libavfilter: add a gblur_vulkan filter

This commit adds a powerful and customizable gblur Vulkan filter,
which provides a maximum 127x127 kernel size of Gaussian Filter.
The size could be adjusted by requirements on quality or performance.

The following command is on how to apply gblur_vulkan filter:

ffmpeg -init_hw_device vulkan -i input.264 -vf hwupload=extra_hw_frames=16,gblur_vulkan,hwdownload,format=yuv420p output.264

Signed-off-by: Wu Jianhua <jianhua.wu@intel.com>

1 files changed, 524 insertions, 0 deletions

diff --git a/libavfilter/vf_gblur_vulkan.c b/libavfilter/vf_gblur_vulkan.c
new file mode 100644
index 0000000000..b9e2c4fbc7
--- /dev/null
+++ b/libavfilter/vf_gblur_vulkan.c
@@ -0,0 +1,524 @@
+/*
+ * copyright (c) 2021 Wu Jianhua <jianhua.wu@intel.com>
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "libavutil/random_seed.h"
+#include "libavutil/opt.h"
+#include "vulkan.h"
+#include "internal.h"
+
+#define CGS 32
+#define GBLUR_MAX_KERNEL_SIZE 127
+
+typedef struct GBlurVulkanContext {
+    FFVulkanContext vkctx;
+    FFVkQueueFamilyCtx qf;
+    FFVkExecContext *exec;
+    FFVulkanPipeline *pl_hor;
+    FFVulkanPipeline *pl_ver;
+    FFVkBuffer params_buf_hor;
+    FFVkBuffer params_buf_ver;
+
+    VkDescriptorImageInfo input_images[3];
+    VkDescriptorImageInfo tmp_images[3];
+    VkDescriptorImageInfo output_images[3];
+    VkDescriptorBufferInfo params_desc_hor;
+    VkDescriptorBufferInfo params_desc_ver;
+
+    int initialized;
+    int size;
+    int planes;
+    int kernel_size;
+    float sigma;
+    float sigmaV;
+    AVFrame *tmpframe;
+} GBlurVulkanContext;
+
+static const char gblur_horizontal[] = {
+    C(0, void gblur(const ivec2 pos, const int index)                                  )
+    C(0, {                                                                             )
+    C(1,     vec4 sum = texture(input_image[index], pos) * kernel[0];                  )
+    C(0,                                                                               )
+    C(1,     for(int i = 1; i < kernel.length(); i++) {                                )
+    C(2,         sum += texture(input_image[index], pos + vec2(i, 0.0)) * kernel[i];   )
+    C(2,         sum += texture(input_image[index], pos - vec2(i, 0.0)) * kernel[i];   )
+    C(1,     }                                                                         )
+    C(0,                                                                               )
+    C(1,     imageStore(output_image[index], pos, sum);                                )
+    C(0, }                                                                             )
+};
+
+static const char gblur_vertical[] = {
+    C(0, void gblur(const ivec2 pos, const int index)                                  )
+    C(0, {                                                                             )
+    C(1,     vec4 sum = texture(input_image[index], pos) * kernel[0];                  )
+    C(0,                                                                               )
+    C(1,     for(int i = 1; i < kernel.length(); i++) {                                )
+    C(2,         sum += texture(input_image[index], pos + vec2(0.0, i)) * kernel[i];   )
+    C(2,         sum += texture(input_image[index], pos - vec2(0.0, i)) * kernel[i];   )
+    C(1,     }                                                                         )
+    C(0,                                                                               )
+    C(1,     imageStore(output_image[index], pos, sum);                                )
+    C(0, }                                                                             )
+};
+
+static inline float gaussian(float sigma, float x)
+{
+    return 1.0 / (sqrt(2.0 * M_PI) * sigma) *
+           exp(-(x * x) / (2.0 * sigma * sigma));
+}
+
+static inline float gaussian_simpson_integration(float sigma, float a, float b)
+{
+    return (b - a) * (1.0 / 6.0) * ((gaussian(sigma, a) +
+           4.0 * gaussian(sigma, (a + b) * 0.5) + gaussian(sigma, b)));
+}
+
+static void init_gaussian_kernel(float *kernel, float sigma, float kernel_size)
+{
+    int x;
+    float sum;
