[Mesa-dev] [PATCH v2 37/41] WIP: radv, ac: implement 16-bit interpolation

Rhys Perry pendingchaos02 at gmail.com
Sat Feb 16 00:30:19 UTC 2019


v2: add to patch series

Signed-off-by: Rhys Perry <pendingchaos02 at gmail.com>
---
 src/amd/common/ac_llvm_build.c           | 33 +++++++++++++++++-------
 src/amd/common/ac_llvm_build.h           |  3 ++-
 src/amd/common/ac_nir_to_llvm.c          | 14 +++++++---
 src/amd/vulkan/radv_nir_to_llvm.c        | 27 ++++++++++++++-----
 src/amd/vulkan/radv_pipeline.c           | 19 ++++++++------
 src/amd/vulkan/radv_shader.h             |  1 +
 src/gallium/drivers/radeonsi/si_shader.c |  2 +-
 7 files changed, 69 insertions(+), 30 deletions(-)

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index dff369aae7f..be2c2251a21 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -937,27 +937,40 @@ ac_build_fs_interp(struct ac_llvm_context *ctx,
 		   LLVMValueRef attr_number,
 		   LLVMValueRef params,
 		   LLVMValueRef i,
-		   LLVMValueRef j)
+		   LLVMValueRef j,
+		   int word)
 {
-	LLVMValueRef args[5];
+	LLVMValueRef args[6];
 	LLVMValueRef p1;
 
 	args[0] = i;
 	args[1] = llvm_chan;
 	args[2] = attr_number;
-	args[3] = params;
-
-	p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
-				ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
+	if (word >= 0) {
+		args[3] = LLVMConstInt(ctx->i1, word, false);
+		args[4] = params;
+		p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1.f16",
+					ctx->f16, args, 5, AC_FUNC_ATTR_READNONE);
+	} else {
+		args[3] = params;
+		p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
+					ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
+	}
 
 	args[0] = p1;
 	args[1] = j;
 	args[2] = llvm_chan;
 	args[3] = attr_number;
-	args[4] = params;
-
-	return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
-				  ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
+	if (word >= 0) {
+		args[4] = LLVMConstInt(ctx->i1, word, false);
+		args[5] = params;
+		return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2.f16",
+					  ctx->f16, args, 6, AC_FUNC_ATTR_READNONE);
+	} else {
+		args[4] = params;
+		return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
+					  ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
+	}
 }
 
 LLVMValueRef
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index 61c9b5e4b6c..655427567c4 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -224,7 +224,8 @@ ac_build_fs_interp(struct ac_llvm_context *ctx,
 		   LLVMValueRef attr_number,
 		   LLVMValueRef params,
 		   LLVMValueRef i,
-		   LLVMValueRef j);
+		   LLVMValueRef j,
+		   int word);
 
 LLVMValueRef
 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index bf7024c68e4..939b8eb13de 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -3120,8 +3120,15 @@ static LLVMValueRef visit_interp(struct ac_nir_context *ctx,
 				LLVMValueRef j = LLVMBuildExtractElement(
 					ctx->ac.builder, interp_param, ctx->ac.i32_1, "");
 
+				/* This fp16 handling isn't technically correct
+				 * but should be correct for the attributes we
+				 * are actually going to use. */
+				bool fp16 = instr->dest.ssa.bit_size == 16;
+				int word = fp16 ? 0 : -1;
 				v = ac_build_fs_interp(&ctx->ac, llvm_chan, attr_number,
-						       ctx->abi->prim_mask, i, j);
+						       ctx->abi->prim_mask, i, j, word);
+				if (fp16)
+					v = ac_build_reinterpret(&ctx->ac, v, ctx->ac.f32);
 			} else {
 				v = ac_build_fs_interp_mov(&ctx->ac, LLVMConstInt(ctx->ac.i32, 2, false),
 							   llvm_chan, attr_number, ctx->abi->prim_mask);
@@ -3134,8 +3141,9 @@ static LLVMValueRef visit_interp(struct ac_nir_context *ctx,
 		result[chan] = LLVMBuildExtractElement(ctx->ac.builder, gather, attrib_idx, "");
 
