radeonsi: do force_persample_interp in shaders for non-trivial cases

Reviewed-by: Michel Dänzer <[email protected]>

3 files changed, 117 insertions, 19 deletions

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index a3df64824c6..32a702fcdf5 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -855,6 +855,56 @@ static int lookup_interp_param_index(unsigned interpolate, unsigned location)
 	}
 }
 
+/* This shouldn't be used by explicit INTERP opcodes. */
+static LLVMValueRef get_interp_param(struct si_shader_context *si_shader_ctx,
+				     unsigned param)
+{
+	struct gallivm_state *gallivm = &si_shader_ctx->radeon_bld.gallivm;
+	unsigned sample_param = 0;
+	LLVMValueRef default_ij, sample_ij, force_sample;
+
+	default_ij = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, param);
+
+	/* If the shader doesn't use center/centroid, just return the parameter.
+	 *
+	 * If the shader only uses one set of (i,j), "si_emit_spi_ps_input" can
+	 * switch between center/centroid and sample without shader changes.
+	 */
+	switch (param) {
+	case SI_PARAM_PERSP_CENTROID:
+	case SI_PARAM_PERSP_CENTER:
+		if (!si_shader_ctx->shader->selector->forces_persample_interp_for_persp)
+			return default_ij;
+
+		sample_param = SI_PARAM_PERSP_SAMPLE;
+		break;
+
+	case SI_PARAM_LINEAR_CENTROID:
+	case SI_PARAM_LINEAR_CENTER:
+		if (!si_shader_ctx->shader->selector->forces_persample_interp_for_linear)
+			return default_ij;
+
+		sample_param = SI_PARAM_LINEAR_SAMPLE;
+		break;
+
+	default:
+		return default_ij;
+	}
+
+	/* Otherwise, we have to select (i,j) based on a user data SGPR. */
+	sample_ij = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, sample_param);
+
+	/* TODO: this can be done more efficiently by switching between
+	 * 2 prologs.
+	 */
+	force_sample = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
+				    SI_PARAM_PS_STATE_BITS);
+	force_sample = LLVMBuildTrunc(gallivm->builder, force_sample,
+				      LLVMInt1TypeInContext(gallivm->context), "");
+	return LLVMBuildSelect(gallivm->builder, force_sample,
+			       sample_ij, default_ij, "");
+}
+
 static void declare_input_fs(
 	struct radeon_llvm_context *radeon_bld,
 	unsigned input_index,
@@ -925,7 +975,7 @@ static void declare_input_fs(
 	if (interp_param_idx == -1)
 		return;
 	else if (interp_param_idx)
-		interp_param = LLVMGetParam(main_fn, interp_param_idx);
+		interp_param = get_interp_param(si_shader_ctx, interp_param_idx);
 
 	/* fs.constant returns the param from the middle vertex, so it's not
 	 * really useful for flat shading. It's meant to be used for custom
@@ -3458,6 +3508,7 @@ static void create_function(struct si_shader_context *si_shader_ctx)
 
 	case TGSI_PROCESSOR_FRAGMENT:
 		params[SI_PARAM_ALPHA_REF] = f32;
+		params[SI_PARAM_PS_STATE_BITS] = i32;
 		params[SI_PARAM_PRIM_MASK] = i32;
 		last_sgpr = SI_PARAM_PRIM_MASK;
 		params[SI_PARAM_PERSP_SAMPLE] = v2i32;
diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h
index 2305b9988b8..b92fa02a171 100644
--- a/src/gallium/drivers/radeonsi/si_shader.h
+++ b/src/gallium/drivers/radeonsi/si_shader.h
@@ -88,6 +88,7 @@ struct radeon_shader_reloc;
 #define SI_SGPR_TCS_OUT_LAYOUT	9  /* TCS & TES only */
 #define SI_SGPR_TCS_IN_LAYOUT	10 /* TCS only */
 #define SI_SGPR_ALPHA_REF	8  /* PS only */
+#define SI_SGPR_PS_STATE_BITS	9  /* PS only */
 
 #define SI_VS_NUM_USER_SGPR	12
 #define SI_LS_NUM_USER_SGPR	13
@@ -95,7 +96,7 @@ struct radeon_shader_reloc;
 #define SI_TES_NUM_USER_SGPR	10
 #define SI_GS_NUM_USER_SGPR	8
 #define SI_GSCOPY_NUM_USER_SGPR	4
-#define SI_PS_NUM_USER_SGPR	9
+#define SI_PS_NUM_USER_SGPR	10
 
 /* LLVM function parameter indices */
 #define SI_PARAM_RW_BUFFERS	0
@@ -148,23 +149,27 @@ struct radeon_shader_reloc;
 
