lavfi: add an avgblur_vulkan filter
authorLynne <dev@lynne.ee>
Sun, 27 Oct 2019 14:47:18 +0000 (14:47 +0000)
committerLynne <dev@lynne.ee>
Tue, 4 Feb 2020 23:19:48 +0000 (23:19 +0000)
This commit adds a fast avgblur Vulkan filter.
This will reset Intel GPUs on Linux due to a known, two-year-old driver bug
(!834 on mesa's gitlab).

configure
libavfilter/Makefile
libavfilter/allfilters.c
libavfilter/vf_avgblur_vulkan.c [new file with mode: 0644]

index efb62ed..293754f 100755 (executable)
--- a/configure
+++ b/configure
@@ -3461,6 +3461,7 @@ ass_filter_deps="libass"
 atempo_filter_deps="avcodec"
 atempo_filter_select="rdft"
 avgblur_opencl_filter_deps="opencl"
+avgblur_vulkan_filter_deps="vulkan libglslang"
 azmq_filter_deps="libzmq"
 blackframe_filter_deps="gpl"
 bm3d_filter_deps="avcodec"
index 217ba22..e3e2f61 100644 (file)
@@ -163,6 +163,7 @@ OBJS-$(CONFIG_ATADENOISE_FILTER)             += vf_atadenoise.o
 OBJS-$(CONFIG_AVGBLUR_FILTER)                += vf_avgblur.o
 OBJS-$(CONFIG_AVGBLUR_OPENCL_FILTER)         += vf_avgblur_opencl.o opencl.o \
                                                 opencl/avgblur.o boxblur.o
+OBJS-$(CONFIG_AVGBLUR_VULKAN_FILTER)         += vf_avgblur_vulkan.o vulkan.o
 OBJS-$(CONFIG_BBOX_FILTER)                   += bbox.o vf_bbox.o
 OBJS-$(CONFIG_BENCH_FILTER)                  += f_bench.o
 OBJS-$(CONFIG_BILATERAL_FILTER)              += vf_bilateral.o
index 747d65a..4acafdd 100644 (file)
@@ -154,6 +154,7 @@ extern AVFilter ff_vf_ass;
 extern AVFilter ff_vf_atadenoise;
 extern AVFilter ff_vf_avgblur;
 extern AVFilter ff_vf_avgblur_opencl;
+extern AVFilter ff_vf_avgblur_vulkan;
 extern AVFilter ff_vf_bbox;
 extern AVFilter ff_vf_bench;
 extern AVFilter ff_vf_bilateral;
diff --git a/libavfilter/vf_avgblur_vulkan.c b/libavfilter/vf_avgblur_vulkan.c
new file mode 100644 (file)
index 0000000..105d753
--- /dev/null
@@ -0,0 +1,406 @@
+/*
+ * 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/opt.h"
+#include "vulkan.h"
+#include "internal.h"
+
+#define CGS 32
+
+typedef struct AvgBlurVulkanContext {
+    VulkanFilterContext vkctx;
+
+    int initialized;
+    FFVkExecContext *exec;
+    VulkanPipeline *pl_hor;
+    VulkanPipeline *pl_ver;
+
+    /* Shader updators, must be in the main filter struct */
+    VkDescriptorImageInfo input_images[3];
+    VkDescriptorImageInfo tmp_images[3];
+    VkDescriptorImageInfo output_images[3];
+
+    int size_x;
+    int size_y;
+    int planes;
+} AvgBlurVulkanContext;
+
+static const char blur_kernel[] = {
+    C(0, shared vec4 cache[DIR(gl_WorkGroupSize) + FILTER_RADIUS*2 + 1];           )
+    C(0,                                                                           )
+    C(0, void distort(const ivec2 pos, const int idx)                              )
+    C(0, {                                                                         )
+    C(1,     const uint cp = DIR(gl_LocalInvocationID) + FILTER_RADIUS;            )
+    C(0,                                                                           )
+    C(1,     cache[cp] = texture(input_img[idx], pos);                             )
+    C(0,                                                                           )
+    C(1,     const ivec2 loc_l = pos - INC(FILTER_RADIUS);                         )
+    C(1,     cache[cp - FILTER_RADIUS] = texture(input_img[idx], loc_l);           )
+    C(0,                                                                           )
