r600: use DX10_CLAMP bit in shader setup
authorRoland Scheidegger <sroland@vmware.com>
Thu, 9 Nov 2017 18:41:29 +0000 (19:41 +0100)
committerRoland Scheidegger <sroland@vmware.com>
Wed, 15 Nov 2017 02:13:46 +0000 (03:13 +0100)
The docs are not very concise in what this really does, however both
Alex Deucher and Nicolai Hähnle suggested this only really affects instructions
using the CLAMP output modifier, and I've confirmed that with the newly
changed piglit isinf_and_isnan test.
So, with this bit set, if an instruction has the CLAMP modifier bit (which
clamps to [0,1]) set, then NaNs will be converted to zero, otherwise the result
will be NaN.
D3D10 would require this, glsl doesn't have modifiers (with mesa
clamp(x,0,1) would get converted to such a modifier) coupled with a
whatever-floats-your-boat specified NaN behavior, but the clamp behavior
should probably always be used (this also matches what a decomposition into
min(1.0, max(x, 0.0)) would do, if min/max also adhere to the ieee spec of
picking the non-nan result).
Some apps may in fact rely on this, as this prevents misrenderings in
This War of Mine since using ieee muls
(ce7a045feeef8cad155f1c9aa07f166e146e3d00), without having to use clamped
rcp opcode, which would also fix this bug there.
radeonsi also seems to set this bit nowadays if I see that righ (albeit the
llvm amdgpu code comment now says "Make clamp modifier on NaN input returns 0"
instead of "Do not clamp NAN to 0" since it was changed, which also looks
a bit misleading).

v2: set it in all shader stages.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103544

Reviewed-by: Dave Airlie <airlied@redhat.com>
src/gallium/drivers/r600/evergreen_state.c
src/gallium/drivers/r600/r600_state.c

index b02d7eeca62acc14f362b14a0e5c69dd4d91f268..7c2dfa092d39c3fe951fe2bd5f3f32af1f9fb869 100644 (file)
@@ -3244,6 +3244,7 @@ void evergreen_update_ps_state(struct pipe_context *ctx, struct r600_pipe_shader
        r600_store_value(cb, /* R_028844_SQ_PGM_RESOURCES_PS */
                         S_028844_NUM_GPRS(rshader->bc.ngpr) |
                         S_028844_PRIME_CACHE_ON_DRAW(1) |
+                        S_028844_DX10_CLAMP(1) |
                         S_028844_STACK_SIZE(rshader->bc.nstack));
        /* After that, the NOP relocation packet must be emitted (shader->bo, RADEON_USAGE_READ). */
 
@@ -3264,6 +3265,7 @@ void evergreen_update_es_state(struct pipe_context *ctx, struct r600_pipe_shader
 
        r600_store_context_reg(cb, R_028890_SQ_PGM_RESOURCES_ES,
                               S_028890_NUM_GPRS(rshader->bc.ngpr) |
+                              S_028890_DX10_CLAMP(1) |
                               S_028890_STACK_SIZE(rshader->bc.nstack));
        r600_store_context_reg(cb, R_02888C_SQ_PGM_START_ES,
                               shader->bo->gpu_address >> 8);
@@ -3326,6 +3328,7 @@ void evergreen_update_gs_state(struct pipe_context *ctx, struct r600_pipe_shader
 
