panfrost/mfbd: Cleanup format code selection
[mesa.git] / src / gallium / drivers / panfrost / pan_job.c
index a7fc5f975cfac0f72286dbe65dc007264da57e82..6838050e57575d0ed636dec999deaeb324948eae 100644 (file)
 #include "pan_context.h"
 #include "util/hash_table.h"
 #include "util/ralloc.h"
+#include "util/u_format.h"
+#include "util/u_pack_color.h"
 
 struct panfrost_job *
 panfrost_create_job(struct panfrost_context *ctx)
 {
-        struct panfrost_job *job = rzalloc(NULL, struct panfrost_job);
+        struct panfrost_job *job = rzalloc(ctx, struct panfrost_job);
 
         job->ctx = ctx;
 
         job->bos = _mesa_set_create(job,
                                     _mesa_hash_pointer,
                                     _mesa_key_pointer_equal);
+
+        job->minx = job->miny = ~0;
+        job->maxx = job->maxy = 0;
+
+        util_dynarray_init(&job->headers, job);
+        util_dynarray_init(&job->gpu_headers, job);
  
         return job;
 }
@@ -89,7 +97,6 @@ panfrost_get_job(struct panfrost_context *ctx,
 
         memcpy(&job->key, &key, sizeof(key));
         _mesa_hash_table_insert(ctx->jobs, &job->key, job);
-        panfrost_job_set_requirements(ctx, job);
 
         return job;
 }
@@ -99,6 +106,12 @@ panfrost_get_job(struct panfrost_context *ctx,
 struct panfrost_job *
 panfrost_get_job_for_fbo(struct panfrost_context *ctx)
 {
+        /* If we're wallpapering, we special case to workaround
+         * u_blitter abuse */
+
+        if (ctx->wallpaper_batch)
+                return ctx->wallpaper_batch;
+
         /* If we already began rendering, use that */
 
         if (ctx->job)
@@ -144,25 +157,20 @@ panfrost_flush_jobs_writing_resource(struct panfrost_context *panfrost,
 void
 panfrost_job_submit(struct panfrost_context *ctx, struct panfrost_job *job)
 {
-        struct pipe_context *gallium = (struct pipe_context *) ctx;
-        struct panfrost_screen *screen = pan_screen(gallium->screen);
         int ret;
 
-        bool has_draws = ctx->draw_count > 0;
+        panfrost_scoreboard_link_batch(job);
+
+        bool has_draws = job->last_job.gpu;
         bool is_scanout = panfrost_is_scanout(ctx);
 
         if (!job)
                 return;
 
-        ret = screen->driver->submit_vs_fs_job(ctx, has_draws, is_scanout);
+        ret = panfrost_drm_submit_vs_fs_job(ctx, has_draws, is_scanout);
 
         if (ret)
                 fprintf(stderr, "panfrost_job_submit failed: %d\n", ret);
-
-        /* Reset job counters */
-        ctx->draw_count = 0;
-        ctx->vertex_job_count = 0;
-        ctx->tiler_job_count = 0;
 }
 
 void
@@ -176,6 +184,137 @@ panfrost_job_set_requirements(struct panfrost_context *ctx,
                 job->requirements |= PAN_REQ_DEPTH_WRITE;
 }
 