+
+    sum = 0;
+    for (x = 0; x < kernel_size; x++) {
+        kernel[x] = gaussian_simpson_integration(sigma, x - 0.5f, x + 0.5f);
+        if (!x)
+            sum += kernel[x];
+        else
+            sum += kernel[x] * 2.0;
+    }
+    /* Normalized */
+    sum = 1.0 / sum;
+    for (x = 0; x < kernel_size; x++) {
+        kernel[x] *= sum;
+    }
+}
+
+static av_cold void init_gaussian_params(GBlurVulkanContext *s)
+{
+    if (!(s->size & 1)) {
+        av_log(s, AV_LOG_WARNING, "kernel size should be odd\n");
+        s->size++;
+    }
+    if (s->sigmaV <= 0)
+        s->sigmaV = s->sigma;
+
+    s->kernel_size = (s->size >> 1) + 1;
+    s->tmpframe = NULL;
+}
+
+static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in)
+{
+    int err;
+    char *kernel_def;
+    uint8_t *kernel_mapped;
+    FFSPIRVShader *shd;
+    GBlurVulkanContext *s = ctx->priv;
+    const int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
+
+    FFVulkanDescriptorSetBinding image_descs[] = {
+        {
+            .name       = "input_image",
+            .type       = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
+            .dimensions = 2,
+            .elems      = planes,
+            .stages     = VK_SHADER_STAGE_COMPUTE_BIT,
+        },
+        {
+            .name       = "output_image",
+            .type       = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
+            .mem_layout = ff_vk_shader_rep_fmt(s->vkctx.output_format),
+            .mem_quali  = "writeonly",
+            .dimensions = 2,
+            .elems      = planes,
+            .stages     = VK_SHADER_STAGE_COMPUTE_BIT,
+        },
+    };
+
+    FFVulkanDescriptorSetBinding buf_desc = {
+        .name        = "data",
+        .type        = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+        .mem_quali   = "readonly",
+        .mem_layout  = "std430",
+        .stages      = VK_SHADER_STAGE_COMPUTE_BIT,
+        .updater     = NULL,
+        .buf_content = NULL,
+    };
+
+    image_descs[0].sampler = ff_vk_init_sampler(ctx, 1, VK_FILTER_LINEAR);
+    if (!image_descs[0].sampler)
+            return AVERROR_EXTERNAL;
+
+    init_gaussian_params(s);
+
+    kernel_def = av_asprintf("float kernel[%i];", s->kernel_size);
+    if (!kernel_def)
+        return AVERROR(ENOMEM);
+
+    buf_desc.buf_content = kernel_def;
+
+    ff_vk_qf_init(ctx, &s->qf, VK_QUEUE_COMPUTE_BIT, 0);
+
+    { /* Create shader for the horizontal pass */
+        image_descs[0].updater = s->input_images;
+        image_descs[1].updater = s->tmp_images;
+        buf_desc.updater = &s->params_desc_hor;
+
+        s->pl_hor = ff_vk_create_pipeline(ctx, &s->qf);
+        if (!s->pl_hor) {
+            err = AVERROR(ENOMEM);
+            goto fail;
+        }
+
+        shd = ff_vk_init_shader(ctx, s->pl_hor, "gblur_compute_hor", image_descs[0].stages);
+        if (!shd) {
+            err = AVERROR(ENOMEM);
+            goto fail;
+        }
+
+        ff_vk_set_compute_shader_sizes(ctx, shd, (int [3]){ CGS, CGS, 1 });
+        RET(ff_vk_add_descriptor_set(ctx, s->pl_hor, shd, image_descs, FF_ARRAY_ELEMS(image_descs), 0));
+        RET(ff_vk_add_descriptor_set(ctx, s->pl_hor, shd, &buf_desc, 1, 0));
+
+        GLSLD(   gblur_horizontal                                         );
+        GLSLC(0, void main()                                              );
+        GLSLC(0, {                                                        );
+        GLSLC(1,     ivec2 size;                                          );
+        GLSLC(1,     const ivec2 pos = ivec2(gl_GlobalInvocationID.xy);   );
+        for (int i = 0; i < planes; i++) {
+            GLSLC(0,                                                      );
+            GLSLF(1,  size = imageSize(output_image[%i]);               ,i);
+            GLSLC(1,  if (IS_WITHIN(pos, size)) {                         );
+            if (s->planes & (1 << i)) {
+                GLSLF(2,      gblur(pos, %i);                           ,i);
+            } else {
+                GLSLF(2, vec4 res = texture(input_image[%i], pos);      ,i);
+                GLSLF(2, imageStore(output_image[%i], pos, res);        ,i);
