Make image builtins r600/llvm-3.9 only

The implementation uses r600 sepcific intrinsics
LLVM-4 switched to _ro_t and _rw_t image types
Portions of the code can be moved back as more targets/llvm versions add image support

Reviewer: Aaron Watry
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 315341

16 files changed, 356 insertions, 0 deletions

diff --git a/libclc/r600/lib/SOURCES_3.9 b/libclc/r600/lib/SOURCES_3.9
new file mode 100644
index 0000000..a44a9ce
--- /dev/null
+++ b/libclc/r600/lib/SOURCES_3.9
@@ -0,0 +1,15 @@
+image/get_image_dim.cl
+image/get_image_width.cl
+image/get_image_height.cl
+image/get_image_depth.cl
+image/get_image_channel_data_type.cl
+image/get_image_channel_order.cl
+image/get_image_attributes_impl.ll
+image/read_imagef.cl
+image/read_imagei.cl
+image/read_imageui.cl
+image/read_image_impl.ll
+image/write_imagef.cl
+image/write_imagei.cl
+image/write_imageui.cl
+image/write_image_impl.ll
diff --git a/libclc/r600/lib/image/get_image_attributes_impl.ll b/libclc/r600/lib/image/get_image_attributes_impl.ll
new file mode 100644
index 0000000..7f1965d
--- /dev/null
+++ b/libclc/r600/lib/image/get_image_attributes_impl.ll
@@ -0,0 +1,87 @@
+%opencl.image2d_t = type opaque
+%opencl.image3d_t = type opaque
+
+declare i32 @llvm.OpenCL.image.get.resource.id.2d(
+  %opencl.image2d_t addrspace(1)*) nounwind readnone
+declare i32 @llvm.OpenCL.image.get.resource.id.3d(
+  %opencl.image3d_t addrspace(1)*) nounwind readnone
+
+declare [3 x i32] @llvm.OpenCL.image.get.size.2d(
+  %opencl.image2d_t addrspace(1)*) nounwind readnone
+declare [3 x i32] @llvm.OpenCL.image.get.size.3d(
+  %opencl.image3d_t addrspace(1)*) nounwind readnone
+
+declare [2 x i32] @llvm.OpenCL.image.get.format.2d(
+  %opencl.image2d_t addrspace(1)*) nounwind readnone
+declare [2 x i32] @llvm.OpenCL.image.get.format.3d(
+  %opencl.image3d_t addrspace(1)*) nounwind readnone
+
+define i32 @__clc_get_image_width_2d(
+                          %opencl.image2d_t addrspace(1)* nocapture %img) #0 {
+  %1 = tail call [3 x i32] @llvm.OpenCL.image.get.size.2d(
+    %opencl.image2d_t addrspace(1)* %img)
+  %2 = extractvalue [3 x i32] %1, 0
+  ret i32 %2
+}
+define i32 @__clc_get_image_width_3d(
+                          %opencl.image3d_t addrspace(1)* nocapture %img) #0 {
+  %1 = tail call [3 x i32] @llvm.OpenCL.image.get.size.3d(
+    %opencl.image3d_t addrspace(1)* %img)
+  %2 = extractvalue [3 x i32] %1, 0
+  ret i32 %2
+}
+
+define i32 @__clc_get_image_height_2d(
+                          %opencl.image2d_t addrspace(1)* nocapture %img) #0 {
+  %1 = tail call [3 x i32] @llvm.OpenCL.image.get.size.2d(
+    %opencl.image2d_t addrspace(1)* %img)
+  %2 = extractvalue [3 x i32] %1, 1
+  ret i32 %2
+}
+define i32 @__clc_get_image_height_3d(
+                          %opencl.image3d_t addrspace(1)* nocapture %img) #0 {
+  %1 = tail call [3 x i32] @llvm.OpenCL.image.get.size.3d(
+    %opencl.image3d_t addrspace(1)* %img)
+  %2 = extractvalue [3 x i32] %1, 1
+  ret i32 %2
+}
+
+define i32 @__clc_get_image_depth_3d(
+                          %opencl.image3d_t addrspace(1)* nocapture %img) #0 {
+  %1 = tail call [3 x i32] @llvm.OpenCL.image.get.size.3d(
+    %opencl.image3d_t addrspace(1)* %img)
+  %2 = extractvalue [3 x i32] %1, 2