+    C(1,     const ivec2 loc_h = pos + INC(DIR(gl_WorkGroupSize));                 )
+    C(1,     cache[cp + DIR(gl_WorkGroupSize)] = texture(input_img[idx], loc_h);   )
+    C(0,                                                                           )
+    C(1,     barrier();                                                            )
+    C(0,                                                                           )
+    C(1,     vec4 sum = vec4(0);                                                   )
+    C(1,     for (int p = -FILTER_RADIUS; p <= FILTER_RADIUS; p++)                 )
+    C(2,         sum += cache[cp + p];                                             )
+    C(0,                                                                           )
+    C(1,     sum /= vec4(FILTER_RADIUS*2 + 1);                                     )
+    C(1,     imageStore(output_img[idx], pos, sum);                                )
+    C(0, }                                                                         )
+};
+
+static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in)
+{
+    int err;
+    SPIRVShader *shd;
+    AvgBlurVulkanContext *s = ctx->priv;
+    const int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
+    VkSampler *sampler = ff_vk_init_sampler(ctx, 1, VK_FILTER_LINEAR);
+
+    VulkanDescriptorSetBinding desc_i[2] = {
+        {
+            .name       = "input_img",
+            .type       = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
+            .dimensions = 2,
+            .elems      = planes,
+            .stages     = VK_SHADER_STAGE_COMPUTE_BIT,
+            .samplers   = DUP_SAMPLER_ARRAY4(*sampler),
+        },
+        {
+            .name       = "output_img",
+            .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,
+        },
+    };
+
+    if (!sampler)
+        return AVERROR_EXTERNAL;
+
+    { /* Create shader for the horizontal pass */
+        desc_i[0].updater = s->input_images;
+        desc_i[1].updater = s->tmp_images;
+
+        s->pl_hor = ff_vk_create_pipeline(ctx);
+        if (!s->pl_hor)
+            return AVERROR(ENOMEM);
+
+        shd = ff_vk_init_shader(ctx, s->pl_hor, "avgblur_compute_hor",
+                                VK_SHADER_STAGE_COMPUTE_BIT);
+
+        ff_vk_set_compute_shader_sizes(ctx, shd, (int [3]){ CGS, 1, 1 });
+
+        RET(ff_vk_add_descriptor_set(ctx, s->pl_hor, shd, desc_i, 2, 0));
+
+        GLSLF(0, #define FILTER_RADIUS (%i)                     ,s->size_x - 1);
+        GLSLC(0, #define INC(x) (ivec2(x, 0))                                 );
+        GLSLC(0, #define DIR(var) (var.x)                                     );
+        GLSLD(   blur_kernel                                                  );
+        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_img[%i]);                     ,i);
+            GLSLC(1,  if (IS_WITHIN(pos, size)) {                             );
+            if (s->planes & (1 << i)) {
+                GLSLF(2, distort(pos, %i);                                  ,i);
+            } else {
+                GLSLF(2, vec4 res = texture(input_img[%i], pos);            ,i);
+                GLSLF(2, imageStore(output_img[%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));
+    }
+
+    { /* Create shader for the vertical pass */
+        desc_i[0].updater = s->tmp_images;
+        desc_i[1].updater = s->output_images;
+
