[Mesa-dev] [PATCH 5/5] radeonsi: implement GL_KHR_blend_equation_advanced
Marek Olšák
maraeo at gmail.com
Fri Mar 23 19:36:03 UTC 2018
From: Marek Olšák <marek.olsak at amd.com>
MSAA is supported using sample shading. Layered rendering and all texture
targets are also supported.
---
docs/features.txt | 2 +-
docs/relnotes/18.1.0.html | 1 +
src/gallium/drivers/radeonsi/si_blit.c | 8 +++
src/gallium/drivers/radeonsi/si_descriptors.c | 87 +++++++++++++++++++++--
src/gallium/drivers/radeonsi/si_get.c | 2 +-
src/gallium/drivers/radeonsi/si_pipe.h | 9 +++
src/gallium/drivers/radeonsi/si_shader.c | 4 +-
src/gallium/drivers/radeonsi/si_shader.h | 3 +
src/gallium/drivers/radeonsi/si_shader_internal.h | 1 +
src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c | 59 +++++++++++++++
src/gallium/drivers/radeonsi/si_state.c | 19 +++--
src/gallium/drivers/radeonsi/si_state.h | 8 +++
src/gallium/drivers/radeonsi/si_state_binning.c | 2 +-
src/gallium/drivers/radeonsi/si_state_shaders.c | 18 +++++
14 files changed, 205 insertions(+), 18 deletions(-)
diff --git a/docs/features.txt b/docs/features.txt
index 5eae34bf0df..d579d245eb4 100644
--- a/docs/features.txt
+++ b/docs/features.txt
@@ -262,21 +262,21 @@ GLES3.1, GLSL ES 3.1 -- all DONE: i965/hsw+, nvc0, r600, radeonsi
Additional functionality not covered above:
glMemoryBarrierByRegion DONE
glGetTexLevelParameter[fi]v - needs updates DONE
glGetBooleani_v - restrict to GLES enums
gl_HelperInvocation support DONE (i965, r600)
GLES3.2, GLSL ES 3.2 -- all DONE: i965/gen9+
GL_EXT_color_buffer_float DONE (all drivers)
- GL_KHR_blend_equation_advanced DONE (i965, nvc0)
+ GL_KHR_blend_equation_advanced DONE (i965, nvc0, radeonsi)
GL_KHR_debug DONE (all drivers)
GL_KHR_robustness DONE (i965, nvc0, radeonsi)
GL_KHR_texture_compression_astc_ldr DONE (freedreno, i965/gen9+)
GL_OES_copy_image DONE (all drivers)
GL_OES_draw_buffers_indexed DONE (all drivers that support GL_ARB_draw_buffers_blend)
GL_OES_draw_elements_base_vertex DONE (all drivers)
GL_OES_geometry_shader DONE (i965/hsw+, nvc0, radeonsi)
GL_OES_gpu_shader5 DONE (all drivers that support GL_ARB_gpu_shader5)
GL_OES_primitive_bounding_box DONE (i965/gen7+, nvc0, radeonsi)
GL_OES_sample_shading DONE (i965, nvc0, r600, radeonsi)
diff --git a/docs/relnotes/18.1.0.html b/docs/relnotes/18.1.0.html
index 3e119078731..a89861d2bda 100644
--- a/docs/relnotes/18.1.0.html
+++ b/docs/relnotes/18.1.0.html
@@ -43,20 +43,21 @@ TBD.
Note: some of the new features are only available with certain drivers.
</p>
<ul>
<li>OpenGL 3.1 with ARB_compatibility on nv50, nvc0, r600, radeonsi, softpipe, llvmpipe, svga</li>
<li>GL_ARB_bindless_texture on nvc0/maxwell+</li>
<li>GL_EXT_semaphore on radeonsi</li>
<li>GL_EXT_semaphore_fd on radeonsi</li>
<li>GL_EXT_shader_framebuffer_fetch on i965 on desktop GL (GLES was already supported)</li>
<li>GL_EXT_shader_framebuffer_fetch_non_coherent on i965</li>
+<li>GL_KHR_blend_equation_advanced on radeonsi</li>
<li>Disk shader cache support for i965 enabled by default</li>
</ul>
<h2>Bug fixes</h2>
<ul>
TBD
</ul>
<h2>Changes</h2>
diff --git a/src/gallium/drivers/radeonsi/si_blit.c b/src/gallium/drivers/radeonsi/si_blit.c
index f1c4f6d1e72..d9d489825f8 100644
--- a/src/gallium/drivers/radeonsi/si_blit.c
+++ b/src/gallium/drivers/radeonsi/si_blit.c
@@ -799,20 +799,28 @@ void si_decompress_textures(struct si_context *sctx, unsigned shader_mask)
si_decompress_resident_textures(sctx);
if (sctx->uses_bindless_images)
si_decompress_resident_images(sctx);
} else if (shader_mask & (1 << PIPE_SHADER_COMPUTE)) {
if (sctx->cs_shader_state.program->uses_bindless_samplers)
si_decompress_resident_textures(sctx);
if (sctx->cs_shader_state.program->uses_bindless_images)
si_decompress_resident_images(sctx);
}
+ if (sctx->ps_uses_fbfetch) {
+ struct pipe_surface *cb0 = sctx->framebuffer.state.cbufs[0];
+ si_decompress_color_texture(sctx,
+ (struct r600_texture*)cb0->texture,
+ cb0->u.tex.first_layer,
+ cb0->u.tex.last_layer);
+ }
+
si_check_render_feedback(sctx);
}
/* Helper for decompressing a portion of a color or depth resource before
* blitting if any decompression is needed.
