radeonsi: use pknorm_i16/u16 and pk_i16/u16 LLVM intrinsics

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

3 files changed, 216 insertions, 113 deletions

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 762351f280..6ede60afe7 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -67,6 +67,7 @@ ac_llvm_context_init(struct ac_llvm_context *ctx, LLVMContextRef context,
 	ctx->f16 = LLVMHalfTypeInContext(ctx->context);
 	ctx->f32 = LLVMFloatTypeInContext(ctx->context);
 	ctx->f64 = LLVMDoubleTypeInContext(ctx->context);
+	ctx->v2i16 = LLVMVectorType(ctx->i16, 2);
 	ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
 	ctx->v3i32 = LLVMVectorType(ctx->i32, 3);
 	ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
@@ -1281,6 +1282,20 @@ LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 				  AC_FUNC_ATTR_READNONE);
 }
 
+LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
+			   LLVMValueRef b)
+{
+	LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
+	return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
+}
+
+LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
+			   LLVMValueRef b)
+{
+	LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
+	return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
+}
+
 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
 			   LLVMValueRef b)
 {
@@ -1506,6 +1521,155 @@ LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 				  AC_FUNC_ATTR_LEGACY);
 }
 
+/* Upper 16 bits must be zero. */
+static LLVMValueRef ac_llvm_pack_two_int16(struct ac_llvm_context *ctx,
+					   LLVMValueRef val[2])
+{
+	return LLVMBuildOr(ctx->builder, val[0],
+			   LLVMBuildShl(ctx->builder, val[1],
+					LLVMConstInt(ctx->i32, 16, 0),
+					""), "");
+}
+
+/* Upper 16 bits are ignored and will be dropped. */
+static LLVMValueRef ac_llvm_pack_two_int32_as_int16(struct ac_llvm_context *ctx,
+						    LLVMValueRef val[2])
+{
+	LLVMValueRef v[2] = {
+		LLVMBuildAnd(ctx->builder, val[0],
+			     LLVMConstInt(ctx->i32, 0xffff, 0), ""),
+		val[1],
+	};
+	return ac_llvm_pack_two_int16(ctx, v);
+}
+
+LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
+				     LLVMValueRef args[2])
+{
+	if (HAVE_LLVM >= 0x0600) {
+		LLVMValueRef res =
+			ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
+					   ctx->v2i16, args, 2,
+					   AC_FUNC_ATTR_READNONE);
+		return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
+	}
+
+	LLVMValueRef val[2];
+
+	for (int chan = 0; chan < 2; chan++) {
+		/* Clamp between [-1, 1]. */
+		val[chan] = ac_build_fmin(ctx, args[chan], ctx->f32_1);
+		val[chan] = ac_build_fmax(ctx, val[chan], LLVMConstReal(ctx->f32, -1));
+		/* Convert to a signed integer in [-32767, 32767]. */
+		val[chan] = LLVMBuildFMul(ctx->builder, val[chan],
+					  LLVMConstReal(ctx->f32, 32767), "");
+		/* If positive, add 0.5, else add -0.5. */
+		val[chan] = LLVMBuildFAdd(ctx->builder, val[chan],
+				LLVMBuildSelect(ctx->builder,
+					LLVMBuildFCmp(ctx->builder, LLVMRealOGE,
+						      val[chan], ctx->f32_0, ""),
+					LLVMConstReal(ctx->f32, 0.5),
+					LLVMConstReal(ctx->f32, -0.5), ""), "");
+		val[chan] = LLVMBuildFPToSI(ctx->builder, val[chan], ctx->i32, "");
+	}
+	return ac_llvm_pack_two_int32_as_int16(ctx, val);
+}
+
+LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
+				     LLVMValueRef args[2])
+{
+	if (HAVE_LLVM >= 0x0600) {
+		LLVMValueRef res =
+			ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
+					   ctx->v2i16, args, 2,
+					   AC_FUNC_ATTR_READNONE);
+		return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
+	}
+
+	LLVMValueRef val[2];
+
+	for (int chan = 0; chan < 2; chan++) {
+		val[chan] = ac_build_clamp(ctx, args[chan]);
+		val[chan] = LLVMBuildFMul(ctx->builder, val[chan],
+					  LLVMConstReal(ctx->f32, 65535), "");
+		val[chan] = LLVMBuildFAdd(ctx->builder, val[chan],
+					  LLVMConstReal(ctx->f32, 0.5), "");
+		val[chan] = LLVMBuildFPToUI(ctx->builder, val[chan],