+  ret i32 %2
+}
+
+define i32 @__clc_get_image_channel_data_type_2d(
+                          %opencl.image2d_t addrspace(1)* nocapture %img) #0 {
+  %1 = tail call [2 x i32] @llvm.OpenCL.image.get.format.2d(
+    %opencl.image2d_t addrspace(1)* %img)
+  %2 = extractvalue [2 x i32] %1, 0
+  ret i32 %2
+}
+define i32 @__clc_get_image_channel_data_type_3d(
+                          %opencl.image3d_t addrspace(1)* nocapture %img) #0 {
+  %1 = tail call [2 x i32] @llvm.OpenCL.image.get.format.3d(
+    %opencl.image3d_t addrspace(1)* %img)
+  %2 = extractvalue [2 x i32] %1, 0
+  ret i32 %2
+}
+
+define i32 @__clc_get_image_channel_order_2d(
+                          %opencl.image2d_t addrspace(1)* nocapture %img) #0 {
+  %1 = tail call [2 x i32] @llvm.OpenCL.image.get.format.2d(
+    %opencl.image2d_t addrspace(1)* %img)
+  %2 = extractvalue [2 x i32] %1, 1
+  ret i32 %2
+}
+define i32 @__clc_get_image_channel_order_3d(
+                          %opencl.image3d_t addrspace(1)* nocapture %img) #0 {
+  %1 = tail call [2 x i32] @llvm.OpenCL.image.get.format.3d(
+    %opencl.image3d_t addrspace(1)* %img)
+  %2 = extractvalue [2 x i32] %1, 1
+  ret i32 %2
+}
+
+attributes #0 = { nounwind readnone alwaysinline }
diff --git a/libclc/r600/lib/image/get_image_channel_data_type.cl b/libclc/r600/lib/image/get_image_channel_data_type.cl
new file mode 100644
index 0000000..2a2478f
--- /dev/null
+++ b/libclc/r600/lib/image/get_image_channel_data_type.cl
@@ -0,0 +1,13 @@
+#include <clc/clc.h>
+
+_CLC_DECL int __clc_get_image_channel_data_type_2d(image2d_t);
+_CLC_DECL int __clc_get_image_channel_data_type_3d(image3d_t);
+
+_CLC_OVERLOAD _CLC_DEF int
+get_image_channel_data_type(image2d_t image) {
+  return __clc_get_image_channel_data_type_2d(image);
+}
+_CLC_OVERLOAD _CLC_DEF int
+get_image_channel_data_type(image3d_t image) {
+  return __clc_get_image_channel_data_type_3d(image);
+}
diff --git a/libclc/r600/lib/image/get_image_channel_order.cl b/libclc/r600/lib/image/get_image_channel_order.cl
new file mode 100644
index 0000000..91e9b89
--- /dev/null
+++ b/libclc/r600/lib/image/get_image_channel_order.cl
@@ -0,0 +1,13 @@
+#include <clc/clc.h>
+
+_CLC_DECL int __clc_get_image_channel_order_2d(image2d_t);
+_CLC_DECL int __clc_get_image_channel_order_3d(image3d_t);
+
+_CLC_OVERLOAD _CLC_DEF int
+get_image_channel_order(image2d_t image) {
+  return __clc_get_image_channel_order_2d(image);
+}
+_CLC_OVERLOAD _CLC_DEF int
+get_image_channel_order(image3d_t image) {
+  return __clc_get_image_channel_order_3d(image);
+}
diff --git a/libclc/r600/lib/image/get_image_depth.cl b/libclc/r600/lib/image/get_image_depth.cl
new file mode 100644
index 0000000..1864645
--- /dev/null
+++ b/libclc/r600/lib/image/get_image_depth.cl
@@ -0,0 +1,8 @@
+#include <clc/clc.h>
+
+_CLC_DECL int __clc_get_image_depth_3d(image3d_t);
+
+_CLC_OVERLOAD _CLC_DEF int
+get_image_depth(image3d_t image) {
+	return __clc_get_image_depth_3d(image);
+}
diff --git a/libclc/r600/lib/image/get_image_dim.cl b/libclc/r600/lib/image/get_image_dim.cl
new file mode 100644
index 0000000..26dbd00
--- /dev/null
+++ b/libclc/r600/lib/image/get_image_dim.cl
@@ -0,0 +1,9 @@
+#include <clc/clc.h>
+
+_CLC_OVERLOAD _CLC_DEF int2 get_image_dim (image2d_t image) {