 /* PS only parameters */
 #define SI_PARAM_ALPHA_REF		4
-#define SI_PARAM_PRIM_MASK		5
-#define SI_PARAM_PERSP_SAMPLE		6
-#define SI_PARAM_PERSP_CENTER		7
-#define SI_PARAM_PERSP_CENTROID		8
-#define SI_PARAM_PERSP_PULL_MODEL	9
-#define SI_PARAM_LINEAR_SAMPLE		10
-#define SI_PARAM_LINEAR_CENTER		11
-#define SI_PARAM_LINEAR_CENTROID	12
-#define SI_PARAM_LINE_STIPPLE_TEX	13
-#define SI_PARAM_POS_X_FLOAT		14
-#define SI_PARAM_POS_Y_FLOAT		15
-#define SI_PARAM_POS_Z_FLOAT		16
-#define SI_PARAM_POS_W_FLOAT		17
-#define SI_PARAM_FRONT_FACE		18
-#define SI_PARAM_ANCILLARY		19
-#define SI_PARAM_SAMPLE_COVERAGE	20
-#define SI_PARAM_POS_FIXED_PT		21
+/* Bits:
+ * 0: force_persample_interp
+ */
+#define SI_PARAM_PS_STATE_BITS		5
+#define SI_PARAM_PRIM_MASK		6
+#define SI_PARAM_PERSP_SAMPLE		7
+#define SI_PARAM_PERSP_CENTER		8
+#define SI_PARAM_PERSP_CENTROID		9
+#define SI_PARAM_PERSP_PULL_MODEL	10
+#define SI_PARAM_LINEAR_SAMPLE		11
+#define SI_PARAM_LINEAR_CENTER		12
+#define SI_PARAM_LINEAR_CENTROID	13
+#define SI_PARAM_LINE_STIPPLE_TEX	14
+#define SI_PARAM_POS_X_FLOAT		15
+#define SI_PARAM_POS_Y_FLOAT		16
+#define SI_PARAM_POS_Z_FLOAT		17
+#define SI_PARAM_POS_W_FLOAT		18
+#define SI_PARAM_FRONT_FACE		19
+#define SI_PARAM_ANCILLARY		20
+#define SI_PARAM_SAMPLE_COVERAGE	21
+#define SI_PARAM_POS_FIXED_PT		22
 
 #define SI_NUM_PARAMS (SI_PARAM_POS_FIXED_PT + 1)
 
@@ -182,6 +187,14 @@ struct si_shader_selector {
 	/* PIPE_SHADER_[VERTEX|FRAGMENT|...] */
 	unsigned	type;
 
+	/* Whether the shader has to use a conditional assignment to
+	 * choose between weights when emulating
+	 * pipe_rasterizer_state::force_persample_interp.
+	 * If false, "si_emit_spi_ps_input" will take care of it instead.
+	 */
+	bool		forces_persample_interp_for_persp;
+	bool		forces_persample_interp_for_linear;
+
 	unsigned	gs_output_prim;
 	unsigned	gs_max_out_vertices;
 	unsigned	gs_num_invocations;
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index aeb28797ace..77c585f958e 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -667,6 +667,34 @@ static void *si_create_shader_state(struct pipe_context *ctx,
 	tgsi_scan_shader(state->tokens, &sel->info);
 	p_atomic_inc(&sscreen->b.num_shaders_created);
 
+	/* First set which opcode uses which (i,j) pair. */
+	if (sel->info.uses_persp_opcode_interp_centroid)
+		sel->info.uses_persp_centroid = true;
+
+	if (sel->info.uses_linear_opcode_interp_centroid)
+		sel->info.uses_linear_centroid = true;
+
+	if (sel->info.uses_persp_opcode_interp_offset ||
+	    sel->info.uses_persp_opcode_interp_sample)
+		sel->info.uses_persp_center = true;
+
+	if (sel->info.uses_linear_opcode_interp_offset ||
+	    sel->info.uses_linear_opcode_interp_sample)
+		sel->info.uses_linear_center = true;
+
+	/* Determine if the shader has to use a conditional assignment when
+	 * emulating force_persample_interp.
+	 */
+	sel->forces_persample_interp_for_persp =
+		sel->info.uses_persp_center +
+		sel->info.uses_persp_centroid +
+		sel->info.uses_persp_sample >= 2;
+
+	sel->forces_persample_interp_for_linear =
+		sel->info.uses_linear_center +
+		sel->info.uses_linear_centroid +
+		sel->info.uses_linear_sample >= 2;
+
 	switch (pipe_shader_type) {
 	case PIPE_SHADER_GEOMETRY:
 		sel->gs_output_prim =
@@ -1100,6 +1128,12 @@ static void si_emit_spi_ps_input(struct si_context *sctx, struct r600_atom *atom
 	radeon_set_context_reg_seq(cs, R_0286CC_SPI_PS_INPUT_ENA, 2);
 	radeon_emit(cs, input_ena);
 	radeon_emit(cs, input_ena);
+
+	if (ps->selector->forces_persample_interp_for_persp ||
+	    ps->selector->forces_persample_interp_for_linear)
+		radeon_set_sh_reg(cs, R_00B030_SPI_SHADER_USER_DATA_PS_0 +
+				      SI_SGPR_PS_STATE_BITS * 4,
+				  sctx->force_persample_interp);
 }
 
 /* Initialize state related to ESGS / GSVS ring buffers */