broadcom/vc4: Add an accelerated path to turn raster R8/RG88 into tiled.
authorEric Anholt <eric@anholt.net>
Tue, 6 Feb 2018 16:43:24 +0000 (16:43 +0000)
committerEric Anholt <eric@anholt.net>
Fri, 9 Mar 2018 17:59:54 +0000 (09:59 -0800)
Drawing a 1080p YV12 video stream generated by MMAL goes from 10.5 FPS to
36.

src/gallium/drivers/vc4/vc4_blit.c
src/gallium/drivers/vc4/vc4_context.c
src/gallium/drivers/vc4/vc4_context.h

index f23f74c8f16987c1ff733d8ea17bcd53738070dd..7543e9d4aa08402ccbee56e156f6c1a60848f19a 100644 (file)
@@ -24,6 +24,7 @@
 #include "util/u_format.h"
 #include "util/u_surface.h"
 #include "util/u_blitter.h"
+#include "nir_builder.h"
 #include "vc4_context.h"
 
 static struct pipe_surface *
@@ -183,9 +184,139 @@ vc4_blitter_save(struct vc4_context *vc4)
                         vc4->fragtex.num_textures, vc4->fragtex.textures);
 }
 
+static void *vc4_get_yuv_vs(struct pipe_context *pctx)
+{
+   struct vc4_context *vc4 = vc4_context(pctx);
+   struct pipe_screen *pscreen = pctx->screen;
+
+   if (vc4->yuv_linear_blit_vs)
+           return vc4->yuv_linear_blit_vs;
+
+   const struct nir_shader_compiler_options *options =
+           pscreen->get_compiler_options(pscreen,
+                                         PIPE_SHADER_IR_NIR,
+                                         PIPE_SHADER_VERTEX);
+
+   nir_builder b;
+   nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_VERTEX, options);
+   b.shader->info.name = ralloc_strdup(b.shader, "linear_blit_vs");
+
+   const struct glsl_type *vec4 = glsl_vec4_type();
+   nir_variable *pos_in = nir_variable_create(b.shader, nir_var_shader_in,
+                                              vec4, "pos");
+
+   nir_variable *pos_out = nir_variable_create(b.shader, nir_var_shader_out,
+                                               vec4, "gl_Position");
+   pos_out->data.location = VARYING_SLOT_POS;
+
+   nir_store_var(&b, pos_out, nir_load_var(&b, pos_in), 0xf);
+
+   struct pipe_shader_state shader_tmpl = {
+           .type = PIPE_SHADER_IR_NIR,
+           .ir.nir = b.shader,
+   };
+
+   vc4->yuv_linear_blit_vs = pctx->create_vs_state(pctx, &shader_tmpl);
+
+   return vc4->yuv_linear_blit_vs;
+}
+
+static void *vc4_get_yuv_fs(struct pipe_context *pctx, int cpp)
+{
+   struct vc4_context *vc4 = vc4_context(pctx);
+   struct pipe_screen *pscreen = pctx->screen;
+   struct pipe_shader_state **cached_shader;
+   const char *name;
+
+   if (cpp == 1) {
+           cached_shader = &vc4->yuv_linear_blit_fs_8bit;
+           name = "linear_blit_8bit_fs";
+   } else {
+           cached_shader = &vc4->yuv_linear_blit_fs_16bit;
+           name = "linear_blit_16bit_fs";
+   }
+
+   if (*cached_shader)
+           return *cached_shader;
+
+   const struct nir_shader_compiler_options *options =
+           pscreen->get_compiler_options(pscreen,
+                                         PIPE_SHADER_IR_NIR,
+                                         PIPE_SHADER_FRAGMENT);
+
+   nir_builder b;
+   nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, options);