+            }
+            GLSLC(1, }                                                    );
+        }
+        GLSLC(0, }                                                        );
+
+        RET(ff_vk_compile_shader(ctx, shd, "main"));
+
+        RET(ff_vk_init_pipeline_layout(ctx, s->pl_hor));
+        RET(ff_vk_init_compute_pipeline(ctx, s->pl_hor));
+
+        RET(ff_vk_create_buf(ctx, &s->params_buf_hor, sizeof(float) * s->kernel_size,
+                             VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT));
+        RET(ff_vk_map_buffers(ctx, &s->params_buf_hor, &kernel_mapped, 1, 0));
+
+        init_gaussian_kernel((float *)kernel_mapped, s->sigma, s->kernel_size);
+
+        RET(ff_vk_unmap_buffers(ctx, &s->params_buf_hor, 1, 1));
+
+        s->params_desc_hor.buffer = s->params_buf_hor.buf;
+        s->params_desc_hor.range  = VK_WHOLE_SIZE;
+
+        ff_vk_update_descriptor_set(ctx, s->pl_hor, 1);
+    }
+
+    { /* Create shader for the vertical pass */
+        image_descs[0].updater = s->tmp_images;
+        image_descs[1].updater = s->output_images;
+        buf_desc.updater = &s->params_desc_ver;
+
+        s->pl_ver = ff_vk_create_pipeline(ctx, &s->qf);
+        if (!s->pl_ver) {
+            err = AVERROR(ENOMEM);
+            goto fail;
+        }
+
+        shd = ff_vk_init_shader(ctx, s->pl_ver, "gblur_compute_ver", image_descs[0].stages);
+        if (!shd) {
+            err = AVERROR(ENOMEM);
+            goto fail;
+        }
+
+        ff_vk_set_compute_shader_sizes(ctx, shd, (int [3]){ CGS, CGS, 1 });
+        RET(ff_vk_add_descriptor_set(ctx, s->pl_ver, shd, image_descs, FF_ARRAY_ELEMS(image_descs), 0));
+        RET(ff_vk_add_descriptor_set(ctx, s->pl_ver, shd, &buf_desc, 1, 0));
+
+        GLSLD(   gblur_vertical                                           );
+        GLSLC(0, void main()                                              );
+        GLSLC(0, {                                                        );
+        GLSLC(1,     ivec2 size;                                          );
+        GLSLC(1,     const ivec2 pos = ivec2(gl_GlobalInvocationID.xy);   );
+        for (int i = 0; i < planes; i++) {
+            GLSLC(0,                                                      );
+            GLSLF(1,  size = imageSize(output_image[%i]);               ,i);
+            GLSLC(1,  if (IS_WITHIN(pos, size)) {                         );
+            if (s->planes & (1 << i)) {
+                GLSLF(2,      gblur(pos, %i);                           ,i);
+            } else {
+                GLSLF(2, vec4 res = texture(input_image[%i], pos);      ,i);
+                GLSLF(2, imageStore(output_image[%i], pos, res);        ,i);
+            }
+            GLSLC(1, }                                                    );
+        }
+        GLSLC(0, }                                                        );
+
+        RET(ff_vk_compile_shader(ctx, shd, "main"));
+
+        RET(ff_vk_init_pipeline_layout(ctx, s->pl_ver));
+        RET(ff_vk_init_compute_pipeline(ctx, s->pl_ver));
+
+        RET(ff_vk_create_buf(ctx, &s->params_buf_ver, sizeof(float) * s->kernel_size,
+                             VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT));
+        RET(ff_vk_map_buffers(ctx, &s->params_buf_ver, &kernel_mapped, 1, 0));
+
+        init_gaussian_kernel((float *)kernel_mapped, s->sigmaV, s->kernel_size);
+
+        RET(ff_vk_unmap_buffers(ctx, &s->params_buf_ver, 1, 1));
+
+        s->params_desc_ver.buffer = s->params_buf_ver.buf;
+        s->params_desc_ver.range  = VK_WHOLE_SIZE;
+
+        ff_vk_update_descriptor_set(ctx, s->pl_ver, 1);
+    }
+
+    RET(ff_vk_create_exec_ctx(ctx, &s->exec, &s->qf));
+
+    s->initialized = 1;
+
+    return 0;
+
+fail:
+    av_free(kernel_def);
+    return err;
+}
+
+static av_cold void gblur_vulkan_uninit(AVFilterContext *avctx)
+{
+    GBlurVulkanContext *s = avctx->priv;
+
+    av_frame_free(&s->tmpframe);
+
+    ff_vk_filter_uninit(avctx);
+    ff_vk_free_buf(avctx, &s->params_buf_hor);
+    ff_vk_free_buf(avctx, &s->params_buf_ver);
+
+    s->initialized = 0;
+}
+
+static int process_frames(AVFilterContext *avctx, AVFrame *outframe, AVFrame *inframe)