* The driver doesn't decompress resources automatically while u_blitter is
* rendering. */
static void si_decompress_subresource(struct pipe_context *ctx,
struct pipe_resource *tex,
unsigned planes, unsigned level,
diff --git a/src/gallium/drivers/radeonsi/si_descriptors.c b/src/gallium/drivers/radeonsi/si_descriptors.c
index 07d1420d8ee..38befa445d2 100644
--- a/src/gallium/drivers/radeonsi/si_descriptors.c
+++ b/src/gallium/drivers/radeonsi/si_descriptors.c
@@ -680,21 +680,21 @@ si_mark_image_range_valid(const struct pipe_image_view *view)
assert(res && res->b.b.target == PIPE_BUFFER);
util_range_add(&res->valid_buffer_range,
view->u.buf.offset,
view->u.buf.offset + view->u.buf.size);
}
static void si_set_shader_image_desc(struct si_context *ctx,
const struct pipe_image_view *view,
bool skip_decompress,
- uint32_t *desc)
+ uint32_t *desc, uint32_t *fmask_desc)
{
struct si_screen *screen = ctx->screen;
struct r600_resource *res;
res = (struct r600_resource *)view->resource;
if (res->b.b.target == PIPE_BUFFER) {
if (view->access & PIPE_IMAGE_ACCESS_WRITE)
si_mark_image_range_valid(view);
@@ -708,27 +708,28 @@ static void si_set_shader_image_desc(struct si_context *ctx,
struct r600_texture *tex = (struct r600_texture *)res;
unsigned level = view->u.tex.level;
unsigned width, height, depth, hw_level;
bool uses_dcc = vi_dcc_enabled(tex, level);
unsigned access = view->access;
/* Clear the write flag when writes can't occur.
* Note that DCC_DECOMPRESS for MSAA doesn't work in some cases,
* so we don't wanna trigger it.
*/
- if (tex->is_depth || tex->resource.b.b.nr_samples >= 2) {
+ if (tex->is_depth ||
+ (!fmask_desc && tex->fmask.size != 0)) {
assert(!"Z/S and MSAA image stores are not supported");
access &= ~PIPE_IMAGE_ACCESS_WRITE;
}
assert(!tex->is_depth);
- assert(tex->fmask.size == 0);
+ assert(fmask_desc || tex->fmask.size == 0);
if (uses_dcc && !skip_decompress &&
(view->access & PIPE_IMAGE_ACCESS_WRITE ||
!vi_dcc_formats_compatible(res->b.b.format, view->format))) {
/* If DCC can't be disabled, at least decompress it.
* The decompression is relatively cheap if the surface
* has been decompressed already.
*/
if (!si_texture_disable_dcc(&ctx->b, tex))
ctx->b.decompress_dcc(&ctx->b.b, tex);
@@ -755,21 +756,21 @@ static void si_set_shader_image_desc(struct si_context *ctx,
hw_level = 0;
}
si_make_texture_descriptor(screen, tex,
false, res->b.b.target,
view->format, swizzle,
hw_level, hw_level,
view->u.tex.first_layer,
view->u.tex.last_layer,
width, height, depth,
- desc, NULL);
+ desc, fmask_desc);
si_set_mutable_tex_desc_fields(screen, tex,
&tex->surface.u.legacy.level[level],
level, level,
util_format_get_blockwidth(view->format),
false, desc);
}
}
static void si_set_shader_image(struct si_context *ctx,
unsigned shader,
@@ -785,21 +786,21 @@ static void si_set_shader_image(struct si_context *ctx,
if (!view || !view->resource) {
si_disable_shader_image(ctx, shader, slot);
return;
}
res = (struct r600_resource *)view->resource;
if (&images->views[slot] != view)
util_copy_image_view(&images->views[slot], view);
- si_set_shader_image_desc(ctx, view, skip_decompress, desc);
+ si_set_shader_image_desc(ctx, view, skip_decompress, desc, NULL);
if (res->b.b.target == PIPE_BUFFER) {
images->needs_color_decompress_mask &= ~(1 << slot);
res->bind_history |= PIPE_BIND_SHADER_IMAGE;
} else {
struct r600_texture *tex = (struct r600_texture *)res;
unsigned level = view->u.tex.level;
if (color_needs_decompression(tex)) {
images->needs_color_decompress_mask |= 1 << slot;
@@ -863,20 +864,91 @@ si_images_update_needs_color_decompress_mask(struct si_images *images)
if (color_needs_decompression(rtex)) {
images->needs_color_decompress_mask |= 1 << i;
} else {
images->needs_color_decompress_mask &= ~(1 << i);
}
}
}
}
+void si_update_ps_colorbuf0_slot(struct si_context *sctx)
+{
+ struct si_buffer_resources *buffers = &sctx->rw_buffers;
+ struct si_descriptors *descs = &sctx->descriptors[SI_DESCS_RW_BUFFERS];
+ unsigned slot = SI_PS_IMAGE_COLORBUF0;
+ struct pipe_surface *surf = NULL;
+
+ /* si_texture_disable_dcc can get us here again. */
+ if (sctx->blitter->running)
+ return;
+
+ /* See whether FBFETCH is used and color buffer 0 is set. */
+ if (sctx->ps_shader.cso &&
+ sctx->ps_shader.cso->info.opcode_count[TGSI_OPCODE_FBFETCH] &&
+ sctx->framebuffer.state.nr_cbufs &&
+ sctx->framebuffer.state.cbufs[0])
+ surf = sctx->framebuffer.state.cbufs[0];
+
+ /* Return if FBFETCH transitions from disabled to disabled. */
+ if (!buffers->buffers[slot] && !surf)
+ return;
+
+ sctx->ps_uses_fbfetch = surf != NULL;
+ si_update_ps_iter_samples(sctx);
+
+ if (surf) {
+ struct r600_texture *tex = (struct r600_texture*)surf->texture;
+ struct pipe_image_view view;
+
+ assert(tex);
+ assert(!tex->is_depth);
+
+ /* Disable DCC, because the texture is used as both a sampler
+ * and color buffer.