+        s->pl_ver = ff_vk_create_pipeline(ctx);
+        if (!s->pl_ver)
+            return AVERROR(ENOMEM);
+
+        shd = ff_vk_init_shader(ctx, s->pl_ver, "avgblur_compute_ver",
+                                VK_SHADER_STAGE_COMPUTE_BIT);
+
+        ff_vk_set_compute_shader_sizes(ctx, shd, (int [3]){ 1, CGS, 1 });
+
+        RET(ff_vk_add_descriptor_set(ctx, s->pl_ver, shd, desc_i, 2, 0));
+
+        GLSLF(0, #define FILTER_RADIUS (%i)                     ,s->size_y - 1);
+        GLSLC(0, #define INC(x) (ivec2(0, x))                                 );
+        GLSLC(0, #define DIR(var) (var.y)                                     );
+        GLSLD(   blur_kernel                                                  );
+        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_img[%i]);                     ,i);
+            GLSLC(1,  if (IS_WITHIN(pos, size)) {                             );
+            if (s->planes & (1 << i)) {
+                GLSLF(2, distort(pos, %i);                                  ,i);
+            } else {
+                GLSLF(2, vec4 res = texture(input_img[%i], pos);            ,i);
+                GLSLF(2, imageStore(output_img[%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));
+    }
+
+    /* Execution context */
+    RET(ff_vk_create_exec_ctx(ctx, &s->exec,
+                              s->vkctx.hwctx->queue_family_comp_index));
+
+    s->initialized = 1;
+
+    return 0;
+
+fail:
+    return err;
+}
+
+static int process_frames(AVFilterContext *avctx, AVFrame *out_f, AVFrame *tmp_f, AVFrame *in_f)
+{
+    int err;
+    AvgBlurVulkanContext *s = avctx->priv;
+    AVVkFrame *in = (AVVkFrame *)in_f->data[0];
+    AVVkFrame *tmp = (AVVkFrame *)tmp_f->data[0];
+    AVVkFrame *out = (AVVkFrame *)out_f->data[0];
+    int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
+
+    for (int i = 0; i < planes; i++) {
+        RET(ff_vk_create_imageview(avctx, &s->input_images[i].imageView, in->img[i],
+                                   av_vkfmt_from_pixfmt(s->vkctx.input_format)[i],
+                                   ff_comp_identity_map));
+
+        RET(ff_vk_create_imageview(avctx, &s->tmp_images[i].imageView, tmp->img[i],
+                                   av_vkfmt_from_pixfmt(s->vkctx.output_format)[i],
+                                   ff_comp_identity_map));
+
+        RET(ff_vk_create_imageview(avctx, &s->output_images[i].imageView, out->img[i],
+                                   av_vkfmt_from_pixfmt(s->vkctx.output_format)[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);
+
+    ff_vk_start_exec_recording(avctx, s->exec);
+
+    for (int i = 0; i < planes; i++) {
+        VkImageMemoryBarrier bar[] = {
+            {
+                .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,
+            },
+        };
+
+        vkCmdPipelineBarrier(s->exec->buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+                             VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0,
+                             0, NULL, 0, NULL, FF_ARRAY_ELEMS(bar), bar);
+
+        in->layout[i]  = bar[0].newLayout;
+        in->access[i]  = bar[0].dstAccessMask;
+
+        tmp->layout[i] = bar[1].newLayout;
+        tmp->access[i] = bar[1].dstAccessMask;
+
+        out->layout[i] = bar[2].newLayout;
+        out->access[i] = bar[2].dstAccessMask;
+    }
+
+    ff_vk_bind_pipeline_exec(avctx, s->exec, s->pl_hor);
+
