panfrost: Use the per-batch fences to wait on the last submitted batch
[mesa.git] / src / gallium / drivers / panfrost / pan_context.h
index cd9a034975d3499b7f9ce46a75dc1f7eaff246be..ce3e0c899a4f99c07fb0a0fa89a8f829c37b9e9c 100644 (file)
@@ -32,6 +32,7 @@
 #include "pan_resource.h"
 #include "pan_job.h"
 #include "pan_blend.h"
+#include "pan_encoder.h"
 
 #include "pipe/p_compiler.h"
 #include "pipe/p_config.h"
@@ -44,6 +45,7 @@
 #include "util/hash_table.h"
 
 #include "midgard/midgard_compile.h"
+#include "compiler/shader_enums.h"
 
 /* Forward declare to avoid extra header dep */
 struct prim_convert_context;
@@ -78,8 +80,16 @@ struct panfrost_query {
         unsigned type;
         unsigned index;
 
-        /* Memory for the GPU to writeback the value of the query */
-        struct panfrost_transfer transfer;
+        union {
+                /* For computed queries. 64-bit to prevent overflow */
+                struct {
+                        uint64_t start;
+                        uint64_t end;
+                };
+
+                /* Memory for the GPU to writeback the value of the query */
+                struct panfrost_transfer transfer;
+        };
 };
 
 struct panfrost_fence {
@@ -87,36 +97,39 @@ struct panfrost_fence {
         int fd;
 };
 
+struct panfrost_streamout {
+        struct pipe_stream_output_target *targets[PIPE_MAX_SO_BUFFERS];
+        uint32_t offsets[PIPE_MAX_SO_BUFFERS];
+        unsigned num_targets;
+};
+
 struct panfrost_context {
         /* Gallium context */
         struct pipe_context base;
 
-        /* Bound job and map of panfrost_job_key to jobs */
-        struct panfrost_job *job;
-        struct hash_table *jobs;
+        /* Compiler context */
+        struct midgard_screen compiler;
+
+        /* Bound job batch and map of panfrost_batch_key to job batches */
+        struct panfrost_batch *batch;
+        struct hash_table *batches;
 
-        /* panfrost_resource -> panfrost_job */
-        struct hash_table *write_jobs;
+        /* Within a launch_grid call.. */
+        const struct pipe_grid_info *compute_grid;
 
         /* Bit mask for supported PIPE_DRAW for this hardware */
         unsigned draw_modes;
 
         struct pipe_framebuffer_state pipe_framebuffer;
+        struct panfrost_streamout streamout;
 
-        struct panfrost_memory cmdstream_persistent;
-        struct panfrost_memory shaders;
-        struct panfrost_memory scratchpad;
-        struct panfrost_memory tiler_heap;
-        struct panfrost_memory varying_mem;
-        struct panfrost_memory tiler_polygon_list;
-        struct panfrost_memory tiler_dummy;
-        struct panfrost_memory depth_stencil_buffer;
-
+        bool active_queries;
+        uint64_t prims_generated;
+        uint64_t tf_prims_generated;
         struct panfrost_query *occlusion_query;
 
         /* Each draw has corresponding vertex and tiler payloads */
-        struct midgard_payload_vertex_tiler payload_vertex;
-        struct midgard_payload_vertex_tiler payload_tiler;
+        struct midgard_payload_vertex_tiler payloads[PIPE_SHADER_TYPES];
 
         /* The fragment shader binary itself is pointed here (for the tripipe) but
          * also everything else in the shader core, including blending, the
@@ -129,33 +142,26 @@ struct panfrost_context {
 
         unsigned vertex_count;
         unsigned instance_count;
+        enum pipe_prim_type active_prim;
 
         /* If instancing is enabled, vertex count padded for instance; if
          * it is disabled, just equal to plain vertex count */
         unsigned padded_count;
 