+   b.shader->info.name = ralloc_strdup(b.shader, name);
+
+   const struct glsl_type *vec4 = glsl_vec4_type();
+   const struct glsl_type *glsl_int = glsl_int_type();
+
+   nir_variable *color_out = nir_variable_create(b.shader, nir_var_shader_out,
+                                                 vec4, "f_color");
+   color_out->data.location = FRAG_RESULT_COLOR;
+
+   nir_variable *pos_in = nir_variable_create(b.shader, nir_var_shader_in,
+                                              vec4, "pos");
+   pos_in->data.location = VARYING_SLOT_POS;
+   nir_ssa_def *pos = nir_load_var(&b, pos_in);
+
+   nir_ssa_def *one = nir_imm_int(&b, 1);
+   nir_ssa_def *two = nir_imm_int(&b, 2);
+
+   nir_ssa_def *x = nir_f2i32(&b, nir_channel(&b, pos, 0));
+   nir_ssa_def *y = nir_f2i32(&b, nir_channel(&b, pos, 1));
+
+   nir_variable *stride_in = nir_variable_create(b.shader, nir_var_uniform,
+                                                 glsl_int, "stride");
+   nir_ssa_def *stride = nir_load_var(&b, stride_in);
+
+   nir_ssa_def *x_offset;
+   nir_ssa_def *y_offset;
+   if (cpp == 1) {
+           nir_ssa_def *intra_utile_x_offset =
+                   nir_ishl(&b, nir_iand(&b, x, one), two);
+           nir_ssa_def *inter_utile_x_offset =
+                   nir_ishl(&b, nir_iand(&b, x, nir_imm_int(&b, ~3)), one);
+
+           x_offset = nir_iadd(&b,
+                               intra_utile_x_offset,
+                               inter_utile_x_offset);
+           y_offset = nir_imul(&b,
+                               nir_iadd(&b,
+                                        nir_ishl(&b, y, one),
+                                        nir_ushr(&b, nir_iand(&b, x, two), one)),
+                               stride);
+   } else {
+           x_offset = nir_ishl(&b, x, two);
+           y_offset = nir_imul(&b, y, stride);
+   }
+
+   nir_intrinsic_instr *load =
+           nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_ubo);
+   load->num_components = 1;
+   nir_ssa_dest_init(&load->instr, &load->dest, load->num_components, 32, NULL);
+   load->src[0] = nir_src_for_ssa(one);
+   load->src[1] = nir_src_for_ssa(nir_iadd(&b, x_offset, y_offset));
+   nir_builder_instr_insert(&b, &load->instr);
+
+   nir_store_var(&b, color_out,
+                 nir_unpack_unorm_4x8(&b, &load->dest.ssa),
+                 0xf);
+
+   struct pipe_shader_state shader_tmpl = {
+           .type = PIPE_SHADER_IR_NIR,
+           .ir.nir = b.shader,
+   };
+
+   *cached_shader = pctx->create_fs_state(pctx, &shader_tmpl);
+
+   return *cached_shader;
+}
+
 static bool
 vc4_yuv_blit(struct pipe_context *pctx, const struct pipe_blit_info *info)
 {
+        struct vc4_context *vc4 = vc4_context(pctx);
         struct vc4_resource *src = vc4_resource(info->src.resource);
         struct vc4_resource *dst = vc4_resource(info->dst.resource);
         bool ok;
@@ -200,6 +331,75 @@ vc4_yuv_blit(struct pipe_context *pctx, const struct pipe_blit_info *info)
         assert(dst->base.format == src->base.format);
         assert(dst->tiled);
 