+					    ctx->i32, "");
+	}
+	return ac_llvm_pack_two_int32_as_int16(ctx, val);
+}
+
+/* The 8-bit and 10-bit clamping is for HW workarounds. */
+LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
+				 LLVMValueRef args[2], unsigned bits, bool hi)
+{
+	assert(bits == 8 || bits == 10 || bits == 16);
+
+	LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
+		bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
+	LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
+		bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
+	LLVMValueRef max_alpha =
+		bits != 10 ? max_rgb : ctx->i32_1;
+	LLVMValueRef min_alpha =
+		bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
+	bool has_intrinsic = HAVE_LLVM >= 0x0600;
+
+	/* Clamp. */
+	if (!has_intrinsic || bits != 16) {
+		for (int i = 0; i < 2; i++) {
+			bool alpha = hi && i == 1;
+			args[i] = ac_build_imin(ctx, args[i],
+						alpha ? max_alpha : max_rgb);
+			args[i] = ac_build_imax(ctx, args[i],
+						alpha ? min_alpha : min_rgb);
+		}
+	}
+
+	if (has_intrinsic) {
+		LLVMValueRef res =
+			ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
+					   ctx->v2i16, args, 2,
+					   AC_FUNC_ATTR_READNONE);
+		return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
+	}
+
+	return ac_llvm_pack_two_int32_as_int16(ctx, args);
+}
+
+/* The 8-bit and 10-bit clamping is for HW workarounds. */
+LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
+				 LLVMValueRef args[2], unsigned bits, bool hi)
+{
+	assert(bits == 8 || bits == 10 || bits == 16);
+
+	LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
+		bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
+	LLVMValueRef max_alpha =
+		bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
+	bool has_intrinsic = HAVE_LLVM >= 0x0600;
+
+	/* Clamp. */
+	if (!has_intrinsic || bits != 16) {
+		for (int i = 0; i < 2; i++) {
+			bool alpha = hi && i == 1;
+			args[i] = ac_build_umin(ctx, args[i],
+						alpha ? max_alpha : max_rgb);
+		}
+	}
+
+	if (has_intrinsic) {
+		LLVMValueRef res =
+			ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
+					   ctx->v2i16, args, 2,
+					   AC_FUNC_ATTR_READNONE);
+		return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
+	}
+
+	return ac_llvm_pack_two_int16(ctx, args);
+}
+
 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
 {
 	assert(HAVE_LLVM >= 0x0600);
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index f95ad6a27c..35f849a986 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -53,6 +53,7 @@ struct ac_llvm_context {
 	LLVMTypeRef f16;
 	LLVMTypeRef f32;
 	LLVMTypeRef f64;
+	LLVMTypeRef v2i16;
 	LLVMTypeRef v2i32;
 	LLVMTypeRef v3i32;
 	LLVMTypeRef v4i32;
@@ -258,6 +259,10 @@ LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
 			   LLVMValueRef b);
 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
 			   LLVMValueRef b);
+LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
+			   LLVMValueRef b);
+LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
+			   LLVMValueRef b);
 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 
@@ -302,6 +307,14 @@ LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 				   struct ac_image_args *a);
 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 				    LLVMValueRef args[2]);
+LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
+				     LLVMValueRef args[2]);
+LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
+				     LLVMValueRef args[2]);
+LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
+				 LLVMValueRef args[2], unsigned bits, bool hi);
+LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
+				 LLVMValueRef args[2], unsigned bits, bool hi);
 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 7b2dc8daf6..1294af3012 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -2388,28 +2388,6 @@ static LLVMValueRef fetch_constant(
 	return bitcast(bld_base, type, buffer_load_const(ctx, bufp, addr));
 }
 