+/* Helper to smear a 32-bit color across 128-bit components */
+
+static void
+pan_pack_color_32(uint32_t *packed, uint32_t v)
+{
+        for (unsigned i = 0; i < 4; ++i)
+                packed[i] = v;
+}
+
+static void
+pan_pack_color_64(uint32_t *packed, uint32_t lo, uint32_t hi)
+{
+        for (unsigned i = 0; i < 4; i += 2) {
+                packed[i + 0] = lo;
+                packed[i + 1] = hi;
+        }
+}
+
+static void
+pan_pack_color(uint32_t *packed, const union pipe_color_union *color, enum pipe_format format)
+{
+        /* Alpha magicked to 1.0 if there is no alpha */
+
+        bool has_alpha = util_format_has_alpha(format);
+        float clear_alpha = has_alpha ? color->f[3] : 1.0f;
+
+        /* Packed color depends on the framebuffer format */
+
+        const struct util_format_description *desc =
+                util_format_description(format);
+
+        if (util_format_is_rgba8_variant(desc)) {
+                pan_pack_color_32(packed,
+                        (float_to_ubyte(clear_alpha) << 24) |
+                        (float_to_ubyte(color->f[2]) << 16) |
+                        (float_to_ubyte(color->f[1]) <<  8) |
+                        (float_to_ubyte(color->f[0]) <<  0));
+        } else if (format == PIPE_FORMAT_B5G6R5_UNORM) {
+                /* First, we convert the components to R5, G6, B5 separately */
+                unsigned r5 = CLAMP(color->f[0], 0.0, 1.0) * 31.0;
+                unsigned g6 = CLAMP(color->f[1], 0.0, 1.0) * 63.0;
+                unsigned b5 = CLAMP(color->f[2], 0.0, 1.0) * 31.0;
+
+                /* Then we pack into a sparse u32. TODO: Why these shifts? */
+                pan_pack_color_32(packed, (b5 << 25) | (g6 << 14) | (r5 << 5));
+        } else if (format == PIPE_FORMAT_B4G4R4A4_UNORM) {
+                /* We scale the components against 0xF0 (=240.0), rather than 0xFF */
+                unsigned r4 = CLAMP(color->f[0], 0.0, 1.0) * 240.0;
+                unsigned g4 = CLAMP(color->f[1], 0.0, 1.0) * 240.0;
+                unsigned b4 = CLAMP(color->f[2], 0.0, 1.0) * 240.0;
+                unsigned a4 = CLAMP(clear_alpha, 0.0, 1.0) * 240.0;
+
+                /* Pack on *byte* intervals */
+                pan_pack_color_32(packed, (a4 << 24) | (b4 << 16) | (g4 << 8) | r4);
+        } else if (format == PIPE_FORMAT_B5G5R5A1_UNORM) {
+                /* Scale as expected but shift oddly */
+                unsigned r5 = round(CLAMP(color->f[0], 0.0, 1.0)) * 31.0;
+                unsigned g5 = round(CLAMP(color->f[1], 0.0, 1.0)) * 31.0;
+                unsigned b5 = round(CLAMP(color->f[2], 0.0, 1.0)) * 31.0;
+                unsigned a1 = round(CLAMP(clear_alpha, 0.0, 1.0)) * 1.0;
+
+                pan_pack_color_32(packed, (a1 << 31) | (b5 << 25) | (g5 << 15) | (r5 << 5));
+        } else {
+                /* Try Gallium's generic default path. Doesn't work for all
+                 * formats but it's a good guess. */
+
+                union util_color out;
+
+                if (util_format_is_pure_integer(format)) {
+                        memcpy(out.ui, color->ui, 16);
+                } else {
+                        util_pack_color(color->f, format, &out);
+                }
+
+                unsigned size = util_format_get_blocksize(format);
+
+                if (size == 1) {
+                        unsigned b = out.ui[0];
+                        unsigned s = b | (b << 8);
+                        pan_pack_color_32(packed, s | (s << 16));
+                } else if (size == 2)
+                        pan_pack_color_32(packed, out.ui[0] | (out.ui[0] << 16));
+                else if (size == 4)
+                        pan_pack_color_32(packed, out.ui[0]);
+                else if (size == 8)
+                        pan_pack_color_64(packed, out.ui[0], out.ui[1]);
+                else if (size == 16)
+                        memcpy(packed, out.ui, 16);
+                else
+                        unreachable("Unknown generic format size packing clear colour");
+        }
+}
+
+void
+panfrost_job_clear(struct panfrost_context *ctx,
+                struct panfrost_job *job,
+                unsigned buffers,
+                const union pipe_color_union *color,
+                double depth, unsigned stencil)
+
+{
+        if (buffers & PIPE_CLEAR_COLOR) {
+                for (unsigned i = 0; i < PIPE_MAX_COLOR_BUFS; ++i) {
+                        if (!(buffers & (PIPE_CLEAR_COLOR0 << i)))
+                                continue;
+
+                        enum pipe_format format = ctx->pipe_framebuffer.cbufs[i]->format;
+                        pan_pack_color(job->clear_color[i], color, format);
+                }
+        }
+
+        if (buffers & PIPE_CLEAR_DEPTH) {
+                job->clear_depth = depth;
+        }
+
+        if (buffers & PIPE_CLEAR_STENCIL) {
+                job->clear_stencil = stencil;
+        }
+
+        job->clear |= buffers;
+
+        /* Clearing affects the entire framebuffer (by definition -- this is
+         * the Gallium clear callback, which clears the whole framebuffer. If
+         * the scissor test were enabled from the GL side, the state tracker
+         * would emit a quad instead and we wouldn't go down this code path) */
+
+        panfrost_job_union_scissor(job, 0, 0,
+                        ctx->pipe_framebuffer.width,
+                        ctx->pipe_framebuffer.height);
+}
+
 void
 panfrost_flush_jobs_reading_resource(struct panfrost_context *panfrost,
                                 struct pipe_resource *prsc)
@@ -207,16 +346,28 @@ panfrost_job_hash(const void *key)
         return _mesa_hash_data(key, sizeof(struct panfrost_job_key));
 }
 
+/* Given a new bounding rectangle (scissor), let the job cover the union of the
+ * new and old bounding rectangles */
+
+void
+panfrost_job_union_scissor(struct panfrost_job *job,
+                unsigned minx, unsigned miny,
+                unsigned maxx, unsigned maxy)
+{
+        job->minx = MIN2(job->minx, minx);
+        job->miny = MIN2(job->miny, miny);
+        job->maxx = MAX2(job->maxx, maxx);
+        job->maxy = MAX2(job->maxy, maxy);
+}
+
 void
 panfrost_job_init(struct panfrost_context *ctx)
 {
-        /* TODO: Don't leak */
-        ctx->jobs = _mesa_hash_table_create(NULL,
+        ctx->jobs = _mesa_hash_table_create(ctx,
                                             panfrost_job_hash,
                                             panfrost_job_compare);
 
-        ctx->write_jobs = _mesa_hash_table_create(NULL,
+        ctx->write_jobs = _mesa_hash_table_create(ctx,
                                             _mesa_hash_pointer,
                                             _mesa_key_pointer_equal);
-
 }