Cycles / CUDA: Increase maximum image textures on GPU.

Instead of 95, we can use 145 images now. This only affects Kepler and above (sm30, sm_35 and sm_50).

This can be increased further if needed, but let's first test if this does not come with a performance impact.

Originally developed during my GSoC 2013.

9 files changed, 139 insertions, 13 deletions

diff --git a/intern/cycles/device/device.h b/intern/cycles/device/device.h
index cbabcb1e20e..bcddd4f73e2 100644
--- a/intern/cycles/device/device.h
+++ b/intern/cycles/device/device.h
@@ -54,6 +54,7 @@ public:
 	bool display_device;
 	bool advanced_shading;
 	bool pack_images;
+	bool extended_images; /* flag for GPU and Multi device */
 	vector<DeviceInfo> multi_devices;
 
 	DeviceInfo()
@@ -64,6 +65,7 @@ public:
 		display_device = false;
 		advanced_shading = true;
 		pack_images = false;
+		extended_images = false;
 	}
 };
 
diff --git a/intern/cycles/device/device_cuda.cpp b/intern/cycles/device/device_cuda.cpp
index 9139a75ef3e..68955211146 100644
--- a/intern/cycles/device/device_cuda.cpp
+++ b/intern/cycles/device/device_cuda.cpp
@@ -1150,6 +1150,7 @@ void device_cuda_info(vector<DeviceInfo>& devices)
 		int major, minor;
 		cuDeviceComputeCapability(&major, &minor, num);
 		info.advanced_shading = (major >= 2);
+		info.extended_images = (major >= 3);
 		info.pack_images = false;
 
 		/* if device has a kernel timeout, assume it is used for display */
diff --git a/intern/cycles/device/device_multi.cpp b/intern/cycles/device/device_multi.cpp
index 59bbf3b9d5a..c866ebaaea2 100644
--- a/intern/cycles/device/device_multi.cpp
+++ b/intern/cycles/device/device_multi.cpp
@@ -328,6 +328,7 @@ static bool device_multi_add(vector<DeviceInfo>& devices, DeviceType type, bool
 
 	info.advanced_shading = with_advanced_shading;
 	info.pack_images = false;
+	info.extended_images = true;
 
 	foreach(DeviceInfo& subinfo, devices) {
 		if(subinfo.type == type) {
@@ -351,6 +352,7 @@ static bool device_multi_add(vector<DeviceInfo>& devices, DeviceType type, bool
 			if(subinfo.display_device)
 				info.display_device = true;
 			info.pack_images = info.pack_images || subinfo.pack_images;
+			info.extended_images = info.extended_images && subinfo.extended_images;
 			num_added++;
 		}
 	}
diff --git a/intern/cycles/kernel/kernel_textures.h b/intern/cycles/kernel/kernel_textures.h
index c8734d67c3b..94115fd388c 100644
--- a/intern/cycles/kernel/kernel_textures.h
+++ b/intern/cycles/kernel/kernel_textures.h
@@ -174,6 +174,61 @@ KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_097)
 KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_098)
 KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_099)
 
+/* Kepler and above */
+#if __CUDA_ARCH__ >= 300
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_100)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_101)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_102)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_103)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_104)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_105)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_106)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_107)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_108)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_109)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_110)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_111)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_112)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_113)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_114)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_115)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_116)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_117)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_118)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_119)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_120)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_121)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_122)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_123)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_124)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_125)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_126)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_127)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_128)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_129)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_130)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_131)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_132)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_133)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_134)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_135)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_136)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_137)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_138)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_139)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_140)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_141)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_142)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_143)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_144)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_145)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_146)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_147)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_148)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_149)
+KERNEL_IMAGE_TEX(uchar4, texture_image_uchar4, __tex_image_150)
+#endif
+
 /* packed image (opencl) */
 KERNEL_TEX(uchar4, texture_uchar4, __tex_image_packed)
 KERNEL_TEX(uint4, texture_uint4, __tex_image_packed_info)
diff --git a/intern/cycles/kernel/svm/svm_image.h b/intern/cycles/kernel/svm/svm_image.h
index 6627786725f..6c658afb9df 100644
--- a/intern/cycles/kernel/svm/svm_image.h
+++ b/intern/cycles/kernel/svm/svm_image.h
@@ -149,8 +149,8 @@ ccl_device float4 svm_image_texture(KernelGlobals *kg, int id, float x, float y,
 	 * - group by size and use a 3d texture, performance impact
 	 * - group into larger texture with some padding for correct lerp
 	 *
-	 * also note that cuda has 128 textures limit, we use 100 now, since
-	 * we still need some for other storage */
+	 * also note that cuda has a textures limit (128 for Fermi, 256 for Kepler),
+	 * and we cannot use all since we still need some for other storage */
 