-/* Upper 16 bits must be zero. */
-static LLVMValueRef si_llvm_pack_two_int16(struct si_shader_context *ctx,
-					   LLVMValueRef val[2])
-{
-	return LLVMBuildOr(ctx->ac.builder, val[0],
-			   LLVMBuildShl(ctx->ac.builder, val[1],
-					LLVMConstInt(ctx->i32, 16, 0),
-					""), "");
-}
-
-/* Upper 16 bits are ignored and will be dropped. */
-static LLVMValueRef si_llvm_pack_two_int32_as_int16(struct si_shader_context *ctx,
-						    LLVMValueRef val[2])
-{
-	LLVMValueRef v[2] = {
-		LLVMBuildAnd(ctx->ac.builder, val[0],
-			     LLVMConstInt(ctx->i32, 0xffff, 0), ""),
-		val[1],
-	};
-	return si_llvm_pack_two_int16(ctx, v);
-}
-
 /* Initialize arguments for the shader export intrinsic */
 static void si_llvm_init_export_args(struct si_shader_context *ctx,
 				     LLVMValueRef *values,
@@ -2417,8 +2395,6 @@ static void si_llvm_init_export_args(struct si_shader_context *ctx,
 				     struct ac_export_args *args)
 {
 	LLVMValueRef f32undef = LLVMGetUndef(ctx->ac.f32);
-	LLVMBuilderRef builder = ctx->ac.builder;
-	LLVMValueRef val[4];
 	unsigned spi_shader_col_format = V_028714_SPI_SHADER_32_ABGR;
 	unsigned chan;
 	bool is_int8, is_int10;
@@ -2452,6 +2428,10 @@ static void si_llvm_init_export_args(struct si_shader_context *ctx,
 	args->out[2] = f32undef;
 	args->out[3] = f32undef;
 
+	LLVMValueRef (*packf)(struct ac_llvm_context *ctx, LLVMValueRef args[2]) = NULL;
+	LLVMValueRef (*packi)(struct ac_llvm_context *ctx, LLVMValueRef args[2],
+			      unsigned bits, bool hi) = NULL;
+
 	switch (spi_shader_col_format) {
 	case V_028714_SPI_SHADER_ZERO:
 		args->enabled_channels = 0; /* writemask */
@@ -2476,113 +2456,59 @@ static void si_llvm_init_export_args(struct si_shader_context *ctx,
 		break;
 
 	case V_028714_SPI_SHADER_FP16_ABGR:
-		args->compr = 1; /* COMPR flag */
-
-		for (chan = 0; chan < 2; chan++) {
-			LLVMValueRef pack_args[2] = {
-				values[2 * chan],
-				values[2 * chan + 1]
-			};
-			LLVMValueRef packed;
-
-			packed = ac_build_cvt_pkrtz_f16(&ctx->ac, pack_args);
-			args->out[chan] = ac_to_float(&ctx->ac, packed);
-		}
+		packf = ac_build_cvt_pkrtz_f16;
 		break;
 
 	case V_028714_SPI_SHADER_UNORM16_ABGR:
-		for (chan = 0; chan < 4; chan++) {
-			val[chan] = ac_build_clamp(&ctx->ac, values[chan]);
-			val[chan] = LLVMBuildFMul(builder, val[chan],
-						  LLVMConstReal(ctx->f32, 65535), "");
-			val[chan] = LLVMBuildFAdd(builder, val[chan],
-						  LLVMConstReal(ctx->f32, 0.5), "");
-			val[chan] = LLVMBuildFPToUI(builder, val[chan],
-						    ctx->i32, "");
-		}
-
-		args->compr = 1; /* COMPR flag */
-		args->out[0] = ac_to_float(&ctx->ac, si_llvm_pack_two_int16(ctx, val));
-		args->out[1] = ac_to_float(&ctx->ac, si_llvm_pack_two_int16(ctx, val+2));
+		packf = ac_build_cvt_pknorm_u16;
 		break;
 
 	case V_028714_SPI_SHADER_SNORM16_ABGR:
-		for (chan = 0; chan < 4; chan++) {
-			/* Clamp between [-1, 1]. */
-			val[chan] = lp_build_emit_llvm_binary(&ctx->bld_base, TGSI_OPCODE_MIN,
-							      values[chan],
-							      LLVMConstReal(ctx->f32, 1));
-			val[chan] = lp_build_emit_llvm_binary(&ctx->bld_base, TGSI_OPCODE_MAX,
-							      val[chan],
-							      LLVMConstReal(ctx->f32, -1));
-			/* Convert to a signed integer in [-32767, 32767]. */
-			val[chan] = LLVMBuildFMul(builder, val[chan],
-						  LLVMConstReal(ctx->f32, 32767), "");
-			/* If positive, add 0.5, else add -0.5. */
-			val[chan] = LLVMBuildFAdd(builder, val[chan],
-					LLVMBuildSelect(builder,
-						LLVMBuildFCmp(builder, LLVMRealOGE,
-							      val[chan], ctx->ac.f32_0, ""),
-						LLVMConstReal(ctx->f32, 0.5),
-						LLVMConstReal(ctx->f32, -0.5), ""), "");
-			val[chan] = LLVMBuildFPToSI(builder, val[chan], ctx->i32, "");
-		}
-
-		args->compr = 1; /* COMPR flag */
-		args->out[0] = ac_to_float(&ctx->ac, si_llvm_pack_two_int32_as_int16(ctx, val));
-		args->out[1] = ac_to_float(&ctx->ac, si_llvm_pack_two_int32_as_int16(ctx, val+2));
+		packf = ac_build_cvt_pknorm_i16;
 		break;
 