+{
+    int err;
+    VkCommandBuffer cmd_buf;
+
+    const VkFormat *input_formats = NULL;
+    const VkFormat *output_formats = NULL;
+    GBlurVulkanContext *s = avctx->priv;
+    FFVulkanFunctions *vk = &s->vkctx.vkfn;
+    AVVkFrame *in = (AVVkFrame *)inframe->data[0];
+    AVVkFrame *tmp = (AVVkFrame *)s->tmpframe->data[0];
+    AVVkFrame *out = (AVVkFrame *)outframe->data[0];
+
+    int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
+
+    ff_vk_start_exec_recording(avctx, s->exec);
+    cmd_buf = ff_vk_get_exec_buf(avctx, s->exec);
+
+    input_formats = av_vkfmt_from_pixfmt(s->vkctx.input_format);
+    output_formats = av_vkfmt_from_pixfmt(s->vkctx.output_format);
+    for (int i = 0; i < planes; i++) {
+        RET(ff_vk_create_imageview(avctx, s->exec, &s->input_images[i].imageView,
+                                   in->img[i],
+                                   input_formats[i],
+                                   ff_comp_identity_map));
+
+        RET(ff_vk_create_imageview(avctx, s->exec, &s->tmp_images[i].imageView,
+                                   tmp->img[i],
+                                   output_formats[i],
+                                   ff_comp_identity_map));
+
+        RET(ff_vk_create_imageview(avctx, s->exec, &s->output_images[i].imageView,
+                                   out->img[i],
+                                   output_formats[i],
+                                   ff_comp_identity_map));
+
+        s->input_images[i].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
+        s->tmp_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL;
+        s->output_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL;
+    }
+
+    ff_vk_update_descriptor_set(avctx, s->pl_hor, 0);
+    ff_vk_update_descriptor_set(avctx, s->pl_ver, 0);
+
+    for (int i = 0; i < planes; i++) {
+        VkImageMemoryBarrier barriers[] = {
+            {
+                .sType                       = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+                .srcAccessMask               = 0,
+                .dstAccessMask               = VK_ACCESS_SHADER_READ_BIT,
+                .oldLayout                   = in->layout[i],
+                .newLayout                   = s->input_images[i].imageLayout,
+                .srcQueueFamilyIndex         = VK_QUEUE_FAMILY_IGNORED,
+                .dstQueueFamilyIndex         = VK_QUEUE_FAMILY_IGNORED,
+                .image                       = in->img[i],
+                .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+                .subresourceRange.levelCount = 1,
+                .subresourceRange.layerCount = 1,
+            },
+            {
+                .sType                       = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+                .srcAccessMask               = 0,
+                .dstAccessMask               = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_SHADER_READ_BIT,
+                .oldLayout                   = tmp->layout[i],
+                .newLayout                   = s->tmp_images[i].imageLayout,
+                .srcQueueFamilyIndex         = VK_QUEUE_FAMILY_IGNORED,
+                .dstQueueFamilyIndex         = VK_QUEUE_FAMILY_IGNORED,
+                .image                       = tmp->img[i],
+                .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+                .subresourceRange.levelCount = 1,
+                .subresourceRange.layerCount = 1,
+            },
+            {
+                .sType                       = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+                .srcAccessMask               = 0,
+                .dstAccessMask               = VK_ACCESS_SHADER_WRITE_BIT,
+                .oldLayout                   = out->layout[i],
+                .newLayout                   = s->output_images[i].imageLayout,
+                .srcQueueFamilyIndex         = VK_QUEUE_FAMILY_IGNORED,
+                .dstQueueFamilyIndex         = VK_QUEUE_FAMILY_IGNORED,
+                .image                       = out->img[i],
