Add support for Radeon VII

Makefile

@@ -59,7 +59,7 @@ endif
 
 # AMD flags
 ROCM_PATH ?= /opt/rocm
-AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-offload-arch -a)
+AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
 HIPCC := $(shell which hipcc 2>/dev/null)
 HIPIFY := $(shell which hipify-perl 2>/dev/null)
 HIPCC_FLAGS = -O3 -march=native -I$(BUILD_DIR)/hip -fno-strict-aliasing
@@ -69,6 +69,10 @@ ifneq ($(filter gfx1100,$(AMDGPU_TARGETS)),)
   USE_HIPBLAS ?= 1
   USE_CK ?= 1
   AMDGPU_TARGETS := gfx1100
+else ifneq ($(filter gfx906,$(AMDGPU_TARGETS)),)
+   WAVEFRONTSIZE64 ?= 1
+   USE_HIPBLAS ?= 1
+   AMDGPU_TARGETS := gfx906
 else ifneq ($(filter gfx90a,$(AMDGPU_TARGETS)),)
   WAVEFRONTSIZE64 ?= 1
   BUILD_XDL ?= 1

README.md

@@ -1,5 +1,7 @@
 # llm.c for AMD devices
-This is a fork of [Andrej Karpathy's llm.c](https://github.com/karpathy/llm.c) with support for AMD's RDNA and CDNA devices.
+This is a fork of [Andrej Karpathy's llm.c](https://github.com/karpathy/llm.c) with support for AMD devices.
+
+It has been tested on Radeon VII (aka gfx906), MI250X (aka gfx90a), and 7900 XTX (aka gfx1100).
 
 ## Performance
 

llmc/mfu.h

@@ -29,6 +29,7 @@ static const PerfData HOPPER = {378.f, 756.f, 756.f, 756.f, 1513.f, 1513.f, 1620
 static const PerfData ADA = {82.6f, 165.2f, 165.2f, 330.3f, 330.3f, 660.6f, 2520.f, 512.f};
 static const PerfData RDNA3 = {61.42f, 122.8f, 122.8f, -1.f, -1.f, -1.f, 2500.f, 384.f};
 static const PerfData CDNA2 = {95.7f, 383.0f, 383.0f, -1.f, -1.f, -1.f, 1690.f, 208.f};
+static const PerfData GCN5 = {10.75f, 10.75f, 21.5f, -1.f, -1.f, -1.f, 1750.f, 60.f};
 
 typedef struct {
     const char* name;
@@ -79,6 +80,7 @@ static GPUEntry gpu_db[] = {
     {"NVIDIA H100 80GB HBM3", &HOPPER, 528, 1830}, // HBM3 = SXM5
     {"Radeon RX 7900 XTX", &RDNA3, 384, 2500},
     {"AMD Instinct MI250X/MI250", &CDNA2, 208, 1690},
+    {"AMD Radeon VII", &GCN5, 60, 1750},
 };
 
 float get_flops_promised(const char* device, int precision_mode) {