+  return (int2)(get_image_width(image), get_image_height(image));
+}
+_CLC_OVERLOAD _CLC_DEF int4 get_image_dim (image3d_t image) {
+  return (int4)(get_image_width(image), get_image_height(image),
+                get_image_depth(image), 0);
+}
diff --git a/libclc/r600/lib/image/get_image_height.cl b/libclc/r600/lib/image/get_image_height.cl
new file mode 100644
index 0000000..80b3640
--- /dev/null
+++ b/libclc/r600/lib/image/get_image_height.cl
@@ -0,0 +1,13 @@
+#include <clc/clc.h>
+
+_CLC_DECL int __clc_get_image_height_2d(image2d_t);
+_CLC_DECL int __clc_get_image_height_3d(image3d_t);
+
+_CLC_OVERLOAD _CLC_DEF int
+get_image_height(image2d_t image) {
+  return __clc_get_image_height_2d(image);
+}
+_CLC_OVERLOAD _CLC_DEF int
+get_image_height(image3d_t image) {
+  return __clc_get_image_height_3d(image);
+}
diff --git a/libclc/r600/lib/image/get_image_width.cl b/libclc/r600/lib/image/get_image_width.cl
new file mode 100644
index 0000000..29e4e94
--- /dev/null
+++ b/libclc/r600/lib/image/get_image_width.cl
@@ -0,0 +1,13 @@
+#include <clc/clc.h>
+
+_CLC_DECL int __clc_get_image_width_2d(image2d_t);
+_CLC_DECL int __clc_get_image_width_3d(image3d_t);
+
+_CLC_OVERLOAD _CLC_DEF int
+get_image_width(image2d_t image) {
+  return __clc_get_image_width_2d(image);
+}
+_CLC_OVERLOAD _CLC_DEF int
+get_image_width(image3d_t image) {
+  return __clc_get_image_width_3d(image);
+}
diff --git a/libclc/r600/lib/image/read_image_impl.ll b/libclc/r600/lib/image/read_image_impl.ll
new file mode 100644
index 0000000..229a252
--- /dev/null
+++ b/libclc/r600/lib/image/read_image_impl.ll
@@ -0,0 +1,46 @@
+%opencl.image2d_t = type opaque
+
+declare <4 x float> @llvm.R600.tex(<4 x float>, i32, i32, i32, i32, i32, i32,
+                                   i32, i32, i32) readnone
+declare i32 @llvm.OpenCL.image.get.resource.id.2d(
+  %opencl.image2d_t addrspace(1)*) nounwind readnone
+declare i32 @llvm.OpenCL.sampler.get.resource.id(i32) readnone
+
+define <4 x float> @__clc_v4f_from_v2f(<2 x float> %v) alwaysinline {
+  %e0 = extractelement <2 x float> %v, i32 0
+  %e1 = extractelement <2 x float> %v, i32 1
+  %res.0 = insertelement <4 x float> undef,  float %e0, i32 0
+  %res.1 = insertelement <4 x float> %res.0, float %e1, i32 1
+  %res.2 = insertelement <4 x float> %res.1, float 0.0, i32 2
+  %res.3 = insertelement <4 x float> %res.2, float 0.0, i32 3
+  ret <4 x float> %res.3
+}
+
+define <4 x float> @__clc_read_imagef_tex(
+    %opencl.image2d_t addrspace(1)* nocapture %img,
+    i32 %sampler, <2 x float> %coord) alwaysinline {
+entry:
+  %coord_v4 = call <4 x float> @__clc_v4f_from_v2f(<2 x float> %coord)
+  %smp_id = call i32 @llvm.OpenCL.sampler.get.resource.id(i32 %sampler)
+  %img_id = call i32 @llvm.OpenCL.image.get.resource.id.2d(
+      %opencl.image2d_t addrspace(1)* %img)
+  %tex_id = add i32 %img_id, 2    ; First 2 IDs are reserved.
+
+  %coord_norm = and i32 %sampler, 1
+  %is_norm = icmp eq i32 %coord_norm, 1
+  br i1 %is_norm, label %NormCoord, label %UnnormCoord
+NormCoord:
+  %data.norm = call <4 x float> @llvm.R600.tex(
+      <4 x float> %coord_v4,
+      i32 0, i32 0, i32 0,        ; Offset.
+      i32 2, i32 %smp_id,
+      i32 1, i32 1, i32 1, i32 1) ; Normalized coords.