+ */
+ si_texture_disable_dcc(&sctx->b, tex);
+
+ if (tex->resource.b.b.nr_samples <= 1 && tex->cmask_buffer) {
+ /* Disable CMASK. */
+ assert(tex->cmask_buffer != &tex->resource);
+ si_eliminate_fast_color_clear(&sctx->b, tex);
+ si_texture_discard_cmask(sctx->screen, tex);
+ }
+
+ view.resource = surf->texture;
+ view.format = surf->format;
+ view.access = PIPE_IMAGE_ACCESS_READ;
+ view.u.tex.first_layer = surf->u.tex.first_layer;
+ view.u.tex.last_layer = surf->u.tex.last_layer;
+ view.u.tex.level = surf->u.tex.level;
+
+ /* Set the descriptor. */
+ uint32_t *desc = descs->list + slot*4;
+ memset(desc, 0, 16 * 4);
+ si_set_shader_image_desc(sctx, &view, true, desc, desc + 8);
+
+ pipe_resource_reference(&buffers->buffers[slot], &tex->resource.b.b);
+ radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
+ &tex->resource, RADEON_USAGE_READ,
+ RADEON_PRIO_SHADER_RW_IMAGE);
+ buffers->enabled_mask |= 1u << slot;
+ } else {
+ /* Clear the descriptor. */
+ memset(descs->list + slot*4, 0, 8*4);
+ pipe_resource_reference(&buffers->buffers[slot], NULL);
+ buffers->enabled_mask &= ~(1u << slot);
+ }
+
+ sctx->descriptors_dirty |= 1u << SI_DESCS_RW_BUFFERS;
+}
+
/* SAMPLER STATES */
static void si_bind_sampler_states(struct pipe_context *ctx,
enum pipe_shader_type shader,
unsigned start, unsigned count, void **states)
{
struct si_context *sctx = (struct si_context *)ctx;
struct si_samplers *samplers = &sctx->samplers[shader];
struct si_descriptors *desc = si_sampler_and_image_descriptors(sctx, shader);
struct si_sampler_state **sstates = (struct si_sampler_state**)states;
@@ -1848,21 +1920,21 @@ static void si_update_bindless_image_descriptor(struct si_context *sctx,
unsigned desc_slot_offset = img_handle->desc_slot * 16;
struct pipe_image_view *view = &img_handle->view;
uint32_t desc_list[8];
if (view->resource->target == PIPE_BUFFER)
return;
memcpy(desc_list, desc->list + desc_slot_offset,
sizeof(desc_list));
si_set_shader_image_desc(sctx, view, true,
- desc->list + desc_slot_offset);
+ desc->list + desc_slot_offset, NULL);
if (memcmp(desc_list, desc->list + desc_slot_offset,
sizeof(desc_list))) {
img_handle->desc_dirty = true;
sctx->bindless_descriptors_dirty = true;
}
}
static void si_update_all_resident_texture_descriptors(struct si_context *sctx)
{
@@ -1914,20 +1986,21 @@ void si_update_all_texture_descriptors(struct si_context *sctx)
continue;
si_set_sampler_view(sctx, shader, i,
samplers->views[i], true);
}
si_update_shader_needs_decompress_mask(sctx, shader);
}
si_update_all_resident_texture_descriptors(sctx);
+ si_update_ps_colorbuf0_slot(sctx);
}
/* SHADER USER DATA */
static void si_mark_shader_pointers_dirty(struct si_context *sctx,
unsigned shader)
{
sctx->shader_pointers_dirty |=
u_bit_consecutive(SI_DESCS_FIRST_SHADER + shader * SI_NUM_SHADER_DESCS,
SI_NUM_SHADER_DESCS);
@@ -2453,21 +2526,21 @@ static uint64_t si_create_image_handle(struct pipe_context *ctx,
if (!view || !view->resource)
return 0;
img_handle = CALLOC_STRUCT(si_image_handle);
if (!img_handle)
return 0;
memset(desc_list, 0, sizeof(desc_list));
si_init_descriptor_list(&desc_list[0], 8, 1, null_image_descriptor);
- si_set_shader_image_desc(sctx, view, false, &desc_list[0]);
+ si_set_shader_image_desc(sctx, view, false, &desc_list[0], NULL);
img_handle->desc_slot = si_create_bindless_descriptor(sctx, desc_list,
sizeof(desc_list));
if (!img_handle->desc_slot) {
FREE(img_handle);
return 0;
}
handle = img_handle->desc_slot;
diff --git a/src/gallium/drivers/radeonsi/si_get.c b/src/gallium/drivers/radeonsi/si_get.c
