Fixed 5xxx bugs

tmrlvi · Jan 18, 2022 · 2ff278a · 2ff278a
1 parent 2427f08
commit 2ff278a
Show file tree

Hide file tree

Showing 2 changed files with 16 additions and 1 deletion.
diff --git a/plugins/opencl/resources/kaspa-opencl.cl b/plugins/opencl/resources/kaspa-opencl.cl
@@ -192,7 +192,7 @@ uint32_t STATIC inline amul4bit(constant uint32_t packed_vec1[32], uint32_t pack
     for (int i=0; i<QUARTER_MATRIX_SIZE; i++) {
         #if __PLATFORM__ == NVIDIA_CUDA && (__COMPUTE_MAJOR__ > 6 || (__COMPUTE_MAJOR__ == 6 && __COMPUTE_MINOR__ >= 1))
         asm("dp4a.u32.u32" " %0, %1, %2, %3;": "=r" (res): "r" (packed_vec1[i]), "r" (packed_vec2[i]), "r" (res));
-        #elif (__FORCE_AMD_V_DOT4_U32_U8__ == 1) || ((__GFXIP_MAJOR__ == 9) && (__GFXIP_MINOR__ == 6 || __GFXIP_MINOR__ == 8)) || ((__GFXIP_MAJOR__ == 10) && (__GFXIP_MINOR__ == 11 || __GFXIP_MINOR__ == 12 || __GFXIP_MINOR__ >= 3)) || __GFXIP_MAJOR__ > 10
+        #elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx1011__) || defined(__gfx1012__) || defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__)
         __asm__("v_dot4_u32_u8" " %0, %1, %2, %3;": "=v" (res): "r" (packed_vec1[i]), "r" (packed_vec2[i]), "r" (res));
         #else
         res += ((constant char4 *)packed_vec1)[i].x*((char4 *)packed_vec2)[i].x;

diff --git a/plugins/opencl/src/worker.rs b/plugins/opencl/src/worker.rs
@@ -204,6 +204,12 @@ impl OpenCLGPUWorker {
                         "",
                     )
                     .unwrap_or_else(|_| panic!("{}::Program::create_and_build_from_binary failed", name)),
+                    "gfx1010" => Program::create_and_build_from_binary(
+                        &context,
+                        &[include_bytes!("../resources/bin/gfx1010_kaspa-opencl.bin")],
+                        "",
+                    )
+                    .unwrap_or_else(|_| panic!("{}::Program::create_and_build_from_binary failed", name)),
                     "gfx1011" => Program::create_and_build_from_binary(
                         &context,
                         &[include_bytes!("../resources/bin/gfx1011_kaspa-opencl.bin")],
@@ -384,6 +390,15 @@ fn from_source(context: &Context, device: &Device, options: &str) -> Result<Prog
         Err(_) => String::new(),
     };
 
+    // Hack to recreate the AMD flags
+    compile_options += &match device.pcie_id_amd() {
+        Ok(_) => {
+            let device_name = device.name().unwrap_or_else(|_| "Unknown".into());
+            format!("-D OPENCL_PLATFORM_AMD -D __{}__", device_name)
+        },
+        Err(_) => String::new(),
+    };
+
     info!("Build OpenCL with {}", compile_options);
 
     Program::create_and_build_from_source(context, PROGRAM_SOURCE, compile_options.as_str())