+        /* Always 1:1 and at the origin */
+        assert(info->src.box.x == 0 && info->dst.box.x == 0);
+        assert(info->src.box.y == 0 && info->dst.box.y == 0);
+        assert(info->src.box.width == info->dst.box.width);
+        assert(info->src.box.height == info->dst.box.height);
+
+        if ((src->slices[info->src.level].offset & 3) ||
+            (src->slices[info->src.level].stride & 3)) {
+                perf_debug("YUV-blit src texture offset/stride misaligned: 0x%08x/%d\n",
+                           src->slices[info->src.level].offset,
+                           src->slices[info->src.level].stride);
+                goto fallback;
+        }
+
+        vc4_blitter_save(vc4);
+
+        /* Create a renderable surface mapping the T-tiled shadow buffer.
+         */
+        struct pipe_surface dst_tmpl;
+        util_blitter_default_dst_texture(&dst_tmpl, info->dst.resource,
+                                         info->dst.level, info->dst.box.z);
+        dst_tmpl.format = PIPE_FORMAT_RGBA8888_UNORM;
+        struct pipe_surface *dst_surf =
+                pctx->create_surface(pctx, info->dst.resource, &dst_tmpl);
+        if (!dst_surf) {
+                fprintf(stderr, "Failed to create YUV dst surface\n");
+                util_blitter_unset_running_flag(vc4->blitter);
+                return false;
+        }
+        dst_surf->width /= 2;
+        if (dst->cpp == 1)
+                dst_surf->height /= 2;
+
+        /* Set the constant buffer. */
+        uint32_t stride = src->slices[info->src.level].stride;
+        struct pipe_constant_buffer cb_uniforms = {
+                .user_buffer = &stride,
+                .buffer_size = sizeof(stride),
+        };
+        pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 0, &cb_uniforms);
+        struct pipe_constant_buffer cb_src = {
+                .buffer = info->src.resource,
+                .buffer_offset = src->slices[info->src.level].offset,
+                .buffer_size = (src->bo->size -
+                                src->slices[info->src.level].offset),
+        };
+        pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 1, &cb_src);
+
+        /* Unbind the textures, to make sure we don't try to recurse into the
+         * shadow blit.
+         */
+        pctx->set_sampler_views(pctx, PIPE_SHADER_FRAGMENT, 0, 0, NULL);
+        pctx->bind_sampler_states(pctx, PIPE_SHADER_FRAGMENT, 0, 0, NULL);
+
+        util_blitter_custom_shader(vc4->blitter, dst_surf,
+                                   vc4_get_yuv_vs(pctx),
+                                   vc4_get_yuv_fs(pctx, src->cpp));
+
+        util_blitter_restore_textures(vc4->blitter);
+        util_blitter_restore_constant_buffer_state(vc4->blitter);
+        /* Restore cb1 (util_blitter doesn't handle this one). */
+        struct pipe_constant_buffer cb_disabled = { 0 };
+        pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 1, &cb_disabled);
+
+        pipe_surface_reference(&dst_surf, NULL);
+
+        return true;
+
+fallback:
         /* Do an immediate SW fallback, since the render blit path
          * would just recurse.
          */
index a9e7ff91f38d478aa78c9141244480281814f771..c1e041d1ef9e9e0e482d751c224701f0769d7fa9 100644 (file)
@@ -115,6 +115,13 @@ vc4_context_destroy(struct pipe_context *pctx)
         pipe_surface_reference(&vc4->framebuffer.cbufs[0], NULL);
         pipe_surface_reference(&vc4->framebuffer.zsbuf, NULL);
 
+        if (vc4->yuv_linear_blit_vs)
+                pctx->delete_vs_state(pctx, vc4->yuv_linear_blit_vs);
+        if (vc4->yuv_linear_blit_fs_8bit)
+                pctx->delete_fs_state(pctx, vc4->yuv_linear_blit_fs_8bit);
+        if (vc4->yuv_linear_blit_fs_16bit)
+                pctx->delete_fs_state(pctx, vc4->yuv_linear_blit_fs_16bit);
+
         vc4_program_fini(pctx);
 
         ralloc_free(vc4);
index d92fcba42c65a59db960a3bf311c2805815ae334..16bebeec40b0de952369026d44f8d9a247815234 100644 (file)
@@ -377,6 +377,10 @@ struct vc4_context {
 
         struct u_upload_mgr *uploader;
 
+        struct pipe_shader_state *yuv_linear_blit_vs;
+        struct pipe_shader_state *yuv_linear_blit_fs_8bit;
+        struct pipe_shader_state *yuv_linear_blit_fs_16bit;
+
         /** @{ Current pipeline state objects */
         struct pipe_scissor_state scissor;
         struct pipe_blend_state *blend;