 	switch(id) {
 		case 0: r = kernel_tex_image_interp(__tex_image_float_000, x, y); break;
@@ -253,7 +253,62 @@ ccl_device float4 svm_image_texture(KernelGlobals *kg, int id, float x, float y,
 		case 97: r = kernel_tex_image_interp(__tex_image_097, x, y); break;
 		case 98: r = kernel_tex_image_interp(__tex_image_098, x, y); break;
 		case 99: r = kernel_tex_image_interp(__tex_image_099, x, y); break;
-		default: 
+
+#if __CUDA_ARCH__ >= 300
+		case 100: r = kernel_tex_image_interp(__tex_image_100, x, y); break;
+		case 101: r = kernel_tex_image_interp(__tex_image_101, x, y); break;
+		case 102: r = kernel_tex_image_interp(__tex_image_102, x, y); break;
+		case 103: r = kernel_tex_image_interp(__tex_image_103, x, y); break;
+		case 104: r = kernel_tex_image_interp(__tex_image_104, x, y); break;
+		case 105: r = kernel_tex_image_interp(__tex_image_105, x, y); break;
+		case 106: r = kernel_tex_image_interp(__tex_image_106, x, y); break;
+		case 107: r = kernel_tex_image_interp(__tex_image_107, x, y); break;
+		case 108: r = kernel_tex_image_interp(__tex_image_108, x, y); break;
+		case 109: r = kernel_tex_image_interp(__tex_image_109, x, y); break;
+		case 110: r = kernel_tex_image_interp(__tex_image_110, x, y); break;
+		case 111: r = kernel_tex_image_interp(__tex_image_111, x, y); break;
+		case 112: r = kernel_tex_image_interp(__tex_image_112, x, y); break;
+		case 113: r = kernel_tex_image_interp(__tex_image_113, x, y); break;
+		case 114: r = kernel_tex_image_interp(__tex_image_114, x, y); break;
+		case 115: r = kernel_tex_image_interp(__tex_image_115, x, y); break;
+		case 116: r = kernel_tex_image_interp(__tex_image_116, x, y); break;
+		case 117: r = kernel_tex_image_interp(__tex_image_117, x, y); break;
+		case 118: r = kernel_tex_image_interp(__tex_image_118, x, y); break;
+		case 119: r = kernel_tex_image_interp(__tex_image_119, x, y); break;
+		case 120: r = kernel_tex_image_interp(__tex_image_120, x, y); break;
+		case 121: r = kernel_tex_image_interp(__tex_image_121, x, y); break;
+		case 122: r = kernel_tex_image_interp(__tex_image_122, x, y); break;
+		case 123: r = kernel_tex_image_interp(__tex_image_123, x, y); break;
+		case 124: r = kernel_tex_image_interp(__tex_image_124, x, y); break;
+		case 125: r = kernel_tex_image_interp(__tex_image_125, x, y); break;
+		case 126: r = kernel_tex_image_interp(__tex_image_126, x, y); break;
+		case 127: r = kernel_tex_image_interp(__tex_image_127, x, y); break;
+		case 128: r = kernel_tex_image_interp(__tex_image_128, x, y); break;
+		case 129: r = kernel_tex_image_interp(__tex_image_129, x, y); break;
+		case 130: r = kernel_tex_image_interp(__tex_image_130, x, y); break;
+		case 131: r = kernel_tex_image_interp(__tex_image_131, x, y); break;
+		case 132: r = kernel_tex_image_interp(__tex_image_132, x, y); break;
+		case 133: r = kernel_tex_image_interp(__tex_image_133, x, y); break;
+		case 134: r = kernel_tex_image_interp(__tex_image_134, x, y); break;
+		case 135: r = kernel_tex_image_interp(__tex_image_135, x, y); break;
+		case 136: r = kernel_tex_image_interp(__tex_image_136, x, y); break;
+		case 137: r = kernel_tex_image_interp(__tex_image_137, x, y); break;
+		case 138: r = kernel_tex_image_interp(__tex_image_138, x, y); break;
+		case 139: r = kernel_tex_image_interp(__tex_image_139, x, y); break;
+		case 140: r = kernel_tex_image_interp(__tex_image_140, x, y); break;
+		case 141: r = kernel_tex_image_interp(__tex_image_141, x, y); break;
+		case 142: r = kernel_tex_image_interp(__tex_image_142, x, y); break;
+		case 143: r = kernel_tex_image_interp(__tex_image_143, x, y); break;
+		case 144: r = kernel_tex_image_interp(__tex_image_144, x, y); break;
+		case 145: r = kernel_tex_image_interp(__tex_image_145, x, y); break;
+		case 146: r = kernel_tex_image_interp(__tex_image_146, x, y); break;
+		case 147: r = kernel_tex_image_interp(__tex_image_147, x, y); break;
+		case 148: r = kernel_tex_image_interp(__tex_image_148, x, y); break;
+		case 149: r = kernel_tex_image_interp(__tex_image_149, x, y); break;
+		case 150: r = kernel_tex_image_interp(__tex_image_150, x, y); break;
+#endif
+
+		default:
 			kernel_assert(0);
 			return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
 	}
diff --git a/intern/cycles/render/image.cpp b/intern/cycles/render/image.cpp
index 8dde642f70b..86755badc42 100644
--- a/intern/cycles/render/image.cpp
+++ b/intern/cycles/render/image.cpp
@@ -59,11 +59,16 @@ void ImageManager::set_osl_texture_system(void *texture_system)
 	osl_texture_system = texture_system;
 }
 