index b4ca5bea943..327d8a28c38 100644
--- a/src/gallium/drivers/radeonsi/si_get.c
+++ b/src/gallium/drivers/radeonsi/si_get.c
@@ -183,20 +183,21 @@ static int si_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
case PIPE_CAP_MEMOBJ:
case PIPE_CAP_LOAD_CONSTBUF:
case PIPE_CAP_INT64:
case PIPE_CAP_INT64_DIVMOD:
case PIPE_CAP_TGSI_CLOCK:
case PIPE_CAP_CAN_BIND_CONST_BUFFER_AS_VERTEX:
case PIPE_CAP_ALLOW_MAPPED_BUFFERS_DURING_EXECUTION:
case PIPE_CAP_TGSI_ANY_REG_AS_ADDRESS:
case PIPE_CAP_SIGNED_VERTEX_BUFFER_OFFSET:
case PIPE_CAP_TGSI_VOTE:
+ case PIPE_CAP_TGSI_FS_FBFETCH:
return 1;
case PIPE_CAP_TGSI_BALLOT:
return HAVE_LLVM >= 0x0500;
case PIPE_CAP_RESOURCE_FROM_USER_MEMORY:
return !SI_BIG_ENDIAN && sscreen->info.has_userptr;
case PIPE_CAP_DEVICE_RESET_STATUS_QUERY:
return (sscreen->info.drm_major == 2 &&
@@ -259,21 +260,20 @@ static int si_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
/* Unsupported features. */
case PIPE_CAP_BUFFER_SAMPLER_VIEW_RGBA_ONLY:
case PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT:
case PIPE_CAP_TGSI_CAN_COMPACT_CONSTANTS:
case PIPE_CAP_USER_VERTEX_BUFFERS:
case PIPE_CAP_FAKE_SW_MSAA:
case PIPE_CAP_TEXTURE_GATHER_OFFSETS:
case PIPE_CAP_VERTEXID_NOBASE:
case PIPE_CAP_PRIMITIVE_RESTART_FOR_PATCHES:
case PIPE_CAP_MAX_WINDOW_RECTANGLES:
- case PIPE_CAP_TGSI_FS_FBFETCH:
case PIPE_CAP_TGSI_MUL_ZERO_WINS:
case PIPE_CAP_UMA:
case PIPE_CAP_POLYGON_MODE_FILL_RECTANGLE:
case PIPE_CAP_POST_DEPTH_COVERAGE:
case PIPE_CAP_TILE_RASTER_ORDER:
case PIPE_CAP_MAX_COMBINED_SHADER_OUTPUT_RESOURCES:
case PIPE_CAP_CONTEXT_PRIORITY_MASK:
return 0;
case PIPE_CAP_FENCE_SIGNAL:
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index dbb04ed7e45..e3d45ef6c3b 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -525,20 +525,21 @@ struct si_context {
unsigned num_vs_blit_sgprs;
uint32_t vs_blit_sh_data[SI_VS_BLIT_SGPRS_POS_TEXCOORD];
/* Vertex and index buffers. */
bool vertex_buffers_dirty;
bool vertex_buffer_pointer_dirty;
struct pipe_vertex_buffer vertex_buffer[SI_NUM_VERTEX_BUFFERS];
/* MSAA config state. */
int ps_iter_samples;
+ bool ps_uses_fbfetch;
bool smoothing_enabled;
/* DB render state. */
unsigned ps_db_shader_control;
unsigned dbcb_copy_sample;
bool dbcb_depth_copy_enabled:1;
bool dbcb_stencil_copy_enabled:1;
bool db_flush_depth_inplace:1;
bool db_flush_stencil_inplace:1;
bool db_depth_clear:1;
@@ -924,11 +925,19 @@ si_htile_enabled(struct r600_texture *tex, unsigned level)
return tex->htile_offset && level == 0;
}
static inline bool
vi_tc_compat_htile_enabled(struct r600_texture *tex, unsigned level)
{
assert(!tex->tc_compatible_htile || tex->htile_offset);
return tex->tc_compatible_htile && level == 0;
}
+static inline unsigned si_get_ps_iter_samples(struct si_context *sctx)
+{
+ if (sctx->ps_uses_fbfetch)
+ return sctx->framebuffer.nr_samples;
+
+ return sctx->ps_iter_samples;
+}
+
#endif
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 08b071e810b..ae98e102eae 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -1919,21 +1919,21 @@ void si_llvm_load_input_fs(
static void declare_input_fs(
struct si_shader_context *ctx,
unsigned input_index,
const struct tgsi_full_declaration *decl,
LLVMValueRef out[4])
{
si_llvm_load_input_fs(ctx, input_index, out);
}
-static LLVMValueRef get_sample_id(struct si_shader_context *ctx)
+LLVMValueRef si_get_sample_id(struct si_shader_context *ctx)
{
return si_unpack_param(ctx, SI_PARAM_ANCILLARY, 8, 4);
}
static LLVMValueRef get_base_vertex(struct ac_shader_abi *abi)
