[Mesa-dev] [PATCH 04/10] radeonsi: rework polygon stippling to use constant buffer instead of texture
Marek Olšák
maraeo at gmail.com
Wed Apr 20 15:47:07 UTC 2016
From: Marek Olšák <marek.olsak at amd.com>
add it to the RW_BUFFERS descriptor array
now the slot masks don't have to have 64 bits
---
src/gallium/drivers/radeonsi/si_descriptors.c | 23 +++++++++
src/gallium/drivers/radeonsi/si_pipe.c | 2 -
src/gallium/drivers/radeonsi/si_pipe.h | 1 -
src/gallium/drivers/radeonsi/si_shader.c | 67 ++++++++++++---------------
src/gallium/drivers/radeonsi/si_state.c | 55 ----------------------
src/gallium/drivers/radeonsi/si_state.h | 8 ++--
6 files changed, 55 insertions(+), 101 deletions(-)
diff --git a/src/gallium/drivers/radeonsi/si_descriptors.c b/src/gallium/drivers/radeonsi/si_descriptors.c
index b8f74f4..194b2eb 100644
--- a/src/gallium/drivers/radeonsi/si_descriptors.c
+++ b/src/gallium/drivers/radeonsi/si_descriptors.c
@@ -1121,6 +1121,26 @@ static void si_desc_reset_buffer_offset(struct pipe_context *ctx,
S_008F04_BASE_ADDRESS_HI(va >> 32);
}
+/* INTERNAL CONST BUFFERS */
+
+static void si_set_polygon_stipple(struct pipe_context *ctx,
+ const struct pipe_poly_stipple *state)
+{
+ struct si_context *sctx = (struct si_context *)ctx;
+ struct pipe_constant_buffer cb = {};
+ unsigned stipple[32];
+ int i;
+
+ for (i = 0; i < 32; i++)
+ stipple[i] = util_bitreverse(state->stipple[i]);
+
+ cb.user_buffer = stipple;
+ cb.buffer_size = sizeof(stipple);
+
+ si_set_constant_buffer(sctx, &sctx->rw_buffers,
+ SI_PS_CONST_POLY_STIPPLE, &cb);
+}
+
/* TEXTURE METADATA ENABLE/DISABLE */
/* CMASK can be enabled (for fast clear) and disabled (for texture export)
@@ -1401,6 +1421,8 @@ void si_emit_graphics_shader_userdata(struct si_context *sctx,
if (sctx->rw_buffers.desc.pointer_dirty) {
si_emit_shader_pointer(sctx, &sctx->rw_buffers.desc,
+ R_00B030_SPI_SHADER_USER_DATA_PS_0, true);
+ si_emit_shader_pointer(sctx, &sctx->rw_buffers.desc,
R_00B130_SPI_SHADER_USER_DATA_VS_0, true);
si_emit_shader_pointer(sctx, &sctx->rw_buffers.desc,
R_00B230_SPI_SHADER_USER_DATA_GS_0, true);
@@ -1478,6 +1500,7 @@ void si_init_all_descriptors(struct si_context *sctx)
sctx->b.b.bind_sampler_states = si_bind_sampler_states;
sctx->b.b.set_shader_images = si_set_shader_images;
sctx->b.b.set_constant_buffer = si_pipe_set_constant_buffer;
+ sctx->b.b.set_polygon_stipple = si_set_polygon_stipple;
sctx->b.b.set_shader_buffers = si_set_shader_buffers;
sctx->b.b.set_sampler_views = si_set_sampler_views;
sctx->b.b.set_stream_output_targets = si_set_streamout_targets;
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index 17d59b6..2a5cf0a 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -61,8 +61,6 @@ static void si_destroy_context(struct pipe_context *context)
for (i = 0; i < Elements(sctx->vgt_shader_config); i++)
si_pm4_delete_state(sctx, vgt_shader_config, sctx->vgt_shader_config[i]);
- if (sctx->pstipple_sampler_state)
- sctx->b.b.delete_sampler_state(&sctx->b.b, sctx->pstipple_sampler_state);
if (sctx->fixed_func_tcs_shader.cso)
sctx->b.b.delete_tcs_state(&sctx->b.b, sctx->fixed_func_tcs_shader.cso);
if (sctx->custom_dsa_flush)
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 48095b0..85bf10f 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -190,7 +190,6 @@ struct si_context {
void *custom_blend_decompress;
void *custom_blend_fastclear;
void *custom_blend_dcc_decompress;
- void *pstipple_sampler_state;
struct si_screen *screen;
struct radeon_winsys_cs *ce_ib;
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 231b2c3..cfea2db 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -5291,15 +5291,14 @@ static void preload_ring_buffers(struct si_shader_context *ctx)
}
static void si_llvm_emit_polygon_stipple(struct si_shader_context *ctx,
- LLVMValueRef param_sampler_views,
+ LLVMValueRef param_rw_buffers,
unsigned param_pos_fixed_pt)
{
struct lp_build_tgsi_context *bld_base =
&ctx->radeon_bld.soa.bld_base;
struct gallivm_state *gallivm = bld_base->base.gallivm;
- struct lp_build_emit_data result = {};
- struct tgsi_full_instruction inst = {};
- LLVMValueRef desc, sampler_index, address[2], pix;
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef slot, desc, offset, row, bit, address[2];
/* Use the fixed-point gl_FragCoord input.