 	}
-	return ac_build_varying_gather_values(&ctx->ac, result, instr->num_components,
-					      var->data.location_frac);
+	LLVMValueRef ret = ac_build_varying_gather_values(&ctx->ac, result, instr->num_components,
+							  var->data.location_frac);
+	return ac_build_reinterpret(&ctx->ac, ret, get_def_type(ctx, &instr->dest.ssa));
 }
 
 static void visit_intrinsic(struct ac_nir_context *ctx,
diff --git a/src/amd/vulkan/radv_nir_to_llvm.c b/src/amd/vulkan/radv_nir_to_llvm.c
index c46eabf3656..49f8d35dd5f 100644
--- a/src/amd/vulkan/radv_nir_to_llvm.c
+++ b/src/amd/vulkan/radv_nir_to_llvm.c
@@ -2051,7 +2051,8 @@ static void interp_fs_input(struct radv_shader_context *ctx,
 			    unsigned attr,
 			    LLVMValueRef interp_param,
 			    LLVMValueRef prim_mask,
-			    LLVMValueRef result[4])
+			    LLVMValueRef result[4],
+			    bool fp16)
 {
 	LLVMValueRef attr_number;
 	unsigned chan;
@@ -2086,7 +2087,10 @@ static void interp_fs_input(struct radv_shader_context *ctx,
 			result[chan] = ac_build_fs_interp(&ctx->ac,
 							  llvm_chan,
 							  attr_number,
-							  prim_mask, i, j);
+							  prim_mask, i, j,
+							  fp16 ? 0 : -1);
+			if (fp16)
+				result[chan] = ac_build_reinterpret(&ctx->ac, result[chan], ctx->ac.f16);
 		} else {
 			result[chan] = ac_build_fs_interp_mov(&ctx->ac,
 							      LLVMConstInt(ctx->ac.i32, 2, false),
@@ -2100,7 +2104,8 @@ static void interp_fs_input(struct radv_shader_context *ctx,
 
 static void
 handle_fs_input_decl(struct radv_shader_context *ctx,
-		     struct nir_variable *variable)
+		     struct nir_variable *variable,
+		     uint64_t *fp16_mask)
 {
 	int idx = variable->data.location;
 	unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
@@ -2110,7 +2115,8 @@ handle_fs_input_decl(struct radv_shader_context *ctx,
 	variable->data.driver_location = idx * 4;
 	mask = ((1ull << attrib_count) - 1) << variable->data.location;
 
-	if (glsl_get_base_type(glsl_without_array(variable->type)) == GLSL_TYPE_FLOAT) {
+	enum glsl_base_type type = glsl_get_base_type(glsl_without_array(variable->type));
+	if (type == GLSL_TYPE_FLOAT || type == GLSL_TYPE_FLOAT16) {
 		unsigned interp_type;
 		if (variable->data.sample)
 			interp_type = INTERP_SAMPLE;
@@ -2120,6 +2126,9 @@ handle_fs_input_decl(struct radv_shader_context *ctx,
 			interp_type = INTERP_CENTER;
 
 		interp = lookup_interp_param(&ctx->abi, variable->data.interpolation, interp_type);
+
+		if (type == GLSL_TYPE_FLOAT16)
+			*fp16_mask |= mask;
 	}
 
 	for (unsigned i = 0; i < attrib_count; ++i)
@@ -2173,8 +2182,9 @@ handle_fs_inputs(struct radv_shader_context *ctx,
 {
 	prepare_interp_optimize(ctx, nir);
 
+	uint64_t fp16_mask = 0;
 	nir_foreach_variable(variable, &nir->inputs)
-		handle_fs_input_decl(ctx, variable);
+		handle_fs_input_decl(ctx, variable, &fp16_mask);
 
 	unsigned index = 0;
 
@@ -2194,11 +2204,14 @@ handle_fs_inputs(struct radv_shader_context *ctx,
 		if (i >= VARYING_SLOT_VAR0 || i == VARYING_SLOT_PNTC ||
 		    i == VARYING_SLOT_PRIMITIVE_ID || i == VARYING_SLOT_LAYER) {
 			interp_param = *inputs;
+			bool fp16 = fp16_mask & (1ull << i);
 			interp_fs_input(ctx, index, interp_param, ctx->abi.prim_mask,
-					inputs);
+					inputs, fp16);
 
 			if (!interp_param)
 				ctx->shader_info->fs.flat_shaded_mask |= 1u << index;
+			if (fp16)
+				ctx->shader_info->fs.fp16_mask |= 1u << index;
 			if (i >= VARYING_SLOT_VAR0)
 				ctx->abi.fs_input_attr_indices[i - VARYING_SLOT_VAR0] = index;
 			++index;
@@ -2210,7 +2223,7 @@ handle_fs_inputs(struct radv_shader_context *ctx,
 
 				interp_param = *inputs;
 				interp_fs_input(ctx, index, interp_param,
-						ctx->abi.prim_mask, inputs);
+						ctx->abi.prim_mask, inputs, false);
 				++index;
 			}
 		} else if (i == VARYING_SLOT_POS) {
diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
index ab56a273a2c..a3260291bce 100644
--- a/src/amd/vulkan/radv_pipeline.c
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -3070,13 +3070,15 @@ radv_pipeline_generate_geometry_shader(struct radeon_cmdbuf *ctx_cs,
 	radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, pipeline->gs_copy_shader);
 }
 