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
/* For non-indexed draws, the base vertex set by the driver
* (for direct draws) or the CP (for indirect draws) is the
@@ -2142,21 +2142,21 @@ void si_load_system_value(struct si_shader_context *ctx,
};
value = lp_build_gather_values(&ctx->gallivm, pos, 4);
break;
}
case TGSI_SEMANTIC_FACE:
value = ctx->abi.front_face;
break;
case TGSI_SEMANTIC_SAMPLEID:
- value = get_sample_id(ctx);
+ value = si_get_sample_id(ctx);
break;
case TGSI_SEMANTIC_SAMPLEPOS: {
LLVMValueRef pos[4] = {
LLVMGetParam(ctx->main_fn, SI_PARAM_POS_X_FLOAT),
LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Y_FLOAT),
LLVMConstReal(ctx->f32, 0),
LLVMConstReal(ctx->f32, 0)
};
pos[0] = lp_build_emit_llvm_unary(&ctx->bld_base,
diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h
index f58978989d4..f598b762e1d 100644
--- a/src/gallium/drivers/radeonsi/si_shader.h
+++ b/src/gallium/drivers/radeonsi/si_shader.h
@@ -534,20 +534,23 @@ struct si_shader_key {
struct {
/* One byte for every input: SI_FIX_FETCH_* enums. */
uint8_t vs_fix_fetch[SI_MAX_ATTRIBS];
union {
uint64_t ff_tcs_inputs_to_copy; /* for fixed-func TCS */
/* When PS needs PrimID and GS is disabled. */
unsigned vs_export_prim_id:1;
struct {
unsigned interpolate_at_sample_force_center:1;
+ unsigned fbfetch_msaa;
+ unsigned fbfetch_is_1D;
+ unsigned fbfetch_layered;
} ps;
} u;
} mono;
/* Optimization flags for asynchronous compilation only. */
struct {
/* For HW VS (it can be VS, TES, GS) */
uint64_t kill_outputs; /* "get_unique_index" bits */
unsigned clip_disable:1;
diff --git a/src/gallium/drivers/radeonsi/si_shader_internal.h b/src/gallium/drivers/radeonsi/si_shader_internal.h
index 1bd52722413..941c6fc736d 100644
--- a/src/gallium/drivers/radeonsi/si_shader_internal.h
+++ b/src/gallium/drivers/radeonsi/si_shader_internal.h
@@ -293,20 +293,21 @@ void si_llvm_emit_store(struct lp_build_tgsi_context *bld_base,
#define NOOP_WAITCNT 0xf7f
#define LGKM_CNT 0x07f
#define VM_CNT 0xf70
LLVMValueRef si_get_indirect_index(struct si_shader_context *ctx,
const struct tgsi_ind_register *ind,
unsigned addr_mul, int rel_index);
LLVMValueRef si_get_bounded_indirect_index(struct si_shader_context *ctx,
const struct tgsi_ind_register *ind,
int rel_index, unsigned num);
+LLVMValueRef si_get_sample_id(struct si_shader_context *ctx);
void si_shader_context_init_alu(struct lp_build_tgsi_context *bld_base);
void si_shader_context_init_mem(struct si_shader_context *ctx);
LLVMValueRef si_load_sampler_desc(struct si_shader_context *ctx,
LLVMValueRef list, LLVMValueRef index,
enum ac_descriptor_type type);
LLVMValueRef si_load_image_desc(struct si_shader_context *ctx,
LLVMValueRef list, LLVMValueRef index,
enum ac_descriptor_type desc_type, bool dcc_off);
diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
index f5fa18fd38a..b90edddf2ef 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
@@ -1866,20 +1866,77 @@ static void si_llvm_emit_txqs(
samples = LLVMBuildLShr(ctx->ac.builder, samples,
LLVMConstInt(ctx->i32, 16, 0), "");
samples = LLVMBuildAnd(ctx->ac.builder, samples,
LLVMConstInt(ctx->i32, 0xf, 0), "");
samples = LLVMBuildShl(ctx->ac.builder, ctx->i32_1,
samples, "");
emit_data->output[emit_data->chan] = samples;
}
+static void si_llvm_emit_fbfetch(const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ struct ac_image_args args = {};
+ LLVMValueRef ptr, image, fmask, addr_vec;
+
+ /* Ignore src0, because KHR_blend_func_extended disallows multiple render
+ * targets.
