radeonsi/compute: Use the compiler's COMPUTE_PGM_RSRC* register values
authorTom Stellard <thomas.stellard@amd.com>
Mon, 16 Nov 2015 20:00:12 +0000 (20:00 +0000)
committerTom Stellard <thomas.stellard@amd.com>
Wed, 25 Nov 2015 16:03:05 +0000 (11:03 -0500)
The compiler has more information and is able to optimize the bits
it sets in these registers.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
CC: <mesa-stable@lists.freedesktop.org>
src/gallium/drivers/radeonsi/si_compute.c
src/gallium/drivers/radeonsi/si_shader.c

index 2d551dd0e6bec90b31c8c339d38e302aa3dc31f6..e134b37b95d6b3e1ac2fe370bb4641c61ae170a0 100644 (file)
 
 #define MAX_GLOBAL_BUFFERS 20
 
-/* XXX: Even though we don't pass the scratch buffer via user sgprs any more
- * LLVM still expects that we specify 4 USER_SGPRS so it can remain compatible
- * with older mesa. */
-#define NUM_USER_SGPRS 4
-
 struct si_compute {
        struct si_context *ctx;
 
@@ -238,7 +233,6 @@ static void si_launch_grid(
        uint64_t kernel_args_va;
        uint64_t scratch_buffer_va = 0;
        uint64_t shader_va;
-       unsigned arg_user_sgpr_count = NUM_USER_SGPRS;
        unsigned i;
        struct si_shader *shader = &program->shader;
        unsigned lds_blocks;
@@ -366,20 +360,7 @@ static void si_launch_grid(
        si_pm4_set_reg(pm4, R_00B830_COMPUTE_PGM_LO, shader_va >> 8);
        si_pm4_set_reg(pm4, R_00B834_COMPUTE_PGM_HI, shader_va >> 40);
 
-       si_pm4_set_reg(pm4, R_00B848_COMPUTE_PGM_RSRC1,
-               /* We always use at least 3 VGPRS, these come from
-                * TIDIG_COMP_CNT.
-                * XXX: The compiler should account for this.
-                */
-               S_00B848_VGPRS((MAX2(3, shader->num_vgprs) - 1) / 4)
-               /* We always use at least 4 + arg_user_sgpr_count.  The 4 extra
-                * sgprs are from TGID_X_EN, TGID_Y_EN, TGID_Z_EN, TG_SIZE_EN
-                * XXX: The compiler should account for this.
-                */
-               |  S_00B848_SGPRS(((MAX2(4 + arg_user_sgpr_count,
-                                       shader->num_sgprs)) - 1) / 8)
-               |  S_00B028_FLOAT_MODE(shader->float_mode))
-               ;
+       si_pm4_set_reg(pm4, R_00B848_COMPUTE_PGM_RSRC1, shader->rsrc1);
 
        lds_blocks = shader->lds_size;
        /* XXX: We are over allocating LDS.  For SI, the shader reports LDS in
@@ -395,17 +376,10 @@ static void si_launch_grid(
 
        assert(lds_blocks <= 0xFF);
 
-       si_pm4_set_reg(pm4, R_00B84C_COMPUTE_PGM_RSRC2,
-               S_00B84C_SCRATCH_EN(shader->scratch_bytes_per_wave > 0)
-               | S_00B84C_USER_SGPR(arg_user_sgpr_count)
-               | S_00B84C_TGID_X_EN(1)
-               | S_00B84C_TGID_Y_EN(1)
-               | S_00B84C_TGID_Z_EN(1)
-               | S_00B84C_TG_SIZE_EN(1)
-               | S_00B84C_TIDIG_COMP_CNT(2)
-               | S_00B84C_LDS_SIZE(lds_blocks)
-               | S_00B84C_EXCP_EN(0))
-               ;
+       shader->rsrc2 &= C_00B84C_LDS_SIZE;
+       shader->rsrc2 |=  S_00B84C_LDS_SIZE(lds_blocks);
+
+       si_pm4_set_reg(pm4, R_00B84C_COMPUTE_PGM_RSRC2, shader->rsrc2);
        si_pm4_set_reg(pm4, R_00B854_COMPUTE_RESOURCE_LIMITS, 0);
 
        si_pm4_set_reg(pm4, R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0,
index 354d0646b995e483da22a44cad31b016e77e65cd..14f12dfb8213fa29e83306ce6ad6c7e22891c632 100644 (file)
@@ -3745,12 +3745,14 @@ void si_shader_binary_read_config(const struct si_screen *sscreen,
                        shader->num_sgprs = MAX2(shader->num_sgprs, (G_00B028_SGPRS(value) + 1) * 8);
                        shader->num_vgprs = MAX2(shader->num_vgprs, (G_00B028_VGPRS(value) + 1) * 4);
                        shader->float_mode =  G_00B028_FLOAT_MODE(value);
+                       shader->rsrc1 = value;
                        break;
                case R_00B02C_SPI_SHADER_PGM_RSRC2_PS:
                        shader->lds_size = MAX2(shader->lds_size, G_00B02C_EXTRA_LDS_SIZE(value));
                        break;
                case R_00B84C_COMPUTE_PGM_RSRC2:
                        shader->lds_size = MAX2(shader->lds_size, G_00B84C_LDS_SIZE(value));
+                       shader->rsrc2 = value;
                        break;
                case R_0286CC_SPI_PS_INPUT_ENA:
                        shader->spi_ps_input_ena = value;