-	case V_028714_SPI_SHADER_UINT16_ABGR: {
-		LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
-			is_int8 ? 255 : is_int10 ? 1023 : 65535, 0);
-		LLVMValueRef max_alpha =
-			!is_int10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
+	case V_028714_SPI_SHADER_UINT16_ABGR:
+		packi = ac_build_cvt_pk_u16;
+		break;
 
-		/* Clamp. */
-		for (chan = 0; chan < 4; chan++) {
-			val[chan] = ac_to_integer(&ctx->ac, values[chan]);
-			val[chan] = lp_build_emit_llvm_binary(&ctx->bld_base, TGSI_OPCODE_UMIN,
-					val[chan],
-					chan == 3 ? max_alpha : max_rgb);
-		}
+	case V_028714_SPI_SHADER_SINT16_ABGR:
+		packi = ac_build_cvt_pk_i16;
+		break;
 
-		args->compr = 1; /* COMPR flag */
-		args->out[0] = ac_to_float(&ctx->ac, si_llvm_pack_two_int16(ctx, val));
-		args->out[1] = ac_to_float(&ctx->ac, si_llvm_pack_two_int16(ctx, val+2));
+	case V_028714_SPI_SHADER_32_ABGR:
+		memcpy(&args->out[0], values, sizeof(values[0]) * 4);
 		break;
 	}
 
-	case V_028714_SPI_SHADER_SINT16_ABGR: {
-		LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
-			is_int8 ? 127 : is_int10 ? 511 : 32767, 0);
-		LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
-			is_int8 ? -128 : is_int10 ? -512 : -32768, 0);
-		LLVMValueRef max_alpha =
-			!is_int10 ? max_rgb : ctx->i32_1;
-		LLVMValueRef min_alpha =
-			!is_int10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
+	/* Pack f16 or norm_i16/u16. */
+	if (packf) {
+		for (chan = 0; chan < 2; chan++) {
+			LLVMValueRef pack_args[2] = {
+				values[2 * chan],
+				values[2 * chan + 1]
+			};
+			LLVMValueRef packed;
 
-		/* Clamp. */
-		for (chan = 0; chan < 4; chan++) {
-			val[chan] = ac_to_integer(&ctx->ac, values[chan]);
-			val[chan] = lp_build_emit_llvm_binary(&ctx->bld_base,
-					TGSI_OPCODE_IMIN,
-					val[chan], chan == 3 ? max_alpha : max_rgb);
-			val[chan] = lp_build_emit_llvm_binary(&ctx->bld_base,
-					TGSI_OPCODE_IMAX,
-					val[chan], chan == 3 ? min_alpha : min_rgb);
+			packed = packf(&ctx->ac, pack_args);
+			args->out[chan] = ac_to_float(&ctx->ac, packed);
 		}
-
 		args->compr = 1; /* COMPR flag */
-		args->out[0] = ac_to_float(&ctx->ac, si_llvm_pack_two_int32_as_int16(ctx, val));
-		args->out[1] = ac_to_float(&ctx->ac, si_llvm_pack_two_int32_as_int16(ctx, val+2));
-		break;
 	}
+	/* Pack i16/u16. */
+	if (packi) {
+		for (chan = 0; chan < 2; chan++) {
+			LLVMValueRef pack_args[2] = {
+				ac_to_integer(&ctx->ac, values[2 * chan]),
+				ac_to_integer(&ctx->ac, values[2 * chan + 1])
+			};
+			LLVMValueRef packed;
 
-	case V_028714_SPI_SHADER_32_ABGR:
-		memcpy(&args->out[0], values, sizeof(values[0]) * 4);
-		break;
+			packed = packi(&ctx->ac, pack_args,
+				       is_int8 ? 8 : is_int10 ? 10 : 16,
+				       chan == 1);
+			args->out[chan] = ac_to_float(&ctx->ac, packed);
+		}
+		args->compr = 1; /* COMPR flag */
 	}
 }