+ */
+
+ /* Load the image descriptor. */
+ STATIC_ASSERT(SI_PS_IMAGE_COLORBUF0 % 2 == 0);
+ ptr = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
+ ptr = LLVMBuildPointerCast(ctx->ac.builder, ptr,
+ ac_array_in_const32_addr_space(ctx->v8i32), "");
+ image = ac_build_load_to_sgpr(&ctx->ac, ptr,
+ LLVMConstInt(ctx->i32, SI_PS_IMAGE_COLORBUF0 / 2, 0));
+
+ LLVMValueRef addr[4];
+ unsigned chan = 0;
+
+ addr[chan++] = si_unpack_param(ctx, SI_PARAM_POS_FIXED_PT, 0, 16);
+
+ if (!ctx->shader->key.mono.u.ps.fbfetch_is_1D)
+ addr[chan++] = si_unpack_param(ctx, SI_PARAM_POS_FIXED_PT, 16, 16);
+
+ /* Get the current render target layer index. */
+ if (ctx->shader->key.mono.u.ps.fbfetch_layered)
+ addr[chan++] = si_unpack_param(ctx, SI_PARAM_ANCILLARY, 16, 11);
+
+ if (ctx->shader->key.mono.u.ps.fbfetch_msaa)
+ addr[chan++] = si_get_sample_id(ctx);
+
+ while (chan < 4)
+ addr[chan++] = LLVMGetUndef(ctx->i32);
+
+ if (ctx->shader->key.mono.u.ps.fbfetch_msaa) {
+ fmask = ac_build_load_to_sgpr(&ctx->ac, ptr,
+ LLVMConstInt(ctx->i32, SI_PS_IMAGE_COLORBUF0_FMASK / 2, 0));
+
+ ac_apply_fmask_to_sample(&ctx->ac, fmask, addr, false);
+ }
+
+ addr_vec = ac_build_gather_values(&ctx->ac, addr, ARRAY_SIZE(addr));
+
+ args.opcode = ac_image_load;
+ args.resource = image;
+ args.addr = addr_vec;
+ args.dmask = 0xf;
+ args.da = ctx->shader->key.mono.u.ps.fbfetch_layered;
+
+ emit_data->output[emit_data->chan] =
+ ac_build_image_opcode(&ctx->ac, &args);
+}
+
static const struct lp_build_tgsi_action tex_action = {
.fetch_args = tex_fetch_args,
.emit = build_tex_intrinsic,
};
/**
* Setup actions for TGSI memory opcode, including texture opcodes.
*/
void si_shader_context_init_mem(struct si_shader_context *ctx)
{
@@ -1898,20 +1955,22 @@ void si_shader_context_init_mem(struct si_shader_context *ctx)
bld_base->op_actions[TGSI_OPCODE_TXF_LZ] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXL] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXL2] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXP] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXQ].fetch_args = txq_fetch_args;
bld_base->op_actions[TGSI_OPCODE_TXQ].emit = txq_emit;
bld_base->op_actions[TGSI_OPCODE_TG4] = tex_action;
bld_base->op_actions[TGSI_OPCODE_LODQ] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXQS].emit = si_llvm_emit_txqs;
+ bld_base->op_actions[TGSI_OPCODE_FBFETCH].emit = si_llvm_emit_fbfetch;
+
bld_base->op_actions[TGSI_OPCODE_LOAD].fetch_args = load_fetch_args;
bld_base->op_actions[TGSI_OPCODE_LOAD].emit = load_emit;
bld_base->op_actions[TGSI_OPCODE_STORE].fetch_args = store_fetch_args;
bld_base->op_actions[TGSI_OPCODE_STORE].emit = store_emit;
bld_base->op_actions[TGSI_OPCODE_RESQ].fetch_args = resq_fetch_args;
bld_base->op_actions[TGSI_OPCODE_RESQ].emit = resq_emit;
tmpl.fetch_args = atomic_fetch_args;
tmpl.emit = atomic_emit;
bld_base->op_actions[TGSI_OPCODE_ATOMUADD] = tmpl;
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index b92ec03a054..16ad1d4f823 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -2891,20 +2891,21 @@ static void si_set_framebuffer_state(struct pipe_context *ctx,
if (!surf->depth_initialized) {
si_init_depth_surface(sctx, surf);
}
if (vi_tc_compat_htile_enabled(zstex, surf->base.u.tex.level))
sctx->framebuffer.DB_has_shader_readable_metadata = true;
si_context_add_resource_size(ctx, surf->base.texture);
}
+ si_update_ps_colorbuf0_slot(sctx);
si_update_poly_offset_state(sctx);
si_mark_atom_dirty(sctx, &sctx->cb_render_state);
si_mark_atom_dirty(sctx, &sctx->framebuffer.atom);
if (sctx->screen->dpbb_allowed)
si_mark_atom_dirty(sctx, &sctx->dpbb_state);
if (sctx->framebuffer.any_dst_linear != old_any_dst_linear)
si_mark_atom_dirty(sctx, &sctx->msaa_config);
@@ -3353,40 +3354,41 @@ static void si_emit_msaa_config(struct si_context *sctx, struct r600_atom *atom)
if (setup_samples > 1) {
/* distance from the pixel center, indexed by log2(nr_samples) */
static unsigned max_dist[] = {
0, /* unused */
4, /* 2x MSAA */
6, /* 4x MSAA */
7, /* 8x MSAA */
8, /* 16x MSAA */
};
unsigned log_samples = util_logbase2(setup_samples);
