int dst_stride_lu,
+                                      int src_stride_ch,
+                                      int dst_stride_ch,
+                                      int height,
+                                      int width,
+                                      int ch,
+                                      int cw)
+{
+
+    int x = get_global_id(0);
+    int y = get_global_id(1);
+    int pad_ch = get_global_size(1)>>1;
+    global unsigned char *dst_u = dst + height * dst_stride_lu;
+    global unsigned char *src_u = src + height * src_stride_lu;
+    global unsigned char *dst_v = dst_u + ch * dst_stride_ch;
+    global unsigned char *src_v = src_u + ch * src_stride_ch;
+    src = y < pad_ch ? src_u : src_v;
+    dst = y < pad_ch ? dst_u : dst_v;
+    y = select(y - pad_ch, y, y < pad_ch);
+    float x_s = x * matrix.x + y * matrix.y + matrix.z;
+    float y_s = x * (-matrix.y) + y * matrix.x + matrix.w;
+    int idx_dst = y * dst_stride_ch + x;
+    unsigned char def;
+
+    if (x < cw && y < ch) {
+        switch (fill) {
+            case 0: //FILL_BLANK
+                def = 0;
+                break;
+            case 1: //FILL_ORIGINAL
+                def = src[y*src_stride_ch + x];
+                break;
+            case 2: //FILL_CLAMP
+                y_s = clipf(y_s, 0, ch - 1);
+                x_s = clipf(x_s, 0, cw - 1);
+                def = src[(int)y_s * src_stride_ch + (int)x_s];
+                break;
+            case 3: //FILL_MIRROR
+                y_s = mirror(y_s, ch - 1);
+                x_s = mirror(x_s, cw - 1);
+                def = src[(int)y_s * src_stride_ch + (int)x_s];
+                break;
+        }
+        switch (interpolate) {
+            case 0: //INTERPOLATE_NEAREST
+                dst[idx_dst] = interpolate_nearest(x_s, y_s, src, cw, ch, src_stride_ch, def);
+                break;
+            case 1: //INTERPOLATE_BILINEAR
+                dst[idx_dst] = interpolate_bilinear(x_s, y_s, src, cw, ch, src_stride_ch, def);
+                break;
+            case 2: //INTERPOLATE_BIQUADRATIC
+                dst[idx_dst] = interpolate_biquadratic(x_s, y_s, src, cw, ch, src_stride_ch, def);
+                break;
+            default:
+                return;
+        }
     }
 }
 );