-void ImageManager::set_extended_image_limits(void)
+void ImageManager::set_extended_image_limits(const DeviceInfo& info)
 {
-	tex_num_images = TEX_EXTENDED_NUM_IMAGES;
-	tex_num_float_images = TEX_EXTENDED_NUM_FLOAT_IMAGES;
-	tex_image_byte_start = TEX_EXTENDED_IMAGE_BYTE_START;
+	if(info.type == DEVICE_CPU) {
+		tex_num_images = TEX_EXTENDED_NUM_IMAGES_CPU;
+		tex_num_float_images = TEX_EXTENDED_NUM_FLOAT_IMAGES;
+		tex_image_byte_start = TEX_EXTENDED_IMAGE_BYTE_START;
+	}
+	else if((info.type == DEVICE_CUDA || info.type == DEVICE_MULTI) && info.extended_images) {
+		tex_num_images = TEX_EXTENDED_NUM_IMAGES_GPU;
+	}
 }
 
 bool ImageManager::set_animation_frame_update(int frame)
diff --git a/intern/cycles/render/image.h b/intern/cycles/render/image.h
index a862ffce5c3..561550fe0d2 100644
--- a/intern/cycles/render/image.h
+++ b/intern/cycles/render/image.h
@@ -17,6 +17,7 @@
 #ifndef __IMAGE_H__
 #define __IMAGE_H__
 
+#include "device.h"
 #include "device_memory.h"
 
 #include "util_string.h"
@@ -27,11 +28,16 @@
 
 CCL_NAMESPACE_BEGIN
 
+/* generic */
 #define TEX_NUM_IMAGES			95
 #define TEX_IMAGE_BYTE_START	TEX_NUM_FLOAT_IMAGES
 
+/* extended gpu */
+#define TEX_EXTENDED_NUM_IMAGES_GPU		145
+
+/* extended cpu */
 #define TEX_EXTENDED_NUM_FLOAT_IMAGES	1024
-#define TEX_EXTENDED_NUM_IMAGES			1024
+#define TEX_EXTENDED_NUM_IMAGES_CPU		1024
 #define TEX_EXTENDED_IMAGE_BYTE_START	TEX_EXTENDED_NUM_FLOAT_IMAGES
 
 /* color to use when textures are not found */
@@ -59,7 +65,7 @@ public:
 
 	void set_osl_texture_system(void *texture_system);
 	void set_pack_images(bool pack_images_);
-	void set_extended_image_limits(void);
+	void set_extended_image_limits(const DeviceInfo& info);
 	bool set_animation_frame_update(int frame);
 
 	bool need_update;
diff --git a/intern/cycles/render/scene.cpp b/intern/cycles/render/scene.cpp
index 4a42b2c9e35..686a1bc406e 100644
--- a/intern/cycles/render/scene.cpp
+++ b/intern/cycles/render/scene.cpp
@@ -63,8 +63,8 @@ Scene::Scene(const SceneParams& params_, const DeviceInfo& device_info_)
 	else
 		shader_manager = ShaderManager::create(this, SceneParams::SVM);
 
-	if (device_info_.type == DEVICE_CPU)
-		image_manager->set_extended_image_limits();
+	/* Extended image limits for CPU and GPUs */
+	image_manager->set_extended_image_limits(device_info_);
 }
 
 Scene::~Scene()
diff --git a/intern/cycles/render/scene.h b/intern/cycles/render/scene.h
index c913b4c59da..0f0bb725823 100644
--- a/intern/cycles/render/scene.h
+++ b/intern/cycles/render/scene.h
@@ -105,8 +105,8 @@ public:
 	/* integrator */
 	device_vector<uint> sobol_directions;
 
-	/* images */
-	device_vector<uchar4> tex_image[TEX_EXTENDED_NUM_IMAGES];
+	/* cpu images */
+	device_vector<uchar4> tex_image[TEX_EXTENDED_NUM_IMAGES_CPU];
 	device_vector<float4> tex_float_image[TEX_EXTENDED_NUM_FLOAT_IMAGES];
 
 	/* opencl images */