+  ret <4 x float> %data.norm
+UnnormCoord:
+  %data.unnorm = call <4 x float> @llvm.R600.tex(
+      <4 x float> %coord_v4,
+      i32 0, i32 0, i32 0,        ; Offset.
+      i32 %tex_id, i32 %smp_id,
+      i32 0, i32 0, i32 0, i32 0) ; Unnormalized coords.
+  ret <4 x float> %data.unnorm
+}
diff --git a/libclc/r600/lib/image/read_imagef.cl b/libclc/r600/lib/image/read_imagef.cl
new file mode 100644
index 0000000..af80ada
--- /dev/null
+++ b/libclc/r600/lib/image/read_imagef.cl
@@ -0,0 +1,14 @@
+#include <clc/clc.h>
+
+_CLC_DECL float4 __clc_read_imagef_tex(image2d_t, sampler_t, float2);
+
+_CLC_OVERLOAD _CLC_DEF float4 read_imagef(image2d_t image, sampler_t sampler,
+                                          int2 coord) {
+  float2 coord_float = (float2)(coord.x, coord.y);
+  return __clc_read_imagef_tex(image, sampler, coord_float);
+}
+
+_CLC_OVERLOAD _CLC_DEF float4 read_imagef(image2d_t image, sampler_t sampler,
+                                          float2 coord) {
+  return __clc_read_imagef_tex(image, sampler, coord);
+}
diff --git a/libclc/r600/lib/image/read_imagei.cl b/libclc/r600/lib/image/read_imagei.cl
new file mode 100644
index 0000000..b973aae
--- /dev/null
+++ b/libclc/r600/lib/image/read_imagei.cl
@@ -0,0 +1,23 @@
+#include <clc/clc.h>
+
+_CLC_DECL float4 __clc_read_imagef_tex(image2d_t, sampler_t, float2);
+
+int4 __clc_reinterpret_v4f_to_v4i(float4 v) {
+  union {
+    int4 v4i;
+    float4 v4f;
+  } res = { .v4f = v};
+  return res.v4i;
+}
+
+_CLC_OVERLOAD _CLC_DEF int4 read_imagei(image2d_t image, sampler_t sampler,
+                                        int2 coord) {
+  float2 coord_float = (float2)(coord.x, coord.y);
+  return __clc_reinterpret_v4f_to_v4i(
+    __clc_read_imagef_tex(image, sampler, coord_float));
+}
+_CLC_OVERLOAD _CLC_DEF int4 read_imagei(image2d_t image, sampler_t sampler,
+                                        float2 coord) {
+  return __clc_reinterpret_v4f_to_v4i(
+    __clc_read_imagef_tex(image, sampler, coord));
+}
diff --git a/libclc/r600/lib/image/read_imageui.cl b/libclc/r600/lib/image/read_imageui.cl
new file mode 100644
index 0000000..ec9836e
--- /dev/null
+++ b/libclc/r600/lib/image/read_imageui.cl
@@ -0,0 +1,23 @@
+#include <clc/clc.h>
+
+_CLC_DECL float4 __clc_read_imagef_tex(image2d_t, sampler_t, float2);
+
+uint4 __clc_reinterpret_v4f_to_v4ui(float4 v) {
+  union {
+    uint4 v4ui;
+    float4 v4f;
+  } res = { .v4f = v};
+  return res.v4ui;
+}
+
+_CLC_OVERLOAD _CLC_DEF uint4 read_imageui(image2d_t image, sampler_t sampler,
+                                          int2 coord) {
+  float2 coord_float = (float2)(coord.x, coord.y);
+  return __clc_reinterpret_v4f_to_v4ui(
+    __clc_read_imagef_tex(image, sampler, coord_float));
+}
+_CLC_OVERLOAD _CLC_DEF uint4 read_imageui(image2d_t image, sampler_t sampler,
+                                          float2 coord) {
+  return __clc_reinterpret_v4f_to_v4ui(
+    __clc_read_imagef_tex(image, sampler, coord));
+}
diff --git a/libclc/r600/lib/image/write_image_impl.ll b/libclc/r600/lib/image/write_image_impl.ll
new file mode 100644
index 0000000..265f5d6
--- /dev/null
+++ b/libclc/r600/lib/image/write_image_impl.ll
@@ -0,0 +1,52 @@
+%opencl.image2d_t = type opaque
+%opencl.image3d_t = type opaque
+
+declare i32 @llvm.OpenCL.image.get.resource.id.2d(
+  %opencl.image2d_t addrspace(1)*) nounwind readnone
+declare i32 @llvm.OpenCL.image.get.resource.id.3d(
+  %opencl.image3d_t addrspace(1)*) nounwind readnone
+
+declare void @llvm.r600.rat.store.typed(<4 x i32> %color, <4 x i32> %coord, i32 %rat_id)
+
+define void @__clc_write_imageui_2d(
+    %opencl.image2d_t addrspace(1)* nocapture %img,
+    <2 x i32> %coord, <4 x i32> %color) #0 {
+
+  ; Coordinate int2 -> int4.