-        union mali_attr attributes[PIPE_MAX_ATTRIBS];
-
-        unsigned varying_height;
-
-        struct mali_single_framebuffer vt_framebuffer_sfbd;
-        struct bifrost_framebuffer vt_framebuffer_mfbd;
-
         /* TODO: Multiple uniform buffers (index =/= 0), finer updates? */
 
         struct panfrost_constant_buffer constant_buffer[PIPE_SHADER_TYPES];
 
-        /* CSOs */
         struct panfrost_rasterizer *rasterizer;
-
-        struct panfrost_shader_variants *vs;
-        struct panfrost_shader_variants *fs;
-
+        struct panfrost_shader_variants *shader[PIPE_SHADER_TYPES];
         struct panfrost_vertex_state *vertex;
 
         struct pipe_vertex_buffer vertex_buffers[PIPE_MAX_ATTRIBS];
         uint32_t vb_mask;
 
+        struct pipe_shader_buffer ssbo[PIPE_SHADER_TYPES][PIPE_MAX_SHADER_BUFFERS];
+        uint32_t ssbo_mask[PIPE_SHADER_TYPES];
+
         struct panfrost_sampler_state *samplers[PIPE_SHADER_TYPES][PIPE_MAX_SAMPLERS];
         unsigned sampler_count[PIPE_SHADER_TYPES];
 
@@ -172,7 +178,7 @@ struct panfrost_context {
          * errors due to unsupported reucrsion */
 
         struct blitter_context *blitter_wallpaper;
-        struct panfrost_job *wallpaper_batch;
+        struct panfrost_batch *wallpaper_batch;
 
         struct panfrost_blend_state *blend;
 
@@ -185,14 +191,8 @@ struct panfrost_context {
         /* True for t6XX, false for t8xx. */
         bool is_t6xx;
 
-        /* If set, we'll require the use of single render-target framebuffer
-         * descriptors (SFBD), for older hardware -- specifically, <T760 hardware, If
-         * false, we'll use the MFBD no matter what. New hardware -does- retain support
-         * for SFBD, and in theory we could flip between them on a per-RT basis, but
-         * there's no real advantage to doing so */
-        bool require_sfbd;
-
-        uint32_t out_sync;
+        /* The out sync fence of the last submitted batch. */
+        struct panfrost_batch_fence *last_out_sync;
 };
 
 /* Corresponds to the CSO */
@@ -211,21 +211,22 @@ struct panfrost_rasterizer {
 
 /* A shader state corresponds to the actual, current variant of the shader */
 struct panfrost_shader_state {
-        struct pipe_shader_state *base;
-
         /* Compiled, mapped descriptor, ready for the hardware */
         bool compiled;
         struct mali_shader_meta *tripipe;
-        mali_ptr tripipe_gpu;
 
         /* Non-descript information */
         int uniform_count;
         bool can_discard;
         bool writes_point_size;
         bool reads_point_coord;
+        bool reads_face;
+        bool reads_frag_coord;
 
         struct mali_attr_meta varyings[PIPE_MAX_ATTRIBS];
         gl_varying_slot varyings_loc[PIPE_MAX_ATTRIBS];
+        struct pipe_stream_output_info stream_output;
+        uint64_t so_mask;
 
         unsigned sysval_count;
         unsigned sysval[MAX_SYSVAL_COUNT];
@@ -235,11 +236,24 @@ struct panfrost_shader_state {
 
         uint16_t point_sprite_mask;
         unsigned point_sprite_upper_left : 1;
+
+        /* Should we enable helper invocations */
+        bool helper_invocations;
+
+        struct panfrost_bo *bo;
 };
 