+ unsigned ps_iter_samples = si_get_ps_iter_samples(sctx);
unsigned log_ps_iter_samples =
- util_logbase2(util_next_power_of_two(sctx->ps_iter_samples));
+ util_logbase2(util_next_power_of_two(ps_iter_samples));
radeon_set_context_reg_seq(cs, R_028BDC_PA_SC_LINE_CNTL, 2);
radeon_emit(cs, sc_line_cntl |
S_028BDC_EXPAND_LINE_WIDTH(1)); /* CM_R_028BDC_PA_SC_LINE_CNTL */
radeon_emit(cs, S_028BE0_MSAA_NUM_SAMPLES(log_samples) |
S_028BE0_MAX_SAMPLE_DIST(max_dist[log_samples]) |
S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples)); /* CM_R_028BE0_PA_SC_AA_CONFIG */
if (sctx->framebuffer.nr_samples > 1) {
radeon_set_context_reg(cs, R_028804_DB_EQAA,
S_028804_MAX_ANCHOR_SAMPLES(log_samples) |
S_028804_PS_ITER_SAMPLES(log_ps_iter_samples) |
S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples) |
S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples) |
S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
S_028804_STATIC_ANCHOR_ASSOCIATIONS(1));
radeon_set_context_reg(cs, R_028A4C_PA_SC_MODE_CNTL_1,
- S_028A4C_PS_ITER_SAMPLE(sctx->ps_iter_samples > 1) |
+ S_028A4C_PS_ITER_SAMPLE(ps_iter_samples > 1) |
sc_mode_cntl_1);
} else if (sctx->smoothing_enabled) {
radeon_set_context_reg(cs, R_028804_DB_EQAA,
S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
S_028804_STATIC_ANCHOR_ASSOCIATIONS(1) |
S_028804_OVERRASTERIZATION_AMOUNT(log_samples));
radeon_set_context_reg(cs, R_028A4C_PA_SC_MODE_CNTL_1,
sc_mode_cntl_1);
}
} else {
@@ -3401,34 +3403,39 @@ static void si_emit_msaa_config(struct si_context *sctx, struct r600_atom *atom)
sc_mode_cntl_1);
}
/* GFX9: Flush DFSM when the AA mode changes. */
if (sctx->screen->dfsm_allowed) {
radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
}
}
+void si_update_ps_iter_samples(struct si_context *sctx)
+{
+ if (sctx->framebuffer.nr_samples > 1)
+ si_mark_atom_dirty(sctx, &sctx->msaa_config);
+ if (sctx->screen->dpbb_allowed)
+ si_mark_atom_dirty(sctx, &sctx->dpbb_state);
+}
+
static void si_set_min_samples(struct pipe_context *ctx, unsigned min_samples)
{
struct si_context *sctx = (struct si_context *)ctx;
if (sctx->ps_iter_samples == min_samples)
return;
sctx->ps_iter_samples = min_samples;
sctx->do_update_shaders = true;
- if (sctx->framebuffer.nr_samples > 1)
- si_mark_atom_dirty(sctx, &sctx->msaa_config);
- if (sctx->screen->dpbb_allowed)
- si_mark_atom_dirty(sctx, &sctx->dpbb_state);
+ si_update_ps_iter_samples(sctx);
}
/*
* Samplers
*/
/**
* Build the sampler view descriptor for a buffer texture.
* @param state 256-bit descriptor; only the high 128 bits are filled in
*/
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index 02659a7a4f3..37887853388 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -210,20 +210,26 @@ enum {
SI_VS_STREAMOUT_BUF1,
SI_VS_STREAMOUT_BUF2,
SI_VS_STREAMOUT_BUF3,
SI_HS_CONST_DEFAULT_TESS_LEVELS,
SI_VS_CONST_INSTANCE_DIVISORS,
SI_VS_CONST_CLIP_PLANES,
SI_PS_CONST_POLY_STIPPLE,
SI_PS_CONST_SAMPLE_POSITIONS,
+ /* Image descriptor of color buffer 0 for KHR_blend_equation_advanced. */
+ SI_PS_IMAGE_COLORBUF0,
+ SI_PS_IMAGE_COLORBUF0_HI,
+ SI_PS_IMAGE_COLORBUF0_FMASK,
+ SI_PS_IMAGE_COLORBUF0_FMASK_HI,
+
SI_NUM_RW_BUFFERS,
};
/* Indices into sctx->descriptors, laid out so that gfx and compute pipelines
* are contiguous:
*
* 0 - rw buffers
* 1 - vertex const and shader buffers
* 2 - vertex samplers and images
* 3 - fragment const and shader buffer
@@ -317,20 +323,21 @@ struct si_buffer_resources {
si_pm4_block_idx(member)); \
} while(0)
/* si_descriptors.c */
void si_set_mutable_tex_desc_fields(struct si_screen *sscreen,
struct r600_texture *tex,
const struct legacy_surf_level *base_level_info,
unsigned base_level, unsigned first_level,
unsigned block_width, bool is_stencil,
uint32_t *state);
+void si_update_ps_colorbuf0_slot(struct si_context *sctx);
void si_get_pipe_constant_buffer(struct si_context *sctx, uint shader,
uint slot, struct pipe_constant_buffer *cbuf);
void si_get_shader_buffers(struct si_context *sctx,
enum pipe_shader_type shader,
uint start_slot, uint count,
struct pipe_shader_buffer *sbuf);