+                .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+                .subresourceRange.levelCount = 1,
+                .subresourceRange.layerCount = 1,
+            },
+        };
+
+        vk->CmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+                               VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0,
+                               0, NULL, 0, NULL, FF_ARRAY_ELEMS(barriers), barriers);
+
+        in->layout[i]  = barriers[0].newLayout;
+        in->access[i]  = barriers[0].dstAccessMask;
+
+        tmp->layout[i] = barriers[1].newLayout;
+        tmp->access[i] = barriers[1].dstAccessMask;
+
+        out->layout[i] = barriers[2].newLayout;
+        out->access[i] = barriers[2].dstAccessMask;
+    }
+
+    ff_vk_bind_pipeline_exec(avctx, s->exec, s->pl_hor);
+
+    vk->CmdDispatch(cmd_buf, FFALIGN(s->vkctx.output_width, CGS)/CGS,
+                    FFALIGN(s->vkctx.output_height, CGS)/CGS, 1);
+
+    ff_vk_bind_pipeline_exec(avctx, s->exec, s->pl_ver);
+
+    vk->CmdDispatch(cmd_buf, FFALIGN(s->vkctx.output_width, CGS)/CGS,
+                    FFALIGN(s->vkctx.output_height, CGS)/CGS, 1);
+
+    ff_vk_add_exec_dep(avctx, s->exec, inframe, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
+    ff_vk_add_exec_dep(avctx, s->exec, outframe, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
+
+    err = ff_vk_submit_exec_queue(avctx, s->exec);
+    if (err)
+        return err;
+
+    ff_vk_qf_rotate(&s->qf);
+
+    return 0;
+fail:
+    ff_vk_discard_exec_deps(avctx, s->exec);
+    return err;
+}
+
+static int gblur_vulkan_filter_frame(AVFilterLink *link, AVFrame *in)
+{
+    int err;
+    AVFrame *out = NULL;
+    AVFilterContext *ctx = link->dst;
+    GBlurVulkanContext *s = ctx->priv;
+    AVFilterLink *outlink = ctx->outputs[0];
+
+    out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
+    if (!out) {
+        err = AVERROR(ENOMEM);
+        goto fail;
+    }
+
+    if (!s->initialized) {
+        RET(init_filter(ctx, in));
+        s->tmpframe = ff_get_video_buffer(outlink, outlink->w, outlink->h);
+        if (!s->tmpframe) {
+            err = AVERROR(ENOMEM);
+            goto fail;
+        }
+    }
+
+    RET(process_frames(ctx, out, in));
+
+    RET(av_frame_copy_props(out, in));
+
+    av_frame_free(&in);
+
+    return ff_filter_frame(outlink, out);
+
+fail:
+    av_frame_free(&in);
+    av_frame_free(&out);
+    av_frame_free(&s->tmpframe);
+
+    return err;
+}
+
+#define OFFSET(x) offsetof(GBlurVulkanContext, x)
+#define FLAGS (AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM)
+static const AVOption gblur_vulkan_options[] = {
+    { "sigma",  "Set sigma",            OFFSET(sigma),  AV_OPT_TYPE_FLOAT, {.dbl = 0.5}, 0.01, 1024.0,                FLAGS },
+    { "sigmaV", "Set vertical sigma",   OFFSET(sigmaV), AV_OPT_TYPE_FLOAT, {.dbl = 0},   0.0,  1024.0,                FLAGS },
+    { "planes", "Set planes to filter", OFFSET(planes), AV_OPT_TYPE_INT,   {.i64 = 0xF}, 0,    0xF,                   FLAGS },
+    { "size",   "Set kernel size",      OFFSET(size),   AV_OPT_TYPE_INT,   {.i64 = 19},  1,    GBLUR_MAX_KERNEL_SIZE, FLAGS },
+    { NULL },
+};
+
+AVFILTER_DEFINE_CLASS(gblur_vulkan);
+
+static const AVFilterPad gblur_vulkan_inputs[] = {
+    {
+        .name         = "default",
+        .type         = AVMEDIA_TYPE_VIDEO,
+        .filter_frame = &gblur_vulkan_filter_frame,
+        .config_props = &ff_vk_filter_config_input,
+    }
+};
+
+static const AVFilterPad gblur_vulkan_outputs[] = {
+    {
+        .name         = "default",
+        .type         = AVMEDIA_TYPE_VIDEO,
+        .config_props = &ff_vk_filter_config_output,
+    }
+};
+
+const AVFilter ff_vf_gblur_vulkan = {
+    .name           = "gblur_vulkan",
+    .description    = NULL_IF_CONFIG_SMALL("Gaussian Blur in Vulkan"),
+    .priv_size      = sizeof(GBlurVulkanContext),
+    .init           = &ff_vk_filter_init,
+    .uninit         = &gblur_vulkan_uninit,
+    FILTER_INPUTS(gblur_vulkan_inputs),
+    FILTER_OUTPUTS(gblur_vulkan_outputs),
+    FILTER_SINGLE_PIXFMT(AV_PIX_FMT_VULKAN),
+    .priv_class     = &gblur_vulkan_class,
+    .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
+};