        r600_store_context_reg(cb, R_028878_SQ_PGM_RESOURCES_GS,
                               S_028878_NUM_GPRS(rshader->bc.ngpr) |
+                              S_028878_DX10_CLAMP(1) |
                               S_028878_STACK_SIZE(rshader->bc.nstack));
        r600_store_context_reg(cb, R_028874_SQ_PGM_START_GS,
                               shader->bo->gpu_address >> 8);
@@ -3366,6 +3369,7 @@ void evergreen_update_vs_state(struct pipe_context *ctx, struct r600_pipe_shader
                               S_0286C4_VS_EXPORT_COUNT(nparams - 1));
        r600_store_context_reg(cb, R_028860_SQ_PGM_RESOURCES_VS,
                               S_028860_NUM_GPRS(rshader->bc.ngpr) |
+                              S_028860_DX10_CLAMP(1) |
                               S_028860_STACK_SIZE(rshader->bc.nstack));
        if (rshader->vs_position_window_space) {
                r600_store_context_reg(cb, R_028818_PA_CL_VTE_CNTL,
@@ -3400,6 +3404,7 @@ void evergreen_update_hs_state(struct pipe_context *ctx, struct r600_pipe_shader
        r600_init_command_buffer(cb, 32);
        r600_store_context_reg(cb, R_0288BC_SQ_PGM_RESOURCES_HS,
                               S_0288BC_NUM_GPRS(rshader->bc.ngpr) |
+                              S_0288BC_DX10_CLAMP(1) |
                               S_0288BC_STACK_SIZE(rshader->bc.nstack));
        r600_store_context_reg(cb, R_0288B8_SQ_PGM_START_HS,
                               shader->bo->gpu_address >> 8);
@@ -3413,6 +3418,7 @@ void evergreen_update_ls_state(struct pipe_context *ctx, struct r600_pipe_shader
        r600_init_command_buffer(cb, 32);
        r600_store_context_reg(cb, R_0288D4_SQ_PGM_RESOURCES_LS,
                               S_0288D4_NUM_GPRS(rshader->bc.ngpr) |
+                              S_0288D4_DX10_CLAMP(1) |
                               S_0288D4_STACK_SIZE(rshader->bc.nstack));
        r600_store_context_reg(cb, R_0288D0_SQ_PGM_START_LS,
                               shader->bo->gpu_address >> 8);
index 0c3315374607e8fb066e1606ee47a94070c0afea..0e266aef422e34fb40e30d2dab30d3e292bc892c 100644 (file)
@@ -2549,6 +2549,12 @@ void r600_update_ps_state(struct pipe_context *ctx, struct r600_pipe_shader *sha
        r600_store_context_reg_seq(cb, R_028850_SQ_PGM_RESOURCES_PS, 2);
        r600_store_value(cb, /* R_028850_SQ_PGM_RESOURCES_PS*/
                         S_028850_NUM_GPRS(rshader->bc.ngpr) |
+       /*
+        * docs are misleading about the dx10_clamp bit. This only affects
+        * instructions using CLAMP dst modifier, in which case they will
+        * return 0 with this set for a NaN (otherwise NaN).
+        */
+                        S_028850_DX10_CLAMP(1) |
                         S_028850_STACK_SIZE(rshader->bc.nstack) |
                         S_028850_UNCACHED_FIRST_INST(ufi));
        r600_store_value(cb, exports_ps); /* R_028854_SQ_PGM_EXPORTS_PS */
@@ -2598,6 +2604,7 @@ void r600_update_vs_state(struct pipe_context *ctx, struct r600_pipe_shader *sha
                               S_0286C4_VS_EXPORT_COUNT(nparams - 1));
        r600_store_context_reg(cb, R_028868_SQ_PGM_RESOURCES_VS,
                               S_028868_NUM_GPRS(rshader->bc.ngpr) |
+                              S_028868_DX10_CLAMP(1) |
                               S_028868_STACK_SIZE(rshader->bc.nstack));
        if (rshader->vs_position_window_space) {
                r600_store_context_reg(cb, R_028818_PA_CL_VTE_CNTL,
@@ -2682,6 +2689,7 @@ void r600_update_gs_state(struct pipe_context *ctx, struct r600_pipe_shader *sha
 
        r600_store_context_reg(cb, R_02887C_SQ_PGM_RESOURCES_GS,
                               S_02887C_NUM_GPRS(rshader->bc.ngpr) |
+                              S_02887C_DX10_CLAMP(1) |
                               S_02887C_STACK_SIZE(rshader->bc.nstack));
        r600_store_context_reg(cb, R_02886C_SQ_PGM_START_GS, 0);
        /* After that, the NOP relocation packet must be emitted (shader->bo, RADEON_USAGE_READ). */
@@ -2696,6 +2704,7 @@ void r600_update_es_state(struct pipe_context *ctx, struct r600_pipe_shader *sha
 
        r600_store_context_reg(cb, R_028890_SQ_PGM_RESOURCES_ES,
                               S_028890_NUM_GPRS(rshader->bc.ngpr) |
+                              S_028890_DX10_CLAMP(1) |
                               S_028890_STACK_SIZE(rshader->bc.nstack));
        r600_store_context_reg(cb, R_028880_SQ_PGM_START_ES, 0);
        /* After that, the NOP relocation packet must be emitted (shader->bo, RADEON_USAGE_READ). */