 /* A collection of varyings (the CSO) */
 struct panfrost_shader_variants {
-        struct pipe_shader_state base;
+        /* A panfrost_shader_variants can represent a shader for
+         * either graphics or compute */
+
+        bool is_compute;
+
+        union {
+                struct pipe_shader_state base;
+                struct pipe_compute_state cbase;
+        };
 
         struct panfrost_shader_state variants[MAX_SHADER_VARIANTS];
         unsigned variant_count;
@@ -265,6 +279,7 @@ struct panfrost_sampler_state {
 struct panfrost_sampler_view {
         struct pipe_sampler_view base;
         struct mali_texture_descriptor hw;
+        bool manual_stride;
 };
 
 static inline struct panfrost_context *
@@ -276,6 +291,9 @@ pan_context(struct pipe_context *pcontext)
 struct pipe_context *
 panfrost_create_context(struct pipe_screen *screen, void *priv, unsigned flags);
 
+void
+panfrost_invalidate_frame(struct panfrost_context *ctx);
+
 void
 panfrost_emit_for_draw(struct panfrost_context *ctx, bool with_vertex_data);
 
@@ -291,44 +309,27 @@ panfrost_flush(
         struct pipe_fence_handle **fence,
         unsigned flags);
 
-bool
-panfrost_is_scanout(struct panfrost_context *ctx);
-
-mali_ptr panfrost_sfbd_fragment(struct panfrost_context *ctx, bool has_draws);
-mali_ptr panfrost_mfbd_fragment(struct panfrost_context *ctx, bool has_draws);
+mali_ptr panfrost_sfbd_fragment(struct panfrost_batch *batch, bool has_draws);
+mali_ptr panfrost_mfbd_fragment(struct panfrost_batch *batch, bool has_draws);
 
 struct bifrost_framebuffer
-panfrost_emit_mfbd(struct panfrost_context *ctx, unsigned vertex_count);
+panfrost_emit_mfbd(struct panfrost_batch *batch, unsigned vertex_count);
 
 struct mali_single_framebuffer
-panfrost_emit_sfbd(struct panfrost_context *ctx, unsigned vertex_count);
+panfrost_emit_sfbd(struct panfrost_batch *batch, unsigned vertex_count);
 
 mali_ptr
-panfrost_fragment_job(struct panfrost_context *ctx, bool has_draws);
-
-void
-panfrost_shader_compile(struct panfrost_context *ctx, struct mali_shader_meta *meta, const char *src, int type, struct panfrost_shader_state *state);
+panfrost_fragment_job(struct panfrost_batch *batch, bool has_draws);
 
 void
-panfrost_pack_work_groups_compute(
-        struct mali_vertex_tiler_prefix *out,
-        unsigned num_x,
-        unsigned num_y,
-        unsigned num_z,
-        unsigned size_x,
-        unsigned size_y,
-        unsigned size_z);
-
-void
-panfrost_pack_work_groups_fused(
-        struct mali_vertex_tiler_prefix *vertex,
-        struct mali_vertex_tiler_prefix *tiler,
-        unsigned num_x,
-        unsigned num_y,
-        unsigned num_z,
-        unsigned size_x,
-        unsigned size_y,
-        unsigned size_z);
+panfrost_shader_compile(
+                struct panfrost_context *ctx,
+                struct mali_shader_meta *meta,
+                enum pipe_shader_ir ir_type,
+                const void *ir,
+                gl_shader_stage stage,
+                struct panfrost_shader_state *state,
+                uint64_t *outputs_written);
 
 /* Instancing */
 
@@ -336,7 +337,7 @@ mali_ptr
 panfrost_vertex_buffer_address(struct panfrost_context *ctx, unsigned i);
 
 void
-panfrost_emit_vertex_data(struct panfrost_job *batch);
+panfrost_emit_vertex_data(struct panfrost_batch *batch);
 
 struct pan_shift_odd {
         unsigned shift;
@@ -352,5 +353,16 @@ panfrost_padded_vertex_count(
 unsigned
 pan_expand_shift_odd(struct pan_shift_odd o);
 
+/* Compute */
+
+void
+panfrost_compute_context_init(struct pipe_context *pctx);
+
+/* Varyings */
+
+void
+panfrost_emit_varying_descriptor(
+        struct panfrost_context *ctx,
+        unsigned vertex_count);
 
 #endif