void si_set_ring_buffer(struct pipe_context *ctx, uint slot,
struct pipe_resource *buffer,
unsigned stride, unsigned num_records,
bool add_tid, bool swizzle,
@@ -388,20 +395,21 @@ si_make_texture_descriptor(struct si_screen *screen,
unsigned width, unsigned height, unsigned depth,
uint32_t *state,
uint32_t *fmask_state);
struct pipe_sampler_view *
si_create_sampler_view_custom(struct pipe_context *ctx,
struct pipe_resource *texture,
const struct pipe_sampler_view *state,
unsigned width0, unsigned height0,
unsigned force_level);
void si_update_fb_dirtiness_after_rendering(struct si_context *sctx);
+void si_update_ps_iter_samples(struct si_context *sctx);
/* si_state_binning.c */
void si_emit_dpbb_state(struct si_context *sctx, struct r600_atom *state);
/* si_state_shaders.c */
bool si_update_shaders(struct si_context *sctx);
void si_init_shader_functions(struct si_context *sctx);
bool si_init_shader_cache(struct si_screen *sscreen);
void si_destroy_shader_cache(struct si_screen *sscreen);
void si_get_active_slot_masks(const struct tgsi_shader_info *info,
diff --git a/src/gallium/drivers/radeonsi/si_state_binning.c b/src/gallium/drivers/radeonsi/si_state_binning.c
index 686701d718f..7138f7a03aa 100644
--- a/src/gallium/drivers/radeonsi/si_state_binning.c
+++ b/src/gallium/drivers/radeonsi/si_state_binning.c
@@ -75,21 +75,21 @@ static struct uvec2 si_get_color_bin_size(struct si_context *sctx,
if (!(cb_target_enabled_4bit & (0xf << (i * 4))))
continue;
struct r600_texture *rtex =
(struct r600_texture*)sctx->framebuffer.state.cbufs[i]->texture;
sum += rtex->surface.bpe;
}
/* Multiply the sum by some function of the number of samples. */
if (nr_samples >= 2) {
- if (sctx->ps_iter_samples >= 2)
+ if (si_get_ps_iter_samples(sctx) >= 2)
sum *= nr_samples;
else
sum *= 2;
}
static const si_bin_size_subtable table[] = {
{
/* One RB / SE */
{
/* One shader engine */
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index 8fe4c04ae79..d7742eafb04 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -1453,20 +1453,37 @@ static inline void si_shader_selector_key(struct pipe_context *ctx,
sel->info.uses_linear_center +
sel->info.uses_linear_centroid +
sel->info.uses_linear_sample > 1;
if (sel->info.opcode_count[TGSI_OPCODE_INTERP_SAMPLE])
key->mono.u.ps.interpolate_at_sample_force_center = 1;
}
}
key->part.ps.epilog.alpha_func = si_get_alpha_test_func(sctx);
+
+ /* ps_uses_fbfetch is true only if the color buffer is bound. */
+ if (sctx->ps_uses_fbfetch) {
+ struct pipe_surface *cb0 = sctx->framebuffer.state.cbufs[0];
+ struct pipe_resource *tex = cb0->texture;
+
+ /* 1D textures are allocated and used as 2D on GFX9. */
+ key->mono.u.ps.fbfetch_msaa = sctx->framebuffer.nr_samples > 1;
+ key->mono.u.ps.fbfetch_is_1D = sctx->b.chip_class != GFX9 &&
+ (tex->target == PIPE_TEXTURE_1D ||
+ tex->target == PIPE_TEXTURE_1D_ARRAY);
+ key->mono.u.ps.fbfetch_layered = tex->target == PIPE_TEXTURE_1D_ARRAY ||
+ tex->target == PIPE_TEXTURE_2D_ARRAY ||
+ tex->target == PIPE_TEXTURE_CUBE ||
+ tex->target == PIPE_TEXTURE_CUBE_ARRAY ||
+ tex->target == PIPE_TEXTURE_3D;
+ }
break;
}
default:
assert(0);
}
if (unlikely(sctx->screen->debug_flags & DBG(NO_OPT_VARIANT)))
memset(&key->opt, 0, sizeof(key->opt));
}
@@ -2419,20 +2436,21 @@ static void si_bind_ps_shader(struct pipe_context *ctx, void *state)
si_mark_atom_dirty(sctx, &sctx->cb_render_state);
if (sctx->screen->has_out_of_order_rast &&
(!old_sel ||
old_sel->info.writes_memory != sel->info.writes_memory ||
old_sel->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL] !=
sel->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL]))
si_mark_atom_dirty(sctx, &sctx->msaa_config);
}
si_set_active_descriptors_for_shader(sctx, sel);
+ si_update_ps_colorbuf0_slot(sctx);
}
static void si_delete_shader(struct si_context *sctx, struct si_shader *shader)
{
if (shader->is_optimized) {
util_queue_drop_job(&sctx->screen->shader_compiler_queue_low_priority,
&shader->ready);
}
util_queue_fence_destroy(&shader->ready);
--
2.15.1
More information about the mesa-dev
mailing list