* Since the stipple pattern is 32x32 and it repeats, just get 5 bits
@@ -5308,29 +5307,21 @@ static void si_llvm_emit_polygon_stipple(struct si_shader_context *ctx,
address[0] = unpack_param(ctx, param_pos_fixed_pt, 0, 5);
address[1] = unpack_param(ctx, param_pos_fixed_pt, 16, 5);
- /* Load the sampler view descriptor. */
- sampler_index = lp_build_const_int32(gallivm, SI_POLY_STIPPLE_SAMPLER);
- desc = get_sampler_desc_custom(ctx, param_sampler_views,
- sampler_index, DESC_IMAGE);
-
- /* Load the texel. */
- inst.Instruction.Opcode = TGSI_OPCODE_TXF;
- inst.Texture.Texture = TGSI_TEXTURE_2D_MSAA; /* = use load, not load_mip */
- result.inst = &inst;
- set_tex_fetch_args(ctx, &result, TGSI_OPCODE_TXF,
- inst.Texture.Texture,
- desc, NULL, address, ARRAY_SIZE(address), 0xf);
- build_tex_intrinsic(&tex_action, bld_base, &result);
-
- /* Kill the thread accordingly. */
- pix = LLVMBuildExtractElement(gallivm->builder, result.output[0],
- lp_build_const_int32(gallivm, 3), "");
- pix = bitcast(bld_base, TGSI_TYPE_FLOAT, pix);
- pix = LLVMBuildFNeg(gallivm->builder, pix, "");
+ /* Load the buffer descriptor. */
+ slot = lp_build_const_int32(gallivm, SI_PS_CONST_POLY_STIPPLE);
+ desc = build_indexed_load_const(ctx, param_rw_buffers, slot);
- lp_build_intrinsic(gallivm->builder, "llvm.AMDGPU.kill",
- LLVMVoidTypeInContext(gallivm->context),
- &pix, 1, 0);
+ /* The stipple pattern is 32x32, each row has 32 bits. */
+ offset = LLVMBuildMul(builder, address[1],
+ LLVMConstInt(ctx->i32, 4, 0), "");
+ row = buffer_load_const(builder, desc, offset, ctx->i32);
+ bit = LLVMBuildLShr(builder, row, address[0], "");
+ bit = LLVMBuildTrunc(builder, bit, ctx->i1, "");
+
+ /* The intrinsic kills the thread if arg < 0. */
+ bit = LLVMBuildSelect(builder, bit, LLVMConstReal(ctx->f32, 0),
+ LLVMConstReal(ctx->f32, -1), "");
+ lp_build_intrinsic(builder, "llvm.AMDGPU.kill", ctx->voidt, &bit, 1, 0);
}
void si_shader_binary_read_config(struct radeon_shader_binary *binary,
@@ -6038,9 +6029,9 @@ int si_compile_tgsi_shader(struct si_screen *sscreen,
if (ctx.is_monolithic && sel->type == PIPE_SHADER_FRAGMENT &&
shader->key.ps.prolog.poly_stipple) {
- LLVMValueRef views = LLVMGetParam(ctx.radeon_bld.main_fn,
- SI_PARAM_SAMPLERS);
- si_llvm_emit_polygon_stipple(&ctx, views,
+ LLVMValueRef list = LLVMGetParam(ctx.radeon_bld.main_fn,
+ SI_PARAM_RW_BUFFERS);
+ si_llvm_emit_polygon_stipple(&ctx, list,
SI_PARAM_POS_FIXED_PT);
}
@@ -6618,17 +6609,17 @@ static bool si_compile_ps_prolog(struct si_screen *sscreen,
/* POS_FIXED_PT is always last. */
unsigned pos = key->ps_prolog.num_input_sgprs +
key->ps_prolog.num_input_vgprs - 1;
- LLVMValueRef ptr[2], views;
+ LLVMValueRef ptr[2], list;
- /* Get the pointer to sampler views. */
- ptr[0] = LLVMGetParam(func, SI_SGPR_SAMPLERS);
- ptr[1] = LLVMGetParam(func, SI_SGPR_SAMPLERS+1);
- views = lp_build_gather_values(gallivm, ptr, 2);
- views = LLVMBuildBitCast(gallivm->builder, views, ctx.i64, "");
- views = LLVMBuildIntToPtr(gallivm->builder, views,
- const_array(ctx.v8i32, SI_NUM_SAMPLERS), "");
+ /* Get the pointer to rw buffers. */
+ ptr[0] = LLVMGetParam(func, SI_SGPR_RW_BUFFERS);
+ ptr[1] = LLVMGetParam(func, SI_SGPR_RW_BUFFERS_HI);
+ list = lp_build_gather_values(gallivm, ptr, 2);
+ list = LLVMBuildBitCast(gallivm->builder, list, ctx.i64, "");
+ list = LLVMBuildIntToPtr(gallivm->builder, list,