-static uint32_t offset_to_ps_input(uint32_t offset, bool flat_shade)
+static uint32_t offset_to_ps_input(uint32_t offset, bool flat_shade, bool fp16)
 {
 	uint32_t ps_input_cntl;
 	if (offset <= AC_EXP_PARAM_OFFSET_31) {
 		ps_input_cntl = S_028644_OFFSET(offset);
 		if (flat_shade)
 			ps_input_cntl |= S_028644_FLAT_SHADE(1);
+		if (fp16 && !flat_shade)
+			ps_input_cntl |= S_028644_FP16_INTERP_MODE(1);
 	} else {
 		/* The input is a DEFAULT_VAL constant. */
 		assert(offset >= AC_EXP_PARAM_DEFAULT_VAL_0000 &&
@@ -3101,7 +3103,7 @@ radv_pipeline_generate_ps_inputs(struct radeon_cmdbuf *ctx_cs,
 	if (ps->info.info.ps.prim_id_input) {
 		unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_PRIMITIVE_ID];
 		if (vs_offset != AC_EXP_PARAM_UNDEFINED) {
-			ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true);
+			ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true, false);
 			++ps_offset;
 		}
 	}
@@ -3111,9 +3113,9 @@ radv_pipeline_generate_ps_inputs(struct radeon_cmdbuf *ctx_cs,
 	    ps->info.info.needs_multiview_view_index) {
 		unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_LAYER];
 		if (vs_offset != AC_EXP_PARAM_UNDEFINED)
-			ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true);
+			ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true, false);
 		else
-			ps_input_cntl[ps_offset] = offset_to_ps_input(AC_EXP_PARAM_DEFAULT_VAL_0000, true);
+			ps_input_cntl[ps_offset] = offset_to_ps_input(AC_EXP_PARAM_DEFAULT_VAL_0000, true, false);
 		++ps_offset;
 	}
 
@@ -3129,21 +3131,21 @@ radv_pipeline_generate_ps_inputs(struct radeon_cmdbuf *ctx_cs,
 
 		vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST0];
 		if (vs_offset != AC_EXP_PARAM_UNDEFINED) {
-			ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, false);
+			ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, false, false);
 			++ps_offset;
 		}
 
 		vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST1];
 		if (vs_offset != AC_EXP_PARAM_UNDEFINED &&
 		    ps->info.info.ps.num_input_clips_culls > 4) {
-			ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, false);
+			ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, false, false);
 			++ps_offset;
 		}
 	}
 
 	for (unsigned i = 0; i < 32 && (1u << i) <= ps->info.fs.input_mask; ++i) {
 		unsigned vs_offset;
-		bool flat_shade;
+		bool flat_shade, fp16;
 		if (!(ps->info.fs.input_mask & (1u << i)))
 			continue;
 
@@ -3155,8 +3157,9 @@ radv_pipeline_generate_ps_inputs(struct radeon_cmdbuf *ctx_cs,
 		}
 
 		flat_shade = !!(ps->info.fs.flat_shaded_mask & (1u << ps_offset));
+		fp16 = !!(ps->info.fs.fp16_mask & (1u << ps_offset));
 
-		ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, flat_shade);
+		ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, flat_shade, fp16);
 		++ps_offset;
 	}
 
diff --git a/src/amd/vulkan/radv_shader.h b/src/amd/vulkan/radv_shader.h
index b67cd2b4f15..f0e9bc249f9 100644
--- a/src/amd/vulkan/radv_shader.h
+++ b/src/amd/vulkan/radv_shader.h
@@ -257,6 +257,7 @@ struct radv_shader_variant_info {
 			unsigned num_interp;
 			uint32_t input_mask;
 			uint32_t flat_shaded_mask;
+			uint32_t fp16_mask;
 			bool can_discard;
 			bool early_fragment_test;
 		} fs;
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index efae02ee91c..c1f82137020 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -1751,7 +1751,7 @@ static LLVMValueRef si_build_fs_interp(struct si_shader_context *ctx,
 		return ac_build_fs_interp(&ctx->ac,
 					  LLVMConstInt(ctx->i32, chan, 0),
 					  LLVMConstInt(ctx->i32, attr_index, 0),
-					  prim_mask, i, j);
+					  prim_mask, i, j, -1);
 	}
 	return ac_build_fs_interp_mov(&ctx->ac,
 				      LLVMConstInt(ctx->i32, 2, 0), /* P0 */
-- 
2.20.1



More information about the mesa-dev mailing list