radeonsi: take LDS into account for compute shader occupancy stats
authorMarek Olšák <marek.olsak@amd.com>
Thu, 1 Dec 2016 02:23:27 +0000 (03:23 +0100)
committerMarek Olšák <marek.olsak@amd.com>
Wed, 7 Dec 2016 17:46:54 +0000 (18:46 +0100)
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
src/gallium/drivers/radeonsi/si_shader.c

index cf87df01ff0e7f3c5269a53c773ccc1cc6f569c4..018c99cf1f606e08d3fae1e79927339feebb0a14 100644 (file)
@@ -6048,19 +6048,21 @@ static void si_shader_dump_disassembly(const struct radeon_shader_binary *binary
 }
 
 static void si_shader_dump_stats(struct si_screen *sscreen,
-                                struct si_shader_config *conf,
-                                unsigned num_inputs,
-                                unsigned code_size,
+                                struct si_shader *shader,
                                 struct pipe_debug_callback *debug,
                                 unsigned processor,
                                 FILE *file)
 {
+       struct si_shader_config *conf = &shader->config;
+       unsigned num_inputs = shader->selector ? shader->selector->info.num_inputs : 0;
+       unsigned code_size = si_get_shader_binary_size(shader);
        unsigned lds_increment = sscreen->b.chip_class >= CIK ? 512 : 256;
        unsigned lds_per_wave = 0;
        unsigned max_simd_waves = 10;
 
        /* Compute LDS usage for PS. */
-       if (processor == PIPE_SHADER_FRAGMENT) {
+       switch (processor) {
+       case PIPE_SHADER_FRAGMENT:
                /* The minimum usage per wave is (num_inputs * 48). The maximum
                 * usage is (num_inputs * 48 * 16).
                 * We can get anything in between and it varies between waves.
@@ -6073,6 +6075,15 @@ static void si_shader_dump_stats(struct si_screen *sscreen,
                 */
                lds_per_wave = conf->lds_size * lds_increment +
                               align(num_inputs * 48, lds_increment);
+               break;
+       case PIPE_SHADER_COMPUTE:
+               if (shader->selector) {
+                       unsigned max_workgroup_size =
+                               si_get_max_workgroup_size(shader);
+                       lds_per_wave = (conf->lds_size * lds_increment) /
+                                      DIV_ROUND_UP(max_workgroup_size, 64);
+               }
+               break;
        }
 
        /* Compute the per-SIMD wave counts. */
@@ -6086,9 +6097,8 @@ static void si_shader_dump_stats(struct si_screen *sscreen,
        if (conf->num_vgprs)
                max_simd_waves = MIN2(max_simd_waves, 256 / conf->num_vgprs);
 
-       /* LDS is 64KB per CU (4 SIMDs), divided into 16KB blocks per SIMD
-        * that PS can use.
-        */
+       /* LDS is 64KB per CU (4 SIMDs), which is 16KB per SIMD (usage above
+        * 16KB makes some SIMDs unoccupied). */
        if (lds_per_wave)
                max_simd_waves = MIN2(max_simd_waves, 16384 / lds_per_wave);
 
@@ -6192,10 +6202,7 @@ void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
                fprintf(file, "\n");
        }
 
-       si_shader_dump_stats(sscreen, &shader->config,
-                            shader->selector ? shader->selector->info.num_inputs : 0,
-                            si_get_shader_binary_size(shader), debug, processor,
-                            file);
+       si_shader_dump_stats(sscreen, shader, debug, processor, file);
 }
 
 int si_compile_llvm(struct si_screen *sscreen,