+    vkCmdDispatch(s->exec->buf, FFALIGN(s->vkctx.output_width, CGS)/CGS,
+                  s->vkctx.output_height, 1);
+
+    ff_vk_bind_pipeline_exec(avctx, s->exec, s->pl_ver);
+
+    vkCmdDispatch(s->exec->buf, s->vkctx.output_width,
+                  FFALIGN(s->vkctx.output_height, CGS)/CGS, 1);
+
+    ff_vk_add_exec_dep(avctx, s->exec, in_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
+    ff_vk_add_exec_dep(avctx, s->exec, out_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
+
+    err = ff_vk_submit_exec_queue(avctx, s->exec);
+    if (err)
+        return err;
+
+fail:
+
+    for (int i = 0; i < planes; i++) {
+        ff_vk_destroy_imageview(avctx, &s->input_images[i].imageView);
+        ff_vk_destroy_imageview(avctx, &s->tmp_images[i].imageView);
+        ff_vk_destroy_imageview(avctx, &s->output_images[i].imageView);
+    }
+
+    return err;
+}
+
+static int avgblur_vulkan_filter_frame(AVFilterLink *link, AVFrame *in)
+{
+    int err;
+    AVFrame *tmp = NULL, *out = NULL;
+    AVFilterContext *ctx = link->dst;
+    AvgBlurVulkanContext *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;
+    }
+
+    tmp = ff_get_video_buffer(outlink, outlink->w, outlink->h);
+    if (!out) {
+        err = AVERROR(ENOMEM);
+        goto fail;
+    }
+
+    if (!s->initialized)
+        RET(init_filter(ctx, in));
+
+    RET(process_frames(ctx, out, tmp, in));
+
+    err = av_frame_copy_props(out, in);
+    if (err < 0)
+        goto fail;
+
+    av_frame_free(&in);
+    av_frame_free(&tmp);
+
+    return ff_filter_frame(outlink, out);
+
+fail:
+    av_frame_free(&in);
+    av_frame_free(&tmp);
+    av_frame_free(&out);
+    return err;
+}
+
+static void avgblur_vulkan_uninit(AVFilterContext *avctx)
+{
+    AvgBlurVulkanContext *s = avctx->priv;
+
+    ff_vk_filter_uninit(avctx);
+
+    s->initialized = 0;
+}
+
+#define OFFSET(x) offsetof(AvgBlurVulkanContext, x)
+#define FLAGS (AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM)
+static const AVOption avgblur_vulkan_options[] = {
+    { "sizeX",  "Set horizontal radius", OFFSET(size_x), AV_OPT_TYPE_INT, {.i64 = 3}, 1, 32, .flags = FLAGS },
+    { "planes", "Set planes to filter (bitmask)", OFFSET(planes), AV_OPT_TYPE_INT, {.i64 = 0xF}, 0, 0xF, .flags = FLAGS },
+    { "sizeY",  "Set vertical radius", OFFSET(size_y), AV_OPT_TYPE_INT, {.i64 = 3}, 1, 32, .flags = FLAGS },
+    { NULL },
+};
+
+AVFILTER_DEFINE_CLASS(avgblur_vulkan);
+
+static const AVFilterPad avgblur_vulkan_inputs[] = {
+    {
+        .name         = "default",
+        .type         = AVMEDIA_TYPE_VIDEO,
+        .filter_frame = &avgblur_vulkan_filter_frame,
+        .config_props = &ff_vk_filter_config_input,
+    },
+    { NULL }
+};
+
+static const AVFilterPad avgblur_vulkan_outputs[] = {
+    {
+        .name = "default",
+        .type = AVMEDIA_TYPE_VIDEO,
+        .config_props = &ff_vk_filter_config_output,
+    },
+    { NULL }
+};
+
+AVFilter ff_vf_avgblur_vulkan = {
+    .name           = "avgblur_vulkan",
+    .description    = NULL_IF_CONFIG_SMALL("Apply avgblur mask to input video"),
+    .priv_size      = sizeof(AvgBlurVulkanContext),
+    .init           = &ff_vk_filter_init,
+    .uninit         = &avgblur_vulkan_uninit,
+    .query_formats  = &ff_vk_filter_query_formats,
+    .inputs         = avgblur_vulkan_inputs,
+    .outputs        = avgblur_vulkan_outputs,
+    .priv_class     = &avgblur_vulkan_class,
+    .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
+};