radeonsi: add missing initialization of registers
authorMarek Olšák <marek.olsak@amd.com>
Tue, 7 Jul 2020 02:05:00 +0000 (22:05 -0400)
committerMarek Olšák <marek.olsak@amd.com>
Wed, 22 Jul 2020 16:08:19 +0000 (12:08 -0400)
(random initial gfx10 commit:)
Fixes: 78cdf9a99f0 - amd/addrlib: add gfx10 support

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5798>

src/gallium/drivers/radeonsi/si_compute.c
src/gallium/drivers/radeonsi/si_state.c

index b3dea26..bd7defd 100644 (file)
@@ -364,10 +364,22 @@ void si_emit_initial_compute_regs(struct si_context *sctx, struct radeon_cmdbuf
       radeon_set_sh_reg_seq(cs, R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2, 2);
       radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
       radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
+
+      /* Disable profiling on compute queues. */
+      if (cs != sctx->gfx_cs || !sctx->screen->info.has_graphics) {
+         radeon_set_sh_reg(cs, R_00B82C_COMPUTE_PERFCOUNT_ENABLE, 0);
+         radeon_set_sh_reg(cs, R_00B878_COMPUTE_THREAD_TRACE_ENABLE, 0);
+      }
    }
 
-   if (sctx->chip_class >= GFX10)
+   if (sctx->chip_class >= GFX10) {
+      radeon_set_sh_reg(cs, R_00B890_COMPUTE_USER_ACCUM_0, 0);
+      radeon_set_sh_reg(cs, R_00B894_COMPUTE_USER_ACCUM_1, 0);
+      radeon_set_sh_reg(cs, R_00B898_COMPUTE_USER_ACCUM_2, 0);
+      radeon_set_sh_reg(cs, R_00B89C_COMPUTE_USER_ACCUM_3, 0);
       radeon_set_sh_reg(cs, R_00B8A0_COMPUTE_PGM_RSRC3, 0);
+      radeon_set_sh_reg(cs, R_00B9F4_COMPUTE_DISPATCH_TUNNEL, 0);
+   }
 
    /* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
     * and is now per pipe, so it should be handled in the
@@ -393,6 +405,15 @@ void si_emit_initial_compute_regs(struct si_context *sctx, struct radeon_cmdbuf
          radeon_set_config_reg(cs, R_00950C_TA_CS_BC_BASE_ADDR, bc_va >> 8);
       }
    }
+
+   /* cs_preamble_state initializes this for the gfx queue, so only do this
+    * if we are on a compute queue.
+    */
+   if (sctx->chip_class >= GFX9 &&
+       (cs != sctx->gfx_cs || !sctx->screen->info.has_graphics)) {
+      radeon_set_uconfig_reg(cs, R_0301EC_CP_COHER_START_DELAY,
+                             sctx->chip_class >= GFX10 ? 0x20 : 0);
+   }
 }
 
 static bool si_setup_compute_scratch_buffer(struct si_context *sctx, struct si_shader *shader,
index feef175..b070104 100644 (file)
@@ -5350,6 +5350,27 @@ void si_init_cs_preamble_state(struct si_context *sctx)
       si_pm4_set_reg(pm4, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
                      S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
       si_pm4_set_reg(pm4, R_030968_VGT_INSTANCE_BASE_ID, 0);
+      si_pm4_set_reg(pm4, R_0301EC_CP_COHER_START_DELAY,
+                     sctx->chip_class >= GFX10 ? 0x20 : 0);
+   }
+
+   if (sctx->chip_class >= GFX10) {
+      si_pm4_set_reg(pm4, R_00B0C8_SPI_SHADER_USER_ACCUM_PS_0, 0);
+      si_pm4_set_reg(pm4, R_00B0CC_SPI_SHADER_USER_ACCUM_PS_1, 0);
+      si_pm4_set_reg(pm4, R_00B0D0_SPI_SHADER_USER_ACCUM_PS_2, 0);
+      si_pm4_set_reg(pm4, R_00B0D4_SPI_SHADER_USER_ACCUM_PS_3, 0);
+      si_pm4_set_reg(pm4, R_00B1C8_SPI_SHADER_USER_ACCUM_VS_0, 0);
+      si_pm4_set_reg(pm4, R_00B1CC_SPI_SHADER_USER_ACCUM_VS_1, 0);
+      si_pm4_set_reg(pm4, R_00B1D0_SPI_SHADER_USER_ACCUM_VS_2, 0);
+      si_pm4_set_reg(pm4, R_00B1D4_SPI_SHADER_USER_ACCUM_VS_3, 0);
+      si_pm4_set_reg(pm4, R_00B2C8_SPI_SHADER_USER_ACCUM_ESGS_0, 0);
+      si_pm4_set_reg(pm4, R_00B2CC_SPI_SHADER_USER_ACCUM_ESGS_1, 0);
+      si_pm4_set_reg(pm4, R_00B2D0_SPI_SHADER_USER_ACCUM_ESGS_2, 0);
+      si_pm4_set_reg(pm4, R_00B2D4_SPI_SHADER_USER_ACCUM_ESGS_3, 0);
+      si_pm4_set_reg(pm4, R_00B4C8_SPI_SHADER_USER_ACCUM_LSHS_0, 0);
+      si_pm4_set_reg(pm4, R_00B4CC_SPI_SHADER_USER_ACCUM_LSHS_1, 0);
+      si_pm4_set_reg(pm4, R_00B4D0_SPI_SHADER_USER_ACCUM_LSHS_2, 0);
+      si_pm4_set_reg(pm4, R_00B4D4_SPI_SHADER_USER_ACCUM_LSHS_3, 0);
    }
 
    sctx->cs_preamble_state = pm4;