+ const_array(ctx.v16i8, SI_NUM_RW_BUFFERS), "");
- si_llvm_emit_polygon_stipple(&ctx, views, pos);
+ si_llvm_emit_polygon_stipple(&ctx, list, pos);
}
/* Interpolate colors. */
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index 305a70b..1f3a5fa 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -3374,60 +3374,6 @@ static void si_set_index_buffer(struct pipe_context *ctx,
/*
* Misc
*/
-static void si_set_polygon_stipple(struct pipe_context *ctx,
- const struct pipe_poly_stipple *state)
-{
- struct si_context *sctx = (struct si_context *)ctx;
- struct pipe_resource *tex;
- struct pipe_sampler_view *view;
- bool is_zero = true;
- bool is_one = true;
- int i;
-
- /* The hardware obeys 0 and 1 swizzles in the descriptor even if
- * the resource is NULL/invalid. Take advantage of this fact and skip
- * texture allocation if the stipple pattern is constant.
- *
- * This is an optimization for the common case when stippling isn't
- * used but set_polygon_stipple is still called by st/mesa.
- */
- for (i = 0; i < Elements(state->stipple); i++) {
- is_zero = is_zero && state->stipple[i] == 0;
- is_one = is_one && state->stipple[i] == 0xffffffff;
- }
-
- if (is_zero || is_one) {
- struct pipe_sampler_view templ = {{0}};
-
- templ.swizzle_r = PIPE_SWIZZLE_ZERO;
- templ.swizzle_g = PIPE_SWIZZLE_ZERO;
- templ.swizzle_b = PIPE_SWIZZLE_ZERO;
- /* The pattern should be inverted in the texture. */
- templ.swizzle_a = is_zero ? PIPE_SWIZZLE_ONE : PIPE_SWIZZLE_ZERO;
-
- view = ctx->create_sampler_view(ctx, NULL, &templ);
- } else {
- /* Create a new texture. */
- tex = util_pstipple_create_stipple_texture(ctx, state->stipple);
- if (!tex)
- return;
-
- view = util_pstipple_create_sampler_view(ctx, tex);
- pipe_resource_reference(&tex, NULL);
- }
-
- ctx->set_sampler_views(ctx, PIPE_SHADER_FRAGMENT,
- SI_POLY_STIPPLE_SAMPLER, 1, &view);
- pipe_sampler_view_reference(&view, NULL);
-
- /* Bind the sampler state if needed. */
- if (!sctx->pstipple_sampler_state) {
- sctx->pstipple_sampler_state = util_pstipple_create_sampler(ctx);
- ctx->bind_sampler_states(ctx, PIPE_SHADER_FRAGMENT,
- SI_POLY_STIPPLE_SAMPLER, 1,
- &sctx->pstipple_sampler_state);
- }
-}
static void si_set_tess_state(struct pipe_context *ctx,
const float default_outer_level[4],
@@ -3590,7 +3536,6 @@ void si_init_state_functions(struct si_context *sctx)
sctx->b.b.texture_barrier = si_texture_barrier;
sctx->b.b.memory_barrier = si_memory_barrier;
- sctx->b.b.set_polygon_stipple = si_set_polygon_stipple;
sctx->b.b.set_min_samples = si_set_min_samples;
sctx->b.b.set_tess_state = si_set_tess_state;
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index e1e7ae0..9875606 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -146,12 +146,8 @@ struct si_shader_data {
uint32_t sh_base[SI_NUM_SHADERS];
};
-/* User sampler views: 0..31
- * Polygon stipple tex: 32
- */
#define SI_NUM_USER_SAMPLERS 32 /* AKA OpenGL textures units per shader */
-#define SI_POLY_STIPPLE_SAMPLER SI_NUM_USER_SAMPLERS
-#define SI_NUM_SAMPLERS (SI_POLY_STIPPLE_SAMPLER + 1)
+#define SI_NUM_SAMPLERS SI_NUM_USER_SAMPLERS
/* User constant buffers: 0..15
* Driver state constants: 16
@@ -182,6 +178,8 @@ enum {
SI_VS_STREAMOUT_BUF2,
SI_VS_STREAMOUT_BUF3,
+ SI_PS_CONST_POLY_STIPPLE,
+
SI_NUM_RW_BUFFERS,
};
--
2.5.0
More information about the mesa-dev
mailing list