+  %e0 = extractelement <2 x i32> %coord, i32 0
+  %e1 = extractelement <2 x i32> %coord, i32 1
+  %coord.0 = insertelement <4 x i32> undef,    i32 %e0, i32 0
+  %coord.1 = insertelement <4 x i32> %coord.0, i32 %e1, i32 1
+  %coord.2 = insertelement <4 x i32> %coord.1, i32 0,  i32 2
+  %coord.3 = insertelement <4 x i32> %coord.2, i32 0,  i32 3
+
+  ; Get RAT ID.
+  %img_id = call i32 @llvm.OpenCL.image.get.resource.id.2d(
+      %opencl.image2d_t addrspace(1)* %img)
+  %rat_id = add i32 %img_id, 1
+
+  ; Call store intrinsic.
+  call void @llvm.r600.rat.store.typed(<4 x i32> %color, <4 x i32> %coord.3, i32 %rat_id)
+  ret void
+}
+
+define void @__clc_write_imagei_2d(
+    %opencl.image2d_t addrspace(1)* nocapture %img,
+    <2 x i32> %coord, <4 x i32> %color) #0 {
+  call void @__clc_write_imageui_2d(
+      %opencl.image2d_t addrspace(1)* nocapture %img,
+      <2 x i32> %coord, <4 x i32> %color)
+  ret void
+}
+
+define void @__clc_write_imagef_2d(
+    %opencl.image2d_t addrspace(1)* nocapture %img,
+    <2 x i32> %coord, <4 x float> %color) #0 {
+  %color.i32 = bitcast <4 x float> %color to <4 x i32>
+  call void @__clc_write_imageui_2d(
+      %opencl.image2d_t addrspace(1)* nocapture %img,
+      <2 x i32> %coord, <4 x i32> %color.i32)
+  ret void
+}
+
+attributes #0 = { alwaysinline }
diff --git a/libclc/r600/lib/image/write_imagef.cl b/libclc/r600/lib/image/write_imagef.cl
new file mode 100644
index 0000000..4483fcf
--- /dev/null
+++ b/libclc/r600/lib/image/write_imagef.cl
@@ -0,0 +1,9 @@
+#include <clc/clc.h>
+
+_CLC_DECL void __clc_write_imagef_2d(image2d_t image, int2 coord, float4 color);
+
+_CLC_OVERLOAD _CLC_DEF void
+write_imagef(image2d_t image, int2 coord, float4 color)
+{
+  __clc_write_imagef_2d(image, coord, color);
+}
diff --git a/libclc/r600/lib/image/write_imagei.cl b/libclc/r600/lib/image/write_imagei.cl
new file mode 100644
index 0000000..394a223
--- /dev/null
+++ b/libclc/r600/lib/image/write_imagei.cl
@@ -0,0 +1,9 @@
+#include <clc/clc.h>
+
+_CLC_DECL void __clc_write_imagei_2d(image2d_t image, int2 coord, int4 color);
+
+_CLC_OVERLOAD _CLC_DEF void
+write_imagei(image2d_t image, int2 coord, int4 color)
+{
+  __clc_write_imagei_2d(image, coord, color);
+}
diff --git a/libclc/r600/lib/image/write_imageui.cl b/libclc/r600/lib/image/write_imageui.cl
new file mode 100644
index 0000000..91344de
--- /dev/null
+++ b/libclc/r600/lib/image/write_imageui.cl
@@ -0,0 +1,9 @@
+#include <clc/clc.h>
+
+_CLC_DECL void __clc_write_imageui_2d(image2d_t image, int2 coord, uint4 color);
+
+_CLC_OVERLOAD _CLC_DEF void
+write_imageui(image2d_t image, int2 coord, uint4 color)
+{
+  __clc_write_imageui_2d(image, coord, color);
+}